Skip to content

The first exam for code agents in the physical world: frozen robot tasks from many benchmark suites, each answered with a program and verified in a separate sandbox. Runs on Harbor.

DexJoCo: instance.npz no longer names its source demonstration; pointer check reads .npz; infra_error for instance mismatch (DexJoCo, Playground quadrupeds) (board 0105) (+8 more) (#190, #188, #197, #204, #200, #199, ...)

Kangrui Wangfeb31cda1e7d
diff --git a/ci_checks/check-agent-pointers.py b/ci_checks/check-agent-pointers.py
index bfa0a25394..74260197da 100755
--- a/ci_checks/check-agent-pointers.py
+++ b/ci_checks/check-agent-pointers.py
@@ -9,7 +9,13 @@ files this check refuses:
datasets, trained policies, project pages, solution videos;
2. a key that locates a source (KEYS): `source_url`, `source_viewer_url`, `demo_url`, `dataset_url`, ...;
3. the name of a demonstration or dataset file (FILES): `*.hdf5`, `*.mcap`, `*.parquet`, `*.rlds`, a Hugging Face
- `datasets/<org>/<name>` id.
+ `datasets/<org>/<name>` id;
+ 4. a key that names the demonstration or episode a file was made from (IDS): `"demo": "<demo id>"`, `demo_id`,
+ `episode_id`, `source_demo`, ... (board 0105: DexJoCo's instance.npz `meta` named the demonstration whose actions
+ are the reference solution).
+Binary files are skipped except numpy archives (`.npz`, a zip of `.npy` members): their string members (`<U`, `S`) and
+any text member are read and scanned like a text file (the members' names are reported as `file.npz[member]`); numeric
+members are not read past their header. Images and other binaries are skipped.
A legitimate mention (a licence notice that must name its source) is listed in ALLOW with its reason. The egress
allowlist is what stops a download; this check keeps the address out of the task. Board 0072: RoboTangle's
`reference_scene.json` carried `source_url` of the upstream demonstration, and one trial downloaded and replayed it.
@@ -18,7 +24,7 @@ is built (docs/LEAKAGE.md).
Usage: check-agent-pointers.py [task dirs...] (default: all of tasks/*)
"""
-import pathlib, re, sys
+import ast, pathlib, re, sys, zipfile
ROOT = pathlib.Path(__file__).resolve().parents[1]
SIDECAR_DIRS = {"sim", "scorer", "rcb_monitor"} # other containers' files under environment/
@@ -31,6 +37,8 @@ HOSTS = re.compile(
KEYS = re.compile(r"""["']?\b((?:source|demo|demonstration|dataset|video|viewer|source_viewer|methods|reference|solution|
checkpoint|ckpt|policy|pretrained|repo|download)[_-]?(?:url|uri|link|repo|repo_id)|repo_id)\b["']?\s*[:=]""",
re.I | re.X)
+IDS = re.compile(r"""["']((?:source_|reference_)?(?:demo|demo_id|demo_name|demonstration|demonstration_id|episode|episode_id|
+ episode_name|trajectory_id|recording|recording_id))["']\s*:\s*["']""", re.I | re.X)
FILES = re.compile(r"""(?<![\w.-])(?:[\w.-]+\.(?:hdf5|h5|mcap|parquet|rlds|tfrecord)\b|datasets/[\w.-]+/[\w.-]+)""", re.I)
# (task-name regex, path regex, matched-text regex, reason)
ALLOW = [
@@ -50,20 +58,58 @@ def agent_files(task: pathlib.Path):
yield p
+def npy_text(f) -> str | None:
+ """The text of one `.npy` member (an open binary stream): a string array's contents, or None for a numeric array
+ (only its header is read). Pickled (object) arrays are read as raw bytes: their strings are stored verbatim."""
+ head = f.read(10)
+ if head[:6] != b"\x93NUMPY":
+ rest = f.read(MAX_BYTES)
+ raw = head + rest
+ return None if b"\0" in raw[:4096] else raw.decode("utf-8", "replace")
+ n = int.from_bytes(head[8:10], "little") if head[6] == 1 else int.from_bytes(head[8:10] + f.read(2), "little")
+ try:
+ descr = str(ast.literal_eval(f.read(n).decode("latin-1")).get("descr", ""))
+ except (ValueError, SyntaxError):
+ return None
+ kind = descr.lstrip("<>|=")[:1]
+ if kind == "U":
+ return f.read(MAX_BYTES).decode("utf-32-be" if descr.startswith(">") else "utf-32-le", "replace").replace("\0", "")
+ if kind in ("S", "O") or descr.startswith("[") or descr.startswith("{"):
+ return f.read(MAX_BYTES).decode("utf-8", "replace").replace("\0", " ")
+ return None
+
+
+def texts(p: pathlib.Path, rel: str):
+ """(name, text) of each readable text in one agent file: the file itself, or a numpy archive's string members."""
+ if p.suffix == ".npz":
+ try:
+ with zipfile.ZipFile(p) as z:
+ for i in z.infolist():
+ if not i.is_dir():
+ with z.open(i) as f:
+ t = npy_text(f)
+ if t:
+ yield f"{rel}[{i.filename}]", t
+ except (zipfile.BadZipFile, OSError, EOFError) as exc:
+ print(f"WARN {rel}: unreadable numpy archive ({exc!r}), not scanned")
+ return
+ if p.stat().st_size > MAX_BYTES:
+ return
+ raw = p.read_bytes()
+ if b"\0" not in raw[:4096]:
+ yield rel, raw.decode("utf-8", "replace")
+
+
def findings(task: pathlib.Path):
for p in agent_files(task):
- if p.stat().st_size > MAX_BYTES:
- continue
- raw = p.read_bytes()
- if b"\0" in raw[:4096]:
- continue
- text = raw.decode("utf-8", "replace").replace("\\/", "/"); rel = p.relative_to(task).as_posix() # JSON may escape "/"
- for kind, rx in (("URL", HOSTS), ("key", KEYS), ("file", FILES)):
- for m in rx.finditer(text):
- hit = m.group(1) if kind == "key" else m.group(0)
- if any(re.search(t, task.name) and re.search(pp, rel) and re.search(h, hit) for t, pp, h, _ in ALLOW):
- continue
- yield rel, text.count("\n", 0, m.start()) + 1, kind, hit[:120]
+ for rel, text in texts(p, p.relative_to(task).as_posix()):
+ text = text.replace("\\/", "/") # JSON may escape "/"
+ for kind, rx in (("URL", HOSTS), ("key", KEYS), ("file", FILES), ("id", IDS)):
+ for m in rx.finditer(text):
+ hit = m.group(1) if kind in ("key", "id") else m.group(0)
+ if any(re.search(t, task.name) and re.search(pp, rel) and re.search(h, hit) for t, pp, h, _ in ALLOW):
+ continue
+ yield rel, text.count("\n", 0, m.start()) + 1, kind, hit[:120]
def main(argv):
diff --git a/ci_checks/check-ci-selftest.py b/ci_checks/check-ci-selftest.py
index 10fbbfd87a..236aaf31cd 100755
--- a/ci_checks/check-ci-selftest.py
+++ b/ci_checks/check-ci-selftest.py
@@ -26,6 +26,8 @@ POINTERS = [sys.executable, str(ROOT / "ci_checks/check-agent-pointers.py")]
IDENTITY = ["bash", str(ROOT / "ci_checks/check-tooling-identity.sh")] # repo-level: runs on a tasks directory (see main)
CUTOFFS = [sys.executable, str(ROOT / "ci_checks/check-cutoffs.py")]
SEEDS = [sys.executable, str(ROOT / "ci_checks/check-seeds.py")]
+DYNAMICS = [sys.executable, str(ROOT / "ci_checks/check-dynamics.py")]
+REALTIME = [sys.executable, str(ROOT / "ci_checks/check-realtime.py")]
STD = "behavior-boxing-books-up-for-storage-i00-standard" # a sim sidecar with a trusted cache volume
PRIV = "behavior-boxing-books-up-for-storage-i00-privileged" # a verifier compose file with a trusted volume
@@ -64,6 +66,18 @@ def append(text: str):
return lambda s: s + text
+def npz_text(member: str, text: str) -> bytes:
+ """A numpy archive (no numpy needed) whose one member is a 0-d unicode array holding `text`, as np.savez writes it."""
+ import io, zipfile
+ head = "{'descr': '<U%d', 'fortran_order': False, 'shape': (), }" % len(text)
+ head += " " * (-(10 + len(head) + 1) % 64) + "\n"
+ buf = io.BytesIO()
+ with zipfile.ZipFile(buf, "w", zipfile.ZIP_DEFLATED) as z:
+ z.writestr(member, b"\x93NUMPY\x01\x00" + len(head).to_bytes(2, "little") + head.encode("latin-1")
+ + text.encode("utf-32-le"))
+ return buf.getvalue()
+
+
COMPOSE, TOML, TEST_COMPOSE = "environment/docker-compose.yaml", "task.toml", "tests/docker-compose.yaml"
TEST_SH, TEST_DF = "tests/test.sh", "tests/Dockerfile"
VOL = r"^ volumes:.*$"
@@ -148,6 +162,12 @@ MUTATIONS = [
("... names the recording its reference replays",
PRIV, "environment/instance.json", sub(r"^\{", '{"source": {"file": "demos/boxing_books_017.hdf5", "demo": 3}, ', count=1, flags=0),
POINTERS, "file `boxing_books_017.hdf5` locates upstream material"),
+ ("... names the demonstration its scene was restored from (the reference replays that demonstration's actions)",
+ PRIV, "environment/instance.json", sub(r"^\{", '{"demo": "boxing_books_demo_17", ', count=1, flags=0),
+ POINTERS, "id `demo` locates upstream material"),
+ ("... in a string member of a numpy archive the agent reads (DexJoCo's instance.npz `meta`, board 0105)",
+ PRIV, "+environment/extra.npz", lambda s: npz_text("meta.npy", '{"source": "x", "demo": "water_plant_demo_10_2026"}'),
+ POINTERS, "environment/extra.npz[meta.npy]:1: id `demo`"),
("a step with a network policy of its own (Harbor would let it override [agent])",
STD, TOML, append('\n[[steps]]\nname = "s1"\n[steps.agent]\nnetwork_mode = "public"\n'), ISOLATION, "a step sets its own network policy"),
("the instruction links the upstream repository",
@@ -274,6 +294,15 @@ MUTATIONS = [
TWIN_P, TOML, sub(r"^allowed = \[(.*)\]$", r"allowed = [\1, 12345]"), SEEDS, "has no frozen instance"),
("another seed's instance in the agent's build context",
TWIN_P, "+environment/seeds/7/instance.json", append("{}"), SEEDS, "a `seeds` directory where the agent's image"),
+ # PROTOCOL §2.14: scene dynamics, from the one table
+ ("a task.toml that contradicts the dynamics table",
+ TWIN_P, TOML, sub(r'^dynamics = "static"', 'dynamics = "dynamic"'), DYNAMICS, "the table says"),
+ ("a task.toml without dynamics",
+ TWIN_P, TOML, sub(r'^dynamics = .*\n', ""), DYNAMICS, "[metadata] dynamics is missing"),
+ # PROTOCOL §4b: a standard-realtime task states its slowdown S
+ ("a standard-realtime task.toml whose slowdown is not its service's",
+ "humanoidbench-loco-walk-i00-standard-realtime", TOML, sub(r"^slowdown = [0-9.]+", "slowdown = 2"), REALTIME,
+ "but its service's RCB_SLOWDOWN default is 1"),
]
@@ -385,7 +414,7 @@ def main() -> int:
return 1
clean[name] = Path(tmp) / "clean" / name
shutil.copytree(src, clean[name])
- for cmd in (ISOLATION, RESOURCES, POINTERS, CUTOFFS, SEEDS):
+ for cmd in (ISOLATION, RESOURCES, POINTERS, CUTOFFS, SEEDS, DYNAMICS):
rc, out = run(cmd, clean[name])
if rc != 0:
print(f"FAIL check-ci-selftest: the unmutated {name} does not pass {Path(cmd[-1]).name}:\n{out[-800:]}")
@@ -398,7 +427,8 @@ def main() -> int:
if r.startswith("+"): # a file the mutation adds
r = r[1:]
(t / r).parent.mkdir(parents=True, exist_ok=True)
- (t / r).write_text(f(""))
+ data = f("")
+ (t / r).write_bytes(data) if isinstance(data, bytes) else (t / r).write_text(data)
continue
(t / r).write_text(f((t / r).read_text()))
except (AssertionError, StopIteration, OSError) as e:
diff --git a/ci_checks/check-cutoffs.py b/ci_checks/check-cutoffs.py
index 0a16fe46c8..35cae58054 100644
--- a/ci_checks/check-cutoffs.py
+++ b/ci_checks/check-cutoffs.py
@@ -17,7 +17,8 @@ scripts/task_cutoffs.toml gives the seconds per benchmark and mode, with optiona
- standard modes: the `sim` service is told the same number (`RCB_WALL_CLOCK_S` in environment/docker-compose.yaml,
what its `/spec` budget and `meta["time_left_s"]` report; PROTOCOL §2.3).
Repository level (no task given): the table itself is usable, has a number for exactly the benchmark-mode pairs that
-have tasks, every exception names a family or a task that exists, every number equals its last recorded value (its
+have tasks, privileged and standard (any mode with `same_cutoff_as` in the mode registry) have one number wherever
+both exist (a benchmark's row, a family's entry, a task and its counterpart), every exception names a family or a task that exists, every number equals its last recorded value (its
newest `[[history]]` entry, else scripts/task_cutoffs.baseline.toml: a change without its entry fails), and
docs/CUTOFFS.md is what `python3 scripts/task_cutoffs.py apply` renders from it.
@@ -160,6 +161,27 @@ def check_repo(tasks: list[Path]) -> list[str]:
for name in sorted(by_task):
if name not in names:
errs.append(f"scripts/task_cutoffs.toml: [task_cutoff_sec] names {name!r}, which is not a task directory")
+ # a benchmark may ship any subset of the modes (PROTOCOL §2.15), but a mode whose registry entry names
+ # `same_cutoff_as` has that mode's number wherever both exist (§2.16; standard = privileged, owner 2026-10-12):
+ # in a benchmark's row, in a family's entry, and for a task and its counterpart of the same base
+ pairs = [(m, d["same_cutoff_as"]) for m, d in tc.MODE_REGISTRY.items() if d["same_cutoff_as"]]
+ for m, other in pairs:
+ for bench, row in sorted(table.items()):
+ if m in row and other in row and row[m] != row[other]:
+ errs.append(f"scripts/task_cutoffs.toml: {bench} gives {other} {row[other]} s and {m} {row[m]} s; the two "
+ "modes have the same cutoff (PROTOCOL §2.12, §2.16)")
+ for bench, rows in sorted(families.items()):
+ for family, row in sorted(rows.items()):
+ if m in row and other in row and row[m] != row[other]:
+ errs.append(f"scripts/task_cutoffs.toml: [family_cutoff_sec] {bench} {family!r} gives {other} "
+ f"{row[other]} s and {m} {row[m]} s; the two modes have the same cutoff")
+ for t in tasks:
+ if tc.mode_of(t) != m:
+ continue
+ o = t.with_name(f"{tc.base_of(t)}-{other}")
+ if (o / "task.toml").is_file() and tc.cutoff_for(t) != tc.cutoff_for(o):
+ errs.append(f"{t.name} has a cutoff of {tc.cutoff_for(t)} s and {o.name} {tc.cutoff_for(o)} s; the two "
+ "modes have the same cutoff (PROTOCOL §2.12, §2.16)")
errs += task_cutoffs.history_errors() # every number equals its last recorded value: a change has its [[history]] entry
doc = ROOT / "docs/CUTOFFS.md"
if not doc.is_file() or doc.read_text(encoding="utf-8") != task_cutoffs.render_doc(tasks):
diff --git a/ci_checks/check-dynamics.py b/ci_checks/check-dynamics.py
new file mode 100644
index 0000000000..d119c7ea85
--- /dev/null
+++ b/ci_checks/check-dynamics.py
@@ -0,0 +1,90 @@
+#!/usr/bin/env python3
+"""Scene dynamics (docs/PROTOCOL.md §2.14): every task says whether its scene is `dynamic` or `static`, and the one table
+says it first. Checks:
+ - scripts/task_dynamics.toml: every entry (a benchmark's row, a family's exception) has `dynamics` in {static, dynamic}
+ and a non-empty `why`; every benchmark of scripts/task_conventions.py BENCHMARKS that has tasks has a row; a
+ family's exception names a benchmark that has a row, and differs from that row (an exception that says what the
+ row says is noise);
+ - every task's task.toml has `[metadata] dynamics`, and it is the table's value for the task (its family's exception,
+ else its benchmark's row);
+ - when run on every task (no arguments): every family exception matches at least one task.
+
+ python3 ci_checks/check-dynamics.py [task dirs] # default: every task
+Fix: edit scripts/task_dynamics.toml, then `python3 scripts/task_conventions.py tasks/<task> ...` (or
+`python3 scripts/task_cutoffs.py apply`, which normalizes every task and refreshes tasks/dataset.toml). Never edit the
+value in a task.toml. Standard library only.
+"""
+from __future__ import annotations
+
+import sys
+import tomllib
+from pathlib import Path
+
+ROOT = Path(__file__).resolve().parents[1]
+sys.path.insert(0, str(ROOT / "scripts"))
+import task_conventions as tc # noqa: E402
+
+
+def table_errors(tasks: list[Path]) -> list[str]:
+ errs = []
+ if not tc.DYNAMICS_FILE.is_file():
+ return [f"{tc.DYNAMICS_FILE.relative_to(ROOT)} is missing"]
+ rows, fams = tc.dynamics_table()
+ entries = [(f'benchmark "{b}"', e) for b, e in rows.items()]
+ entries += [(f'family "{b}" / "{f}"', e) for b, fs in fams.items() for f, e in fs.items()]
+ for where, e in entries:
+ if not isinstance(e, dict) or e.get("dynamics") not in tc.DYNAMICS:
+ errs.append(f"{where}: dynamics must be one of {tc.DYNAMICS}")
+ elif not str(e.get("why", "")).strip():
+ errs.append(f"{where}: give the reason (`why`)")
+ labels = {label for _, label in tc.BENCHMARKS}
+ for b in rows:
+ if b not in labels:
+ errs.append(f'benchmark "{b}" is not a label of scripts/task_conventions.py BENCHMARKS')
+ for b, fs in fams.items():
+ if b not in rows:
+ errs.append(f'family exceptions for "{b}", which has no benchmark row')
+ continue
+ for f, e in fs.items():
+ if isinstance(e, dict) and e.get("dynamics") == rows[b].get("dynamics"):
+ errs.append(f'family "{b}" / "{f}" says what the benchmark row says: remove it')
+ for b in sorted({tc.bench_of(t) for t in tasks} - set(rows)):
+ errs.append(f'no row for benchmark "{b or "?"}" (a task directory no BENCHMARKS prefix matches has no benchmark)')
+ return [f"FAIL scripts/task_dynamics.toml: {e}" for e in errs]
+
+
+def task_errors(task: Path) -> list[str]:
+ md = tomllib.loads((task / "task.toml").read_text()).get("metadata", {})
+ want = tc.dynamics_for(task)
+ have = md.get("dynamics")
+ if have is None:
+ return [f"FAIL {task.name}: [metadata] dynamics is missing (run scripts/task_conventions.py on it)"]
+ if have not in tc.DYNAMICS:
+ return [f"FAIL {task.name}: [metadata] dynamics = {have!r}, not one of {tc.DYNAMICS}"]
+ if want is not None and have != want:
+ return [f"FAIL {task.name}: [metadata] dynamics = {have!r}, the table says {want!r} (scripts/task_dynamics.toml; "
+ "run scripts/task_conventions.py on it)"]
+ return []
+
+
+def main(argv: list[str]) -> int:
+ every = not argv
+ tasks = sorted(p.parent for p in (ROOT / "tasks").glob("*/task.toml")) if every else \
+ [Path(a) for a in argv if (Path(a) / "task.toml").is_file()]
+ errs = table_errors(tasks)
+ for t in tasks:
+ errs += task_errors(t)
+ if every:
+ _, fams = tc.dynamics_table()
+ seen = {(tc.bench_of(t), tc.family_of(t)) for t in tasks}
+ errs += [f'FAIL scripts/task_dynamics.toml: family "{b}" / "{f}" matches no task'
+ for b, fs in fams.items() for f in fs if (b, f) not in seen]
+ for e in errs:
+ print(e)
+ if not errs:
+ print(f"check-dynamics: {len(tasks)} task(s) OK")
+ return 1 if errs else 0
+
+
+if __name__ == "__main__":
+ sys.exit(main(sys.argv[1:]))
diff --git a/ci_checks/check-instance-copies.sh b/ci_checks/check-instance-copies.sh
index 44d47cf00e..7157a53899 100755
--- a/ci_checks/check-instance-copies.sh
+++ b/ci_checks/check-instance-copies.sh
@@ -4,38 +4,36 @@
# privileged task's tests/instance.{json,npz}, and the copy the task runs from (privileged: environment/; standard and
# standard-realtime: the sim sidecar's environment/sim/) are byte-identical, and neither exists without the other.
# RoboPaint's privileged environment/ copy is reduced by design (check-robopaint-leak.sh), so its scorer sidecar's
-# environment/scorer/ copy is compared instead. A standard or standard-realtime task's twin <base>-privileged must
-# exist, except for the benchmarks PROTOCOL §4b exempts by owner decision (NO_TWIN below; the same list:
-# scripts/task_conventions.py NO_TWIN): they have no privileged mode and nothing to compare. STANDARD_ONLY: standard tasks
-# without a twin that carry their own frozen instance (PROTOCOL §4.8, physics that does not replay bit for bit: graded by
-# the live episode); their generator checks that copy against the benchmark's frozen instance (mk_tasks.py --check).
+# environment/scorer/ copy is compared instead.
+# A benchmark ships any subset of the three modes (PROTOCOL §2.15, owner 2026-10-14). A standard or standard-realtime
+# task without its twin <base>-privileged is accepted when scripts/task_conventions.py no_twin_reason() gives a reason:
+# its benchmark has no privileged mode (no privileged number in scripts/task_cutoffs.toml), or the task is listed in
+# NO_TWIN (e.g. RoboDojo's six standard-only tasks, PROTOCOL §4.8). Such a task carries its own frozen instance, if it
+# has one: its sim sidecar's copy and its verifier's tests/ copy, where both exist, are byte-identical (a generator's
+# --check compares them with the benchmark's frozen instance). Any other missing twin fails.
# Usage: check-instance-copies.sh tasks/<task-dir>
set -uo pipefail
t="${1%/}"; n="$(basename "$t")"; rc=0
-NO_TWIN=(roboolympics-) # PROTOCOL §4b: RoboOlympics (owner, 2026-10-08), standard-realtime only
-STANDARD_ONLY=(robodojo-fold-clothes-i00-standard robodojo-fold-clothes-random-i00-standard robodojo-make-toast-random-i00-standard robodojo-pour-by-language-i00-standard robodojo-pour-liquid-into-cup-i00-standard robodojo-pour-liquid-into-cup-random-i00-standard) # PROTOCOL §4.8 exception (owner, 2026-10-11)
-case "$n" in
- *-standard-realtime) base="${n%-standard-realtime}"; ref="$base-privileged/tests"; copy="environment/sim" ;; # a sim sidecar too
- *-standard) base="${n%-standard}"; ref="$base-privileged/tests"; copy="environment/sim" ;;
- robopaint-*) base=""; ref="$n/tests"; copy="environment/scorer" ;;
- *) base=""; ref="$n/tests"; copy="environment" ;;
-esac
-if [ -n "$base" ]; then
- for s in "${STANDARD_ONLY[@]}"; do # while its twin still exists (until the tasks change lands) the pair rules below apply
- if [ "$n" = "$s" ] && [ ! -e "$(dirname "$t")/$base-privileged" ]; then
- [ -f "$t/$copy/instance.json" ] || { echo "FAIL $t: a standard-only task carries its frozen instance ($copy/instance.json)"; exit 1; }
- exit 0
- fi
+root="$(cd "$(dirname "$0")/.." && pwd)"
+# the mode registry (PROTOCOL §2.16; run_all.sh reads it once): "<mode> <sidecar 0|1> <twin mode or ->", longest first
+modes="${RCB_MODES_TABLE:-$(python3 "$root/scripts/task_conventions.py" modes)}"
+base=""; ref="$n/tests"; copy="environment"; twin=""
+case "$n" in robopaint-*) copy="environment/scorer" ;; esac
+while read -r m sidecar tw; do
+ [ -n "$m" ] && [[ "$n" == *"-$m" ]] || continue
+ if [ "$tw" != "-" ]; then base="${n%-$m}"; twin="$tw"; ref="$base-$tw/tests"; fi
+ [ "$sidecar" = 1 ] && copy="environment/sim"
+ break
+done <<< "$modes"
+if [ -n "$base" ] && [ ! -d "$(dirname "$t")/$base-$twin" ]; then
+ why="$(python3 -c 'import sys; from pathlib import Path; sys.path.insert(0, sys.argv[1]); import task_conventions as tc; print(tc.no_twin_reason(Path(sys.argv[2])) or "")' "$root/scripts" "$t")" \
+ || { echo "FAIL $t: scripts/task_conventions.py no_twin_reason() failed"; exit 1; }
+ [ -n "$why" ] || { echo "FAIL $t: its twin $base-$twin is missing; its benchmark has a $twin mode, so a task without a twin is listed in scripts/task_conventions.py NO_TWIN with the reason (PROTOCOL §2.15)"; exit 1; }
+ for f in instance.json instance.npz; do # no twin: its own two copies, where both exist, agree
+ a="$t/tests/$f"; b="$t/$copy/$f"
+ [ -f "$a" ] && [ -f "$b" ] && ! cmp -s "$a" "$b" && { echo "FAIL $t: $copy/$f and tests/$f differ"; rc=1; }
done
- for p in "${NO_TWIN[@]}"; do
- if [[ "$n" == "$p"* ]]; then
- for f in instance.json instance.npz; do # no twin, so no instance copy to keep in step with one
- [ -f "$t/$copy/$f" ] && { echo "FAIL $t: $copy/$f without a privileged twin (PROTOCOL §4b NO_TWIN)"; rc=1; }
- done
- exit $rc
- fi
- done
- [ -d "$(dirname "$t")/$base-privileged" ] || { echo "FAIL $t: its privileged twin $base-privileged is missing (PROTOCOL §2.2)"; exit 1; }
+ exit $rc
fi
for f in instance.json instance.npz; do
a="$(dirname "$t")/$ref/$f"; b="$t/$copy/$f"
diff --git a/ci_checks/check-protocol.py b/ci_checks/check-protocol.py
index 620682d044..3a1cf0cb96 100755
--- a/ci_checks/check-protocol.py
+++ b/ci_checks/check-protocol.py
@@ -16,8 +16,11 @@ from pathlib import Path
ROOT = Path(__file__).resolve().parent.parent
REGISTRY = "ghcr.io/mll-lab-nu/eai-"
-PROTOCOLS = {"eai-standard/2.1", "eai-standard/2.2", "eai-standard/2.4"} # standard mode (2.4: the operations capability, issue #89)
-REALTIME_PROTOCOLS = {"eai-standard/2.3"} # standard-realtime mode (PROTOCOL §4b): 2.3 = the realtime capability
+sys.path.insert(0, str(ROOT / "scripts"))
+import task_conventions as tc # noqa: E402 (the mode registry, PROTOCOL §2.16)
+# the `protocol` values each mode may declare: standard 2.1 / 2.2 / 2.4 (2.4: the operations capability, issue #89),
+# standard-realtime 2.3 (the realtime capability, §4b), privileged none
+MODE_PROTOCOLS = {m: set(d["protocols"]) for m, d in tc.MODE_REGISTRY.items()}
PERSONAL_PATH = re.compile(rb"/(home|Users)/[A-Za-z0-9_.-]+/")
TEXT_SUFFIXES = {".md", ".toml", ".py", ".sh", ".yaml", ".yml", ".json", ".jsonl", ".txt", ".cfg", ".xml", ".urdf", ""}
@@ -53,19 +56,21 @@ def check_task(task: Path) -> None:
meta = tomllib.loads((task / "task.toml").read_text()).get("metadata", {})
mode = meta.get("mode")
if mode is None:
- warn("missing `mode` (privileged | standard | standard-realtime)", name)
- # whole suffixes only: `-standard-realtime` does not end in `-standard`, and `mode` is compared exactly
- realtime = mode == "standard-realtime" or name.endswith("-standard-realtime")
- standard = not realtime and (mode == "standard" or name.endswith("-standard"))
+ warn("missing `mode` (" + " | ".join(tc.MODES) + ")", name)
+ # whole suffixes only (`-standard-realtime` does not end in `-standard`), and `mode` is compared exactly
+ try:
+ named = tc.mode_of(task)
+ except ValueError:
+ named = None
if name.endswith("-realtime") and not name.endswith("-standard-realtime") or mode == "realtime":
warn("mode name `realtime` / `-realtime` (the third mode is `standard-realtime`, PROTOCOL D12)", name)
proto = meta.get("protocol")
- if standard and proto not in PROTOCOLS:
- warn(f"standard-mode `protocol` not in {sorted(PROTOCOLS)}", f"{name} ({proto})")
- if realtime and proto not in REALTIME_PROTOCOLS:
- warn(f"standard-realtime `protocol` not in {sorted(REALTIME_PROTOCOLS)}", f"{name} ({proto})")
- if not standard and not realtime and proto is not None:
- warn("`protocol` set on a privileged task", name)
+ for m in {mode, named} & set(tc.MODES):
+ allowed = MODE_PROTOCOLS[m]
+ if allowed and proto not in allowed:
+ warn(f"{m}-mode `protocol` not in {sorted(allowed)}", f"{name} ({proto})")
+ if not allowed and proto is not None:
+ warn(f"`protocol` set on a {m} task", name)
oracle = meta.get("oracle")
if oracle is None:
diff --git a/ci_checks/check-realtime.py b/ci_checks/check-realtime.py
new file mode 100644
index 0000000000..6cb334f6c8
--- /dev/null
+++ b/ci_checks/check-realtime.py
@@ -0,0 +1,76 @@
+#!/usr/bin/env python3
+"""Standard-realtime tasks (docs/PROTOCOL.md §4b): every one states its slowdown S and allows resets. Checks, for every
+task whose directory ends in `-standard-realtime`:
+ - its sim service sets `RCB_SLOWDOWN` with a default (`"${RCB_SLOWDOWN:-S}"`, S > 0) in environment/docker-compose.yaml;
+ - `[metadata] slowdown` in task.toml is that S (written by scripts/task_conventions.py `normalize`);
+ - its service sets `RCB_RESET_BUDGET` with a default of at least 1 (owner, 2026-10-14: every standard-realtime task
+ allows resets, 10 by default), unless its benchmark is in RESET_EXCEPTIONS (PROTOCOL §4b, per-benchmark exceptions);
+ - its instruction states S, as `S = <S>` (the generator's text; the README's "Real time" row and task.toml say it too),
+ unless its benchmark is in TEXT_GRANDFATHERED (the twins that existed before the rule; their S is in /spec.timing).
+
+ python3 ci_checks/check-realtime.py [task dirs] # default: every task
+Standard library only.
+"""
+from __future__ import annotations
+
+import re
+import sys
+import tomllib
+from pathlib import Path
+
+ROOT = Path(__file__).resolve().parents[1]
+sys.path.insert(0, str(ROOT / "scripts"))
+import task_conventions as tc # noqa: E402
+
+# PROTOCOL §4b: benchmarks whose standard-realtime tasks default to no reset (owner decisions)
+RESET_EXCEPTIONS = {
+ "BiGym 2.0": "its realtime twins stay as they are (owner, 2026-10-14, D2); budget.reset = 0 since 2026-10-08",
+ "RoboOlympics": "a match is one episode (owner, 2026-10-08)",
+}
+# Realtime twins made before 2026-10-14 (owner D2: kept as they are): their instructions need not quote S
+TEXT_GRANDFATHERED = {"HumanoidBench", "BiGym 2.0", "RoboOlympics"}
+
+
+def check(task: Path) -> list[str]:
+ errs = []
+ compose = tc.read_text(task / "environment/docker-compose.yaml") or ""
+ m = tc.SLOWDOWN_RE.search(compose)
+ if not m:
+ return [f'{task.name}: its sim service needs `RCB_SLOWDOWN: "${{RCB_SLOWDOWN:-S}}"` in environment/docker-compose.yaml']
+ S = float(m.group(1))
+ if not S > 0:
+ errs.append(f"{task.name}: RCB_SLOWDOWN's default must be > 0 (got {m.group(1)})")
+ md = tomllib.loads((task / "task.toml").read_text()).get("metadata", {})
+ if "slowdown" not in md:
+ errs.append(f"{task.name}: [metadata] slowdown is missing (run scripts/task_conventions.py on the task)")
+ elif float(md["slowdown"]) != S:
+ errs.append(f"{task.name}: [metadata] slowdown = {md['slowdown']}, but its service's RCB_SLOWDOWN default is {S:g}")
+ bench = tc.bench_of(task)
+ r = tc.RESET_RE.search(compose)
+ if bench not in RESET_EXCEPTIONS and (r is None or int(r.group(1)) < 1):
+ errs.append(f"{task.name}: a standard-realtime task allows resets (RCB_RESET_BUDGET default >= 1; 10 by default, "
+ f"owner 2026-10-14), got {'none' if r is None else r.group(1)}")
+ if bench not in TEXT_GRANDFATHERED:
+ text = tc.read_text(task / "instruction.md") or ""
+ if not re.search(rf"\bS = {re.escape(tc.slowdown_text(S))}\b", text):
+ errs.append(f"{task.name}: instruction.md must state the slowdown as `S = {tc.slowdown_text(S)}`")
+ return errs
+
+
+def main(argv: list[str]) -> int:
+ dirs = [Path(a) for a in argv] or sorted(p for p in (ROOT / "tasks").iterdir() if (p / "task.toml").is_file())
+ tasks = [d for d in dirs if d.name.endswith("-standard-realtime")]
+ errs = [e for t in tasks for e in check(t)]
+ if errs:
+ print("FAIL standard-realtime tasks (PROTOCOL §4b: the slowdown S stated, resets allowed):")
+ for e in errs[:60]:
+ print(" -", e)
+ if len(errs) > 60:
+ print(f" ... {len(errs) - 60} more")
+ return 1
+ print(f"check-realtime: {len(tasks)} standard-realtime task(s) OK")
+ return 0
+
+
+if __name__ == "__main__":
+ sys.exit(main(sys.argv[1:]))
diff --git a/ci_checks/check-seeds.py b/ci_checks/check-seeds.py
index 66cd48532c..ffd42bae85 100755
--- a/ci_checks/check-seeds.py
+++ b/ci_checks/check-seeds.py
@@ -34,12 +34,8 @@ def seeds_table(task: Path) -> tuple[dict, dict]:
def holder(task: Path) -> Path:
"""The task whose tests/ holds the pair's instances: the privileged twin of a standard task, else the task."""
- mode = tc.mode_of(task)
- if mode in tc.SIDECAR_MODES:
- twin = task.with_name(tc.base_of(task) + "-privileged")
- if twin.is_dir():
- return twin
- return task
+ twin = tc.twin_of(task) # the mode registry's twin mode (PROTOCOL §2.16)
+ return twin if twin is not None and twin.is_dir() else task
def instance_kinds(d: Path) -> set[str]:
diff --git a/ci_checks/check-standard-agent-image.sh b/ci_checks/check-standard-agent-image.sh
index 1da357eb1f..85c7bf9d75 100755
--- a/ci_checks/check-standard-agent-image.sh
+++ b/ci_checks/check-standard-agent-image.sh
@@ -6,7 +6,11 @@
# Usage: check-standard-agent-image.sh tasks/<task-dir>
set -uo pipefail
t="${1%/}"
-case "$(basename "$t")" in *-standard|*-standard-realtime) ;; *) exit 0 ;; esac
+# the sidecar modes of the mode registry (PROTOCOL §2.16; run_all.sh reads it once): standard, standard-realtime
+modes="${RCB_MODES_TABLE:-$(python3 "$(dirname "$0")/../scripts/task_conventions.py" modes)}"
+n="$(basename "$t")"; side=0
+while read -r m sidecar _tw; do [ -n "$m" ] && [[ "$n" == *"-$m" ]] && { side="$sidecar"; break; }; done <<< "$modes"
+[ "$side" = 1 ] || exit 0
python3 - "$t" <<'PY'
import json, pathlib, posixpath, re, sys
t = pathlib.Path(sys.argv[1]); rc = 0
diff --git a/ci_checks/check-standard-client.sh b/ci_checks/check-standard-client.sh
index d4c22ba5a3..eeb1c71d0c 100755
--- a/ci_checks/check-standard-client.sh
+++ b/ci_checks/check-standard-client.sh
@@ -15,6 +15,7 @@ root="$(cd "$(dirname "$0")/.." && pwd)"
python3 - "${1:-$root/tasks}" "$root/scripts/eai_standard" <<'PY'
import os, re, shlex, sys, tomllib
tasks, src = sys.argv[1], sys.argv[2]
+sys.path.insert(0, os.path.dirname(os.path.abspath(src))); import task_conventions as tc # the mode registry (§2.16)
env = dict(l.split("=", 1) for l in open(os.path.join(src, "client.env")).read().splitlines() if "=" in l and not l.startswith("#"))
IMAGE = env["CLIENT_IMAGE"].strip()
req_path = os.path.join(src, env.get("CLIENT_REQUIREMENTS", "client-requirements.txt").strip())
@@ -75,7 +76,7 @@ for t in sorted(os.listdir(tasks)):
if not os.path.isfile(tt):
continue
meta = tomllib.load(open(tt, "rb")).get("metadata", {})
- if not (meta.get("mode") in ("standard", "standard-realtime") or t.endswith(("-standard", "-standard-realtime"))):
+ if not (meta.get("mode") in tc.SIDECAR_MODES or t.endswith(tuple("-" + m for m in tc.SIDECAR_MODES))):
continue
probs, froms = [], []
for sub in ("environment", "tests"):
diff --git a/ci_checks/check-task-fields.sh b/ci_checks/check-task-fields.sh
index dcd9c63c8d..ae30a53042 100755
--- a/ci_checks/check-task-fields.sh
+++ b/ci_checks/check-task-fields.sh
@@ -1,6 +1,6 @@
#!/usr/bin/env bash
# task.toml: name matches folder, required metadata (docs/PROTOCOL.md §2.3, docs/TAXONOMY.md), deliverable dir, agent timeout cap (with per-benchmark exceptions).
-t="$1"; python3 - "$t" <<'PY'
+t="$1"; python3 - "$t" "$(cd "$(dirname "$0")/.." && pwd)" <<'PY'
import sys, tomllib, pathlib
t = pathlib.Path(sys.argv[1]); d = tomllib.load(open(t / "task.toml", "rb")); rc = 0
def fail(m):
@@ -10,12 +10,17 @@ m = d.get("metadata", {})
for k in ["author_name", "task_type", "simulator", "robot", "family", "expert_time_estimate_hours"]:
if k not in m: fail(f"[metadata].{k} missing")
if m.get("task_type") not in ("trajectory", "policy"): fail("task_type must be trajectory|policy")
-# Protocol v1.1.1 (§2.2, §2.3, D12): the directory ends in its mode (-privileged, -standard, -standard-realtime) and
-# `mode` says the same; no `tier`. The suffix is the whole mode name, longest first: -standard-realtime is not -realtime.
-MODES = ("standard-realtime", "privileged", "standard")
+# Protocol v1.1.1 (§2.2, §2.3, D12): the directory ends in its mode (-privileged, -standard, -standard-realtime, or a mode
+# added to the registry, §2.16) and `mode` says the same; no `tier`. The suffix is the whole mode name, longest first:
+# -standard-realtime is not -realtime. The modes: RCB_MODES_TABLE (run_all.sh) or scripts/task_conventions.py MODE_REGISTRY.
+import os
+table = os.environ.get("RCB_MODES_TABLE", "")
+if not table:
+ sys.path.insert(0, str(pathlib.Path(sys.argv[2]) / "scripts")); import task_conventions as tc; table = tc.modes_table()
+MODES = tuple(l.split()[0] for l in table.splitlines() if l.strip()) # longest first
suffix = next((x for x in MODES if t.name.endswith("-" + x)), t.name.rsplit("-", 1)[-1])
-if suffix not in MODES: fail("directory name must end in -privileged, -standard or -standard-realtime")
-if m.get("mode") not in MODES: fail("[metadata].mode must be privileged|standard|standard-realtime")
+if suffix not in MODES: fail("directory name must end in " + ", ".join("-" + x for x in MODES))
+if m.get("mode") not in MODES: fail("[metadata].mode must be " + "|".join(MODES))
elif m["mode"] != suffix: fail(f"[metadata].mode = {m['mode']!r} but the directory ends in -{suffix}")
if "tier" in m: fail("[metadata].tier was removed in protocol v1.0 (use mode)")
if "/app/output/" not in d.get("artifacts", []): fail('top-level artifacts must include "/app/output/"')
diff --git a/ci_checks/check-tooling-identity.sh b/ci_checks/check-tooling-identity.sh
index 7c569a7402..c93fd087cc 100755
--- a/ci_checks/check-tooling-identity.sh
+++ b/ci_checks/check-tooling-identity.sh
@@ -35,7 +35,9 @@ set -uo pipefail
root="$(cd "$(dirname "$0")/.." && pwd)"
python3 - "${1:-$root/tasks}" "$root/scripts/eai_standard" <<'PY'
import collections, hashlib, os, re, sys, tomllib
+from pathlib import Path
tasks, src = sys.argv[1], sys.argv[2]
+sys.path.insert(0, os.path.dirname(os.path.abspath(src))); import task_conventions as tc # the mode registry (§2.16)
MODULE = {"behavior": "rcb_behavior.py", "robolab": "rcb_robolab.py", "robowits": "rcb_robowits.py", "maniguard": "rcb_maniguard.py",
"molmospaces": "rcb_molmospaces.py", "myosuite": "rcb_myosuite.py",
"dexjoco": "rcb_dexjoco.py", "roboarc": "rcb_topogym.py", "bigym2": "rcb_bigym2.py", "yamlab": "rcb_yamlab.py", "robodojo": "rcb_robodojo.py",
@@ -130,8 +132,7 @@ for t in sorted(os.listdir(tasks)):
bad.append(f"{t}: the sim service needs `restart: on-failure:N` with N >= 3 (crash recovery)")
if not re.search(r"^\s*command:\s*[\"']?sh /srv/run\.sh[\"']?\s*$", blk, re.M):
bad.append(f"{t}: the sim service must run `sh /srv/run.sh` (it exits non-zero when the service crashes)")
- mode = ("standard-realtime" if t.endswith("-standard-realtime") else "standard" if t.endswith("-standard")
- else "privileged")
+ mode = tc.mode_of(Path(t))
own_fam, module = next((v for k, v in FAMILY.items() if t.startswith(k)), (None, None))
fam = "dextoolbench" if os.path.isfile(os.path.join(tasks, t, "environment/sim/server_mj.py")) else own_fam
if fam:
diff --git a/ci_checks/check-unit-tests.sh b/ci_checks/check-unit-tests.sh
index f13e0deee2..943e985273 100755
--- a/ci_checks/check-unit-tests.sh
+++ b/ci_checks/check-unit-tests.sh
@@ -43,6 +43,7 @@ if python3 -c "import numpy, requests" 2>/dev/null; then
robolab/tests/test_tooling_units.py robotwin/tests/test_gripper_state.py robotwin/tests/test_standard_service.py \
robopaint_strict/tests/test_server_requests.py roboarc/tests/test_server_requests.py \
mujoco-playground/tests/test_standard_2_1_servers.py dexjoco/tests/test_standard_server.py \
+ dexjoco/tests/test_grade_privileged.py \
myosuite/tests/test_standard_server.py robowits/tests/test_server_frames.py \
robodojo/tests/test_privileged_grading.py; do
run 900 "$t"
diff --git a/ci_checks/check-verifier-isolation.sh b/ci_checks/check-verifier-isolation.sh
index 564990b9f3..00863f812f 100755
--- a/ci_checks/check-verifier-isolation.sh
+++ b/ci_checks/check-verifier-isolation.sh
@@ -229,7 +229,14 @@ OWN_NETWORK = ("network_mode", "networks")
# Files of the task or the host handed to a container by another road than `volumes:`; `build` replaces Harbor's own
# build of main (context: environment/) or a sidecar's pinned image.
HANDS_FILES = ("secrets", "configs", "env_file", "build")
-mode = next((m for m in ("standard-realtime", "privileged", "standard") if t.name.endswith("-" + m)), "")
+# the mode registry (PROTOCOL §2.16): RCB_MODES_TABLE from run_all.sh, "<mode> <sidecar 0|1> <twin>" longest first
+import os
+_table = os.environ.get("RCB_MODES_TABLE", "")
+if not _table:
+ sys.path.insert(0, str(root / "scripts")); import task_conventions as _tc; _table = _tc.modes_table()
+_modes = [l.split() for l in _table.splitlines() if l.strip()]
+mode = next((m for m, _s, _t in _modes if t.name.endswith("-" + m)), "")
+sidecar_mode = any(m == mode and s == "1" for m, s, _t in _modes)
if env_c.is_file():
services, bad = compose_services(env_c)
for b in bad: fail(f"environment/docker-compose.yaml: {b} (ci_checks/check-verifier-isolation.sh reads no YAML: keep to the block form)")
@@ -254,7 +261,7 @@ if env_c.is_file():
continue
if svc == "main" and src.endswith("_trusted"): fail(f"the agent container mounts trusted volume {src}")
if svc != "main" and not src.endswith("_trusted"): fail(f"sidecar {svc} mounts {src}, which is not a *_trusted volume")
- if mode in ("standard", "standard-realtime"):
+ if sidecar_mode:
# the verifier's evidence comes from a sidecar (a channel the agent cannot write), never from the agent's container
sidecars = [s for s in services if s not in ("main", "rcb_monitor")]
for a in d.get("artifacts") or []:
diff --git a/ci_checks/run_all.sh b/ci_checks/run_all.sh
index 8a8c91ae4a..4850ec51e5 100755
--- a/ci_checks/run_all.sh
+++ b/ci_checks/run_all.sh
@@ -6,6 +6,9 @@ set -uo pipefail
# the umask the repository is written with; the checkout itself must be made with umask 022 too.
umask 022
root="$(cd "$(dirname "$0")/.." && pwd)"; rc=0
+# the mode registry (PROTOCOL §2.16, scripts/task_conventions.py MODE_REGISTRY), read once for the per-task shell checks
+RCB_MODES_TABLE="$(python3 "$root/scripts/task_conventions.py" modes)" || { echo "FAIL scripts/task_conventions.py modes"; exit 1; }
+export RCB_MODES_TABLE
tasks=("$@"); [ ${#tasks[@]} -eq 0 ] && tasks=("$root"/tasks/*/)
for t in "${tasks[@]}"; do
t="${t%/}"; [ -f "$t/task.toml" ] || continue
@@ -31,6 +34,10 @@ if [ $# -eq 0 ]; then python3 "$root/ci_checks/check-resources.py" || rc=1; else
# instruction.md, the README row and the standard modes' sim service all say its number, and nothing else states the limit
if [ $# -eq 0 ]; then python3 "$root/ci_checks/check-cutoffs.py" || rc=1; else python3 "$root/ci_checks/check-cutoffs.py" "$@" || rc=1; fi
if [ $# -eq 0 ]; then python3 "$root/ci_checks/check-seeds.py" || rc=1; else python3 "$root/ci_checks/check-seeds.py" "$@" || rc=1; fi
+# PROTOCOL §2.14: scene dynamics, `[metadata] dynamics` from the one table (scripts/task_dynamics.toml)
+if [ $# -eq 0 ]; then python3 "$root/ci_checks/check-dynamics.py" || rc=1; else python3 "$root/ci_checks/check-dynamics.py" "$@" || rc=1; fi
+# PROTOCOL §4b: a standard-realtime task states its slowdown S ([metadata] slowdown, the instruction) and allows resets
+if [ $# -eq 0 ]; then python3 "$root/ci_checks/check-realtime.py" || rc=1; else python3 "$root/ci_checks/check-realtime.py" "$@" || rc=1; fi
# The task protocol (docs/PROTOCOL.md): the protocol version per mode, oracle fields, the image registry, no results/, no personal paths
if [ $# -eq 0 ]; then python3 "$root/ci_checks/check-protocol.py" || rc=1; else python3 "$root/ci_checks/check-protocol.py" "$@" || rc=1; fi
# PROTOCOL §2.10 (v1.2): the resource monitor sidecar and the verifier step, wired by scripts/rcb_monitor/inject.py
diff --git a/docs/CUTOFFS.md b/docs/CUTOFFS.md
index f436394b38..5471133a03 100644
--- a/docs/CUTOFFS.md
+++ b/docs/CUTOFFS.md
@@ -23,7 +23,7 @@ Every task gives the agent a fixed wall-clock time, its **cutoff**: Harbor's `[a
| MetaWorld+ | 1800 s (30 min) | 1800 s (30 min) | — | 50 / 50 / — |
| MolmoSpaces | 3600 s (60 min) | 3600 s (60 min) | — | 8 / 8 / — |
| MuJoCo Playground | 1800 s (30 min) | 1800 s (30 min) | — | 13 / 13 / — |
-| MyoSuite | 3600 s (60 min) | 3600 s (60 min) | — | 22 / 22 / — |
+| MyoSuite | 3600 s (60 min) | 3600 s (60 min) | 3600 s (60 min) | 22 / 22 / 1 |
| Panda Dynamic Scenes | 3600 s (60 min) | 3600 s (60 min) | — | 25 / 25 / — |
| RoboARC | 10800 s (180 min) | 10800 s (180 min) | — | 94 / 94 / — |
| RoboCasa | 3600 s (60 min) | 3600 s (60 min) | — | 29 / 29 / — |
@@ -39,7 +39,7 @@ Every task gives the agent a fixed wall-clock time, its **cutoff**: Harbor's `[a
| VLABench | 1800 s (30 min) | 1800 s (30 min) | — | 36 / 36 / — |
| YAMLab | 3600 s (60 min) | 3600 s (60 min) | — | 10 / 10 / — |
-1810 task directories. P / S / RT: privileged, standard, standard-realtime. A cutoff is a whole number of seconds, at most 14400 (PROTOCOL D8; RoboTangle's exception: 21600).
+1811 task directories. P / S / RT: privileged, standard, standard-realtime. A cutoff is a whole number of seconds, at most 14400 (PROTOCOL D8; RoboTangle's exception: 21600).
## Exceptions (per family, per task)
@@ -116,3 +116,4 @@ The table started on 2026-10-11 with the number every task had on that day; each
| 2026-10-13 | Panda Dynamic Scenes, privileged, standard | (none) -> 3600 s | proposal #177 (D8 default) | the benchmark's first tasks: the protocol's default; a scene builds in seconds and simulates at 60-150 control steps per second, so the hour is the agent's own work |
| 2026-10-13 | RoboTangle, privileged, standard | 7200 s -> (removed) | owner | removed: the benchmark is renamed RoboARC (next entry) and its Genesis tasks are deleted; RoboARC's MuJoCo release replaces them, and results recorded under this row are not carried over |
| 2026-10-13 | RoboARC, privileged, standard | (none) -> 10800 s | owner | new row: RoboARC, the MuJoCo release that replaces RoboTangle's Genesis tasks (prepared as RoboTangle-New, then RoboTangle; their entries, 0 -> 7200 on 2026-10-12 and 7200 -> 10800 on 2026-10-13, are folded into this one): 3 h wall clock for every family, within the default cap |
+| 2026-10-14 | MyoSuite, standard-realtime | (none) -> 3600 s | owner (board 0112: a realtime twin's cutoff equals its standard twin's) | the first standard-realtime task (table_tennis): the standard twin's number |
diff --git a/docs/LEAKAGE.md b/docs/LEAKAGE.md
index 29c9aaad1f..a0806139f7 100644
--- a/docs/LEAKAGE.md
+++ b/docs/LEAKAGE.md
@@ -20,7 +20,7 @@ a reward function, a success predicate, a trained policy, a demonstration set.
| Earlier Docker layers | a layer is an immutable tarball; a later `rm` does not unwrite it | download and remove in the **same `RUN`**; verify with `docker save` |
| `.git` | `git show` reconstructs every file ever committed | delete it in that same `RUN` |
| The network | upstream is public and pinned, one `curl` away | `[agent] network_mode = "allowlist"` of the model-API hosts (`scripts/set_agent_egress.py`) in **every task** since 2026-10-10 (PROTOCOL §2.11, [CI]); since 2026-10-11 the list also has the package registries (PyPI, the Debian / Ubuntu apt archives), which are not where solutions are published, with exceptions: some PyPI distributions ship a benchmark's scripted experts, trained policies or task sources (the table in PROTOCOL §2.11; nothing blocks them, a trial that installs one is looked at). Until then six benchmarks were open, and trials fetched upstream task sources with their reference solution, a recorded demonstration and robot models from their own shell (board 0072); a pointer inside the image (a `source_url`) is what led one of them there |
-| Instance metadata and upstream docs | a frozen instance that records where it came from (a `source_url`, a demonstration file, the expert's module) hands the agent the address of its solution: RoboTangle's `reference_scene.json` named the upstream demonstration, YAMLab's `instance.json` the recording its oracle replays, RoboLab's image shipped a recorded successful episode of one task; upstream READMEs name the public demonstration datasets and checkpoints | keep provenance in `scripts/<bench>/` and the README, never in a file under `environment/` the agent reads ([CI] `check-agent-pointers.py`); grep a new image for `https?://`, `*.hdf5` and recorded episodes before publishing it; the allowlist covers what an upstream checkout's documentation still names |
+| Instance metadata and upstream docs | a frozen instance that records where it came from (a `source_url`, a demonstration file, the expert's module) hands the agent the address of its solution: RoboTangle's `reference_scene.json` named the upstream demonstration, YAMLab's `instance.json` the recording its oracle replays, DexJoCo's `instance.npz` (`meta`) the demonstration whose actions are the reference, RoboLab's image shipped a recorded successful episode of one task; upstream READMEs name the public demonstration datasets and checkpoints | keep provenance in `scripts/<bench>/` and the README, never in a file under `environment/` the agent reads ([CI] `check-agent-pointers.py`); grep a new image for `https?://`, `*.hdf5` and recorded episodes before publishing it; the allowlist covers what an upstream checkout's documentation still names |
| The model API | tools that run at the provider, reached through an allowed model-API host: Codex's `web_search` and Claude Code's `WebSearch`; under a ChatGPT login also the account's connectors and apps (238 tools on 2026-10-10: web open and search, GitHub, mail, drives), image generation and plugins | **closed**: `scripts/run_agent_batch.sh` passes Codex `--ak web_search=disabled` and Claude Code `--ak disallowed_tools=WebSearch,WebFetch` unless `WEB_SEARCH=allowed` (since 2026-10-02; before that it was open unless a run's own agent config turned it off, audit below); the runner starts Codex without connectors, apps, plugins and image generation (the same switch: those features set to false in the run's `config.toml` by `scripts/codex_tools_off.py`; 2026-10-10, after 21 trials had called connectors: Codex 0.160 calls them from its `exec` tool as `tools.mcp__codex_apps__*`, so grep the transcript for that name); and with the usage gateway on, the gateway is the container's only egress and forwards model-API calls only (PROTOCOL §2.11) |
| The model-API key | Harbor's native agents put it in the agent's container. Codex: `OPENAI_API_KEY` in its environment and in `/tmp/codex-secrets/auth.json` (linked from `$CODEX_HOME/auth.json`), removed only after the run; Claude Code: `ANTHROPIC_API_KEY` (or its OAuth token) in its environment. With the model API allowed, the agent can spend it on any model, a search-enabled one included, which undoes the row above | **narrowed 2026-10-10**: with the usage gateway on, the container reaches its provider only through the gateway's model-API paths (no connector, file or conversation endpoint, whatever the agent runs with the key or the login), and a model request is forwarded only if every tool it declares is one the agent runs itself (a web search, image generation or remote MCP tool, or an unreadable body, is refused). Still open: the key can be spent on another model of the same provider (a gateway that holds the key is ruled out by D11): one key per batch with a spending cap (`docs/RUNNING.md`), and read the transcript for direct API calls |
| The model's memory | the benchmark's source is public, so it is in pretraining data: Opus 5.5 recalled H1's 19 actuator ranges exactly and the sit reward's seat-height band, with no tool | nothing in the harness removes it; report it with the result |
diff --git a/docs/PROTOCOL.md b/docs/PROTOCOL.md
index f409fdd285..1c168aca67 100644
--- a/docs/PROTOCOL.md
+++ b/docs/PROTOCOL.md
@@ -7,7 +7,7 @@ cap above the host's CPU count is clamped by the runner); enforcement is off by
the calibration batch has confirmed the limits. Also on 2026-10-10 and 2026-10-11 (owner decisions, board 0072): every task's
agent egress is the model APIs and the package registries, every instruction states the same honour rule about looking
solutions up (§2.11), and the usage gateway forwards model-API calls only (D11, §5). On 2026-10-11 (owner): §4.8's exception for
-physics that does not replay bit for bit (such a task is standard only and graded by the live episode; RoboDojo's six). Also on 2026-10-11 (owner, board 0080): the agent's time cutoffs live in one table, and every instruction ends with its cutoff and the request not to cheat in one format (§2.12; 2026-10-12: at the end of the instruction, and the table takes exceptions per family and per task). On 2026-10-12 (owner): §3's and §4.8's exception for grading by one replay instead of two (RoboDojo). Earlier versions, whose rules are all part of this text: v1.0
+physics that does not replay bit for bit (such a task is standard only and graded by the live episode; RoboDojo's six). Also on 2026-10-11 (owner, board 0080): the agent's time cutoffs live in one table, and every instruction ends with its cutoff and the request not to cheat in one format (§2.12; 2026-10-12: at the end of the instruction, and the table takes exceptions per family and per task). On 2026-10-14 (owner, boards 0106 / 0112): scene dynamics (§2.15), and every standard-realtime task states its slowdown S and allows resets, 10 by default (§4b). On 2026-10-12 (owner): §3's and §4.8's exception for grading by one replay instead of two (RoboDojo). On 2026-10-14 (owner, board 0107 E6): a benchmark ships any subset of the three modes, a general rule that replaces the per-benchmark exceptions for a missing twin (§2.15; where privileged and standard both exist, their cutoffs stay equal), and new modes may be added by pull request, through one registry of modes (§2.16). Earlier versions, whose rules are all part of this text: v1.0
(2026-09-30: the two modes and their names), v1.1 (2026-10-08: the standard-realtime mode, §4b), v1.1.4 (compute
resources, §2.9), v1.1.5 (resource limits, §2.10).
The tasks in the repository follow it except for the open items in §7. The wire format of standard and
@@ -51,18 +51,18 @@ Unchanged: the task name prefix `robot-coding-bench/`, the client and tool file
| # | Topic | Decision |
|---|---|---|
-| D1 | Standard-mode protocol | Only `eai-standard/2.x`. Additive extensions are accepted through a protocol-change PR (2.2 is accepted once its services implement crash recovery). Reset is only an optional capability (`budget.reset = N`, default 0; results with N > 0 are reported separately). Documented per-benchmark exceptions (§4.5): HumanoidBench standard defaults to N = 50 (owner, 2026-10-07), MyoSuite standard to N = 10 (owner, 2026-10-09); BiGym2 and KinDER standard to N = 10 (owner, 2026-10-11); DexToolBench standard to N = 20 (owner, 2026-10-13). The standard-realtime mode's `eai-standard/2.3` (§4b) is 2.2 plus the realtime capability. `eai-standard/2.4` (issue #89) is 2.3 plus the operations capability; standard tasks may use it. No step caps. |
+| D1 | Standard-mode protocol | Only `eai-standard/2.x`. Additive extensions are accepted through a protocol-change PR (2.2 is accepted once its services implement crash recovery). Reset is only an optional capability (`budget.reset = N`, default 0; results with N > 0 are reported separately). Documented per-benchmark exceptions (§4.5): HumanoidBench standard defaults to N = 50 (owner, 2026-10-07), MyoSuite standard to N = 10 (owner, 2026-10-09); BiGym2 and KinDER standard to N = 10 (owner, 2026-10-11); DexToolBench standard to N = 20 (owner, 2026-10-13); MuJoCo Playground's quadruped standard tasks to N = 10 (owner, 2026-10-14). The standard-realtime mode's `eai-standard/2.3` (§4b) is 2.2 plus the realtime capability. `eai-standard/2.4` (issue #89) is 2.3 plus the operations capability; standard tasks may use it. No step caps. |
| D2 | Tasks without an oracle | Allowed, with mandatory human review (§2.7). |
| D3 | Results | Not in the repository: agent results (per-task READMEs, videos, timelines, batch tables) are published on the results site and the Hugging Face dataset; a task directory has no `results/`. |
| D4 | Run id | `<harness>-<model>-<effort>`, model token with underscores (`codex-gpt6_luna-xhigh`); provider, tier and settings go in metadata. |
| D5 | Images | One registry (EAI organization), images named `eai-<name>` and versioned in `images/<name>/VERSION`; every image that may be redistributed is public, the rest are `BUILD_ONLY`. One image per **benchmark**, not per simulator: benchmarks on the same simulator may need different versions of it and of its dependencies (e.g. RoboCasa, RoboCasa365 and RoboCasa-GR1 on robosuite), so each may have its own image; an image is shared only where the dependency set is identical. |
| D6 | PR granularity | One benchmark per PR; changes to shared files go in a separate PR reviewed by a maintainer. |
| D7 | World-frame information in instructions | Decided per benchmark, consistent within it, documented in the benchmark's doc. |
-| D8 | Agent time budget (the cutoff) | 3600 s by default, at most 14400 s; longer budgets are justified, uniform within a benchmark unless the cutoff table gives a family or a task an exception (§2.12), and results are grouped by budget. Documented per-benchmark exceptions to the cap (§2.3): RoboTangle 21600 s (owner, 2026-10-08; its tasks' cutoff is 7200 s since 2026-10-12). The numbers live in one table, `scripts/task_cutoffs.toml`, per benchmark and mode with optional exceptions per family and per task, and every instruction states its own in one format, in its last section (§2.12; owner, 2026-10-11 and 2026-10-12). |
+| D8 | Agent time budget (the cutoff) | 3600 s by default, at most 14400 s; longer budgets are justified, uniform within a benchmark unless the cutoff table gives a family or a task an exception (§2.12), always the same in a task's privileged and standard versions (§2.15), and results are grouped by budget. Documented per-benchmark exceptions to the cap (§2.3): RoboTangle 21600 s (owner, 2026-10-08; its tasks' cutoff is 7200 s since 2026-10-12). The numbers live in one table, `scripts/task_cutoffs.toml`, per benchmark and mode with optional exceptions per family and per task, and every instruction states its own in one format, in its last section (§2.12; owner, 2026-10-11 and 2026-10-12). |
| D9 | Shared code | Shared sources are copied into tasks by the generator; nothing at run time imports repository code outside the task directory. |
| D10 | Token budget | Per trial: **100M input tokens, 10M output tokens** by default; exceeding either fails the trial. |
| D11 | Harness | Official results use Harbor's native agents only; no agent adapters, and no gateway that holds credentials or changes requests. The usage gateway (`scripts/usage_gateway/`) is allowed: it forwards model-API requests unchanged, records token usage, and refuses what a trial may not have: a request past the token budget (§5) and, since 2026-10-10, every request that is not a model-API call (§2.11). No tool that runs outside the trial's container is offered to the model (§2.11): official runs turn the harness's web search off (Codex `web_search = "disabled"`, Claude Code `WebSearch` and `WebFetch` disallowed; `scripts/run_agent_batch.sh` `WEB_SEARCH=disabled`, the default) and start the harness without account connectors, apps, plugins, image generation and cloud agents (Codex: with the same switch the runner sets those features to false in the run's `config.toml`, `scripts/codex_tools_off.py`). These run at the model provider, so no egress allowlist stops them (2026-10-02: Codex with a ChatGPT login searched GitHub by default; 2026-10-10: the same login handed the model 238 account connector tools, and 21 trials called them) |
-| D12 | Mode naming | `mode ∈ {privileged, standard, standard-realtime}` (the third mode since v1.1, #86; named `standard-realtime` since v1.1.1, owner 2026-10-08, #97: it is standard mode's robot as a service on a clock that does not stop); the `tier` field is removed; every task directory ends in its mode: `-privileged`, `-standard` or `-standard-realtime`. |
+| D12 | Mode naming | `mode ∈ {privileged, standard, standard-realtime}` (the third mode since v1.1, #86; named `standard-realtime` since v1.1.1, owner 2026-10-08, #97: it is standard mode's robot as a service on a clock that does not stop); the `tier` field is removed; every task directory ends in its mode: `-privileged`, `-standard` or `-standard-realtime`. More modes may be added by pull request, through the registry of §2.16 (owner, 2026-10-14). |
## 1. Checklist (every task)
@@ -116,8 +116,8 @@ Unchanged: the task name prefix `robot-coding-bench/`, the client and tool file
| Item | Level | Basis |
|---|---|---|
| Privileged tasks: `<base>-privileged`, where `<base>` is `<benchmark>-<family>-i<NN>` (`metaworldplus-<env>-i00`, `kinder-<family>-i00`, `humanoidbench-<category>-<task>-i00`) or an upstream name kept as is (`robotwin_2_0_<task>`, `dextoolbench_<cat>_<obj>_<task>`, `mujoco_playground_<env>`) | MUST | suffix [CI] |
-| Standard tasks: `<base>-standard`, sharing the privileged twin's frozen instance and `instance_seed`, and its default and allowed seeds (§2.13) (a standard-only task under the §4.8 exception for non-reproducible physics has no twin: its sidecar carries the frozen instance) | MUST | suffix [CI]; shared instance [current] |
-| Standard-realtime tasks: `<base>-standard-realtime`, sharing the privileged twin's frozen instance and `instance_seed`, and its default and allowed seeds (§2.13); its twin is `<base>-privileged` (a benchmark that has standard-realtime tasks usually has `<base>-standard` too) | MAY (the mode) / MUST (the name) | suffix and instance [CI] [new] (v1.1) |
+| Standard tasks: `<base>-standard`, sharing the privileged twin's frozen instance and `instance_seed`, and its default and allowed seeds (§2.13), where the task has a twin (a task without one, §2.15, carries its own frozen instance in its sidecar) | MUST | suffix [CI]; shared instance [current] |
+| Standard-realtime tasks: `<base>-standard-realtime`, sharing the privileged twin's frozen instance and `instance_seed`, and its default and allowed seeds (§2.13), where the task has a twin; its twin is `<base>-privileged` (which modes a benchmark ships: §2.15) | MAY (the mode) / MUST (the name) | suffix and instance [CI] [new] (v1.1) |
| Versions or durations in task names (`-v3-`, `-60min-`) | MUST NOT | [new] — versions are dataset digests |
### 2.3 `task.toml`
@@ -132,6 +132,7 @@ Unchanged: the task name prefix `robot-coding-bench/`, the client and tool file
| `simulator`, `robot` from the vocabularies in [`TAXONOMY.md`](TAXONOMY.md) (extend by PR) | MUST | [current]; CI check [new] |
| `instance_seed` (where present, equal to the default seed, §2.13); benchmark-specific fields registered in TAXONOMY | MAY | [current] |
| `[metadata.seeds]`: `default`, `allowed`, `oracle` (§2.13) | MUST | [CI] (2026-10-13) |
+| `[metadata] dynamics ∈ {static, dynamic}` (§2.14), from `scripts/task_dynamics.toml` | MUST | [CI] (2026-10-14) |
| Top-level `artifacts` includes `"/app/output/"` | MUST | [CI] |
| `[agent].timeout_sec` set, ≤ 14400 (per-benchmark exceptions below); 3600 unless justified (uniform within a benchmark); the number is the cutoff table's for the task (its benchmark and mode, unless the table gives its family or the task an exception), written by `scripts/task_conventions.py` (§2.12) | MUST / SHOULD / MUST | [CI] / [new] / [CI] |
| `[verifier].environment_mode = "separate"` | MUST | [CI] |
@@ -225,7 +226,7 @@ Harbor builds the task images locally from the task directory on every run; only
| Item | Level | Basis |
|---|---|---|
| Reference solutions only under `solution/` (mounted for the oracle agent only) | MUST | [current] |
-| A solution, or a pointer to one, anywhere the agent can reach (task image, base image, tooling, docs in the container). A pointer is anything that locates the solution of this task: a URL, a dataset or file name, a demonstration or episode id, the name of the upstream expert that produced the reference. Provenance is kept in `scripts/<bench>/` and the README, never in a file under `environment/` the agent reads | MUST NOT | [CI] for the task's own files (`ci_checks/check-agent-pointers.py`: hosts of code, data and video, `*_url` keys, dataset and demonstration file names); [current] for images, reviewed when they are built. Upstream documentation inside a simulator image that names the benchmark's public datasets or checkpoints is not removed file by file: the allowlist of §2.11 makes them unreachable. Open (2026-10-10): RoboTangle's image (docs/benchmarks/robotangle.md) |
+| A solution, or a pointer to one, anywhere the agent can reach (task image, base image, tooling, docs in the container). A pointer is anything that locates the solution of this task: a URL, a dataset or file name, a demonstration or episode id, the name of the upstream expert that produced the reference. Provenance is kept in `scripts/<bench>/` and the README, never in a file under `environment/` the agent reads | MUST NOT | [CI] for the task's own files (`ci_checks/check-agent-pointers.py`: hosts of code, data and video, `*_url` keys, dataset and demonstration file names, keys that name a demonstration or episode (`"demo": ...`, `episode_id`); text files and the string members of `.npz` archives, since 2026-10-14 (board 0105: DexJoCo's `instance.npz` named its demonstration)); [current] for images, reviewed when they are built. Upstream documentation inside a simulator image that names the benchmark's public datasets or checkpoints is not removed file by file: the allowlist of §2.11 makes them unreachable. Open (2026-10-10): RoboTangle's image (docs/benchmarks/robotangle.md) |
| Upstream experts scrubbed from images (e.g. RoboTwin `play_once`) | MUST | [current] |
| Agent egress allowlisted in `task.toml` (the model APIs and, since 2026-10-11, the package registries), in every task (§2.11) | MUST | [CI] (2026-10-10; until then a SHOULD "when a solution is publicly downloadable", and RoboTwin, MuJoCo Playground's four Aloha / Panda pairs, RoboWits, RoboLab, BEHAVIOR-1K and RoboTangle were open) |
| `nop` scores 0 | MUST | [current] |
@@ -620,7 +621,7 @@ own, seed 0 names the task's one frozen instance.
| Item | Level | Basis |
|---|---|---|
| `[metadata.seeds]` in `task.toml`: `default` (an integer); `allowed`, the seeds a trial may run from, `default` among them: a list of distinct integers (`[0, 3, 7]`), a range with both ends included (`{ from = 0, to = 100 }`; TOML's `[0, 100]` would be the two seeds 0 and 100), or `"any"` (every seed; TOML has no null); `oracle` (the seeds the reference solution under `solution/` is for, a list of allowed seeds: not empty for `oracle = "full"` or `"partial"`, empty for `"none"`); `build`, `"frozen"` (the default: every seed in `allowed` is a committed instance) or `"runtime"` (the task builds a seed's instance itself; required for a range and for `"any"`, which cannot all be committed). A generator writes what it has; `scripts/task_conventions.py` completes the rest: `default` = `instance_seed` (0 without one), `allowed = [default]`, `oracle = [default]` or `[]` | MUST | [CI] (`ci_checks/check-seeds.py`) |
-| `instance_seed`, where present, equals `default`; a standard or standard-realtime task has its privileged twin's `default`, `allowed` and `build` | MUST | [CI] |
+| `instance_seed`, where present, equals `default`; a standard or standard-realtime task has its privileged twin's `default`, `allowed` and `build` (where it has a twin, §2.15) | MUST | [CI] |
| The default seed's frozen instance is the task's instance, where §2.2 and §3 put it. With `build = "frozen"`, every other allowed seed's frozen instance is `tests/seeds/<seed>/instance.json` (or `.npz`, the same kind as the default's) of the task that holds the pair's instance (the privileged twin; a task without one: itself); `tests/seeds/` holds nothing else (a `"runtime"` task may commit some seeds' instances there too) | MUST | [CI] |
| `build = "runtime"`: given a file `seed` (the integer) next to its instance, the task's environment builds that seed's instance deterministically, and its verifier builds the same instance from its own copy of the file (never from anything the agent's container wrote) and grades on it; the same seed gives the same start hash in a fresh process. The benchmark's doc says how, and its generator's tests show it on a sample of seeds | MUST | [new] |
| Another seed's instance, or a `seeds` directory, anywhere under `environment/` (where the agent's image, or a sidecar's that the agent's container could share, is built from); in privileged mode the agent reads the instance of its own trial's seed only, as before | MUST NOT | [CI] |
@@ -633,6 +634,80 @@ own, seed 0 names the task's one frozen instance.
Until a benchmark's generator writes more, every task allows its default seed only (`allowed = [default]`): the 2026-10-13
migration changed `task.toml` only (the instances, the instructions and the READMEs are unchanged).
+### 2.14 Scene dynamics (2026-10-14, board 0106; merged with #188)
+
+The owner, 2026-10-14: "every task runs in standard mode by default; a subset, the dynamic tasks, gets a
+standard-realtime mode as well". A task is **dynamic** when something in its scene other than the robot can do work by
+a force other than the robot's own and gravity ("非机器人力和重力原因下能够做功的"): a body that starts with a velocity
+(a thrown, served, rolling or sliding object; a pendulum already swinging), an actuated or scripted body (a conveyor, a
+motor-driven rail or turntable, a scripted push, a timer that moves a part), an opponent or another agent. Everything
+else is **static**: the world waits while the agent thinks, and objects move only because the robot moves them or
+gravity pulls them. Decided with it (board 0106, owner 2026-10-14):
+
+- settling under gravity after reset, cloth / liquid / granular material the robot moves, a falling domino chain the
+ robot started: static (gravity does the work);
+- passive springs, joint stiffness and elastic material that only give back what the robot put in (a spring button, a
+ self-closing latch, a springboard, an elastic rope): static;
+- a world change applied once and instantly when the robot comes near (LIBERO-MAX's on-proximity relocations): static;
+ a change on a clock, or a mechanism that keeps moving bodies after the robot started it (a venue that launches a
+ ball at t = 2 s, a goalkeeper that moves after the kick, a timer dial that turns back by itself): dynamic;
+- visual-only or state-only change (a faucet's water and a burner's flame drawn as a site, cooking and freezing
+ states, an on/off flag): static;
+- an episode horizon or step cap is not a force: static;
+- a robot that must balance actively (a floating-base humanoid or legged robot, a hovering drone) does not make a task
+ dynamic by itself (owner, 2026-10-14: HumanoidBench needs no realtime mode by this definition).
+
+| Item | Level | Basis |
+|---|---|---|
+| `[metadata] dynamics ∈ {static, dynamic}` in `task.toml`, the value of the one table `scripts/task_dynamics.toml` for the task (its family's exception, else its benchmark's row; every entry gives its reason, `why`), written by `scripts/task_conventions.py` (`normalize`); never edited in a task, a template or a generator | MUST | [CI] (`ci_checks/check-dynamics.py`) |
+| A new benchmark adds its row (and its families' exceptions) to the table in the pull request that adds its tasks | MUST | [CI] |
+| The same family has the same value in every mode (the field describes the scene, not the mode) | MUST | [CI] (the table is keyed by family) |
+| A `dynamic` task may have a `-standard-realtime` twin (§4b); a `static` one needs none | MAY | [new] (owner, 2026-10-14) |
+
+### 2.15 Modes a benchmark ships (owner, 2026-10-14)
+
+A benchmark may ship **any subset** of the three modes: privileged only, standard only, standard-realtime only, or any
+combination, and a family within a benchmark may lack a mode its benchmark has (RoboDojo's six standard-only tasks,
+§4.8). No owner decision is needed for the choice; what a benchmark leaves out is stated in its doc. This general rule
+replaces the per-benchmark exceptions that allowed a missing twin before (RoboOlympics without privileged mode, §4b,
+owner 2026-10-08; RoboDojo's six, §4.8, owner 2026-10-11); those benchmarks are unchanged and conform.
+
+| Item | Level | Basis |
+|---|---|---|
+| The modes a benchmark ships are the modes of its row in the cutoff table (§2.12), which has a number for exactly the modes that have tasks | MUST | [CI] (`ci_checks/check-cutoffs.py`) |
+| Where both exist for a task, its privileged and standard versions have **the same cutoff**: one number in the benchmark's row, in a family's entry, and for a task and its standard twin (owner, 2026-10-12) | MUST | [CI] (`ci_checks/check-cutoffs.py`, 2026-10-14) |
+| Twins, where both modes exist for a family, share one frozen instance, success rule and seeds (§2.2, §2.13) | MUST | [CI] |
+| A standard or standard-realtime task without a privileged twin carries its own frozen instance, if it has one, in its sidecar, and its verifier's copy (where it keeps one) is byte-identical; its README's Twin row says why there is none. In a benchmark with no privileged mode that is automatic; in a benchmark that has one, the task is listed with its reason in `scripts/task_conventions.py` `NO_TWIN` (a missing twin that is neither fails CI, so a deleted or misnamed twin is caught) | MUST | [CI] (`ci_checks/check-instance-copies.sh`) |
+| A privileged task without a standard twin needs nothing more (its README's Twin row says "privileged only") | MAY | [current] |
+| Results are reported per mode (§5); a benchmark's results state which modes it ships, and a mode it does not ship is "not offered", not a failure | SHOULD | [new] |
+
+### 2.16 Adding a mode (owner, 2026-10-14)
+
+The three modes (privileged §3, standard §4, standard-realtime §4b) are not a closed list: anyone may propose a new mode
+by pull request. The modes live in one registry, `MODE_REGISTRY` in `scripts/task_conventions.py`; everything in the
+repository that parses or branches on a mode reads it (`mode_of`, `base_of`, `twin_of`, the limits' scope, the cutoff
+table's columns, `docs/CUTOFFS.md`, and the checks under `ci_checks/`, which get it once per run from
+`python3 scripts/task_conventions.py modes`). An entry gives the mode's name (= its directory suffix and `[metadata]
+mode`), its column label, its PROTOCOL section, whether the simulator runs in a trusted `sim` sidecar, whom the resource
+limits bind, the `protocol` values its tasks may declare, its twin mode (the task of the same base that holds the frozen
+instance and seeds) and the mode whose cutoff it must equal, if any.
+
+| A new-mode pull request contains | Level | Basis |
+|---|---|---|
+| A name, which is also the task-directory suffix (`<base>-<name>`) and the `[metadata] mode` value: lower case, words joined by `-`, not a suffix of another mode's name nor ending in one (`-standard-realtime` is recognised by its whole suffix, longest first), no version or duration (§2.2) | MUST | [CI] (`scripts/task_conventions.py` refuses a registry that breaks it; `ci_checks/check-task-fields.sh` checks every task's suffix and `mode`) |
+| A PROTOCOL section of its own, like §3 / §4 / §4b, defining: what the agent gets (its container, the interface, privileged information or not, the wire protocol and its version if any); what is graded (the deliverable or the service's record, the success rule, replays and determinism); the time rule (what the cutoff counts, how a service reports it, §2.12) and the reset rule (none by default, §4.5); and the failure semantics, which failures are infrastructure (`infra_error`, never the agent's 0) and which are the agent's | MUST | [new] |
+| Its `MODE_REGISTRY` entry, its `CUTOFF_NOTE` (the second paragraph of `## Time limit` for the mode; the module refuses a mode without one) and, where it has a protocol, its reference implementation (as §6 requires of a protocol change) | MUST | [CI] |
+| CI support: the existing checks pass for its tasks through the registry; a rule specific to the mode is a new check (or a branch keyed on a registry field, never on a hard-coded mode name list), with a `ci_checks/check-ci-selftest.py` case that breaks a real task of the mode and must fail | MUST | [CI] |
+| At least one benchmark using it: its tasks in the benchmark's own pull request (shared files first, as for a new benchmark), with the oracle and `nop` evidence of §2.7 in the new mode | MUST | [new] |
+| `docs/TAXONOMY.md` (`mode`, and `protocol` if any), `docs/REVIEWING.md` if reviewers check something new, and the results: the new mode is reported separately from the others (§5) | MUST | [new] |
+| The runner side (batch_eval_live and the results site) is told the new suffix in the same change; it reads the registry or follows it | SHOULD | [new] |
+
+**Cutoffs of a new mode.** A benchmark that ships the mode has a number for it in its row of the cutoff table (§2.12),
+like any mode, with its `[[history]]` entry; the column appears in `docs/CUTOFFS.md` from the registry. Whether it must
+equal another mode's number is the registry's `same_cutoff_as`, decided in the mode's pull request: standard has
+`same_cutoff_as = "privileged"` (owner, 2026-10-12, CI-enforced), standard-realtime has none (today its numbers equal the
+others everywhere). A new mode is free unless its pull request says otherwise.
+
## 3. Privileged mode (`mode = "privileged"`, `task_type = "trajectory"`)
| Item | Level | Basis |
@@ -644,7 +719,7 @@ migration changed `task.toml` only (the instances, the instructions and the READ
| Exception, one replay: a benchmark whose tasks are measured to replay bit for bit may be graded by ONE fresh-process replay, to save grading time (`success` = that replay's success). Nothing is compared then, and `reward.json` says so: `deterministic = 1` with `determinism_checked = 0` and `replays = 1`. The replay must start from the frozen state, shown by the recorded initial-state hash: a rebuilt instance that hashes differently gives no verdict (`instance_mismatch` with `infra_error`). The benchmark's doc states the rule, the measurement behind it and what it gives up (a trajectory whose outcome differs between fresh processes is no longer caught). RoboDojo, its 48 privileged tasks ([`benchmarks/robodojo.md`](benchmarks/robodojo.md)) | MAY, by owner decision | [new] (owner, 2026-10-12) |
| Success judged on the state at the end of the replay ("success at any step" needs a stated reason) | SHOULD | [current] |
| The frozen instance copied into both `environment/` and `tests/` with its expected initial-state hash, checked on rebuild | SHOULD | [current] |
-| A hash check before moving to new hardware. Reference hardware is an RTX 4090: `place_a2b_left`, `place_a2b_right`, `place_object_scale` and `place_object_stand` rebuild to a different state on an L40S, and grading there reports `instance_mismatch` with `infra_error` (privileged: also when no trajectory was handed in; standard: the service's `reason`) | SHOULD | [current] |
+| A hash check before moving to new hardware; a host that rebuilds a different state is graded `instance_mismatch` with `infra_error` (privileged: also when no trajectory was handed in; standard: the service's `reason`). A scene build must not depend on the host: no unsorted directory listing, no wall clock. RoboTwin's `place_a2b_left`, `place_a2b_right`, `place_object_scale` and `place_object_stand` drew their object from an unsorted `glob` (directory order) and never rebuilt off the reference workstation; since 2026-10-14 the tooling sorts it and their instances are re-frozen (board 0113) | SHOULD | [current] |
| Unlimited resets, privileged state, planning primitives, self-replay | MAY | [current] |
| Tool API shape | task-defined | [current] (recommended: `Session` with `reset` / `replay` / `status` / `save_trajectory`) |
| Self-reported cost (resets, simulation steps) in `agent_*` fields | SHOULD | [current] |
@@ -660,7 +735,7 @@ Wire details (endpoints, RCB1 frames, `meta`, error codes, crash recovery) are i
|---|---|---|
| eai-standard/2.1 | 2026-09-30 | the base (formerly `rcb-limited/2.1`): no reset; `/submit` action chunks, `/observe`; binary RCB1 responses; 82-column proprioception, optional depth noise; crash recovery. Served by the standard tasks of RoboTwin, DexToolBench, MuJoCo Playground (manipulation), KinDER, HumanoidBench, MetaWorld+, VLABench, the three RoboCasa benchmarks, LIBERO-MAX, BiGym 2.0 and RoboFunOlympics |
| eai-standard/2.2 | 2026-10-01 | additive extension (§6, [`STANDARD_MODE.md`](STANDARD_MODE.md) Part 2) with crash recovery; served by BEHAVIOR-1K, RoboLab, RoboWits, RoboPaint, MolmoSpaces, MyoSuite, DexJoCo, YAMLab, RoboDojo, LabUtopia, AutoBio, ManiGuard, Panda Dynamic Scenes and MuJoCo Playground's quadruped tasks |
-| eai-standard/2.3 | 2026-10-08 | 2.2 + the realtime capability (§4b; [`STANDARD_MODE.md`](STANDARD_MODE.md) Part 3, "Standard-realtime mode"); every 2.2 addition stays a MAY. Served by standard-realtime tasks only (HumanoidBench, RoboOlympics, BiGym 2.0); standard tasks stay on 2.1 / 2.2. Issue #86 |
+| eai-standard/2.3 | 2026-10-08 | 2.2 + the realtime capability (§4b; [`STANDARD_MODE.md`](STANDARD_MODE.md) Part 3, "Standard-realtime mode"); every 2.2 addition stays a MAY. Served by standard-realtime tasks only (HumanoidBench, RoboOlympics, BiGym 2.0, MyoSuite); standard tasks stay on 2.1 / 2.2. Issue #86 |
| eai-standard/2.4 | 2026-10-08 | 2.3 + the operations capability ([`STANDARD_MODE.md`](STANDARD_MODE.md) Part 4): operation ids and request deadlines (408 / 409), `GET /operations/<id>`, a service-owned call watchdog, a physics-tick ledger, fsynced journals; every earlier addition, the realtime capability included, stays a MAY. Served by standard tasks (RoboTangle); reference implementation `scripts/eai_standard/v2_4/`. Issue #89 |
Standard services report `eai-standard/2.1`, `eai-standard/2.2` or `eai-standard/2.4`, standard-realtime services `eai-standard/2.3`; the shipped clients
@@ -754,7 +829,7 @@ results must state the reset budget and are compared only with results under the
`budget.reset = 50`, HumanoidBench's value for the same reason. A sport is attempts: a jump, a throw or a kick is over
in a second or two and ends at the first foul or fall, and the agent has no simulator to practise in. `RCB_RESET_BUDGET=0`
still runs one attempt ([`benchmarks/robofunolympics.md`](benchmarks/robofunolympics.md)).
-- **MyoSuite** (owner, 2026-10-09; standard tasks): `budget.reset = 10`. MyoSuite's own termination ends most episodes
+- **MyoSuite** (owner, 2026-10-09; standard tasks; its standard-realtime twin has the mode's default, 10, too): `budget.reset = 10`. MyoSuite's own termination ends most episodes
within a second or two of simulated time (a dropped object, an object off the reference motion), and the agent has no
simulator to practise in, so a single episode is one blind attempt at muscle-level control. `RCB_RESET_BUDGET=0` still
runs one episode; MyoSuite standard results recorded before 2026-10-09 ran with 0 (and saw the wrong `joint_effort`)
@@ -778,6 +853,13 @@ results must state the reset budget and are compared only with results under the
(`images/mujoco/limited/server_mj.py`) implements the capability as above; its other tasks (MuJoCo Playground ALOHA,
Panda) keep 0. `RCB_RESET_BUDGET=0` still runs one episode; DexToolBench standard results recorded before this change
ran with 0 and are reported separately ([`benchmarks/dextoolbench.md`](benchmarks/dextoolbench.md)).
+- **MuJoCo Playground, quadrupeds** (owner, 2026-10-14, board 0113; the nine `mujoco_playground_{go1,spot,barkour}_*`
+ standard tasks): `budget.reset = 10`. A fall, a failed turn or a wrong first chunk used to end the trial. A reset
+ restores the frozen start (`rcb_quadruped.Session.reset()`: the same state hash and proprioception). The quadrupeds'
+ 2.2 service (`scripts/mujoco-playground/shared/standard/server.py`) implements the capability as above, as MyoSuite's
+ does (its instruction calls `POST /reset` through `Robot._post`). The ALOHA and Panda standard tasks keep 0.
+ `RCB_RESET_BUDGET=0` still runs one episode; quadruped standard results recorded before this change ran with 0 and
+ are reported separately ([`benchmarks/mujoco-playground.md`](benchmarks/mujoco-playground.md)).
### 4.6 Errors (no physics executed)
`400` invalid input (type, dimension, NaN/inf, `dt`/`seconds` invalid, chunk too long, frame limit, malformed body)
@@ -807,7 +889,7 @@ results must state the reset budget and are compared only with results under the
| Needing a custom agent adapter to finish an episode | MUST NOT | [new] |
| A fresh replay decides: `success = replay_success ∧ deterministic` | MUST | [current] |
| `∧ live_success` when a live predicate exists | SHOULD | [current] |
-| Exception, physics that does not replay bit for bit: a task measured not to (its frozen instance records it and why, e.g. `"determinism": "verdict"`; the benchmark's doc names the tasks and the measurement) is graded by the live episode the trusted service ran, `success = live_success`, read from the sim sidecar's own record of that episode (`final.json`; nothing the agent's container writes is read); the fresh replays still run and are reported (`replay_success`, `state_identical`, `deterministic`) but never decide, a replay without a result does not void the live verdict, and `reward.json` says `graded_by_live = 1`. A missing live record stays `infra_error`; an episode resumed after a crash (§4.7) is graded on its live state after the resume, so `reward.json` reports `resumes` and `resume_state_mismatch` and results state them. Such a task is standard-only: an open-loop trajectory of it cannot be graded reproducibly, so it has no privileged twin (`ci_checks/check-instance-copies.sh` and `scripts/task_conventions.py` list exactly these tasks). The verifier still grades `final.json` only, so §2.6's same verdict in a fresh process holds. RoboDojo: 6 tasks (two cloth, three liquid, `make_toast_random`) | MAY, by owner decision | [new] (owner, 2026-10-11) |
+| Exception, physics that does not replay bit for bit: a task measured not to (its frozen instance records it and why, e.g. `"determinism": "verdict"`; the benchmark's doc names the tasks and the measurement) is graded by the live episode the trusted service ran, `success = live_success`, read from the sim sidecar's own record of that episode (`final.json`; nothing the agent's container writes is read); the fresh replays still run and are reported (`replay_success`, `state_identical`, `deterministic`) but never decide, a replay without a result does not void the live verdict, and `reward.json` says `graded_by_live = 1`. A missing live record stays `infra_error`; an episode resumed after a crash (§4.7) is graded on its live state after the resume, so `reward.json` reports `resumes` and `resume_state_mismatch` and results state them. Such a task is standard-only: an open-loop trajectory of it cannot be graded reproducibly, so its family ships no privileged task (§2.15; `scripts/task_conventions.py` `NO_TWIN` lists these tasks with that reason). The verifier still grades `final.json` only, so §2.6's same verdict in a fresh process holds. RoboDojo: 6 tasks (two cloth, three liquid, `make_toast_random`) | MAY, by owner decision | [new] (owner, 2026-10-11) |
| Exception, one replay (the same decision as §3's): a benchmark whose tasks are measured to replay bit for bit may replay the recorded episode ONCE in a fresh process, and one successful replay is enough: `success = live_success ∧ replay_success` (the live predicate is then required, not SHOULD: it is what guards against a replay that succeeds where the episode did not, as the replay guards against a live success that does not repeat). `deterministic` is still measured, the replay's end state against the live episode's (= `state_identical`, `live_matches_replay`; `determinism_checked = 1`, `replays = 1`), and reported, but does not decide. A replay without a result, or one that did not start from the frozen state, stays `infra_error`. A task graded by the live episode (the exception above) records one replay instead of two. RoboDojo, its 54 standard tasks ([`benchmarks/robodojo.md`](benchmarks/robodojo.md)) | MAY, by owner decision | [new] (owner, 2026-10-12) |
| The verifier grades `final.json` only, without network access | MUST | [current] |
@@ -832,11 +914,11 @@ agent latency counts, the agent's own program may act), on the 2.x endpoints and
| Grading replays the targets the simulator **applied** at each control step (the journal), so the replay equals the live run whatever the timing; latency and jitter are recorded (`ticks.f64`; `max_lag_s`, `stalls`, `held_steps`, `late_steps` in `final.json`) and never decide the verdict | MUST | [new] |
| How well the clock was kept is evidence in `final.json`, and the grader repeats it in `reward.json` as numbers (a missing one is 0): `clock_s` (wall seconds the clock ran, over the trial), `max_lag_s` (the furthest a control step ran behind its time), `stalls` (the times the service, more than 0.5 s behind, dropped the debt instead of fast-forwarding the robot), `stalled_s` (the wall time dropped that way: the robot stood still while its clock should have run), `stalled_share` = `stalled_s` / `clock_s`, `held_steps`, `late_steps`, and what the requests cost: `requests_refused` (answered 4xx), `request_busy_s` (simulator-thread time spent on requests). They decide no verdict: a trial whose clock stalled is flagged by these fields, not failed or voided (whether a stalled trial counts is an open owner decision) | SHOULD | [new] (v1.2-rc: review R2 M1; before, only RoboOlympics' grader exported `stalls`, and `max_lag_s` never passed 0.5) |
| What a request can cost the clock is bounded by the service: the body is read, decoded and checked off the simulator's thread, so a refused or malformed request never reaches it (a body with more than 1 MiB of JSON text is parsed by a separate process; the 64 MB limit of §4 stands); behind the clock, control steps run before waiting requests, which are served once it has caught up (they queue behind each other: an `/observe` may take longer to answer, the robot's time does not slow down), and a rendering request that is a stall by itself (more than 0.5 s) holds the next one back four times as long, so requests take at most a fifth of the clock's time; and the cameras are rendered at most once per control step however many requests and watches ask | SHOULD | [new] (v1.2-rc: review R2 M1) |
-| Slowdown S: `RCB_SLOWDOWN`, default **1** (real time; per-benchmark exceptions below: RoboOlympics 10). Results with S ≠ 1 are reported separately from S = 1 and state S; official results use S = 1, or the benchmark's documented S | MUST | [new] |
-| Reset budget: **no reset by default** (`budget.reset = 0`), as in standard mode; per-benchmark exceptions only as in §4.5 (HumanoidBench: 50; MyoSuite has no standard-realtime tasks). The reset capability is §4.5's; `RCB_RESET_BUDGET` overrides it; results state the budget and are compared only under the same budget | MUST | [new] (v1.1.2, #102: v1.1 and v1.1.1 wrongly made HumanoidBench's 50 the mode's default) |
+| Slowdown S: `RCB_SLOWDOWN`, its default set per task and **stated in the task**: `[metadata] slowdown` in `task.toml` (written by `scripts/task_conventions.py` from the sim service's `RCB_SLOWDOWN` default), the README's "Real time" row and the instruction (`S = <S>`). A twin made for a `dynamic` family (§2.14) takes S from the real-time factor RTF of its own service, measured on the reference workstation (robot seconds per wall second with the clock running flat out, `RCB_SLOWDOWN` ≪ 1, CPU-capped as in its compose file): **S = 1 when RTF ≥ 3**, else S ≈ 1.5 / RTF rounded up, so that the clock needs at most two thirds of what the simulator can do (owner, 2026-10-14, board 0112); the benchmark's doc records the measurement. S = 1 is real time; RoboOlympics' 10 is its own (below). Results with S ≠ 1 are reported separately from S = 1 and state S; official results use the task's S | MUST | [CI] (`ci_checks/check-realtime.py`, 2026-10-14) |
+| Reset budget: every standard-realtime task allows resets, **10 by default** (`budget.reset = 10`; owner, 2026-10-14: "all realtime tasks allow resets"): an episode on a clock that does not stop is one attempt, often over in seconds, and the agent has no simulator to practise in. Per-benchmark exceptions only as in §4.5: HumanoidBench 50; BiGym 2.0 0 (its twins stay as they were, owner 2026-10-14); RoboOlympics 0 (a match is one episode). The reset capability is §4.5's; `RCB_RESET_BUDGET` overrides it; results state the budget and are compared only under the same budget | MUST | [CI] (`ci_checks/check-realtime.py`, 2026-10-14; before: no reset by default, v1.1.2, #102) |
| A simulator crash is an infrastructure failure, never the agent's: past `RCB_MAX_RESUMES` restarts the trial ends `sim_crashed` (no call spans a crash: `/submit` answers at once), a failed resume `resume_failed`, and a simulator error that ends the graded episode (`reason = "error"`) — all with `infra_error = 1` | MUST | [new] (owner, 2026-10-08) |
| The agent's controller runs in the agent's own container and streams targets through the client; no agent code runs next to the simulator | MUST | [new] |
-| A standard-realtime task shares its privileged twin's frozen instance and success rule (its verifier grades the sidecar's record, as in standard mode; a benchmark without a privileged mode only by the exception below); its instruction says that time does not stop | MUST | [new] |
+| A standard-realtime task shares its privileged twin's frozen instance and success rule where it has one (its verifier grades the sidecar's record, as in standard mode; a task without a twin, §2.15, carries its own frozen instance and success rule); its instruction says that time does not stop | MUST | [new] |
**Per-benchmark exceptions.** A benchmark's standard-realtime tasks may differ from a default or a MUST of this section
only by owner decision, recorded here and in `docs/benchmarks/<bench>.md` (and, where CI checks the rule, in its
@@ -849,9 +931,8 @@ settings.
- slowdown: `RCB_SLOWDOWN` defaults to **10**, RoboOlympics' official slowdown of these sports (a rally is over in a
few seconds, faster than a model's turn); it stays an override, and the results state S and are reported
separately (as every S ≠ 1).
- - no privileged twin: the sports measure play from camera images against an opponent; there is no trajectory
- deliverable. `ci_checks/check-instance-copies.sh` and `scripts/task_conventions.py` (`NO_TWIN`) exempt exactly
- these tasks.
+ - no privileged twin (not an exception since 2026-10-14: a benchmark ships any subset of the modes, §2.15): the
+ sports measure play from camera images against an opponent; there is no trajectory deliverable.
The breaks between points are part of the sports' rules, not a protocol feature: the clock never stops (the match
stands still for a bounded time while the control steps go on, and the agent's `ready` command ends a break early).
diff --git a/docs/REVIEWING.md b/docs/REVIEWING.md
index b550d2b493..2e88a92263 100644
--- a/docs/REVIEWING.md
+++ b/docs/REVIEWING.md
@@ -9,9 +9,9 @@ Guide for maintainers. The contract is [PROTOCOL.md](PROTOCOL.md); contributors
2. **Static checks.** `bash ci_checks/run_all.sh` must pass.
3. **Generator.** For generated tasks, read the generator and the shared sources, not 500 copies. Re-run it and check
that `git diff tasks/` is empty.
-4. **Sample tasks.** Read at least two tasks per mode in this order: `instruction.md` → `tests/` → `solution/` →
+4. **Sample tasks.** Read at least two tasks per mode the benchmark ships in this order: `instruction.md` → `tests/` → `solution/` →
`environment/` → `task.toml`.
-5. **Runs.** Oracle and `nop` on a sample (≥ 2 per mode); for a new simulator or family, a determinism check (two
+5. **Runs.** Oracle and `nop` on a sample (≥ 2 per mode shipped); for a new simulator or family, a determinism check (two
fresh replays, same hash) on the hardware used for evaluation.
6. **Adversarial trial.** For new benchmarks, one agent trial instructed to make the verifier pass without solving the
task. Any success blocks the merge until the verifier is fixed.
@@ -32,6 +32,19 @@ Guide for maintainers. The contract is [PROTOCOL.md](PROTOCOL.md); contributors
the recording-only usage gateway (PROTOCOL D11)?
- Are images pinned, published (or marked `BUILD_ONLY`), and shared with the other tasks of the same simulator?
- Names end in their mode, `-privileged` / `-standard` / `-standard-realtime`; README in the standard layout; no `results/` in task directories.
+- Modes (PROTOCOL §2.15): a benchmark may ship any subset of privileged, standard and standard-realtime; a missing
+ mode is not a reason to block, but the benchmark's doc says which modes it ships and why the others are left out.
+ Where a family has twins, do they share the frozen instance, success rule and seeds, and is the cutoff of privileged
+ and standard one number? A standard or standard-realtime task without a twin: does it carry its own frozen instance,
+ and, in a benchmark that has a privileged mode, is it listed in `scripts/task_conventions.py` `NO_TWIN` with the
+ reason (a deleted or misnamed twin must not pass silently)?
+- A new mode (PROTOCOL §2.16): does the pull request have everything §2.16 lists: the name / suffix and `mode` value,
+ its own PROTOCOL section (what the agent gets, what is graded, the time and reset rules, which failures are
+ infrastructure and which the agent's), the `MODE_REGISTRY` entry and `CUTOFF_NOTE`, CI support keyed on the registry
+ (no new hard-coded list of mode names anywhere) with a self-test case, at least one benchmark using it, the TAXONOMY
+ entry, and a stated cutoff relation (`same_cutoff_as`, or free)? Is the runner side (batch_eval_live, the results
+ site) told the new suffix? A new mode is an owner-visible protocol change: one pull request of its own, shared files
+ only, merged before the benchmark that uses it.
- Time cutoff (PROTOCOL §2.12): is the benchmark's number in `scripts/task_cutoffs.toml`, under the benchmark's label of
`scripts/task_conventions.py` `BENCHMARKS` (a new benchmark: the label first, in its shared-files pull request; the
row with its tasks) (3600 s unless the pull
diff --git a/docs/STANDARD_MODE.md b/docs/STANDARD_MODE.md
index 5dccacdb16..b918452ba0 100644
--- a/docs/STANDARD_MODE.md
+++ b/docs/STANDARD_MODE.md
@@ -209,12 +209,13 @@ camera poses are in the robot's starting frame); see [docs/benchmarks/](benchmar
tables. A request body over 64 MB gets 413 before it is decoded.
### Optional reset capability
-Off by default (`budget.reset = 0`): one trial is one episode, in standard and standard-realtime mode alike. Exception, by
+Off by default in standard mode (`budget.reset = 0`): one trial is one episode. In standard-realtime mode it is on, 10
+by default (owner, 2026-10-14; PROTOCOL §4b: HumanoidBench 50, BiGym 2.0 and RoboOlympics 0). Exceptions in standard mode, by
owner decision (PROTOCOL §4.5): HumanoidBench's standard tasks (2026-10-07) and standard-realtime tasks (2026-10-08) default
to `budget.reset = 50`, MyoSuite's standard tasks (2026-10-09) to `budget.reset = 10`; BiGym2's and KinDER's standard
-tasks (2026-10-11) to `budget.reset = 10`; DexToolBench's standard tasks (2026-10-13) to `budget.reset = 20`. A run may set `RCB_RESET_BUDGET=N`
-on the service (implemented by the KinDER, HumanoidBench, BiGym2, MyoSuite and MuJoCo (`server_mj.py`: DexToolBench, Playground ALOHA / Panda) services; MyoSuite's 2.2 client has no `reset()`,
-so its instruction calls `POST /reset` through `Robot._post`). Then `/spec.budget.reset = N`, `/spec.reset` describes it, and
+tasks (2026-10-11) to `budget.reset = 10`; DexToolBench's standard tasks (2026-10-13) to `budget.reset = 20`; MuJoCo Playground's quadruped standard tasks (2026-10-14) to `budget.reset = 10`. A run may set `RCB_RESET_BUDGET=N`
+on the service (implemented by the KinDER, HumanoidBench, BiGym2, MyoSuite and MuJoCo (`server_mj.py`: DexToolBench, Playground ALOHA / Panda) services and the Playground quadrupeds' 2.2 service; the 2.2 client has no `reset()`,
+so MyoSuite's and the quadrupeds' instructions call `POST /reset` through `Robot._post`). Then `/spec.budget.reset = N`, `/spec.reset` describes it, and
`POST /reset` (RCB1 or JSON `{cameras, video}`, no physics) abandons the running episode and starts a new one from the
same frozen state, `N` times in all; `/status` and `meta` add `episode`, `resets_left` and `control_steps_total`. After a
success or once the agent's time is up, `/reset` returns 409, as it does when the budget is spent. Only the last
@@ -550,13 +551,14 @@ for `include_depth` and ignore the new parts and keys).
Time does not stop. The wire protocol of the standard-realtime mode (PROTOCOL §4b, v1.1; protocol-change issue #86), served by
standard-realtime tasks only (`<base>-standard-realtime`, `mode = "standard-realtime"`, `protocol = "eai-standard/2.3"`; implemented by HumanoidBench's,
-`scripts/humanoidbench/shared/realtime/rcb_realtime.py`, which RoboOlympics' tasks use unchanged). 2.3 is 2.2 plus the realtime capability below; every 2.2
+`scripts/humanoidbench/shared/realtime/rcb_realtime.py`, which RoboOlympics', BiGym 2.0's and MyoSuite's tasks use unchanged). 2.3 is 2.2 plus the realtime capability below; every 2.2
addition stays a MAY, so a 2.1 service with this capability is a 2.3 service. Same endpoints, RCB1 frames, crash
recovery and grading as 2.1. `/spec.timing` declares it: `realtime: true`, `slowdown` S (`RCB_SLOWDOWN`, default 1 = real
time; results with S ≠ 1 are reported separately from S = 1), `control_dt`, `wall_s_per_control_step`, `max_ahead_steps`.
-Defaults of the mode: `budget.reset = 0`, as in standard mode (the optional reset capability of Part 1; `RCB_RESET_BUDGET` overrides it;
-HumanoidBench's per-benchmark exception is 50, PROTOCOL §4.5), and S = 1 (RoboOlympics' per-benchmark exception is 10,
-PROTOCOL §4b).
+Defaults of the mode: `budget.reset = 10` (since 2026-10-14; before, 0 as in standard mode: the optional reset capability of
+Part 1; `RCB_RESET_BUDGET` overrides it; HumanoidBench's per-benchmark exception is 50, BiGym 2.0's and RoboOlympics' 0,
+PROTOCOL §4b), and S stated per task (`[metadata] slowdown`, RCB_SLOWDOWN's default: 1 unless the service's measured
+real-time factor is below 3, PROTOCOL §4b; RoboOlympics' is 10).
- **Clock.** An episode starts paused at its starting state. Its clock starts at its first `/submit` or `/observe`
with `seconds > 0` (after a reset, and after a crash recovery, the episode is paused again) and runs one control step
every `control_dt` × S seconds of wall time, whatever the agent does, until the episode ends.
diff --git a/docs/TAXONOMY.md b/docs/TAXONOMY.md
index 1ae95fbbeb..f031edfee4 100644
--- a/docs/TAXONOMY.md
+++ b/docs/TAXONOMY.md
@@ -6,11 +6,13 @@ Known values; a new value is added by pull request. See [PROTOCOL.md](PROTOCOL.m
|---|---|
| `author_name` | the author's name (an organization may go in `author_org`) |
| `task_type` | `trajectory` · `policy` |
-| `mode` | `privileged` · `standard` (formerly `unlimited` / `limited`) · `standard-realtime` (protocol v1.1; named so since v1.1.1, PROTOCOL §4b) |
+| `mode` | `privileged` · `standard` (formerly `unlimited` / `limited`) · `standard-realtime` (protocol v1.1; named so since v1.1.1, PROTOCOL §4b); the list is `MODE_REGISTRY` in `scripts/task_conventions.py`, and a new mode is added by pull request (PROTOCOL §2.16) |
| `protocol` | standard and standard-realtime mode only. Standard-realtime: `eai-standard/2.3`. Standard: `eai-standard/2.1` · `eai-standard/2.2` (BEHAVIOR-1K, RoboLab, RoboWits, RoboPaint, ManiGuard, MolmoSpaces, MyoSuite, DexJoCo, MuJoCo Playground's quadruped tasks, YAMLab, RoboDojo, LabUtopia, AutoBio, Panda Dynamic Scenes) · `eai-standard/2.4` (the operations capability: RoboARC). All are specified in [STANDARD_MODE.md](STANDARD_MODE.md) |
| `simulator` | `robosuite` · `robotwin` (RoboTwin 2.0 = SAPIEN 3 + CuRobo) · `mujoco` · `kinder` (KinDER: MuJoCo and PyBullet) · `humanoidbench` (HumanoidBench: MuJoCo + dm_control) · `robowits` (RoboWits on Genesis) · `robolab` (NVIDIA RoboLab on Isaac Lab) · `labutopia` (LabUtopia on Isaac Sim) · `behavior` (BEHAVIOR-1K: OmniGibson on Isaac Sim) · `maniskill` (ManiSkill 3 / SAPIEN: RoboPaint; Panda Dynamic Scenes, ManiSkill 3.0.1 / SAPIEN 3.0.3 with CPU PhysX at 1000 Hz) · `vlabench` (VLABench: MuJoCo + dm_control) · `maniguard` (ManiGuard: BEHAVIOR-1K 3.7.2, OmniGibson on Isaac Sim 4.5) · `molmospaces` (Ai2 MolmoSpaces: MuJoCo) · `myosuite` (MyoSuite: musculoskeletal models on MuJoCo) · `dexjoco` (DexJoCo: MuJoCo 3.4 + its gymnasium envs) · `bigym2` (BiGym 2.0: MuJoCo 3.8.1 with NVIDIA's GR00T-WBC lower-body policy in every step) · `roboarc` (RoboARC, formerly TopoGym and RoboTangle: MuJoCo 3.14.0, CPU float64, with the blanket cloth on a deterministic CUDA step; Nyx RGB. Its Genesis 1.2.1 release, withdrawn on 2026-10-13, used `robotangle`) · `yamlab` (YAMLab on Isaac Lab / Isaac Sim 5.1, PhysX on the CPU) · `robodojo` (RoboDojo's simulation benchmark: its Isaac Lab 2.3 fork on Isaac Sim 5.1, PhysX with GPU dynamics, cuRobo IK) · `autobio` (AutoBio: MuJoCo 3.3 with AutoBio's own thread, detent and liquid-surface mechanisms) — the simulator, not the image: benchmarks on one simulator may use different images and versions (PROTOCOL D5) |
| `robot` | `franka-panda` · `aloha-agilex` · `aloha-2` (ALOHA 2, MuJoCo Menagerie: MuJoCo Playground, RoboARC) · `kuka-iiwa14-sharpa-ha4` · `tidybot-plus-plus` · `unitree-h1` · `unitree-h1hand` (H1 with two Shadow hands) · `sawyer` (Rethink Robotics Sawyer, as in Meta-World) · `marvin-bimanual` (RoboWits) · `franka-robotiq-2f85` (a Franka arm with a Robotiq 2F-85, the DROID setup: RoboLab, MolmoSpaces' FR3, RoboARC's Panda) · `r1pro` (BEHAVIOR-1K) · `franka-panda-omron` (RoboCasa's PandaOmron: Franka Panda on an Omron mobile base with a lift torso) · `fourier-gr1` (Fourier GR1 humanoid: two arms, waist and two Fourier dexterous hands, fixed base; RoboCasa-GR1) · `franka-panda-longfinger` (ManiGuard: Franka Panda with long fin-ray fingers) · `rainbow-rby1` (Rainbow Robotics RB-Y1: holonomic base, 6-joint torso, two 7-joint arms with parallel grippers; MolmoSpaces) · `myohand` (MyoSuite's MyoHand: a musculoskeletal human forearm and hand, 39 muscles) · `myoarm` (MyoSuite's MyoArm: a musculoskeletal human shoulder, arm and hand, 34 or 63 muscles) · `myoarm-torso` (MyoSuite's table-tennis player: torso, right arm and hand on a sliding pelvis) · `myoarm-mpl` (MyoSuite's MyoArm next to the Modular Prosthetic Limb, a motorized arm) · `myolegs` (MyoSuite's MyoLegs: a free-standing musculoskeletal torso and both legs) · `myoleg-osl` (MyoSuite's MyoLeg with an Open-Source Leg knee-ankle prosthesis) · `unitree-go1` (Unitree Go1 quadruped; MuJoCo Playground) · `boston-dynamics-spot` (Boston Dynamics Spot quadruped; MuJoCo Playground) · `google-barkour-vb` (Google Barkour vB quadruped; MuJoCo Playground) · `franka-allegro` (Franka Panda with a 16-DoF Allegro hand; DexJoCo) · `franka-allegro-bimanual` (two of them side by side; DexJoCo) · `franka-panda-bimanual` (two stock Franka Panda (FER) arms with their standard hands, facing each other or side by side at one table, each with a rigid task tool on its hand where the task has one; Panda Dynamic Scenes) · `kuka-iiwa14` (KUKA LBR iiwa 14 with a tool, here an air hockey mallet; RoboOlympics) · `unitree-g1` (Unitree G1 29-DoF humanoid with its rubber hands, a switchable grip on each palm; RoboOlympics' robot library with Unitree's limits and motor curves; RoboFunOlympics) · `unitree-h1-2` (Unitree H1-2, the adult-size 27-DoF humanoid, with the same grip; RoboOlympics' robot library; RoboFunOlympics' tasks whose official venue needs an adult-size body) · `unitree-g1-dex1` (Unitree G1 29-DoF humanoid with Dex1-1 two-finger grippers; legs and waist driven by GR00T-WBC; BiGym 2.0) · `i2rt-yam-bimanual` (two I2RT YAM arms, 6 joints each, with parallel two-finger grippers on a workstation table; YAMLab) · `arx-x5-bimanual` (two ARX X5 arms, 6 joints each, with parallel two-finger grippers at a table; RoboDojo) · `aloha-2-arm` (one arm of ALOHA 2 with its parallel gripper, fixed to a bench; AutoBio, whose two-arm tasks use `aloha-2`) · `ur5e-robotiq-2f85` (a UR5e with a Robotiq 2F-85; AutoBio) · `ur5e-dexhand-2f85-bimanual` (two UR5e arms, one with a DexHand021 dexterous hand that holds a pipette, one with a Robotiq 2F-85; AutoBio) |
| `family` | free-form slug naming the scene/goal, shared by its instances (e.g. `stack_blocks_two`, `dextoolbench/brush/blue_brush/sweep_forward`) |
+| `dynamics` | `static` · `dynamic` — whether something in the scene other than the robot does work by a force other than the robot's and gravity (PROTOCOL §2.14); written from `scripts/task_dynamics.toml`, never by hand |
+| `slowdown` | standard-realtime tasks only: the slowdown S their service starts with (the world runs at 1/S of real speed; PROTOCOL §4b); written from the sim service's `RCB_SLOWDOWN` default, never by hand |
| `oracle` | `full` (oracle scores 1) · `partial` (oracle reaches part of the goal) · `none` — see PROTOCOL §2.7 |
| `oracle_evidence` | required unless `oracle = "full"`: where the evidence for solvability is (PR, results site, upstream record) |
| `expert_time_estimate_hours` | number |
diff --git a/docs/benchmarks/autobio.md b/docs/benchmarks/autobio.md
index 7d7de65a27..3404d8ed7d 100644
--- a/docs/benchmarks/autobio.md
+++ b/docs/benchmarks/autobio.md
@@ -176,7 +176,9 @@ they still succeed (the success step moves by at most 22 control steps).
not its transfer: the check is on the tip's position relative to the surface and on the plunger command.
- **The rotor check holds only while the tube is held, in this instance.** After AutoBio's own expert lets go, the tube
rests 5.0 to 5.1 mm from the check's reference point, just outside its 5 mm; the check holds for 0.41 s while the
- tube is being lowered. Under rule `first` that moment is the success.
+ tube is being lowered. Under rule `first` that moment is the success. Both instructions state this since 2026-10-14, with the
+ standard text's missing height window (0.955-0.961 m) and the fact that the reference is computed at the episode's start
+ (a rotor turned by the agent moves the slot away from it).
- **Contact with a named finger.** `pickup` and `screw_loose` look for contact between the cap and the first arm's
left finger; a grasp between both fingers satisfies it, and the instructions say so.
- **AutoBio's own evaluator shows a frozen mixer display** after its first two frames (the helper described under
diff --git a/docs/benchmarks/dexjoco.md b/docs/benchmarks/dexjoco.md
index cb364d06b6..f2dc17429f 100644
--- a/docs/benchmarks/dexjoco.md
+++ b/docs/benchmarks/dexjoco.md
@@ -38,7 +38,11 @@ from a public human demonstration, in both modes (`scripts/dexjoco/`).
global RNG and its restorers edit the MuJoCo model (table height, fixed bodies), so the instance is a snapshot, not a
seed: `instance.npz` holds the full physics state (`mjSTATE_INTEGRATION`), mocap targets, controls, every numeric
`MjModel` array a task can change, and the env's Python-side task state (counters, flags). Its hash is checked
- whenever a task rebuilds it (privileged verifier, standard service, every replay).
+ whenever a task rebuilds it (privileged verifier, standard service, every replay). Its `meta` entry never says which
+ demonstration the scene came from (the reference is that demonstration's actions, and the agent reads
+ `/app/instance.npz`; PROTOCOL §2.7): `rcb_dexjoco.save_instance` refuses such keys and values, `generate.py` refuses
+ an instance that carries them, and `ci_checks/check-agent-pointers.py` reads the archive's string members. The
+ demonstration is named in `facts.json` (`demo`) and the task READMEs, which the agent does not read.
- **Action.** DexJoCo's own row per control step (50 Hz, 10 physics substeps of 2 ms): per arm the tcp target pose
(xyz, quaternion wxyz, world frame), which DexJoCo's operational-space controller tracks, and 16 Allegro joint targets.
Single arm 23 values, bimanual 46 (right first; the left hand's fingers in DexJoCo's mirrored order).
@@ -80,7 +84,9 @@ from a public human demonstration, in both modes (`scripts/dexjoco/`).
`rcb_dexjoco.Session` on the frozen instance: unlimited exact resets, live object poses (`s.objects()`), DexJoCo's state
dict, thin primitives (tcp-target moves, hand-joint interpolation, hold), rendering, recording and self-replay. The
deliverable is `/app/output/trajectory.npz`; the verifier replays it twice in a fresh process from its own copy of the
-instance: success = both replays succeed ∧ same final-state hash ∧ the instance rebuilt exactly.
+instance: success = both replays succeed ∧ same final-state hash. An instance that does not rebuild exactly (its hash
+differs: the image no longer builds the scene) is no verdict: nothing is replayed, `instance_mismatch`, `grader_error`
+and `infra_error` = 1.
- **Hand-in cap.** A trajectory has at most 50,000 rows (`rcb_dexjoco.MAX_TRAJECTORY_ROWS`: 1000 s of robot time; the
longest reference has 1085). `save_trajectory` refuses a longer one and the verifier scores it 0 (`bad_trajectory`).
The cap is what the verifier can replay: the slowest family (`bimanual_hanoi`) runs about 140 control steps/s, so two
@@ -163,6 +169,22 @@ python3 scripts/dexjoco/generate.py && python3 scripts/dexjoco/generate.py --che
```
## Change log
+- 2026-10-14 (review of board 0105, B-M3 and B6-3; the scenes, the success checks and the references are unchanged, so
+ the tasks keep their names):
+ - all 22 tasks: the committed instances (`instances/<family>/instance.npz` and its copies in `environment/`,
+ `tests/` and `environment/sim/`) had `meta` = `{"source": "DexJoCo-Datasets-Raw demo initial state …", "demo":
+ "<demonstration id>", "dexjoco": "8d23b0f"}`. The reference solution is that demonstration's own actions, so a
+ privileged agent could read at `/app/instance.npz` where a solution was (PROTOCOL §2.7); `ci_checks/check-agent-
+ pointers.py` skipped binary files and did not see it. `meta` is now `{"dexjoco": "8d23b0f"}`, rewritten by
+ `make_instances.py --strip-meta` (every other array is written back as loaded; all 11 rebuild to their recorded
+ instance sha). The tooling refuses to save such a `meta`, the generator refuses such an instance, and the pointer
+ check reads `.npz` string members and keys that name a demonstration or episode.
+ - privileged `tests/grade.py`: an instance that does not rebuild exactly is no verdict (`instance_mismatch`,
+ `grader_error`, `infra_error` = 1, nothing replayed); before, it was the agent's 0 (`instance_ok = 0`,
+ `infra_error = 0`). `scripts/dexjoco/tests/test_grade_privileged.py` covers both.
+ - Past results: under the egress allowlist the demonstration itself was out of reach (Hugging Face is not on it), so
+ a privileged result stands unless the trial had open network and fetched it. A past privileged 0 with
+ `instance_ok = 0` was an infrastructure failure.
- 2026-10-10 (review of 2026-10-10, board 0067; verifiers, the collect hook and the tooling's hand-in checks only: the
instances, the success checks and the reference trajectories are unchanged, so the tasks keep their names):
- standard: `final.json` missing / unreadable / without a verdict is `infra_error = 1` (a missing file used to be the
diff --git a/docs/benchmarks/libero-max.md b/docs/benchmarks/libero-max.md
index 7d4d8863ec..1e030092e7 100644
--- a/docs/benchmarks/libero-max.md
+++ b/docs/benchmarks/libero-max.md
@@ -84,6 +84,19 @@ tasks, each with one online event, in both modes: 64 cases, 128 tasks.
finger positions, joint positions and velocities, last command. The agent is told that the scene may change once
(what kinds of change), not which change or when. No object poses, no goal, no success signal.
+### Notes on single cases (2026-10-14, board 0113)
+- A move event (`target_relocation`, `receptacle_relocation`) sets one entity's position (`libero_backend._move_object`):
+ what rests in or on a moved receptacle stays behind, and an entity that lands on another is pushed apart by the
+ physics. Both instructions say so.
+- Four cases of `LIVING_ROOM_SCENE5_put_the_white_mug_on_the_left_plate...`: LIBERO's left plate (`plate_1`, in
+ `plate_left_region`) is on the robot's right; the instructions say that left and right are as seen facing the robot.
+- `STUDY_SCENE1_..._back_compartment_of_the_caddy` with the mug (LIBERO-PRO's task variant replaces the book): the back
+ compartment is the narrow back cell, about 5.5 cm across, narrower than the mug; LIBERO's `In` (origin in the region's
+ box, contact with the caddy) holds for a mug resting on its opening. The instructions say so; the check is unchanged.
+- The `initial_pose_position_angle` cases of `pick_up_the_black_bowl_{on,next_to}_the_cookie_box...` shift the target bowl
+ 0.17 m in x and -0.06 m in y and turn it 45 degrees, so the sentence's spatial relation no longer holds; the
+ instructions say which bowl is meant.
+
## Collect budget (standard mode)
When the agent's container is gone, the collect hook (`environment/sim/finalize.sh`) signals the service, which ends the
episode and replays it in a fresh process (`lmx_tool.py --replay`, one simulator per process). That replay is the
diff --git a/docs/benchmarks/mujoco-playground.md b/docs/benchmarks/mujoco-playground.md
index bd542f931c..4e975584aa 100644
--- a/docs/benchmarks/mujoco-playground.md
+++ b/docs/benchmarks/mujoco-playground.md
@@ -67,7 +67,7 @@ ALOHA and Franka Panda, scripted IK references, `eai-standard/2.1`) predate it a
| mode | the agent gets | resets | graded |
|---|---|---|---|
| privileged | the simulator (`rcb_quadruped.Session`) and its full state | unlimited | its `trajectory.npz`, replayed in two fresh processes |
-| standard (2.2) | cameras (Go1, Spot: four that follow the robot; Barkour: one that follows it and its head camera; RGB, optional depth), joints, IMU (gyro, accelerometer, gravity's direction, velocimeter), foot contacts; the feet's FK/IK in the trunk frame | none | the service's record of the one episode: live check, two fresh-process replays, determinism |
+| standard (2.2) | cameras (Go1, Spot: four that follow the robot; Barkour: one that follows it and its head camera; RGB, optional depth), joints, IMU (gyro, accelerometer, gravity's direction, velocimeter), foot contacts; the feet's FK/IK in the trunk frame | 10 (`RCB_RESET_BUDGET`, from 2026-10-14; before: none) | the service's record of the last episode: live check, two fresh-process replays, determinism |
## Reference solutions
- **Go1 joystick:** MuJoCo Playground's own Go1 joystick policy (`experimental/sim2sim/onnx/go1_policy.onnx`, the one its
@@ -163,6 +163,12 @@ fresh-process replay of its record ends in the live state hash. Generator: `pyth
installed; the PyPI packages `playground` and `envpool` carry Playground's task sources, PROTOCOL §2.11). Scenes, instances, tooling and pass rules
are unchanged; results from before this date were obtained with open egress. The quadruped pairs were already on the
allowlist.
+- 2026-10-14 (review of board 0105, B1-m2), the nine quadruped privileged tasks, `tests/grade.py` only: a replay
+ process that died or timed out (`replay_failed`), a rebuilt start that does not match the instance's hash
+ (`instance_mismatch`) and an exception of the verifier write `infra_error = 1` beside `grader_error = 1` (no verdict
+ on the agent, PROTOCOL: infrastructure vs agent failure); before, `grader_error` alone, which counted as the agent's
+ 0. Every `reward.json` carries `infra_error`. Each task README has the entry. Scenes, instances, pass rules and
+ references are unchanged; `scripts/mujoco-playground/tests/test_grade_privileged.py` covers the three paths.
- 2026-10-13 (board 0086), `mujoco_playground_aloha_single_peg_insertion`, both modes: **the success rule's rear
bound.** `rcb_aloha.Session.check` required `0 <= t <= 0.75` for the peg tip's axial parameter, and the texts said
"t ≤ 0.75, t = 0 at the rear wall, at least 25 % of the depth inside". Site `socket_rear` is at the tube's centre:
@@ -198,6 +204,20 @@ fresh-process replay of its record ends in the live state hash. Generator: `pyth
`reward.json` the fields `resets` and `reset_budget`. DexToolBench's standard twins set 20 resets (owner decision,
[`dextoolbench.md`](dextoolbench.md)); the two ALOHA and two Panda standard tasks set none: `/reset` is a 404 and
`budget.reset` 0, so their episodes, observations and verdicts are unchanged.
+- 2026-10-14 (owner decision C1, board 0113), the nine quadruped standard tasks: **10 resets** (PROTOCOL §4.5). The
+ quadrupeds' service (`shared/standard/server.py`) gained the optional reset capability as MyoSuite's has it:
+ `RCB_RESET_BUDGET` (the compose files' default 10; 0 restores the single episode), `POST /reset` (JSON or RCB1
+ `{cameras, video_fps, depth}`, no physics) abandons the running or ended episode and restarts from the frozen start,
+ 409 after a success, once the agent's time is up and once the budget is spent; only the last episode is graded
+ (`trajectory.npz` holds its rows; the collect hook's two fresh-process replays are unchanged); `meta` / `/status` add
+ `episode`, `resets_left`, `control_steps_total`; `final.json` and `reward.json` add `resets`, `reset_budget`,
+ `episodes`. Crash recovery replays the journal with the committed resets (checkpoint `persist`: `resets`,
+ `reset_at`). Instruction (rules, `r._post("/reset", ...)`), task.toml description and README updated. Checked in
+ eai-mujoco:0.1.3 (CPU): Go1Getup and BarkourJoystick, 10 resets with random chunks and a service crash between them,
+ each reset back to the start's state hash and proprioception, the 11th refused, the oracle in episode 11 -> reward 1
+ (trajectory = its rows only, deterministic), an idle episode 11 -> reward 0, a reset after a success refused.
+ Scene, instances, success checks and the grader's pass rule unchanged; earlier quadruped standard results ran with
+ one episode and are reported separately.
## Files
| | |
diff --git a/docs/benchmarks/myosuite.md b/docs/benchmarks/myosuite.md
index d085fe5466..b736469b07 100644
--- a/docs/benchmarks/myosuite.md
+++ b/docs/benchmarks/myosuite.md
@@ -1,7 +1,9 @@
# MyoSuite
Twenty-two [MyoSuite](https://github.com/MyoHub/myosuite) tasks (2.12.2, Apache-2.0) on MuJoCo 3.6.0, CPU only, in both
-modes: 44 task directories, `tasks/myosuite-<family>-i00-{privileged,standard}`. The robot is a musculoskeletal model of a
+modes: 44 task directories, `tasks/myosuite-<family>-i00-{privileged,standard}`, and one standard-realtime twin
+(`tasks/myosuite-table-tennis-i00-standard-realtime`, since 2026-10-14: table_tennis is the benchmark's one dynamic
+family). The robot is a musculoskeletal model of a
human body: the MyoHand (forearm and hand, 39 muscles), the MyoArm (34 or 63 muscles), MyoSuite's torso-and-arm table
tennis player, the MyoArm next to a prosthetic MPL arm, and the MyoLegs (with or without a prosthetic leg). The main
command is **one excitation per muscle** (a few families add non-muscle actuators: a positioned wrist, a sliding pelvis,
@@ -74,13 +76,17 @@ locomotion challenges and by hand_pose); elbow and finger pose / reach (near-dup
step budget (50-200 steps) is lifted, so the wall clock is the only budget. Why the holds: a first version counted 0.25 s,
and agents passed it by sweeping through the goal (or, in baoding, by parking the balls where the moving targets sweep
past); a 1 s hold of a fixed goal, and one whole orbit in baoding, require really reaching and keeping it.
+ Baoding's targets are a precomputed trajectory of 1000 points (`baoding_v1.create_goal_trajectory`), so since
+ 2026-10-14 its episode ends as a failure after 1000 control steps (25 s) if it has not succeeded (before, the step after
+ it raised an IndexError: a 500 in standard mode, board 0086).
- **Success, MyoChallenge (`event`)**: the challenge's own `solved` is the success event and ends the episode (goal scored,
opponent tagged, ball returned onto the opponent's half, object placed, track finished); its own failure conditions
and time limits apply (soccer 10 s, chase_tag 20 s, bimanual 10 s, table_tennis 3 s; osl_run has none, so the gym step
limit being lifted leaves it unlimited). Deviations from the official evaluation: bimanual's evaluation counts a path
when `solved` holds for more than 5 steps and its contact check passes, but the env ends the episode one step after the
first `solved`; we count the first `solved` with the contact check. osl_run's official score is the distance reached
- within its 1000-step horizon (5 s); we require finishing the 29 m track, without a time limit.
+ within its 1000-step horizon (5 s); we require finishing the 29 m track, without a time limit. The prosthesis is the
+ right leg (`osl_knee_angle_r`, `osl_ankle_angle_r`; the texts said left until 2026-10-14).
- **Success, MyoDex (`track`)**: MyoDex defines a dense reward only (`solved` is always 0). We count a success when the
whole reference motion is reproduced: the object's displacement since the start within 3 cm of the reference's
displacement since its start (and, for the four families where orientation is the point of the skill, its rotation
@@ -267,6 +273,37 @@ that score 0; apple_lift's first (exploit) trajectory scores 0 under the final r
- Determinism of set 3 before building: three replays in one process and one in a fresh process gave the same final
state for random action sequences in every family.
+## Standard-realtime (`eai-standard/2.3`, since 2026-10-14)
+The owner, 2026-10-14 (board 0107 D5, board 0112): the dynamic families (PROTOCOL §2.14, `scripts/task_dynamics.toml`)
+get a `-standard-realtime` twin. In MyoSuite that is **table_tennis** only (the ball is served at 5.6 m/s at reset;
+every other family's objects start at rest, and chase_tag's opponent stands still in its frozen instance).
+- **Service.** The shared realtime layer of HumanoidBench, BiGym 2.0 and RoboOlympics, unchanged
+ (`shared/realtime/rcb_realtime.py`, `rcb_service.py`, `rcb_client.py`, `grade.py`: byte-identical copies of
+ `scripts/humanoidbench/shared/`; `ci_checks/check-tooling-identity.sh`), with `shared/realtime/server.py`, which
+ implements `rcb_service.Sim` on `rcb_myosuite.Session`: the standard twin's frozen instance, action (`actuators`: 273
+ muscles and the two pelvis slides, nothing clipped), success rule (the challenge's own event), proprioception
+ (`q_`, `qd_`, `act_`, `len_`, `effort_`) and cameras (images and depth; no pose of a camera that rides on a body of
+ the scene). Not served: the 2.2 additions (`/kinematics`, per-frame camera poses, depth modes), all MAYs in 2.3. The
+ grade is the live episode, a replay in a fresh `Session` in the same process (MuJoCo on the CPU repeats bit for bit)
+ and their final states equal; the clock's record is evidence only.
+- **Rendering on the GPU.** On a clock that does not stop, a render is time the simulator thread cannot step, and an
+ OSMesa frame of this scene takes 1.3 s (1373 visual geoms, 207k mesh vertices; 640 x 480 or 320 x 240 alike) against
+ 5 ms with EGL on an RTX 4090. The twin therefore reserves a GPU for its `sim` service (`graphics` capability); the image
+ (`eai-myosuite:0.1.1`, unchanged) has no NVIDIA EGL ICD manifest and sets `MUJOCO_GL=osmesa`, so `server.py` writes the
+ manifest to `/tmp`, probes EGL in a throwaway process and falls back to OSMesa where it fails (`RCB_GL` forces one;
+ `/spec` `gl_backend` says which).
+- **Slowdown S = 1, resets 10.** Real-time factor of the service (ws-1, 32 cores at load 30-60, the service capped
+ at 4 CPUs; board 0112, 2026-10-14): 2.08 with the realtime clock flat out (`RCB_SLOWDOWN=0.01`, 11 episodes, 13.3 s
+ of robot time) and 2.6-2.7 for the bare physics; below 3, so PROTOCOL §4b's S = ceil(1.5 / 2.08) = 1. At S = 1 a
+ controller at the control rate (`cameras=[]`) with a camera image every 0.25 s ran 11 episodes (10 resets) with no
+ stall and a largest lag of 0.23 s. An episode that does nothing ends after 121 control steps (1.21 s): the serve
+ passes. The reset budget is the mode's default, 10; a reset takes about 1 s (MyoSuite's env reset).
+- **Checked at S = 1** (board 0112): a successful privileged episode (GPT-6 Luna xhigh, 82 rows) planned from step 0
+ in one `submit(..., at=0)` succeeds live and in the fresh replay (deterministic, no stall); streamed one row per call
+ from the agent's side it succeeded too; nop (no request) ends `timeout`, not graded a success; the 11 failed
+ episodes above replay to the live state. The cutoff is the standard twin's (3600 s).
+
## Files
-`scripts/myosuite/`: `generate.py`, `make_facts.py`, `instances/`, `oracles/`, `shared/` (README there);
+`scripts/myosuite/`: `generate.py`, `make_facts.py`, `instances/`, `oracles/`, `shared/` (README there; `shared/realtime/`
+the standard-realtime twin's service);
`images/myosuite/` (Dockerfile, VERSION).
diff --git a/docs/benchmarks/robodojo.md b/docs/benchmarks/robodojo.md
index 7a5b744957..740df7f8ef 100644
--- a/docs/benchmarks/robodojo.md
+++ b/docs/benchmarks/robodojo.md
@@ -201,6 +201,19 @@ each, and one call captures at most 3000 images (cameras × frames, per-frame de
`max_frames_per_call`): a `joint_position` chunk or an `/observe` that would exceed it is a 400; an `ee_pose` call, whose
length is only known once its IK is solved, keeps executing and stops filming (`meta["frames_capped"]`).
+## Known issues (board 0086, 2026-10-14)
+- **Joint limits in `/spec` and `robot_spec()`** are the articulation's soft limits as Isaac Lab reports them (about
+ +-10 rad for several joints). Two trials found the arm's joint4 stopping at about +-1.78 rad. Not verified in the
+ simulator here, so neither the instructions nor the limits have been changed.
+- **Privileged `s.reset()`** in a long-lived process is not always reliable after an articulated object moved ("articulation
+ not found", or "All scene Unstable Error"); a fresh process (as the verifier uses) always builds the frozen start.
+- **`pack_objects_into_box(_random)`**: "front sides facing left" is the assets' own axis convention (the checkpoint
+ point's -x, the shoe's root +x), which images do not show; every standard run failed it.
+- Fixed in the instructions on 2026-10-14 (text only, RoboDojo's checks unchanged): `sort_nesting_dolls_by_size(_random)`
+ had the size order reversed (the config's comments call index mod 5 = 4 the largest doll; the dolls' measured sizes say
+ it is the smallest); `swap_blocks`' button rests at about 0.79 of its travel, below the 0.9 / 0.95 its stages and press
+ counter need; `hang_mugs(_random)`' orientation conditions use the first functional point of each tag.
+
## Declarations
- **Tooling: thin wrappers only.** Joint moves, gripper commands, hold, RoboDojo's IK, plus the frozen instance, state,
recording, replay and rendering. Task-level skills are never added.
diff --git a/docs/benchmarks/robolab.md b/docs/benchmarks/robolab.md
index 14109c8313..573a969367 100644
--- a/docs/benchmarks/robolab.md
+++ b/docs/benchmarks/robolab.md
@@ -219,6 +219,14 @@ sees it and still ends the episode with a 500.
privileged trajectory may hold at most that many rows.
- An `ee_pose` submit has no frame cap: a long one at a high `video_fps` keeps many frames in memory (the client's
defaults, 5 fps and mp4, stay small).
+- RoboLab's success terms are upstream's and stay as they are (owner, 2026-10-14: "the original simulator decides"),
+ including two that are looser than their words (board 0086): `require_gripper_detached` can read an object that is
+ still pinched or wedged in the gripper as detached, and "inside" extends without limit along the container's up axis
+ (the instructions state the latter). `tools_picking_drill`'s term names the left bin while the drill starts in the
+ right bin; the instruction says so (2026-10-14).
+- Contact can make the physics diverge (joint readings of 1e6 and more, objects thrown off the table) after a grasp
+ that looks ordinary; in standard mode that episode is lost (board 0086 saw it once in about 350 trials). A
+ privileged `s.reset()` does not restore the start bit for bit; the verifier replays in a fresh process.
## Licences
RoboLab: Apache-2.0 (code and assets); Isaac Lab: BSD-3-Clause on GitHub, its pip wheels and Isaac Sim 5.1 are NVIDIA
diff --git a/docs/benchmarks/roboolympics.md b/docs/benchmarks/roboolympics.md
index ad42665678..e961889f42 100644
--- a/docs/benchmarks/roboolympics.md
+++ b/docs/benchmarks/roboolympics.md
@@ -67,7 +67,7 @@ what it heard, the serve); never the ball's or the opponent's state, nor the see
| | PROTOCOL §4b | RoboOlympics | why |
|---|---|---|---|
| slowdown S | `RCB_SLOWDOWN` default 1 | default **10**, overridable | RoboOlympics' official slowdown of all three sports: a model call takes seconds and a rally is over in a few, so at S = 1 play would be reflexes, not reasoning. Results state S; S ≠ 1 is reported separately anyway (§4b, §5) |
-| privileged twin | `<base>-privileged` shares the instance | none | the sports measure play from camera images against an opponent; there is no trajectory deliverable to hand in. CI exempts `roboolympics-*` explicitly (`ci_checks/check-instance-copies.sh`, `scripts/task_conventions.py` `NO_TWIN`) |
+| privileged twin | `<base>-privileged` shares the instance | none | the sports measure play from camera images against an opponent; there is no trajectory deliverable to hand in. Since 2026-10-14 a benchmark may ship any subset of the modes (PROTOCOL §2.15): RoboOlympics has no privileged number in `scripts/task_cutoffs.toml`, so `ci_checks/check-instance-copies.sh` expects no twin |
The reset budget is §4b's default, 0: one match is one episode. Everything else is `eai-standard/2.3` as specified.
diff --git a/docs/benchmarks/robowits.md b/docs/benchmarks/robowits.md
index 08de992066..68ca316a89 100644
--- a/docs/benchmarks/robowits.md
+++ b/docs/benchmarks/robowits.md
@@ -232,6 +232,9 @@ licensed to whoever downloads them, never in the image or the repository (`image
runs, and `ee_pose` rows that take longer than counted (up to 400 control steps each) stop adding frames there
(`meta["frames_skipped"]`; `scripts/robowits/tests/test_server_frames.py`). Until 2026-10-10 there was no limit: a
long chunk or watch with per-frame video or depth of three cameras could exhaust the sidecar's memory.
+- `ball_into_bottle`: the red bottle's base collides as a thin shell (about 3.5 mm), so a ball that falls fast down the
+ bottle can pass through it onto the table (board 0086: 9 of 16 slightly perturbed replays of one pour). The
+ instruction says so since 2026-10-14; a thicker collision base would change the frozen scene and has not been made.
- During the standard service's in-process grading replay, a late `/status` reports the replay's counter.
- The predicates are RoboWits' own and some are looser than their task text (the criterion sections say where); the
review checks each family's predicate (PROTOCOL §2.7).
diff --git a/scripts/autobio/families.json b/scripts/autobio/families.json
index 3a38fdecb6..4850f9504c 100644
--- a/scripts/autobio/families.json
+++ b/scripts/autobio/families.json
@@ -118,13 +118,17 @@
"robot": "ur5e-robotiq-2f85",
"robot_text": "a **UR5e** arm (6 joints) with a Robotiq 2F-85 gripper, fixed to the bench",
"hours": 3.0,
- "goal": "Load a second 1.5 mL tube into the centrifuge's rotor, into the slot diametrically opposite the tube that is already in it (slot 8 of 30, `s.task.slot_id`; the tube in the rotor is in slot 23). The tube to load stands in the plate on the bench. Success at the first control step where the tube's body is within 5 mm of its seated position in that slot (`s.task.final_tar_tubepose`) and its height is between 0.955 and 0.961 m.",
- "goal_std": "Load a second 1.5 mL tube into the centrifuge's rotor, into the slot diametrically opposite the tube that is already in it (the rotor has 30 slots). The tube to load stands in the plate on the bench. Success the moment the tube is seated in that slot: within 5 mm of its seated position.",
+ "goal": "Load a second 1.5 mL tube into the centrifuge's rotor, into the slot diametrically opposite the tube that is already in it (slot 8 of 30, `s.task.slot_id`; the tube in the rotor is in slot 23). The tube to load stands in the plate on the bench. Success at the first control step where the tube's body is within 5 mm of its seated position in that slot (`s.task.final_tar_tubepose`) and its height is between 0.955 and 0.961 m. That reference position is computed when the episode starts: turning the rotor moves the slot away from it. In this instance a tube that has been let go comes to rest 5.0 to 5.1 mm from it, just outside the 5 mm (AutoBio's own expert's tube included), so the check holds while the tube is still being lowered into the slot; the episode ends at that moment.",
+ "goal_std": "Load a second 1.5 mL tube into the centrifuge's rotor, into the slot diametrically opposite the tube that is already in it (the rotor has 30 slots). The tube to load stands in the plate on the bench. Success the moment the tube is seated in that slot: its body within 5 mm of its seated position there, and its body's height within a fixed 6 mm window (AutoBio's test; standard mode gives no world heights). That seated position is computed when the episode starts: turning the rotor moves the slot away from it, so leave the rotor where it is. In this instance a tube that has been let go comes to rest just outside the 5 mm (5.0 to 5.1 mm away, AutoBio's own expert's tube included), so the test holds while the tube is still being lowered into the slot, held by the gripper; the episode ends in success at that moment.",
"scene": "On the bench: a centrifuge (an Eppendorf 5430 model) with its lid open and a 30-slot rotor holding one 1.5 mL tube, and a well plate with a second 1.5 mL tube standing in it.",
"inventory": [
"centrifuge (Eppendorf 5430 model), lid open, with a 30-slot rotor",
"two 1.5 mL tubes with screw caps (one in the rotor, one in the plate)",
"60-well plate, blue"
- ]
+ ],
+ "change_log": {
+ "std": "- 2026-10-14 (board 0113, from board 0086): the instruction now states the whole of AutoBio's test (before, the standard text left out its 6 mm height window): the reference position is computed when the episode starts (turning the rotor moves the slot away from it), and in this instance a released tube rests 5.0 to 5.1 mm from it, so the test holds while the tube is still being lowered (before, this was only in `docs/benchmarks/autobio.md`, Known limits). Text only: the scene, the frozen instance, the reference solution, AutoBio's check and the verifier are unchanged, so the task keeps its name.\n",
+ "priv": "- 2026-10-14 (board 0113, from board 0086): the instruction now states the whole of AutoBio's test: the reference position is computed when the episode starts (turning the rotor moves the slot away from it), and in this instance a released tube rests 5.0 to 5.1 mm from it, so the test holds while the tube is still being lowered (before, this was only in `docs/benchmarks/autobio.md`, Known limits). Text only: the scene, the frozen instance, the reference solution, AutoBio's check and the verifier are unchanged, so the task keeps its name.\n"
+ }
}
}
diff --git a/scripts/autobio/generate.py b/scripts/autobio/generate.py
index 7aa5bede7c..2b146588b8 100644
--- a/scripts/autobio/generate.py
+++ b/scripts/autobio/generate.py
@@ -541,7 +541,7 @@ Generated by `scripts/autobio/generate.py` from `scripts/autobio/instances/{fam}
### Change Log
{CHANGE_LOG[std]}
-"""
+{t.get("change_log", {}).get("std" if std else "priv", "")}"""
def make(fam: str, out: Path) -> tuple[Path, Path]:
diff --git a/scripts/dexjoco/generate.py b/scripts/dexjoco/generate.py
index 19e5f97203..482d2b2b63 100644
--- a/scripts/dexjoco/generate.py
+++ b/scripts/dexjoco/generate.py
@@ -12,6 +12,7 @@ The texts are the templates below. docs/benchmarks/dexjoco.md describes the benc
"""
from __future__ import annotations
+import ast
import filecmp
import json
import re
@@ -19,6 +20,7 @@ import shutil
import stat
import sys
import tempfile
+import zipfile
from pathlib import Path
HERE = Path(__file__).resolve().parent
@@ -157,6 +159,33 @@ def dash(fam: str) -> str:
return fam.replace("_", "-")
+# What an instance's `meta` must not carry (PROTOCOL §2.7: the agent reads instance.npz, and the reference solution is the
+# actions of the demonstration the instance was restored from), read from the tooling: its one source.
+_TOOL = (SHARED / "rcb_dexjoco.py").read_text()
+PROVENANCE_KEYS = ast.literal_eval(re.search(r"^PROVENANCE_KEYS = (\(.*\))$", _TOOL, re.M).group(1))
+PROVENANCE_VALUE = re.compile(ast.literal_eval(re.search(r"^PROVENANCE_VALUE = re\.compile\((r\".*\"), re\.I\)$", _TOOL, re.M).group(1)), re.I)
+
+
+def instance_meta(path: Path) -> dict:
+ """The `meta` entry of an instance.npz (a 0-d numpy unicode array holding JSON), read without numpy."""
+ with zipfile.ZipFile(path) as z:
+ raw = z.read("meta.npy")
+ assert raw[:6] == b"\x93NUMPY", path
+ n, off = (int.from_bytes(raw[8:10], "little"), 10) if raw[6] == 1 else (int.from_bytes(raw[8:12], "little"), 12)
+ descr = ast.literal_eval(raw[off:off + n].decode("latin-1"))["descr"]
+ assert descr.lstrip("<>=|").startswith("U"), (path, descr)
+ return json.loads(raw[off + n:].decode("utf-32-be" if descr.startswith(">") else "utf-32-le").rstrip("\0"))
+
+
+def check_instance(path: Path) -> None:
+ """Refuse an instance whose `meta` says where its scene came from (the rule of rcb_dexjoco.check_instance_meta)."""
+ meta = instance_meta(path)
+ bad = [k for k, v in meta.items() if any(p in str(k).lower() for p in PROVENANCE_KEYS) or PROVENANCE_VALUE.search(json.dumps(v))]
+ if bad:
+ raise SystemExit(f"{path}: meta {bad} names where the scene came from, and the agent reads instance.npz (PROTOCOL "
+ "§2.7): provenance belongs in facts.json; make_instances.py --strip-meta rewrites it")
+
+
def facts(fam: str) -> dict:
return json.loads((HERE / "instances" / fam / "facts.json").read_text())
@@ -288,8 +317,10 @@ limited to the model APIs.
## Verification
Open-loop replay of `/app/output/trajectory.npz` from `tests/instance.npz`, twice in a fresh process in the verifier's own
-sandbox (no network); pass iff both episodes end in success, their final-state hashes agree and the rebuilt instance
-matches its recorded hash (`tests/grade.py`); plus `/logs/verifier/replay.mp4`, rendered after the verdict is written.
+sandbox (no network); pass iff both episodes end in success and their final-state hashes agree (`tests/grade.py`); plus
+`/logs/verifier/replay.mp4`, rendered after the verdict is written. A rebuilt instance that does not match its recorded
+hash (the image no longer rebuilds the scene) replays nothing and gives no verdict: `instance_mismatch`, `grader_error`
+and `infra_error` = 1, not the agent's 0.
A file that is not a readable trajectory (not an npz, no `task` / `actions`, a wrong shape, NaN) or has more than
{MAX_ROWS:,} rows scores 0 as `bad_trajectory`: the cap is what two replays of the slowest family fit into the verifier's
{PRIV_VERIFIER_TIMEOUT} s. A verifier that runs fewer than 4 OpenBLAS threads (fewer than 4 CPUs visible to it) replays
@@ -311,6 +342,18 @@ real-time sleep, episode cap and hanoi's wall-clock auto-reset are disabled in t
### Provenance
Generated by `scripts/dexjoco/generate.py` from `scripts/dexjoco/instances/{fam}/` (`make_instances.py`).
+
+### Change Log
+- 2026-10-14 (review of board 0105, B-M3 and B6-3): `environment/instance.npz` and `tests/instance.npz` no longer name
+ the demonstration the scene was restored from: their `meta` entry had `source` (the dataset) and `demo` (the
+ demonstration's id), and the reference solution is that demonstration's own actions, so the agent could read at
+ `/app/instance.npz` where a solution was (PROTOCOL §2.7). Only `meta` changed: every array of the scene is the same
+ (instance sha `{F["instance_sha"]}`), so the scene, the pass rule and the reference are unchanged and the task keeps its
+ name. The tooling (`rcb_dexjoco.py`, both copies) refuses to save an instance whose `meta` says where its scene came
+ from. `tests/grade.py`: a rebuilt instance that does not match its recorded hash is now no verdict
+ (`instance_mismatch`, `grader_error`, `infra_error` = 1, nothing replayed); before, it was the agent's 0 with
+ `infra_error` = 0 and `instance_ok` = 0. Past results: the demonstration itself was out of reach under the egress
+ allowlist (Hugging Face is not on it); a past 0 with `instance_ok` = 0 was an infrastructure failure.
"""
@@ -454,6 +497,14 @@ def std_changes(F: dict) -> str:
" pose, which standard mode must not give (PROTOCOL §4.5). It now gives images and depth only: no mount,\n"
" `extrinsic_cv` / `cam2world_gl` null in `meta[\"cameras\"]`, NaN per frame. Standard-mode results of this task\n"
" from before this date could read the object's pose and are reported separately.")
+ out.append(
+ "- 2026-10-14 (review of board 0105, B-M3): `environment/sim/instance.npz` and `environment/sim/rcb_dexjoco.py`. The\n"
+ " frozen instance's `meta` entry no longer names the demonstration the scene was restored from (`source`, `demo`;\n"
+ " the reference solution is that demonstration's actions), and the tooling refuses to save an instance that does.\n"
+ " The sidecar's files are out of the agent's reach, so nothing reached the agent from this copy; it changed with\n"
+ " the privileged twin's, from the one committed instance. Every array of the scene is the same (instance sha\n"
+ f" `{F['instance_sha']}`): the scene, the pass rule and the grader are unchanged, so the task keeps its name and earlier\n"
+ " results stand.")
out.append(
"- 2026-10-13 (owner, boards 0085 and 0087): the sim service's clock (`environment/sim/rcb_limited.py`,\n"
" `environment/sim/server.py`, the docstring of `environment/rcb_client.py`). `meta[\"time_s\"]` and\n"
@@ -727,6 +778,7 @@ def render(root: Path) -> list[Path]:
for fam in FAMILIES:
F = facts(fam)
inst = HERE / "instances" / fam
+ check_instance(inst / "instance.npz")
# privileged
t = root / f"dexjoco-{dash(fam)}-i00-privileged"
if t.exists():
diff --git a/scripts/dexjoco/instances/bimanual_assembly/instance.npz b/scripts/dexjoco/instances/bimanual_assembly/instance.npz
index d8c094224d..97687642a1 100644
Binary files a/scripts/dexjoco/instances/bimanual_assembly/instance.npz and b/scripts/dexjoco/instances/bimanual_assembly/instance.npz differ
diff --git a/scripts/dexjoco/instances/bimanual_hanoi/instance.npz b/scripts/dexjoco/instances/bimanual_hanoi/instance.npz
index 0f1065cea8..552afa8a8c 100644
Binary files a/scripts/dexjoco/instances/bimanual_hanoi/instance.npz and b/scripts/dexjoco/instances/bimanual_hanoi/instance.npz differ
diff --git a/scripts/dexjoco/instances/bimanual_microwave_cook/instance.npz b/scripts/dexjoco/instances/bimanual_microwave_cook/instance.npz
index d07075df98..db769fb23f 100644
Binary files a/scripts/dexjoco/instances/bimanual_microwave_cook/instance.npz and b/scripts/dexjoco/instances/bimanual_microwave_cook/instance.npz differ
diff --git a/scripts/dexjoco/instances/bimanual_photograph/instance.npz b/scripts/dexjoco/instances/bimanual_photograph/instance.npz
index 6e5ca5b70e..0aacad571e 100644
Binary files a/scripts/dexjoco/instances/bimanual_photograph/instance.npz and b/scripts/dexjoco/instances/bimanual_photograph/instance.npz differ
diff --git a/scripts/dexjoco/instances/bimanual_unlock_ipad/instance.npz b/scripts/dexjoco/instances/bimanual_unlock_ipad/instance.npz
index 8bc4e3d0c4..a8a691967b 100644
Binary files a/scripts/dexjoco/instances/bimanual_unlock_ipad/instance.npz and b/scripts/dexjoco/instances/bimanual_unlock_ipad/instance.npz differ
diff --git a/scripts/dexjoco/instances/click_mouse/instance.npz b/scripts/dexjoco/instances/click_mouse/instance.npz
index 274a4b7352..ad7d5c22cf 100644
Binary files a/scripts/dexjoco/instances/click_mouse/instance.npz and b/scripts/dexjoco/instances/click_mouse/instance.npz differ
diff --git a/scripts/dexjoco/instances/fold_glasses/instance.npz b/scripts/dexjoco/instances/fold_glasses/instance.npz
index ac1baebfb1..1467404f75 100644
Binary files a/scripts/dexjoco/instances/fold_glasses/instance.npz and b/scripts/dexjoco/instances/fold_glasses/instance.npz differ
diff --git a/scripts/dexjoco/instances/hammer_nail/instance.npz b/scripts/dexjoco/instances/hammer_nail/instance.npz
index 443d63038d..0ef696d692 100644
Binary files a/scripts/dexjoco/instances/hammer_nail/instance.npz and b/scripts/dexjoco/instances/hammer_nail/instance.npz differ
diff --git a/scripts/dexjoco/instances/pick_bucket/instance.npz b/scripts/dexjoco/instances/pick_bucket/instance.npz
index 473da0693d..eabd7c2ed1 100644
Binary files a/scripts/dexjoco/instances/pick_bucket/instance.npz and b/scripts/dexjoco/instances/pick_bucket/instance.npz differ
diff --git a/scripts/dexjoco/instances/pinch_tongs/instance.npz b/scripts/dexjoco/instances/pinch_tongs/instance.npz
index fb46289594..22b9c055b9 100644
Binary files a/scripts/dexjoco/instances/pinch_tongs/instance.npz and b/scripts/dexjoco/instances/pinch_tongs/instance.npz differ
diff --git a/scripts/dexjoco/instances/water_plant/instance.npz b/scripts/dexjoco/instances/water_plant/instance.npz
index c71cd35bcb..fd7054e2c5 100644
Binary files a/scripts/dexjoco/instances/water_plant/instance.npz and b/scripts/dexjoco/instances/water_plant/instance.npz differ
diff --git a/scripts/dexjoco/make_instances.py b/scripts/dexjoco/make_instances.py
index 53f21c2d1f..091c25e22d 100644
--- a/scripts/dexjoco/make_instances.py
+++ b/scripts/dexjoco/make_instances.py
@@ -14,6 +14,13 @@ Runs INSIDE the base image (DexJoCo + MuJoCo); writes scripts/dexjoco/instances/
--demo /demos/water_plant/<demo>/replay.zarr
# without --demo: keep the committed instance and oracle, check them and refresh facts.json only
docker run ... python scripts/dexjoco/make_instances.py --all
+ # --strip-meta: rewrite only the committed instances' `meta` without provenance (same arrays, same instance sha)
+ docker run ... python scripts/dexjoco/make_instances.py --all --strip-meta
+
+The agent reads instance.npz (/app/instance.npz), and the reference solution is the demonstration's own actions, so the
+instance's `meta` never names the demonstration or its dataset (PROTOCOL §2.7; rcb_dexjoco.check_instance_meta refuses
+it, generate.py refuses an instance that carries it). Provenance lives in facts.json (`demo`) and the tasks' READMEs,
+which the agent does not read. Until 2026-10-14 (board 0105) the meta named the demo.
Instance = the task's env after reset with the initial object poses and table height of the demonstration restored
(DexJoCo's own state restorer, demo state[0]). The reference = the demonstration's actions replayed through the tooling
@@ -69,8 +76,7 @@ def build_from_demo(fam: str, demo: str, out: Path) -> None:
s.base.reset()
restore_initial_state(s.env, fam, s.cfg, np.asarray(root["data"]["state"][0]).ravel())
s._snapshot = s.snapshot(); s._episode_reset()
- s.save_instance(out / "instance.npz", {"source": "DexJoCo-Datasets-Raw demo initial state (dexjoco.tasks.state_restorers)",
- "demo": FAMILIES[fam][1], "dexjoco": "8d23b0f"})
+ s.save_instance(out / "instance.npz", {"dexjoco": "8d23b0f"}) # no provenance: the agent reads it (docstring)
A = np.asarray(root["data"]["action"], dtype=np.float64)
s2 = R.Session(fam, instance=out / "instance.npz")
st = s2.replay(A)
@@ -80,6 +86,26 @@ def build_from_demo(fam: str, demo: str, out: Path) -> None:
# facts() rounds the rows to float32 (and checks they still succeed) if they are not float32-exact
+def strip_meta(fam: str, out: Path) -> None:
+ """Rewrite the committed instance with its `meta` cut to what is not provenance; every other array is written back
+ as it was loaded, and the rebuilt state must still hash to the recorded instance sha."""
+ p = out / "instance.npz"
+ with np.load(p, allow_pickle=False) as d:
+ arrs = {k: d[k] for k in d.files}
+ meta = json.loads(str(arrs["meta"]))
+ keep = {k: v for k, v in meta.items() if k == "dexjoco"}
+ R.check_instance_meta(keep)
+ if keep != meta:
+ np.savez_compressed(p, **{**arrs, "meta": np.asarray(json.dumps(keep))})
+ with np.load(p, allow_pickle=False) as d:
+ assert d.files == list(arrs) and all(np.array_equal(d[k], arrs[k]) and d[k].dtype == arrs[k].dtype
+ for k in arrs if k != "meta"), fam
+ s = R.Session(fam, instance=p)
+ if not s.instance_exact:
+ raise SystemExit(f"{fam}: the rewritten instance does not rebuild (state sha {s.state_sha()} != {s.instance_sha})")
+ print(json.dumps({"family": fam, "meta": keep, "was": sorted(meta), "instance_sha": s.instance_sha}), flush=True)
+
+
def scene_cameras(s) -> list:
"""Cameras that are mounted on, track or are aimed at a body that is neither fixed in the world nor moved by the
robot's own joints alone (a mocap body, or a body behind an object's joint: in bimanual_photograph `cam_pos_z`, the
@@ -162,11 +188,15 @@ def facts(fam: str, out: Path) -> dict:
def main() -> None:
ap = argparse.ArgumentParser()
ap.add_argument("families", nargs="*"); ap.add_argument("--all", action="store_true"); ap.add_argument("--demo")
+ ap.add_argument("--strip-meta", action="store_true")
a = ap.parse_args()
fams = sorted(FAMILIES) if a.all else a.families
for fam in fams:
out = HERE / "instances" / fam
out.mkdir(parents=True, exist_ok=True)
+ if a.strip_meta:
+ strip_meta(fam, out)
+ continue
if a.demo:
build_from_demo(fam, a.demo, out)
f = facts(fam, out)
diff --git a/scripts/dexjoco/shared/grade.py b/scripts/dexjoco/shared/grade.py
index c754f999ad..3cfbab9c42 100644
--- a/scripts/dexjoco/shared/grade.py
+++ b/scripts/dexjoco/shared/grade.py
@@ -6,8 +6,10 @@ write Harbor rewards.
Writes /logs/verifier/reward.txt (1/0), reward.json (numeric fields only; Harbor rejects strings), a replay video
/logs/verifier/replay.mp4 (front camera + wrist camera) and replay.last.png. Replays `--repeat` times in this fresh
process; success requires every replay to end its episode in success (DexJoCo's own rule: the episode ends at the first
-success or task failure) AND all final-state hashes to agree (MuJoCo on the CPU is bit-deterministic), and the frozen
-instance rebuilt here from tests/instance.npz must match its recorded hash.
+success or task failure) AND all final-state hashes to agree (MuJoCo on the CPU is bit-deterministic). The frozen
+instance rebuilt here from tests/instance.npz must match its recorded hash; one that does not (the image no longer
+rebuilds the scene) is no verdict on the agent: `instance_mismatch`, `grader_error` and `infra_error` = 1, nothing
+replayed.
A file the agent got wrong is the agent's 0, `bad_trajectory` (not `grader_error`): not an npz, pickled or object
arrays, no `task` or `actions` entry, a wrong shape, NaN/inf, or more than rcb_dexjoco.MAX_TRAJECTORY_ROWS rows (the
@@ -119,6 +121,10 @@ def main() -> int:
print("too few CPUs: " + rcb_dexjoco.threads_message(threads), file=sys.stderr)
write(False, {"grader_error": 1, "infra_error": 1, "too_few_cpus": 1, "n_actions": n_rows, **threads, **cost}); return 0
s = Session(a.family, instance=a.instance)
+ if not s.instance_exact: # the image no longer rebuilds the frozen scene: not the agent's 0 (infra, no verdict)
+ print(f"instance mismatch: rebuilt state sha {s.state_sha()} != recorded {s.instance_sha}", file=sys.stderr)
+ write(False, {"grader_error": 1, "infra_error": 1, "instance_mismatch": 1, "instance_ok": 0,
+ "n_actions": n_rows, **threads, **cost}); return 0
if acts.shape[1] != s.action_dim:
write(False, {"bad_trajectory": 1, "n_actions": n_rows, **cost}); return 0
t0 = time.time()
@@ -130,10 +136,9 @@ def main() -> int:
write(False, {"bad_trajectory": 1, "n_actions": n_rows, **cost}); return 0
all_success = all(o["success"] for o in outs)
deterministic = len({o["state_sha"] for o in outs}) == 1
- instance_ok = bool(s.instance_exact)
- ok = all_success and deterministic and instance_ok
+ ok = all_success and deterministic
out = {"success": int(ok), "all_success": int(all_success), "deterministic": int(deterministic),
- "instance_ok": int(instance_ok), "n_actions": int(len(acts)), "sim_steps": int(outs[0]["steps"]),
+ "instance_ok": 1, "n_actions": int(len(acts)), "sim_steps": int(outs[0]["steps"]),
"task_failure": int(bool(outs[0]["reason"]) and not outs[0]["success"]),
"replay_wall_s": round(time.time() - t0, 1), "grader_error": 0, "video_rendered": 0, **threads, **cost}
write(ok, out) # the verdict is on disk before the video; the video only adds video_rendered
diff --git a/scripts/dexjoco/shared/rcb_dexjoco.py b/scripts/dexjoco/shared/rcb_dexjoco.py
index f2e2db848f..8eb183fe2a 100644
--- a/scripts/dexjoco/shared/rcb_dexjoco.py
+++ b/scripts/dexjoco/shared/rcb_dexjoco.py
@@ -54,6 +54,7 @@ import argparse
import copy
import hashlib
import json
+import re
import sys
import time
from pathlib import Path
@@ -222,6 +223,21 @@ def water_aim_deg(model, data) -> float:
return float(np.degrees(np.arccos(np.clip(c, -1.0, 1.0))))
+# instance.npz is in the agent's container (/app/instance.npz), and a task's reference solution is the actions of the
+# recording its instance was restored from: its `meta` must not say where the scene came from (PROTOCOL §2.7, no pointer
+# to a solution where the agent can reach). save_instance refuses such keys, and values that name a recording.
+PROVENANCE_KEYS = ("source", "demo", "dataset", "episode", "recording", "reference", "solution", "zarr", "url")
+PROVENANCE_VALUE = re.compile(r"demo|episode|dataset|recording|replay\.zarr|huggingface|https?://", re.I)
+
+
+def check_instance_meta(meta: dict) -> None:
+ """ValueError if an instance's `meta` says where its scene came from (keys or values; see PROVENANCE_KEYS)."""
+ bad = [k for k, v in meta.items() if any(p in str(k).lower() for p in PROVENANCE_KEYS)
+ or PROVENANCE_VALUE.search(json.dumps(v))]
+ if bad:
+ raise ValueError(f"instance meta {bad} names where the scene came from: the agent reads instance.npz")
+
+
def _log(*a: Any) -> None:
print("[rcb_dexjoco]", *a, file=sys.stderr, flush=True)
@@ -349,6 +365,7 @@ class Session:
return h.hexdigest()[:16]
def save_instance(self, path: str | Path, meta: dict | None = None) -> None:
+ check_instance_meta(meta or {})
snap = self.snapshot()
np.savez_compressed(path, task=self.task, physics=snap["physics"], mocap_pos=snap["mocap_pos"],
**{"model__" + k: v for k, v in snap["model"].items()},
@@ -803,15 +820,14 @@ def cmd_make_instance(a) -> None:
restore_initial_state(s.env, a.task, s.cfg, state0)
s._snapshot = s.snapshot()
s._episode_reset()
- meta = {"source": "DexJoCo-Datasets-Raw demo, initial state restored with dexjoco.tasks.state_restorers",
- "demo": a.demo_name or os.path.basename(os.path.dirname(os.path.abspath(a.demo))), "dexjoco": "8d23b0f"}
- s.save_instance(a.out, meta)
- _log("instance", a.out, "sha", s.state_sha())
+ demo = a.demo_name or os.path.basename(os.path.dirname(os.path.abspath(a.demo)))
+ s.save_instance(a.out, {"dexjoco": "8d23b0f"}) # no provenance in the instance (check_instance_meta)
+ _log("instance", a.out, "sha", s.state_sha(), "from", demo)
if a.oracle_out:
A = np.asarray(root["data"]["action"], dtype=np.float64)
s2 = Session(a.task, instance=a.out)
st = s2.replay(A)
- s2.save_trajectory(A[:st["replayed_rows"]], a.oracle_out, {"source": "demo actions", "demo": meta["demo"]})
+ s2.save_trajectory(A[:st["replayed_rows"]], a.oracle_out, {"source": "demo actions", "demo": demo})
print(json.dumps({"oracle": a.oracle_out, **{k: st[k] for k in ("success", "reason", "replayed_rows", "rows", "state_sha")}}))
diff --git a/scripts/dexjoco/tests/test_grade_privileged.py b/scripts/dexjoco/tests/test_grade_privileged.py
new file mode 100644
index 0000000000..eea8af6abe
--- /dev/null
+++ b/scripts/dexjoco/tests/test_grade_privileged.py
@@ -0,0 +1,127 @@
+"""The DexJoCo privileged verifier (shared/grade.py) on a stand-in Session, and the frozen instances' `meta`; needs
+numpy, no MuJoCo.
+
+ python3 scripts/dexjoco/tests/test_grade_privileged.py
+
+- A rebuilt instance that does not match its recorded hash is no verdict on the agent: `instance_mismatch`,
+ `grader_error` and `infra_error` = 1, nothing replayed (board 0105, B6-3; before, the agent's 0 with infra_error 0).
+- The agent reads instance.npz (/app/instance.npz) and every reference is the actions of the demonstration the instance
+ was restored from: no committed instance names it (board 0105, B-M3); the tooling refuses to save such a `meta`, and
+ the generator refuses an instance that carries one.
+"""
+import json
+import os
+import shutil
+import subprocess
+import sys
+import tempfile
+import unittest
+from pathlib import Path
+
+import numpy as np
+
+HERE = Path(__file__).resolve().parent
+DEXJOCO = HERE.parent
+GRADE = DEXJOCO / "shared" / "grade.py"
+
+# what grade.py uses of rcb_dexjoco; STUB_EXACT = 0: the image no longer rebuilds the frozen instance
+STUB_TOOL = '''import os
+SESSION_LOG = ""
+MAX_TRAJECTORY_ROWS = 50_000
+REPLAYS = []
+
+
+class PrimitiveError(RuntimeError):
+ pass
+
+
+def blas_threads():
+ return {"visible_cpus": 8, "openblas_num_threads_env": 4, "blas_threads": 4, "threads_ok": 1}
+
+
+def threads_message(t):
+ return "stub"
+
+
+class Session:
+ action_dim = 23
+
+ def __init__(self, task, instance=None):
+ self.instance_exact = os.environ["STUB_EXACT"] == "1"
+ self.instance_sha = "recorded"
+
+ def state_sha(self):
+ return "recorded" if self.instance_exact else "other"
+
+ def replay(self, acts):
+ open(os.environ["STUB_REPLAYS"], "a").write("replay\\n")
+ ok = os.environ.get("STUB_SUCCESS", "1") == "1"
+ return {"success": ok, "state_sha": "end", "steps": len(acts), "reason": "success" if ok else None,
+ "replayed_rows": len(acts)}
+
+ def close(self):
+ pass
+'''
+
+
+class InstanceMismatch(unittest.TestCase):
+ def grade(self, exact: bool, success: bool = True) -> tuple[str, dict, int]:
+ with tempfile.TemporaryDirectory() as d:
+ d = Path(d); log = d / "log"
+ shutil.copy(GRADE, d / "grade.py")
+ (d / "rcb_dexjoco.py").write_text(STUB_TOOL)
+ np.savez(d / "trajectory.npz", task="water_plant", actions=np.zeros((5, 23)))
+ r = subprocess.run([sys.executable, "-B", str(d / "grade.py"), "--family", "water_plant", "--instance", "x",
+ "--traj", str(d / "trajectory.npz"), "--no-video"], capture_output=True, text=True, cwd=d,
+ env={**os.environ, "RCB_VERIFIER_LOG_DIR": str(log), "STUB_EXACT": str(int(exact)),
+ "STUB_SUCCESS": str(int(success)), "STUB_REPLAYS": str(d / "replays")})
+ self.assertEqual(r.returncode, 0, r.stderr[-400:])
+ n = len((d / "replays").read_text().splitlines()) if (d / "replays").exists() else 0
+ return (log / "reward.txt").read_text().strip(), json.loads((log / "reward.json").read_text()), n
+
+ def test_exact_instance_success(self):
+ txt, rw, n = self.grade(True)
+ self.assertEqual((txt, rw["success"], rw["infra_error"], rw["grader_error"], rw["instance_ok"], n), ("1", 1, 0, 0, 1, 2), rw)
+
+ def test_exact_instance_failure_is_the_agents_0(self):
+ txt, rw, n = self.grade(True, success=False)
+ self.assertEqual((txt, rw["success"], rw["infra_error"], rw["grader_error"]), ("0", 0, 0, 0), rw)
+
+ def test_an_instance_that_does_not_rebuild_is_no_verdict(self):
+ txt, rw, n = self.grade(False)
+ self.assertEqual((txt, rw["success"], rw["infra_error"], rw["grader_error"], rw["instance_mismatch"],
+ rw["instance_ok"], n), ("0", 0, 1, 1, 1, 0, 0), rw)
+
+
+class NoProvenance(unittest.TestCase):
+ def test_no_committed_instance_names_its_demonstration(self):
+ for d in sorted((DEXJOCO / "instances").iterdir()):
+ with np.load(d / "instance.npz", allow_pickle=False) as z:
+ meta = json.loads(str(z["meta"]))
+ self.assertEqual(meta, {"dexjoco": "8d23b0f"}, d.name)
+
+ def test_the_tooling_refuses_a_meta_with_provenance(self):
+ sys.path.insert(0, str(DEXJOCO / "shared"))
+ import rcb_dexjoco as R # numpy only at import (MuJoCo is imported by a Session)
+ R.check_instance_meta({"dexjoco": "8d23b0f"})
+ for bad in ({"demo": "x"}, {"source": "x"}, {"note": "water_plant_demo_10_2026-02-05"},
+ {"origin": "DexJoCo-Datasets-Raw episode 3"}, {"demo_id": 3}):
+ with self.assertRaises(ValueError, msg=bad):
+ R.check_instance_meta(bad)
+
+ @unittest.skipIf(sys.version_info < (3, 12), "generate.py needs Python 3.12 (f-strings)")
+ def test_the_generator_refuses_an_instance_with_provenance(self):
+ sys.path.insert(0, str(DEXJOCO))
+ import generate as G
+ with tempfile.TemporaryDirectory() as d:
+ p = Path(d) / "instance.npz"
+ np.savez_compressed(p, task="water_plant", physics=np.zeros(3), meta=json.dumps({"dexjoco": "8d23b0f"}))
+ self.assertEqual(G.instance_meta(p), {"dexjoco": "8d23b0f"})
+ G.check_instance(p)
+ np.savez_compressed(p, task="water_plant", meta=json.dumps({"source": "DexJoCo-Datasets-Raw", "demo": "d"}))
+ with self.assertRaises(SystemExit):
+ G.check_instance(p)
+
+
+if __name__ == "__main__":
+ unittest.main()
diff --git a/scripts/libero-max/generate.py b/scripts/libero-max/generate.py
index db58f6685c..47f977b7f8 100644
--- a/scripts/libero-max/generate.py
+++ b/scripts/libero-max/generate.py
@@ -138,6 +138,48 @@ def nums(v, fmt="{:g}") -> str:
return "(" + ", ".join(fmt.format(round(float(x), 3)) for x in v) + ")"
+# A LIBERO-MAX move is a jump of one entity (libero_backend._move_object sets that object's position, nothing else).
+TELEPORT = ("it jumps there alone: what already rests in or on it stays where it was, and where it lands on another object "
+ "the physics pushes the two apart")
+# Where a case's sentence and its scene need a word more (board 0086 / 0113): the note follows the task sentence in both
+# instructions, and its Change Log entry ends both READMEs. LIBERO-PRO's own files (bddl, the caddy's asset) are the source.
+_PLATES = ("Left and right are LIBERO's, as seen from the front of the table, facing the robot: the left plate (`plate_1`, "
+ "the goal's) is the one on the robot's right, the right plate (`plate_2`) the one on the robot's left "
+ "(LIBERO-PRO's scene file puts them in `plate_left_region` / `plate_right_region`, 0.3 m to either side).")
+_CADDY = ("The caddy has a large cell at each end and, between them, two narrow cells: the front one and the back one, which "
+ "runs along the caddy's tall wall. The back compartment is that narrow back cell (`desk_caddy_1_back_contain_region`, "
+ "about 5.5 cm across): the mug (about 9 cm wide) does not fit down into it. LIBERO's In tests only the mug's origin "
+ "against the region's box (its top is about 15 cm above the table) and its contact with the caddy, so a mug resting "
+ "on top of the back cell's opening, its origin just inside the top of that box, counts; a mug in any other cell "
+ "does not.")
+_BOWL_MOVED = ("LIBERO-PRO's initial-pose variant moved the target bowl (`akita_black_bowl_1`, the goal's) 0.17 m away from the "
+ "robot and 0.06 m to the robot's right of where the sentence puts it, and turned it 45 degrees")
+CASE_NOTES = {
+ "pro-camera-view-angle-10-t04-i08-camera_shift-d1-p195": _PLATES,
+ "pro-camera-view-angle-10-t04-i09-illumination_switch-d0-p195": _PLATES,
+ "pro-position-10-t04-i00-visual_theme_switch-d1-p195": _PLATES,
+ "pro-semantic-10-t04-i09-visual_theme_switch-d1-p195": _PLATES,
+ "pro-task-10-t05-i00-illumination_switch-d1-p195": _CADDY,
+ "pro-task-10-t05-i04-camera_shift-d1-p195": _CADDY,
+ "pro-initial-pose-position-angle-spatial-t03-i00-sensor_noise_onset-d1-p195": (
+ _BOWL_MOVED + ": it starts on the table, not on the cookies box (no bowl is on the box). The bowl to move is still that "
+ "one; the other black bowl stands on top of the wooden cabinet."),
+ "pro-initial-pose-position-angle-spatial-t06-i01-target_relocation-d0-p195": (
+ _BOWL_MOVED + ": it no longer stands right next to the cookies box. The bowl to move is still that one; the other "
+ "black bowl stands on the stove."),
+}
+CHANGE_NOTE = ("- 2026-10-14 (board 0113, from board 0086): the instruction adds a note after the task sentence, where the "
+ "sentence alone misleads in this LIBERO-PRO variant: {what}. Text only: the instance, the event, the reference "
+ "solution, LIBERO's success check and the verifier are unchanged, so the task keeps its name.\n")
+CHANGE_WHAT = {_PLATES: "which plate is the left one (LIBERO's left is as seen facing the robot, the robot's right)",
+ _CADDY: "which cell is the caddy's back compartment, that the mug does not fit into it, and what LIBERO's In accepts",
+ }
+CHANGE_TELEPORT = ("- 2026-10-14 (board 0113, from board 0086): the instruction says that a moved {what} jumps alone: what "
+ "rests in or on it stays where it was (LIBERO-MAX moves one entity; board 0086 saw a basket jump through "
+ "a can already in it). Text only: the instance, the event, the reference solution, LIBERO's success check "
+ "and the verifier are unchanged, so the task keeps its name.\n")
+
+
def event_text(change_type: str, ch: dict) -> tuple[str, str]:
"""(short words, what happens) for a LIBERO-MAX change."""
obj = f"`{ch.get('object')}`"
@@ -149,7 +191,7 @@ def event_text(change_type: str, ch: dict) -> tuple[str, str]:
if change_type == "receptacle_relocation":
return (
"receptacle relocation",
- f"the receptacle {obj} is moved {ch['distance_m']:g} m, direction {nums(ch['direction_xy'])}",
+ f"the receptacle {obj} is moved {ch['distance_m']:g} m, direction {nums(ch['direction_xy'])}; {TELEPORT}",
)
if change_type == "obstacle_insertion":
return (
@@ -334,13 +376,21 @@ def make(case: dict, name: str, out: Path) -> tuple[Path, Path]:
"n_actions": rep["live"]["control_steps"],
"event_step": rep["live"]["event_step"],
"live_sha": rep["live"]["state_sha"][:12],
+ "case_note": (" " + CASE_NOTES[case["case_id"]]) if case["case_id"] in CASE_NOTES else "",
}
+ note = CASE_NOTES.get(case["case_id"])
+ what = CHANGE_WHAT.get(note, "the target bowl is not where the sentence puts it, and which bowl is meant") if note else None
+ priv_log = CHANGE_NOTE.format(what=what) if note else ""
+ std_log = priv_log
+ if case["scenario"]["change_type"] == "receptacle_relocation":
+ priv_log += CHANGE_TELEPORT.format(what="receptacle")
+ std_log += CHANGE_TELEPORT.format(what="object or receptacle") # the standard text is the same for every case
p = out / f"{name}-privileged"
reset_dir(p)
put(p / "task.toml", render("task.toml", values))
put(p / "instruction.md", render("instruction.md", values))
- put(p / "README.md", render("README.md", values))
+ put(p / "README.md", render("README.md", {**values, "change_log_extra": priv_log}))
put(p / "environment/Dockerfile", PRIV_ENV_DOCKERFILE)
put(p / "environment/lmx_tool.py", src=SHARED / "lmx_tool.py")
put(p / "environment/instance.json", src=fz / "instance.json")
@@ -357,7 +407,7 @@ def make(case: dict, name: str, out: Path) -> tuple[Path, Path]:
reset_dir(s)
put(s / "task.toml", render("standard/task.toml", values))
put(s / "instruction.md", render("standard/instruction.md", values))
- put(s / "README.md", render("standard/README.md", values))
+ put(s / "README.md", render("standard/README.md", {**values, "change_log_extra": std_log}))
put(s / "environment/Dockerfile", STD_ENV_DOCKERFILE)
put(s / "environment/docker-compose.yaml", compose(instruction, episode))
put(s / "environment/rcb_client.py", src=SHARED / "standard/rcb_client.py")
diff --git a/scripts/libero-max/templates/README.md b/scripts/libero-max/templates/README.md
index f1f5ee185e..ad0412550b 100644
--- a/scripts/libero-max/templates/README.md
+++ b/scripts/libero-max/templates/README.md
@@ -52,3 +52,4 @@ Generated by `scripts/libero-max/generate.py` from `scripts/libero-max/cases.jso
it executes any action). Before, such a trial read as an agent failure with `grader_error = 1` only.
- A replay that records a video prints its verdict before it finishes the video, and a failing video no longer fails
the replay; `video_rendered` is 1 only when the video was finished.
+{{change_log_extra}}
\ No newline at end of file
diff --git a/scripts/libero-max/templates/instruction.md b/scripts/libero-max/templates/instruction.md
index 0b7174e3c4..39a0e5e592 100644
--- a/scripts/libero-max/templates/instruction.md
+++ b/scripts/libero-max/templates/instruction.md
@@ -1,6 +1,6 @@
# LIBERO-MAX · {{suite}} · {{event_words}} — instance {{inst}}
-A **Franka Panda** arm (LIBERO on robosuite / MuJoCo) at a table. **Task: {{instruction}}**
+A **Franka Panda** arm (LIBERO on robosuite / MuJoCo) at a table. **Task: {{instruction}}**{{case_note}}
The scene is one frozen LIBERO-MAX case: a LIBERO-PRO variant of a {{suite}} task ({{pro_category}}), with one online
event: the first time the gripper comes within {{trigger_m}} m of `{{trigger_object}}`, LIBERO-MAX changes the scene
once ({{event_detail}}). You have the full simulator state, so you can see the change happen (`s.state()["event_fired"]`).
diff --git a/scripts/libero-max/templates/standard/README.md b/scripts/libero-max/templates/standard/README.md
index 707bbbf7db..8dcd537264 100644
--- a/scripts/libero-max/templates/standard/README.md
+++ b/scripts/libero-max/templates/standard/README.md
@@ -62,3 +62,4 @@ and the video can neither change nor cost it (`docs/benchmarks/libero-max.md`, "
/reset` ends the episode as `error` (`sim_error = 1`), not `reset` (reachable only when a run grants resets; the
default is none). The scene, instance, pass rule and every verdict reached before are unchanged, so the task keeps
its name.
+{{change_log_extra}}
\ No newline at end of file
diff --git a/scripts/libero-max/templates/standard/instruction.md b/scripts/libero-max/templates/standard/instruction.md
index 485466b6e4..60c9469f57 100644
--- a/scripts/libero-max/templates/standard/instruction.md
+++ b/scripts/libero-max/templates/standard/instruction.md
@@ -1,12 +1,13 @@
# LIBERO-MAX · {{suite}} · {{event_words}} — instance {{inst}} — STANDARD MODE (robot as a service)
You control a simulated **Franka Panda** arm (LIBERO on robosuite / MuJoCo) **as if it were real hardware**: the
-simulator runs in a service you cannot open, and you reach it only through a client. **Task: {{instruction}}**
+simulator runs in a service you cannot open, and you reach it only through a client. **Task: {{instruction}}**{{case_note}}
The scene can change while you work: once during the episode, when your gripper first gets close to one of the task's
objects, the scene changes (as in LIBERO-MAX): an object or a receptacle may be moved, an obstacle or a few other
objects may appear, a camera may shift, or the lighting, the colours or the image noise may change. Nothing tells you
-when; the task and its success check stay the same, so keep checking what the cameras show.
+when; the task and its success check stay the same, so keep checking what the cameras show. A moved object or
+receptacle jumps there alone: what already rests in or on it stays where it was.
## Rules (real-robot rules)
- **One episode, no reset.** The episode starts when the service comes up. It ends when the task succeeds, after
diff --git a/scripts/mujoco-playground/generate.py b/scripts/mujoco-playground/generate.py
index 976fa9e65e..1d643a8835 100644
--- a/scripts/mujoco-playground/generate.py
+++ b/scripts/mujoco-playground/generate.py
@@ -118,6 +118,7 @@ ENVS = {
"xml": "locomotion/go1/xmls/scene_mjx_flat_terrain.xml"},
}
SETTLE, LIN_TOL, YAW_TOL, HOLD, TILT, POSE_HOLD, POSE_TOL, GAIT_TOL = 25, 0.2, 0.3, 25, 10, 100, 30, 0.05
+RESET_BUDGET = 10 # standard mode's default reset budget (owner, 2026-10-14; PROTOCOL §4.5, docs/benchmarks/mujoco-playground.md)
PRIV_ENV_DOCKERFILE = f"""# MuJoCo Playground quadruped task (privileged): plain MuJoCo with Playground's and Menagerie's assets
# (images/mujoco/prepare_assets.sh), the task's tooling and its frozen instance.
@@ -179,6 +180,7 @@ services:
RCB_STARTUP_TIMEOUT_S: "300" # the scene builds in about a second
PYTHONPATH: /srv
RCB_SESSION_LOG: "" # the sidecar's own Session must not write the agent-side interaction log
+ RCB_RESET_BUDGET: "${{RCB_RESET_BUDGET:-{RESET_BUDGET}}}" # Playground quadrupeds standard: {RESET_BUDGET} resets (PROTOCOL §4.5); 0: one episode
"""
@@ -247,7 +249,8 @@ def values(env: str, inst: dict) -> dict:
"kp": e["kp"], "kd": e["kd"], "sim_dt": e["sim_dt"], "dt": f"{e.get('dt', 0.02):g}",
"substeps": round(e.get("dt", 0.02) / e["sim_dt"]), "xml": e["xml"], "gains": gains(e),
"cameras": CAMERAS[e["robot"]], "n_rows": fa["reference_rows"], "selection": selection_text(env), "author": AUTHOR,
- "org": ORG, "allowed_hosts": ", ".join(json.dumps(h) for h in HOSTS),
+ "org": ORG, "allowed_hosts": ", ".join(json.dumps(h) for h in HOSTS), "resets": RESET_BUDGET,
+ "episodes": RESET_BUDGET + 1,
}
if e["kind"] == "joystick":
rough = "Rough" in env
@@ -440,13 +443,26 @@ STD_CHANGES_R2 = [
"`wait_ready()` raise `EpisodeOver` at once; before, nothing answered and they waited out their timeout (up to "
"an hour). The adapter, the collect hook, the grader, the scene and the pass rule are unchanged.",
]
+STD_CHANGES_RESETS = [
+ f"2026-10-14 (owner decision C1, board 0113): {RESET_BUDGET} resets (PROTOCOL §4.5). The service "
+ "(`environment/sim/server.py`) gained the optional reset capability: `RCB_RESET_BUDGET` (the compose file's default "
+ f"{RESET_BUDGET}; 0 restores the old single episode), `POST /reset` abandons the running or ended episode and starts a "
+ "new one from the same frozen start (no physics), refused (409) after a success, once the agent's time is up and once "
+ "the budget is spent; only the last episode is graded (`trajectory.npz` holds its rows, replayed twice from the "
+ "frozen start as before); `/status` and every `meta` add `episode`, `resets_left` and `control_steps_total`; "
+ "`final.json` and `reward.json` add `resets`, `reset_budget` and `episodes`. A crash resumes the current episode "
+ "(the journal is replayed with the committed resets). The instruction's rules, the client example (`POST /reset` "
+ "through `Robot._post`) and this README say so. The scene, the instance, the success check and the grader's pass "
+ "rule are unchanged; results recorded before this change ran with one episode and are reported separately.",
+]
def make(env: str, out: Path) -> tuple[Path, Path]:
inst_path = HERE / "instances" / f"{env}.json"
inst = json.loads(inst_path.read_text())
v = values(env, inst)
- v["change_log_std"] = "\n".join(f"- {c}" for c in STD_CHANGES + (STD_CHANGES_BARKOUR if "Barkour" in env else []) + STD_CHANGES_R2)
+ v["change_log_std"] = "\n".join(f"- {c}" for c in STD_CHANGES + (STD_CHANGES_BARKOUR if "Barkour" in env else []) + STD_CHANGES_R2
+ + STD_CHANGES_RESETS)
oracle = HERE / "oracles" / f"{env}.npy"
p = out / f"{v['name']}-privileged"
diff --git a/scripts/mujoco-playground/shared/grade.py b/scripts/mujoco-playground/shared/grade.py
index c75c326868..1667d090c8 100644
--- a/scripts/mujoco-playground/shared/grade.py
+++ b/scripts/mujoco-playground/shared/grade.py
@@ -6,6 +6,11 @@ Two replays, each in its own fresh process (`rcb_quadruped.py replay`: the froze
task's success check on the whole trajectory). success = both replays succeed and end in the same state. Writes
/logs/verifier/reward.txt (1/0) and reward.json (numbers only); then a third replay renders replay.mp4, after the
verdict is on disk (it never changes it).
+
+Failures that are not the agent's (docs/PROTOCOL.md: infra vs agent failure) carry `grader_error` = 1 and `infra_error`
+= 1, so they are no verdict on the agent: a replay process that died or timed out, an image that no longer rebuilds the
+frozen start (`instance_mismatch`), and any exception of the verifier itself. A file the agent got wrong is the agent's
+0 (`bad_trajectory`, infra_error 0). Every reward.json carries success, deterministic, grader_error and infra_error.
"""
import argparse
import json
@@ -22,8 +27,12 @@ TOOL = os.path.join(os.path.dirname(os.path.abspath(__file__)), "rcb_quadruped.p
REPLAY_TIMEOUT_S = 300 # a 1000-step replay takes a few seconds; the video replay renders 4 views every 2nd step
+BASE = {"success": 0, "deterministic": 0, "grader_error": 0, "infra_error": 0}
+
+
def write(ok: bool, out: dict) -> None:
os.makedirs(LOG_DIR, exist_ok=True)
+ out = {**BASE, **out}
open(os.path.join(LOG_DIR, "reward.txt"), "w").write("1\n" if ok else "0\n")
json.dump(out, open(os.path.join(LOG_DIR, "reward.json"), "w"))
print(json.dumps(out), flush=True)
@@ -96,11 +105,11 @@ def main() -> int:
"n_actions": int(acts.shape[0]) if acts.ndim == 2 else 0, "grader_error": 0, **cost})
return 0
runs = [replay(a.instance, a.traj, None), replay(a.instance, a.traj, None)]
- if any("error" in r and not r.get("bad_trajectory") for r in runs):
- write(False, {"success": 0, "deterministic": 0, "grader_error": 1, **cost})
+ if any("error" in r and not r.get("bad_trajectory") for r in runs): # the replay process died / timed out
+ write(False, {"success": 0, "deterministic": 0, "grader_error": 1, "infra_error": 1, "replay_failed": 1, **cost})
return 0
if not all(r.get("instance_ok") for r in runs): # the image no longer builds the frozen start: not the agent's 0
- write(False, {"success": 0, "deterministic": 0, "grader_error": 1, "instance_mismatch": 1, **cost})
+ write(False, {"success": 0, "deterministic": 0, "grader_error": 1, "infra_error": 1, "instance_mismatch": 1, **cost})
return 0
all_success = all(bool(r.get("success")) for r in runs)
deterministic = runs[0].get("state_sha") == runs[1].get("state_sha")
@@ -116,7 +125,7 @@ def main() -> int:
write(ok, out) # the verdict, before any video
except Exception: # noqa: BLE001
traceback.print_exc(file=sys.stderr)
- write(False, {"success": 0, "deterministic": 0, "grader_error": 1, **cost})
+ write(False, {"success": 0, "deterministic": 0, "grader_error": 1, "infra_error": 1, **cost})
return 0
try: # the video, for humans: it never changes the verdict
if replay(a.instance, a.traj, os.path.join(LOG_DIR, "replay.mp4")).get("video"):
diff --git a/scripts/mujoco-playground/shared/standard/grade.py b/scripts/mujoco-playground/shared/standard/grade.py
index 41ce495421..3d4c23ad7e 100644
--- a/scripts/mujoco-playground/shared/standard/grade.py
+++ b/scripts/mujoco-playground/shared/standard/grade.py
@@ -77,6 +77,9 @@ write(
"sim_error": int(d.get("reason") == "error"),
"sim_startups": int(d.get("sim_startups") or 0),
"resumes": int(d.get("resumes") or 0),
+ "resets": int(d.get("resets") or 0), # PROTOCOL §4.5: resets the agent used; only the last episode is graded
+ "reset_budget": int(d.get("reset_budget") or 0),
+ "episodes": int(d.get("episodes") or 1),
"resume_state_mismatch": int(d.get("resume_state_mismatch") or 0),
"resume_max_joint_dev": num("resume_max_joint_dev"),
},
diff --git a/scripts/mujoco-playground/shared/standard/server.py b/scripts/mujoco-playground/shared/standard/server.py
index 108eca3cae..cad8920259 100644
--- a/scripts/mujoco-playground/shared/standard/server.py
+++ b/scripts/mujoco-playground/shared/standard/server.py
@@ -1,16 +1,20 @@
"""eai-standard/2.2 sim service for MuJoCo Playground's quadruped tasks (docs/STANDARD_MODE.md): the adapter of the
shared layer rcb_limited.py; the simulator is rcb_quadruped.Session (plain MuJoCo on the CPU).
-HTTP surface: GET /spec, GET /status, POST /observe, POST /submit, POST /kinematics. Every call is logged to
-/data/episode.jsonl.
+HTTP surface: GET /spec, GET /status, POST /observe, POST /submit, POST /kinematics, and POST /reset when the run grants
+resets (RCB_RESET_BUDGET, the optional reset capability of docs/STANDARD_MODE.md / PROTOCOL §4.5; the quadrupeds'
+standard tasks default to 10 in their compose file, owner 2026-10-14). Every call is logged to /data/episode.jsonl.
The episode is the frozen instance's: one control step = the instance's control period (Go1, Spot: 1/50 s; Barkour:
1/100 s). A joystick episode ends after the whole command schedule
(the horizon), when the robot meets Playground's termination rule (it falls over: a failure), when the agent submits
`done`, or when its time is over; success = the task's check on the steps executed. A getup episode ends the moment the
-check holds (success), after the horizon, on `done` or on timeout.
+check holds (success), after the horizon, on `done` or on timeout. With a reset budget, POST /reset abandons the running
+(or ended) episode and starts a new one from the frozen instance; only the last episode is graded, and a success ends
+the trial.
Crash recovery (rcb_limited.Episode, resume class `exact`: MuJoCo on the CPU repeats bit for bit in a fresh process):
-every executed row is journaled; after a crash the service rebuilds the episode and replays the journal. When the
+every executed row is journaled; after a crash the service rebuilds the episode and replays the journal, resetting where
+the committed resets did (their journal rows are checkpointed, `persist`). When the
episode ends the service writes /data/trajectory.npz + /data/episode.json; the collect hook (finalize.sh) makes it exit
and replays the trajectory in two fresh processes."""
@@ -31,18 +35,21 @@ INSTANCE = os.environ.get("RCB_INSTANCE", os.path.join(os.path.dirname(os.path.a
WALL_CLOCK_S = float(os.environ.get("RCB_WALL_CLOCK_S", "3600"))
MAX_CHUNK_S = float(os.environ.get("RCB_MAX_CHUNK_S", "10"))
MAX_WATCH_S = float(os.environ.get("RCB_MAX_WATCH_S", "30"))
+RESET_BUDGET = max(0, int(os.environ.get("RCB_RESET_BUDGET") or 0)) # the compose file's default: 10 (PROTOCOL §4.5)
CAM_W, CAM_H = 320, 240
LOCK = threading.Lock()
PID_FILE = os.path.join(DATA, "server.pid")
E = {"ready": False, "step": 0, "calls": 0, "done": False, "reason": None, "success": False, "t0": time.time(), "error": None,
- "t_first": None} # t_first: the agent's first call (its clock starts there; kept through a crash by the checkpoints)
+ "t_first": None, # t_first: the agent's first call (its clock starts there; kept through a crash by the checkpoints)
+ "resets": 0, # resets done (the episode is resets + 1)
+ "reset_at": []} # the journal row at which each reset happened: a resume resets there again
os.environ["RCB_SESSION_LOG"] = ""
os.environ["RCB_AGENT_TRACE_DIR"] = ""
import rcb_quadruped as Q # noqa: E402 (first: it picks the rendering backend before mujoco loads)
import mujoco # noqa: E402
-EP = L.Episode(DATA, dim=Q.N, resume="exact", persist=("t_first",))
+EP = L.Episode(DATA, dim=Q.N, resume="exact", persist=("t_first", "resets", "reset_at"))
EP.startup() # start count, back-off after failed starts, the startup watchdog until scene_built()
S = Q.Session(INSTANCE)
ROWS: list = [] # every executed row (a resume replays them first): trajectory.npz
@@ -153,13 +160,15 @@ def schedule_text() -> list:
"wz_radps": round(c["wz"], 4)} for c in S.inst["commands"]]
+ONE = "one episode, no reset" if not RESET_BUDGET else (
+ f"an episode (with up to {RESET_BUDGET} resets: POST /reset starts a new one; only the last is graded)")
TASK_EPISODE = {
- "joystick": f"one episode, no reset: it ends after {S.horizon} control steps (the {S.horizon * DT:g} s command "
+ "joystick": f"{ONE}: it ends after {S.horizon} control steps (the {S.horizon * DT:g} s command "
f"schedule), when the robot meets its termination rule ({S.spec.get('term_text')}: a failure), when you "
"submit `done`, or when your time is over",
- "getup": f"one episode, no reset: it ends the moment the robot has stood upright for {Q.HOLD_STEPS} control steps in a "
+ "getup": f"{ONE}: it ends the moment the robot has stood upright for {Q.HOLD_STEPS} control steps in a "
f"row (success), after {S.horizon} control steps, when you submit `done`, or when your time is over",
- "stand_on_two": f"one episode, no reset: it ends the moment the pose has been held for {Q.POSE_HOLD_STEPS} control steps "
+ "stand_on_two": f"{ONE}: it ends the moment the pose has been held for {Q.POSE_HOLD_STEPS} control steps "
"in a row (success), when the robot meets the termination rule (a failure: its trunk's up axis below "
f"-0.25, or {'a leg' if S.spec.get('hind_leg_contact') else 'a front leg'}'s thigh, calf or hip on the "
f"floor), after {S.horizon} control steps, when you "
@@ -204,7 +213,7 @@ SPEC = {
"world": {"frame": "z up; the robot starts where the episode puts it (no world coordinates are given)", "up": "+z",
"units": {"length": "m", "angle": "rad", "time": "s"}, "quaternion_order": "wxyz",
"camera_axes": "intrinsic_cv / extrinsic_cv: OpenCV camera axes (x right, y down, z forward); cam2world_gl: OpenGL axes"},
- "budget": {"reset": 0, "steps": None, "wall_clock_s": WALL_CLOCK_S}, # + "clock" (rcb_limited.CLOCK_NOTE): what time_left_s counts from
+ "budget": {"reset": RESET_BUDGET, "steps": None, "wall_clock_s": WALL_CLOCK_S}, # + "clock" (rcb_limited.CLOCK_NOTE): what time_left_s counts from
"max_control_steps": S.horizon,
"episode": TASK_EPISODE,
"command_schedule": schedule_text(),
@@ -223,6 +232,15 @@ SPEC = {
"depth": {"modes": list(L.DEPTH_MODES), "default": "none", "kind": "z-depth", "unit": "m", "invalid": 0,
"max_m": L.DEPTH_MAX_M, "frames_encoding": "uint16 mm, zlib", "noise": L.DEPTH_NOISE},
}
+if RESET_BUDGET:
+ SPEC["reset"] = (
+ f"POST /reset (JSON or RCB1 {{cameras, video_fps, depth}}; no physics) abandons the running episode, or follows "
+ f"one that has ended, and starts a new one from the same frozen starting state, {RESET_BUDGET} times in all; it "
+ "returns the new episode's first frames and proprioception. Only the last episode is graded: an episode you "
+ "reset away is never graded, a success ends the trial (409 after it), and when your time is up the episode then "
+ "running is the one graded. The wall clock never restarts. /status and every meta add `episode`, `resets_left` "
+ "and `control_steps_total` (all episodes); `control_steps` counts the current episode's."
+ )
L.check_spec(SPEC) # a spec that misses a 2.2 field or names a forbidden one never starts the episode
EP.scene_built()
@@ -277,6 +295,7 @@ def summary() -> dict:
"reason": E["reason"], "done": E["done"], "env": S.env, "protocol": L.PROTOCOL, "n_actions": len(ROWS),
"fell_at": chk.get("fell_at"), "worst_lin_err": chk.get("worst_lin_err"), "worst_yaw_err": chk.get("worst_yaw_err"),
"tilt_deg": chk.get("tilt_deg"), "imu_height": chk.get("imu_height"), "pose_angle_deg": chk.get("pose_angle_deg"),
+ "episodes": E["resets"] + 1, "resets": E["resets"], "reset_budget": RESET_BUDGET, "control_steps_total": EP.rows,
**EP.final_fields()}
@@ -294,7 +313,8 @@ def finish(reason: str) -> None:
E["success"] = bool(chk["success"])
except Exception: # noqa: BLE001
traceback.print_exc()
- EP.log({"ev": "end", "reason": reason, "step": E["step"], "control_steps": ctrl_steps(), "success": bool(E["success"])})
+ EP.log({"ev": "end", "reason": reason, "step": E["step"], "control_steps": ctrl_steps(), "success": bool(E["success"]),
+ "episode": E["resets"] + 1, "journal_rows": EP.rows})
dump()
@@ -388,7 +408,7 @@ def meta_common(cams) -> dict:
"sim_time_s": round(n * DT, 3), "time_s": round(_agent_time(), 2),
"time_left_s": max(0.0, round(WALL_CLOCK_S - _agent_time(), 1)),
"proprio": {"layout": PROPRIO_LAYOUT, "now": proprio_row().tolist()}, "done": E["done"], "reason": E["reason"],
- "cameras": {c: camera_info(c) for c in cams}}
+ "cameras": {c: camera_info(c) for c in cams}, **reset_counters()}
if S.kind == "joystick":
out["command_now"] = dict(zip(("vx_mps", "vy_mps", "wz_radps"), S.command_now().round(4).tolist()))
return out
@@ -478,10 +498,54 @@ def do_observe(meta: dict):
return out, rec.parts(video if seconds > 0 else ("frames" if video == "mp4" else video), fps)
+def reset_counters() -> dict:
+ if not RESET_BUDGET:
+ return {}
+ left = 0 if (E["success"] or E.get("exiting")) else RESET_BUDGET - E["resets"]
+ return {"episode": E["resets"] + 1, "resets_left": left, "control_steps_total": EP.rows}
+
+
def status() -> dict:
_touch()
return {"step": E["step"], "calls": E["calls"], "control_steps": ctrl_steps(), "done": E["done"], "reason": E["reason"],
- "elapsed_s": round(time.time() - E["t0"], 1)}
+ "elapsed_s": round(time.time() - E["t0"], 1), **reset_counters()}
+
+
+class Refused(Exception):
+ """A /reset the service turns down (409): nothing changes."""
+
+ def __init__(self, body: dict):
+ super().__init__(body.get("error"))
+ self.body = body
+
+
+def do_reset(meta: dict):
+ """POST /reset: a new episode from the frozen instance (the running one, if any, is abandoned, never graded)."""
+ meta = _request(meta)
+ cams, fps = cameras_of(meta), video_fps(meta)
+ L.depth_mode(meta) # a bad depth value is a 400 before anything changes
+ if E["success"]:
+ raise Refused({"error": "trial finished: an episode succeeded", "reason": E["reason"], "resets_left": 0})
+ if E["resets"] >= RESET_BUDGET:
+ raise Refused({"error": "reset budget exhausted", "reason": "reset budget exhausted", "budget": RESET_BUDGET,
+ "resets_left": 0})
+ abandoned = {"episode": E["resets"] + 1, "control_steps": ctrl_steps(), "done": E["done"], "reason": E["reason"]}
+ if not E["done"]: # the abandoned episode's end, for the log (never graded: its rows leave the trajectory)
+ EP.log({"ev": "end", "reason": "reset", "step": E["step"], "control_steps": ctrl_steps(), "success": False,
+ "episode": E["resets"] + 1, "journal_rows": EP.rows})
+ S.reset()
+ ROWS.clear()
+ E["resets"] += 1
+ E["reset_at"] = list(E["reset_at"]) + [EP.rows]
+ E.update(done=False, reason=None, success=False, t_end=None, episode_control_steps=None, dumped=False, check=None)
+ EP._ck = None # the state changed without a journal row: the checkpoint after this call must hash it anew
+ E["calls"] += 1
+ EP.log({"ev": "reset", "call": E["calls"], "episode": E["resets"] + 1, "journal_rows": EP.rows,
+ "abandoned": abandoned, "state_sha": S.state_sha()})
+ rec = recorder(cams, fps, meta)
+ rec.tick(force=True, last=True)
+ out = {**meta_common(cams), "reset_id": E["resets"], "ok": True}
+ return out, rec.parts("frames", fps)
def end_from_signal() -> None:
@@ -498,20 +562,106 @@ def end_from_signal() -> None:
os._exit(0)
+# A resume replays the journal from the frozen instance and resets where the committed resets did: their journal rows
+# are in the last checkpoint (`persist`), the one Episode.begin resumes from.
+_CPS = [e for e in EP.events() if e.get("ev") == "checkpoint"]
+RESUME = {"at": list((_CPS[-1].get("persist") or {}).get("reset_at") or []) if _CPS else [], "i": 0}
+
+
+def _resume_resets() -> None:
+ while RESUME["at"] and int(RESUME["at"][0]) <= RESUME["i"]:
+ RESUME["at"].pop(0)
+ S.reset()
+ ROWS.clear()
+
+
def replay_row(row) -> None:
"""A journaled row during a resume."""
+ _resume_resets()
S.step(row)
ROWS.append(np.asarray(row, dtype=np.float64))
+ RESUME["i"] += 1
-EP.begin(state=E, step=replay_row, state_sha=S.state_sha, proprio=proprio_row, spec=SPEC, on_end=end_from_signal,
- start_event={"env": S.env, "seed": S.inst.get("seed"), "cameras": CAMS})
+def resume_sha() -> str:
+ _resume_resets() # a reset after the last replayed row (or before the first)
+ return S.state_sha()
+
+
+RESUMED = EP.begin(state=E, step=replay_row, state_sha=resume_sha, proprio=proprio_row, spec=SPEC, on_end=end_from_signal,
+ start_event={"env": S.env, "seed": S.inst.get("seed"), "cameras": CAMS})
+if RESUMED and E["resets"]: # Episode.begin looks for an "end" of any episode: only the current one's counts
+ _ends = [e for e in EP.events() if e.get("ev") == "end" and int(e.get("episode") or 1) == E["resets"] + 1
+ and e.get("reason") != "reset" and int(e.get("journal_rows") or 0) <= EP.rows]
+ if _ends:
+ E.update(done=True, reason=_ends[-1].get("reason"), success=bool(_ends[-1].get("success")),
+ t_end=float(_ends[-1]["ts"]), episode_control_steps=int(_ends[-1].get("control_steps") or ctrl_steps()))
+ else:
+ E.update(done=False, reason=None, success=False, t_end=None, episode_control_steps=None)
if E["done"]:
dump() # resumed after the episode had ended: the record is rewritten from the replay
E["ready"] = True
write_final()
open(PID_FILE, "w").write(str(os.getpid()))
print(f"{L.PROTOCOL} MuJoCo Playground quadruped sim ready ({S.env}; startup {EP.n_startups}, resumes {EP.resumes})", flush=True)
-L.serve(L.Service(spec=SPEC, lock=LOCK, state=E, submit=do_submit, observe=do_observe, kinematics=do_kinematics, status=status,
- on_submit_error=lambda exc: finish("error"), log=EP.log, episode=EP, state_sha=S.state_sha,
- proprio=proprio_row), port=int(os.environ.get("RCB_PORT", "8000")))
+
+
+def make_handler(svc: L.Service):
+ """The shared 2.2 handler, plus POST /reset when the run grants resets (the shared layer has no reset)."""
+ base = L.make_handler(svc)
+
+ class H(base):
+ def do_POST(self):
+ if self.path != "/reset" or not RESET_BUDGET:
+ return super().do_POST()
+ try:
+ meta, _ = self._body()
+ except Exception as exc: # noqa: BLE001
+ return self._json(413 if "too large" in str(exc) else 400, {"error": str(exc)[:200]})
+ with LOCK:
+ if E.get("exiting"):
+ return self._json(409, {"error": "trial finished: your time is up", "reason": E.get("reason"),
+ "resets_left": 0})
+ cid = EP.call_begin(self.path, self.connection)
+ out = blobs = err = None
+ try:
+ out, blobs = do_reset(meta)
+ except Refused as exc:
+ err = (409, exc.body)
+ except ValueError as exc:
+ err = (400, {"error": str(exc)})
+ except Exception as exc: # noqa: BLE001
+ traceback.print_exc()
+ E["error"] = repr(exc)
+ EP.log({"ev": "error", "path": self.path, "error": repr(exc)})
+ try:
+ finish("error")
+ except Exception: # noqa: BLE001
+ traceback.print_exc()
+ err = (500, {"error": "simulator error", "detail": repr(exc)[:200]})
+ finally:
+ if out is not None or (err and err[0] == 500):
+ try:
+ EP.checkpoint(E, S.state_sha, proprio_row)
+ except Exception: # noqa: BLE001
+ traceback.print_exc()
+ EP.call_end(cid, self.path)
+ if err is not None:
+ return self._json(*err)
+ return self._rcb1({**out, "stopped": None}, blobs)
+
+ return H
+
+
+def serve(svc: L.Service, port: int) -> None:
+ """L.serve with this service's handler (the same spec defaults and check, and L.bind: a port still held when the
+ service restarts is waited for, not a crash)."""
+ svc.spec.setdefault("on_service_error", L.ON_SERVICE_ERROR)
+ svc.spec.setdefault("max_call_wall_s", EP.max_call_wall_s)
+ L.check_spec(svc.spec)
+ L.bind(port, make_handler(svc)).serve_forever()
+
+
+serve(L.Service(spec=SPEC, lock=LOCK, state=E, submit=do_submit, observe=do_observe, kinematics=do_kinematics, status=status,
+ on_submit_error=lambda exc: finish("error"), log=EP.log, episode=EP, state_sha=S.state_sha,
+ proprio=proprio_row), port=int(os.environ.get("RCB_PORT", "8000")))
diff --git a/scripts/mujoco-playground/templates/README.md b/scripts/mujoco-playground/templates/README.md
index af89772958..0c6e975cde 100644
--- a/scripts/mujoco-playground/templates/README.md
+++ b/scripts/mujoco-playground/templates/README.md
@@ -29,3 +29,11 @@ succeed and end in the same state (`deterministic`). {{check_short}} `nop`: no t
### Provenance
Generated by `scripts/mujoco-playground/generate.py` from `scripts/mujoco-playground/instances/{{env}}.json`
(`make_instances.py`, `select_instances.py`) and `scripts/mujoco-playground/oracles/{{env}}.npy` (`oracle.py`).
+
+### Change Log
+- 2026-10-14 (review of board 0105, B1-m2): `tests/grade.py` only. A replay process that died or ran out of time, a
+ rebuilt start that does not match the instance's hash (`instance_mismatch`) and an exception of the verifier itself
+ now write `infra_error = 1` beside `grader_error = 1`: no verdict on the agent. Before, they wrote `grader_error = 1`
+ alone, which counted as the agent's 0. Every `reward.json` now carries `infra_error`. The scene, the instance, the
+ pass rule and the reference are unchanged, so the task keeps its name; a past 0 with `grader_error = 1` was an
+ infrastructure failure.
diff --git a/scripts/mujoco-playground/templates/standard/README.md b/scripts/mujoco-playground/templates/standard/README.md
index 39fd614824..291968f8db 100644
--- a/scripts/mujoco-playground/templates/standard/README.md
+++ b/scripts/mujoco-playground/templates/standard/README.md
@@ -2,14 +2,14 @@
## Task
MuJoCo Playground `{{env}}` ({{robot_name}}) on the same frozen instance as `{{name}}-privileged`: {{task}}
-{{task_detail}} Standard mode (eai-standard/2.2): the agent drives the robot through a client in one episode, with only
-the robot's own sensors and four cameras that follow it.
+{{task_detail}} Standard mode (eai-standard/2.2): the agent drives the robot through a client, with {{resets}} resets (only
+the last episode is graded), the robot's own sensors and the cameras that follow it.
## Difficulty
-Same instance and success check as `{{name}}-privileged`. In standard mode the agent loses resets, the simulator and
+Same instance and success check as `{{name}}-privileged`. In standard mode the agent loses free resets, the simulator and
the world state: it sees camera images and proprioception (joints, IMU, foot contacts), acts in joint position chunks,
-and gets one episode, so it has to close the loop on the robot's own sensors and recover from its own mistakes without
-starting over.
+and gets {{resets}} resets (`RCB_RESET_BUDGET`, the compose file's default; PROTOCOL §4.5), so it has to close the loop
+on the robot's own sensors; only the last episode is graded.
## Reference solution
`solution/oracle_replay.py` streams the privileged reference's rows (`solution/oracle.npy`, {{n_rows}} rows, float32-exact)
diff --git a/scripts/mujoco-playground/templates/standard/instruction.md b/scripts/mujoco-playground/templates/standard/instruction.md
index 1fa1868cd8..d6148be489 100644
--- a/scripts/mujoco-playground/templates/standard/instruction.md
+++ b/scripts/mujoco-playground/templates/standard/instruction.md
@@ -2,14 +2,22 @@
**Task:** {{task}}
-You control the {{robot_name}} quadruped — **as if it were real hardware** — in one episode of MuJoCo Playground's
+You control the {{robot_name}} quadruped — **as if it were real hardware** — in MuJoCo Playground's
`{{env}}` environment (plain MuJoCo 3.3.7 with Playground's scene, commit 4057c14). The simulator runs in a service you
-cannot open. You get one client and one episode.
+cannot open. You get one client and up to {{episodes}} episodes.
{{episode_md_std}}
## Rules (real-robot rules)
-- **No reset.** One episode. You see four cameras that follow the robot (RGB, depth if you ask) and the robot's own
+- **Episodes and resets.** You have **{{resets}} resets** (`spec["budget"]["reset"]`, which is authoritative: if it
+ says 0, there is no reset and the one episode is the run). A reset (`POST /reset`, below) puts the robot back on the
+ same frozen starting state and starts a new episode, either abandoning the running one or after one has ended; it
+ takes no simulated time and returns the start's images and proprioception. Once an episode has ended, `submit`
+ raises `EpisodeOver` until you reset. **Only the last episode is graded**: an episode you reset away is never graded,
+ a success ends the run at once (no reset after it), and when your time is up the episode then running is the one
+ graded. The wall clock is the run's and never restarts; `meta["episode"]` and `meta["resets_left"]` count the
+ episodes (`meta["control_steps"]` is the current episode's).
+- **No privileged state.** You see the cameras that follow the robot (RGB, depth if you ask) and the robot's own
sensors: joint positions, velocities and torques, the IMU (gyro, accelerometer, gravity's direction, the trunk's
velocity), foot contacts. No world positions, no success check.
- **The episode ends** {{std_end}}
@@ -18,8 +26,8 @@ cannot open. You get one client and one episode.
`get_obs(seconds=...)`, which holds the current targets). `meta["time_left_s"]` is the service's count of the limit,
from your first request to the service: it never runs out before your time does, so plan by it.
- Nothing to hand in: the service records the episode and grades it. Leave your script at `/app/output/gen.py`.
-- **If the service crashes**, a call fails with a connection error: it restarts and resumes the SAME episode (never a
- reset) in the state after the last call it completed. `r.wait_ready()` waits for it; then check `r.status()`.
+- **If the service crashes**, a call fails with a connection error: it restarts and resumes the SAME episode (a crash
+ is never a reset and costs none) in the state after the last call it completed. `r.wait_ready()` waits for it; then check `r.status()`.
- Nothing is put into your context automatically: images come back as arrays in Python objects.
## Client (`/opt/tools/rcb_client.py`)
@@ -32,7 +40,8 @@ res = r.submit("joint_position", A, dt={{dt}}) # A: (T, 12) rows dt apart; res
obs = r.get_obs(seconds=1.0, video="mp4") # hold the current targets and watch for 1 s (max 30)
r.fk(q); r.ik({"{{foot0}}": [x, y, z, 1, 0, 0, 0]}, seed=q) # feet in the trunk frame <-> joints (no physics)
r.submit("done", []) # end the episode now
-r.status() # calls, control steps, done, reason, elapsed time
+r.status() # calls, control steps, done, reason, elapsed time, episode, resets_left
+obs = r._post("/reset", {"cameras": ["track"]}, [], 900) # POST /reset: a new episode (the 2.2 client has no reset())
```
Cameras: {{cameras}}; `spec["cameras"]` gives each one's mounting and intrinsics, and every response carries its pose. Proprioception (`spec["proprio_groups"]`): `joint_pos`, `joint_vel`, `joint_effort` (12 each, action
order), `imu_gyro`, `imu_accelerometer`, `imu_gravity`, `imu_linvel` (3 each, IMU frame: x forward, y left, z up),
diff --git a/scripts/mujoco-playground/templates/standard/task.toml b/scripts/mujoco-playground/templates/standard/task.toml
index 1103d99c70..a485d5ce17 100644
--- a/scripts/mujoco-playground/templates/standard/task.toml
+++ b/scripts/mujoco-playground/templates/standard/task.toml
@@ -16,7 +16,7 @@ artifacts = [
[task]
name = "robot-coding-bench/{{name}}-standard"
-description = "STANDARD MODE (robot as a service, eai-standard/2.2): MuJoCo Playground {{env}} ({{robot_name}}) on the same frozen instance as {{name}}-privileged: {{task_toml}} The agent only gets a client: four cameras that follow the robot (RGB, optional depth), the robot's joints and IMU, joint position chunks at 50 Hz, the feet's kinematics; one episode, no reset, no world state; the sim service records the episode, replays it in fresh processes and grades it."
+description = "STANDARD MODE (robot as a service, eai-standard/2.2): MuJoCo Playground {{env}} ({{robot_name}}) on the same frozen instance as {{name}}-privileged: {{task_toml}} The agent only gets a client: four cameras that follow the robot (RGB, optional depth), the robot's joints and IMU, joint position chunks at 50 Hz, the feet's kinematics; {{resets}} resets (only the last episode is graded), no world state; the sim service records the episode, replays it in fresh processes and grades it."
authors = [{ name = "{{author}}" }]
keywords = ["robotics", "mujoco", "mujoco-playground", "{{robot_tax}}", "quadruped", "locomotion", "{{kind}}", "code-as-policy", "trajectory", "standard-mode"]
diff --git a/scripts/mujoco-playground/tests/test_grade_privileged.py b/scripts/mujoco-playground/tests/test_grade_privileged.py
index 326818a1ad..99bd0ecc14 100644
--- a/scripts/mujoco-playground/tests/test_grade_privileged.py
+++ b/scripts/mujoco-playground/tests/test_grade_privileged.py
@@ -1,9 +1,15 @@
-"""The privileged quadruped verifier (shared/grade.py) and the agent's session log: standard library only.
+"""The privileged quadruped verifier (shared/grade.py): the agent's session log, and infrastructure vs agent failure.
+Standard library only.
python3 scripts/mujoco-playground/tests/test_grade_privileged.py
The log is self-reported cost, written in the agent's container: whatever bytes it holds, the verdict is written as
without it (a log that was not UTF-8 used to end the verifier before any reward file).
+
+The replays run `rcb_quadruped.py replay` in fresh processes; here a stub of it (next to a copy of grade.py) answers as
+the real one does on each path, and a stub numpy reads the trajectory (a JSON file). A replay process that dies, or a
+start that does not rebuild, is no verdict on the agent (`infra_error` = 1, board 0105); the agent's own failure is a 0
+with `infra_error` = 0.
"""
import json
import os
@@ -34,8 +40,76 @@ class SessionLog(unittest.TestCase):
self.assertEqual(r.returncode, 0, r.stderr[-400:])
reward = json.loads((log / "reward.json").read_text())
self.assertEqual((log / "reward.txt").read_text().strip(), "0")
- self.assertEqual((reward["missing_trajectory"], reward["grader_error"], reward["agent_sessions"],
- reward["agent_resets"]), (1, 0, 2, 1), reward)
+ self.assertEqual((reward["missing_trajectory"], reward["grader_error"], reward["infra_error"],
+ reward["agent_sessions"], reward["agent_resets"]), (1, 0, 0, 2, 1), reward)
+
+
+STUB_NUMPY = """import json
+float64 = float
+class _A:
+ def __init__(self, rows):
+ self.rows = rows; self.ndim = 2; self.shape = (len(rows), len(rows[0]) if rows else 0)
+ def __len__(self):
+ return len(self.rows)
+class _D(dict):
+ files = property(lambda self: list(self))
+def load(path, allow_pickle=False):
+ return _D(json.load(open(path)))
+def asarray(x, dtype=None):
+ return x if isinstance(x, _A) else _A(x)
+def zeros(shape):
+ return _A([])
+def isfinite(a):
+ return True
+def all(x):
+ return bool(x)
+"""
+# the replay tool: STUB_MODE = ok (success, same state), fail (the agent's 0), crash (no result: the process died),
+# mismatch (the start did not rebuild), bad (the rows were refused: bad_trajectory)
+STUB_TOOL = """import json, os, sys
+a = sys.argv; mode = os.environ["STUB_MODE"]
+if mode == "crash":
+ sys.exit(3)
+out = {"instance_ok": mode != "mismatch", "success": mode in ("ok", "mismatch"), "state_sha": "s", "video": False}
+if mode == "bad":
+ out.update(success=False, bad_trajectory=1, error="PrimitiveError()")
+json.dump(out, open(a[a.index("--out") + 1], "w"))
+"""
+
+
+class InfraVsAgent(unittest.TestCase):
+ def grade(self, mode: str) -> tuple[str, dict]:
+ with tempfile.TemporaryDirectory() as d:
+ d = Path(d); log = d / "log"
+ (d / "grade.py").write_text(GRADE.read_text().replace('"/logs/agent/rcb/rcb_session_log.jsonl"', repr(str(d / "x")))
+ .replace('"/app/output/rcb_session_log.jsonl"', repr(str(d / "x"))))
+ (d / "numpy.py").write_text(STUB_NUMPY)
+ (d / "rcb_quadruped.py").write_text(STUB_TOOL)
+ (d / "instance.json").write_text(json.dumps({"horizon": 1000}))
+ (d / "trajectory.npz").write_text(json.dumps({"actions": [[0.0] * 12] * 5}))
+ r = subprocess.run([sys.executable, "-B", str(d / "grade.py"), "--instance", str(d / "instance.json"),
+ "--traj", str(d / "trajectory.npz")], capture_output=True, text=True, cwd=d,
+ env={**os.environ, "RCB_VERIFIER_LOG_DIR": str(log), "STUB_MODE": mode,
+ "PYTHONPATH": str(d)})
+ self.assertEqual(r.returncode, 0, r.stderr[-400:])
+ return (log / "reward.txt").read_text().strip(), json.loads((log / "reward.json").read_text())
+
+ def test_success(self):
+ txt, rw = self.grade("ok")
+ self.assertEqual((txt, rw["success"], rw["deterministic"], rw["infra_error"], rw["grader_error"]), ("1", 1, 1, 0, 0), rw)
+
+ def test_the_agents_failure_is_its_0(self):
+ for mode in ("fail", "bad"):
+ txt, rw = self.grade(mode)
+ self.assertEqual((txt, rw["success"], rw["infra_error"], rw["grader_error"]), ("0", 0, 0, 0), (mode, rw))
+
+ def test_a_replay_process_without_a_result_is_no_verdict(self):
+ txt, rw = self.grade("crash")
+ self.assertEqual((txt, rw["success"], rw["infra_error"], rw["grader_error"], rw["replay_failed"]), ("0", 0, 1, 1, 1), rw)
+
+ def test_a_start_that_does_not_rebuild_is_no_verdict(self):
+ txt, rw = self.grade("mismatch")
+ self.assertEqual((txt, rw["success"], rw["infra_error"], rw["grader_error"], rw["instance_mismatch"]), ("0", 0, 1, 1, 1), rw)
if __name__ == "__main__":
diff --git a/scripts/myosuite/generate.py b/scripts/myosuite/generate.py
index fe9bad422f..9cc35de53b 100644
--- a/scripts/myosuite/generate.py
+++ b/scripts/myosuite/generate.py
@@ -1,5 +1,6 @@
#!/usr/bin/env python3
-"""Generate the MyoSuite tasks: tasks/myosuite-<family>-i00-privileged and ...-standard (10 families x 2 modes).
+"""Generate the MyoSuite tasks: tasks/myosuite-<family>-i00-privileged and ...-standard (every family), and
+...-standard-realtime for the families whose scene is dynamic (REALTIME; PROTOCOL §4b, §2.14).
python3 scripts/myosuite/generate.py # every family in instances/
python3 scripts/myosuite/generate.py --check # exit 1 if tasks/ differs from what this would write
@@ -36,7 +37,7 @@ AUTHOR = "Chi Gui"
# The agent's time cutoff per mode: scripts/task_cutoffs.toml, the one table (PROTOCOL §2.12; why not the default:
# docs/benchmarks/myosuite.md). No copy of the number here: normalize writes it into task.toml, the instruction's first
# paragraphs, the README row and the sim service's wall clock.
-CUTOFF = {m: cutoff_sec("MyoSuite", m) for m in ("privileged", "standard")}
+CUTOFF = {m: cutoff_sec("MyoSuite", m) for m in ("privileged", "standard", "standard-realtime")}
# The standard service's bound on the images of one call, quoted from the source: the texts state the enforced value.
MAX_FRAMES = int(re.search(r"^MAX_FRAMES = (\d+)", (SHARED / "standard" / "server.py").read_text(), re.M).group(1))
RESET_BUDGET = 10 # standard mode's default reset budget (owner, 2026-10-09; PROTOCOL §4.5, docs/benchmarks/myosuite.md)
@@ -54,7 +55,7 @@ BODY = {
"a musculoskeletal right arm and hand next to a motorized prosthetic left arm (the Modular Prosthetic Limb)"),
"legs": ("MyoLegs", "myolegs", "a free-standing musculoskeletal model of a human torso and both legs"),
"osl": ("MyoLeg + OSL", "myoleg-osl",
- "a free-standing musculoskeletal model of a human whose left leg is an Open-Source Leg knee-ankle prosthesis"),
+ "a free-standing musculoskeletal model of a human whose right leg is an Open-Source Leg knee-ankle prosthesis"),
}
DEFAULT_CAMERA = {"table_tennis": "default", "bimanual": "front_view", "soccer": "agent_view", "chase_tag": "topview",
"osl_run": "side_view", **{f"dex_{k}": "center_cam" for k in ("apple_lift", "cylinder_pass", "bunny_inspect",
@@ -214,6 +215,344 @@ services:
"""
+# ------------------------------------------------------------------------------------------- standard-realtime twins
+# The dynamic families (scripts/task_dynamics.toml; PROTOCOL §2.14) get a `-standard-realtime` twin (owner, 2026-10-14,
+# board 0112): the same frozen instance and success rule, served by shared/realtime/server.py (the rcb_service.Sim
+# adapter on rcb_myosuite.Session) behind the shared eai-standard/2.3 layer (rcb_realtime.py + rcb_service.py, copied
+# unchanged from scripts/humanoidbench/shared/; check-tooling-identity.sh). Per family: the slowdown S its service runs
+# at, from the real-time factor RTF measured on the reference workstation (PROTOCOL §4b: S = 1 when RTF >= 3, else
+# ~1.5 / RTF rounded up), and how it was measured (docs/benchmarks/myosuite.md, "Standard-realtime").
+REALTIME = {
+ "table_tennis": {
+ "slowdown": 1,
+ "rtf": "2.1-2.7",
+ "measured": "2026-10-14, board 0112, ws-1 (32 cores, load 30-60), the service capped at 4 CPUs: 2.08 with "
+ "the realtime service's clock flat out (RCB_SLOWDOWN = 0.01, 11 episodes, 13.3 s of robot time), 2.6-2.7 "
+ "for the bare physics; S = ceil(1.5 / 2.08) = 1. At S = 1 a controller at the control rate with a "
+ "camera image every 0.25 s ran 11 episodes with no stall (largest lag 0.23 s), and an EGL frame takes "
+ "5 ms (an OSMesa one 1.3 s)",
+ "evidence": "a successful episode of the privileged twin (GPT-6 Luna xhigh, trial "
+ "`myosuite-table-tennis-i00-privil__SMNgbHk`, 82 rows) planned from step 0 in one "
+ "`submit(..., at=0)` at S = 1 succeeds live and in the fresh replay, with no stall (board 0112, "
+ "2026-10-14); the standard twin's evidence holds for the same instance and rule",
+ "rally": "the serve reaches you and the rally is decided within about 1.2 s of robot time after the clock starts "
+ "(an episode that does nothing ends after 121 control steps)",
+ },
+}
+RT_REVIEW_PR = "201" # the pull request that adds the standard-realtime twins (human review, PROTOCOL §2.7)
+RT_RESET_BUDGET = 10 # PROTOCOL §4b: every standard-realtime task allows resets, 10 by default (owner, 2026-10-14)
+RT = SHARED / "realtime"
+RT_ENV_DOCKERFILE = f"""# Standard-realtime mode: the agent container has NO simulator and no MyoSuite code, only Python and the realtime client
+# (eai-standard/2.3), on the shared pinned client base with the shared pins (scripts/eai_standard/client.env).
+FROM {CLIENT_IMAGE}
+COPY client-requirements.txt /tmp/client-requirements.txt
+RUN pip install --no-cache-dir -r /tmp/client-requirements.txt
+COPY rcb_client.py rcb1.py /opt/tools/
+ENV PYTHONPATH=/opt/tools
+RUN mkdir -p /app/output
+WORKDIR /app
+"""
+
+
+def rt_compose(fam: str) -> str:
+ S = REALTIME[fam]["slowdown"]
+ return f"""# Standard-realtime mode (PROTOCOL §4b): the simulator runs in the `sim` sidecar (the agent cannot open it), on its own clock
+# once an episode starts (sim/rcb_realtime.py, the shared eai-standard/2.3 layer, with sim/server.py, MyoSuite's adapter);
+# the agent only reaches port 8000. No `expose`/`ports` on sim: under Harbor's egress control it shares main's network
+# namespace (localhost:8000).
+# GPU on sim: the physics is MuJoCo on the CPU, but every image the agent asks for is rendered here, and on a clock that
+# does not stop a render is time the simulator cannot step: EGL on the GPU takes ~5 ms a frame, OSMesa ~1.3 s
+# (table_tennis, 640x480). `graphics` is the capability that mounts the EGL driver (server.py writes its ICD manifest).
+# Crash recovery: compose restarts the service, which resumes the same episode, paused, never a reset.
+# Knobs: ${{NAME:-default}} takes NAME from the shell that runs harbor. The defaults are the task's configuration;
+# results under other settings are reported separately.
+services:
+ main:
+ depends_on: [sim]
+ sim:
+ image: {IMAGE}
+ volumes:
+ - ./sim:/srv:ro
+ working_dir: /tmp
+ command: python /srv/rcb_realtime.py
+ restart: on-failure:6 # N >= 3 (PROTOCOL §4.1); a resume continues the episode, paused
+ environment:
+ RCB_INSTANCE: /srv/instance.npz
+ RCB_TASK_INFO: /srv/task.json
+ RCB_DATA: /data
+ RCB_WALL_CLOCK_S: "{CUTOFF["standard-realtime"]}"
+ RCB_STARTUP_TIMEOUT_S: "300" # the scene builds in a few seconds
+ PYTHONPATH: /srv
+ RCB_SESSION_LOG: "" # the sidecar's own Session must not write the agent-side interaction log
+ RCB_SLOWDOWN: "${{RCB_SLOWDOWN:-{S}}}" # S = {S}: the world runs at 1/S of real speed (this family's measured RTF; PROTOCOL §4b)
+ RCB_RESET_BUDGET: "${{RCB_RESET_BUDGET:-{RT_RESET_BUDGET}}}" # {RT_RESET_BUDGET} resets (PROTOCOL §4b default); 0: one episode
+ RCB_GL: "${{RCB_GL:-auto}}" # auto: EGL on the GPU when it works, else OSMesa
+ NVIDIA_DRIVER_CAPABILITIES: compute,graphics,utility
+ deploy:
+ resources:
+ reservations:
+ devices:
+ - driver: nvidia
+ count: 1
+ capabilities: [gpu]
+"""
+
+
+def rt_instruction(fam: str, f: dict) -> str:
+ b, rt = BODY[f["body"]], REALTIME[fam]
+ S = rt["slowdown"]
+ hz = round(1 / f["dt"])
+ ms = round(f["dt"] * 1000, 1)
+ ms = int(ms) if ms == int(ms) else ms
+ atype = "muscle" if not f["actuators"] else "actuators"
+ root = (", the pose and velocity of your free-standing root (the pelvis: `root_*`)" if f["free_root"] else "")
+ return f"""# MyoSuite · {fam} — STANDARD-REALTIME MODE (robot as a service; time does not stop)
+
+You control the **{b[0]}**, {b[2]}, **as if it were real**: the simulator (MyoSuite 2.12.2 on MuJoCo) runs in a service
+you cannot open, you reach it only through a client, and like the real world it does not wait for you: once an episode
+is running, the body and the ball move while you think. You act only through its actuators ({f["n_muscles"]} muscles{
+f" and {len(f['actuators'])} other actuators" if f["actuators"] else ""}).
+**Task:** {f["goal"]}
+
+{f["scene"]} Here the rally is fast: {rt["rally"]}.
+
+**Slowdown: S = {S}.** The world runs at 1/S of real speed: one control step ({ms} ms of robot time) every S x {ms} ms of
+wall-clock time{" (real time)" if S == 1 else ""}. `spec()["timing"]["slowdown"]` states it.
+
+## Rules (real-world rules)
+- **Time does not stop.** An episode starts paused, on the frozen starting state, with every command at -1 (muscles
+ relaxed). Its clock starts at your first `submit` (or `get_obs(seconds=s)`) and from then on the simulator runs at
+ {hz} Hz of robot time (S = {S}), one control step every {ms * S:g} ms of wall-clock time, whether or not you send anything, until
+ the episode ends. Between your commands it executes what you planned, then holds the last commands you sent; nothing
+ pauses for you, not your thinking, not your tool calls.
+- **You are far too slow to steer it turn by turn.** One turn of yours (a model response, a tool call) takes seconds;
+ the whole rally takes about a second. Drive the body with a program: a controller of your own (Python in your
+ container, with numpy and the client) that plans or streams the commands, started in the background
+ (`nohup python /app/output/gen.py > /app/output/run.log 2>&1 &`) or in the foreground if it ends with the episode. Use
+ your turns to start episodes, read what happened, and improve the program.
+- **Episodes and resets.** An episode ends at the task's own success event, at its failure conditions (see the task),
+ when your time is up, or on a simulator error. You have **{RT_RESET_BUDGET} resets** (`spec()["budget"]["reset"]`,
+ which is authoritative: if it says 0, there is no reset and the one episode is the run). `r.reset()` puts the body,
+ the ball and the scene back on the same frozen starting state and starts a new episode, paused until your next
+ `submit`, either abandoning the running episode or after one has ended. Once an episode has ended, `submit` raises
+ `EpisodeOver` (its message gives `resets_left`) until you reset. **Only the last episode is graded**: an episode you
+ reset away is never graded, a success ends the run at once (no reset after it), and when your time is up the episode
+ then running is the one graded. The wall clock is the run's and never restarts; `meta["episode"]` and
+ `meta["resets_left"]` count the episodes.
+- **No privileged state, no score.** You see what the cameras see (RGB, and depth on request) and the body's own
+ proprioception: joint positions and velocities, muscle activations and lengths, the actuators' generalized force on
+ each joint{root}. No ball, paddle, table or goal poses, no success check: the service judges the episode and only
+ tells you that it ended and why (`meta["done"]`, `meta["reason"]`: `"success"`, `"error"` after a simulator error,
+ else `"failed"`). `spec()` gives the body's own facts: the actuators, the joints and their limits, each camera's
+ intrinsics and mounting.
+- There is nothing to hand in: the service records every control step it ran, replays the last episode in a fresh
+ simulator and grades the replay. Keep the controller that drives the body at `/app/output/gen.py` for the record:
+ save it early and keep it current.
+- Do not stop early: while the episode runs, or a reset is left, the run can still pass.
+
+## Time and streaming
+- `submit` answers at once, with the body as it is now: the rows you sent have not run yet. They become the plan from
+ control step `at` (an episode step number; by default the next step), replacing whatever was planned from `at` on and
+ keeping what is planned before it. When the plan runs out, the last commands are held. `res.meta["control_steps"]`
+ is the number of steps run so far in this episode (so the next step has that number), `res.meta["proprio"]["now"]`
+ the proprioception now, `res.meta["plan_end_step"]` where the plan now ends.
+- Open loop: plan a whole movement at once (`dt` a multiple of {f["dt"]:g} s; a chunk covers at most 10 s), from step
+ `at = 0` for the episode's start: `r.submit("{atype}", rows, dt={f["dt"]:g}, interpolate="hold", at=0, cameras=[])`
+ starts the clock and runs exactly those rows from the first step; `at` may be up to
+ `spec()["timing"]["max_ahead_steps"]` steps ahead.
+- Closed loop: send one row per control step, `r.submit("{atype}", a[None], cameras=[])`, computed from the
+ proprioception the previous call returned. `cameras=[]` skips the images (the fastest call).
+- Latency: a reading is a few milliseconds old when it reaches you, and a row takes effect at the next control step at
+ the earliest. A row whose step has already run when the request arrives is dropped (`res.meta["late_steps"]` counts
+ them); the rest of its chunk applies from the next step.
+- `get_obs()` reads the body now; `get_obs(seconds=s)` (at most 30) waits while s seconds of robot time run and returns
+ what the cameras saw meanwhile (the body keeps executing its plan). `spec()["timing"]` states all this; `r.status()`
+ shows `running` and `lag_s`.
+
+## Client (`/opt/tools/rcb_client.py`, protocol `eai-standard/2.3`, standard-realtime mode)
+Your whole interface to the body. Its docstring is the full reference (`python -c "import rcb_client;
+help(rcb_client)"`); the calls and the endpoints they use:
+
+| call | endpoint | what it does |
+|---|---|---|
+| `r = Robot()` | | connects (`http://sim:8000`, then `http://localhost:8000`) and waits until the scene is ready |
+| `r.spec()` | `GET /spec` | the body's facts: the actuators' order (`action_types["{atype}"]["layout"]`), `control_dt` ({f["dt"]:g} s), `timing`, the cameras, `proprio_layout`, `budget` |
+| `r.status()` | `GET /status` | `calls`, `rows`, `control_steps`, `done`, `reason`, `elapsed_s`, `running`, `planned_steps`, `lag_s`, and `episode`, `resets_left`, `control_steps_total` (all episodes) |
+| `r.submit("{atype}", A, dt={f["dt"]:g}, interpolate="linear", at=None, cameras=None)` | `POST /submit` | plans `A`, rows of {f["nu"]} commands `dt` seconds apart (a chunk covers at most 10 s), from step `at`; answers at once; the first one starts the episode's clock |
+| `r.get_obs(cameras=None, seconds=0, include_depth=False, video="frames")` | `POST /observe` | the body now, or what the cameras saw over the next `seconds` |
+| `r.reset(cameras=None, include_depth=False)` | `POST /reset` | a new episode from the frozen starting state, paused (at most {RT_RESET_BUDGET} in all); returns the start's frames |
+
+The answers are binary (RCB1) and come back decoded: `res.meta` (the counters, `done`, `reason`, `proprio["now"]`),
+`res.last[camera]` (the last frame), `res.frames` or `res.video` (what the cameras saw during a watch), `res.proprio`
+(proprioception rows). A malformed request raises `ValueError` and changes nothing.
+```python
+from rcb_client import Robot
+import numpy as np
+r = Robot()
+rows = -np.ones((80, {f["nu"]})) # 0.8 s of commands, one row per control step
+res = r.submit("{atype}", rows, dt={f["dt"]:g}, interpolate="hold", at=0, cameras=[]) # starts the clock
+while not r.status()["done"]:
+ ... # or stream: r.submit("{atype}", a[None], cameras=[]) each step
+```
+Cameras: {", ".join(f"`{c}`" for c in f["cameras"])} (640x480, default `{default_camera(fam, f)}`), rendered on the
+service's GPU in a few milliseconds each; `cameras=None` means every camera, so name the ones you need.
+
+Each proprioception row (`meta["proprio"]["now"]`, `res.proprio`) has, in `spec()["proprio_layout"]` order: `q_<joint>`,
+`qd_<joint>`, `act_<muscle>`, `len_<muscle>` (m), `effort_<joint>` (the actuators' generalized force on the joint){
+", `root_*`" if f["free_root"] else ""}.
+
+### Action type `{atype}` ({f["nu"]})
+{actuation(f)} Columns in `spec()["action_types"]["{atype}"]["layout"]` order; rows `dt` apart, `interpolate="linear"`
+(default: a straight line from the commands in force) or `"hold"`; the client sends them as float32. The control rate
+is {hz} Hz ({f["frame_skip"]} physics substeps).
+"""
+
+
+def rt_task_toml(fam: str, f: dict) -> str:
+ b, rt = BODY[f["body"]], REALTIME[fam]
+ oracle, ev = "none", rt["evidence"] + f"; human review in PR #{RT_REVIEW_PR}"
+ first = f["goal"].split(". ")[0]
+ keywords = ["robotics", "myosuite", "mujoco", "musculoskeletal", b[1], "trajectory", fam, "standard-realtime-mode"]
+ desc = (f"STANDARD-REALTIME MODE (robot as a service on a clock that does not stop, eai-standard/2.3, slowdown S = "
+ f"{rt['slowdown']}): MyoSuite {fam} ({b[0]}, {f['n_muscles']} muscles) on the same frozen instance as "
+ f"myosuite-{dash(fam)}-i00-privileged: {first}. Once an episode starts the body and the ball move whatever the "
+ "agent does, so the agent plans or streams actuator commands from a program of its own through the client "
+ f"(cameras, the body's joints and muscles); {RT_RESET_BUDGET} resets, only the last episode graded; no object "
+ "state; the sim service records every executed control step, replays the last episode in a fresh simulator and grades it.")
+ oracle_lines = f'oracle = "{oracle}"\n' + (f"oracle_evidence = {toml_str(ev)}\n" if ev else "")
+ return f"""# Generated by scripts/myosuite/generate.py -- edit the generator, the shared sources or the instance facts, not this.
+schema_version = "1.3"
+artifacts = [
+ {{ source = "/data/final.json", service = "sim" }},
+ {{ source = "/data/episode.jsonl", service = "sim" }},
+ {{ source = "/data/trajectory.npz", service = "sim" }},
+ {{ source = "/data/trajectory_live.f64", service = "sim" }},
+ {{ source = "/data/replay.mp4", service = "sim" }},
+ {{ source = "/data/replay.last.png", service = "sim" }},
+ {{ source = "/data/ticks.f64", service = "sim" }},
+ {{ source = "/data/calls.jsonl", service = "sim" }},
+ {{ source = "/data/submitted.f64", service = "sim" }},
+ {{ source = "/data/startups.jsonl", service = "sim" }},
+ "/app/output/",
+]
+
+[task]
+name = "robot-coding-bench/myosuite-{dash(fam)}-i00-standard-realtime"
+description = {toml_str(desc)}
+authors = [{{ name = "{AUTHOR}" }}]
+keywords = [{", ".join(toml_str(k) for k in keywords)}]
+
+[metadata]
+author_name = "{AUTHOR}"
+task_type = "trajectory"
+simulator = "myosuite" # MyoSuite 2.12.2 on MuJoCo 3.6.0 (CPU physics; EGL rendering on the GPU)
+robot = "{b[1]}"
+family = "{fam}"
+mode = "standard-realtime"
+protocol = "eai-standard/2.3"
+{oracle_lines}instance_seed = 0 # {f['env_id']}: seed(0), then reset(seed=0)
+expert_time_estimate_hours = {f['hours'] + 2.0}
+
+[verifier]
+timeout_sec = 900.0
+environment_mode = "separate"
+network_mode = "no-network"
+
+# The agent's time is up: stop whatever it left running in its container. Harbor stops that container only after
+# collecting from it (~12 s), and a script still running there would keep driving the body past the deadline.
+[[verifier.collect]]
+service = "main"
+command = "kill -9 -1 2>/dev/null; exit 0"
+timeout_sec = 30
+
+# Then the episode ends: SIGUSR1 to the service (re-sent every 30 s, in case it was restarting), which grades it in a
+# fresh simulator. The verdict ("finalized": true, set before any video is rendered) may take until 840 s; then the
+# hook waits for the replay video ("video_pending") until 870 s -- a video that is not ready by then is lost, the
+# verdict is not.
+[[verifier.collect]]
+service = "sim"
+command = "t0=$(date +%s); while [ ! -f /data/server.pid ] && [ $(( $(date +%s) - t0 )) -lt 120 ]; do sleep 2; done; i=0; while [ $(( $(date +%s) - t0 )) -lt 840 ] && ! grep -q '\\"finalized\\": true' /data/final.json 2>/dev/null; do [ $((i % 15)) -eq 0 ] && kill -USR1 $(cat /data/server.pid 2>/dev/null) 2>/dev/null; i=$((i + 1)); sleep 2; done; while [ $(( $(date +%s) - t0 )) -lt 870 ] && grep -q '\\"video_pending\\": true' /data/final.json 2>/dev/null; do sleep 2; done; cat /data/final.json"
+timeout_sec = 900
+
+[agent]
+timeout_sec = {CUTOFF["standard-realtime"]}.0
+# The agent reaches the model APIs and the package registries, nothing else (scripts/set_agent_egress.py HOSTS), as in
+# every MyoSuite task (PROTOCOL §2.11).
+network_mode = "allowlist"
+allowed_hosts = [{", ".join(toml_str(h) for h in HOSTS)}]
+
+[environment]
+build_timeout_sec = 1800.0
+cpus = 4
+memory_mb = 8192
+"""
+
+
+def rt_readme(fam: str, f: dict) -> str:
+ b, rt = BODY[f["body"]], REALTIME[fam]
+ name = f"myosuite-{dash(fam)}-i00-standard-realtime"
+ S = rt["slowdown"]
+ task = (f"MyoSuite `{f['env_id']}` ({b[0]}, {f['n_muscles']} muscles) on the same frozen instance as "
+ f"`myosuite-{dash(fam)}-i00-privileged`: {f['goal']}\n{f['scene']} Standard-realtime mode (eai-standard/2.3, "
+ f"slowdown S = {S}): once an episode's clock starts the body and the ball move whatever the agent does; it "
+ f"drives the body through the realtime client from a program of its own, with only its own proprioception "
+ f"and the scene's cameras, and {RT_RESET_BUDGET} resets (only the last episode is graded).")
+ diff = (f"Same instance and success rule as the standard twin (first-pass rating: {f['difficulty']}, harder here). "
+ f"The scene is dynamic (scripts/task_dynamics.toml: the ball is served with an initial velocity): "
+ f"{rt['rally']}, so nothing can be steered call by call; the agent has to plan or stream the muscle "
+ f"commands from a program, and learn the body and the serve from {RT_RESET_BUDGET + 1} real attempts.")
+ ref = ("None in the repository. Solvability evidence: a successful episode of the privileged twin (82 rows, GPT-6 Luna "
+ "xhigh) planned from step 0 through this service at S = 1 succeeded live and in the fresh replay, with no "
+ "stall (board 0112, 2026-10-14)." if fam == "table_tennis" else f"None. Solvability evidence: {rt['evidence']}.")
+ ver = ("The sim service (`environment/sim/rcb_realtime.py` + `rcb_service.py`, the shared eai-standard/2.3 layer, with "
+ "`environment/sim/server.py`, MyoSuite's adapter) judges the episode live and records every control step it "
+ "ran (`trajectory_live.f64`); after the agent's time the collect hook stops the agent's programs and signals "
+ "the service, which replays the graded episode in a fresh MyoSuite session and writes `final.json`; "
+ "`tests/grade.py` (separate sandbox, no network) passes iff the live episode succeeded, the replay succeeds "
+ "and ends in the same state. The clock's record (`stalls`, `max_lag_s`, ...) is copied into `reward.json` "
+ "and decides nothing. `nop`: the episode never starts (its clock starts at the first submit), reward 0.")
+ details = [
+ ("Upstream", "MyoSuite 2.12.2 (Apache-2.0) on MuJoCo 3.6.0, the models that ship in its wheel"),
+ ("Slowdown", f"S = {S} (`RCB_SLOWDOWN`'s default): measured RTF {rt['rtf']} ({rt['measured']})"),
+ ("Service", "the realtime layer of HumanoidBench, BiGym 2.0 and RoboOlympics (`rcb_realtime.py`, `rcb_service.py`, "
+ "unchanged) with `server.py`, which implements `rcb_service.Sim` on `rcb_myosuite.Session` (the "
+ "standard twin's scene, action, success rule and replay); cameras " + ", ".join(f"`{c}`" for c in f["cameras"])
+ + " (640 x 480, EGL on the GPU)"),
+ ("Resources", "4 CPUs and a GPU for rendering (the physics is on the CPU)"),
+ ("Reset budget", f"{RT_RESET_BUDGET} (`RCB_RESET_BUDGET`, compose default; PROTOCOL §4b): `POST /reset` starts a new "
+ "episode from the frozen instance, paused; only the last episode is graded, a success ends the "
+ "trial; results under another budget or slowdown are reported separately"),
+ ]
+ rows = "\n".join(f"| {k} | {v} |" for k, v in details)
+ return f"""# {name}
+
+## Task
+{task}
+
+## Difficulty
+{diff}
+
+## Reference solution
+{ref}
+
+## Verification
+{ver}
+
+## Details
+| | |
+|---|---|
+{rows}
+
+### Provenance
+Generated by `scripts/myosuite/generate.py` from `scripts/myosuite/instances/{fam}.npz` and `{fam}.json`
+(`make_facts.py`) and `scripts/myosuite/shared/realtime/`; see `docs/benchmarks/myosuite.md`, "Standard-realtime".
+
+### Change Log
+- 2026-10-14 (owner, board 0112): added. The family's scene is dynamic (PROTOCOL §2.14), so it gets a standard-realtime
+ twin with S = {S} and {RT_RESET_BUDGET} resets.
+"""
+
+
def dash(f: str) -> str:
return f.replace("_", "-")
@@ -559,6 +898,21 @@ R3_PRIV = """\
"""
+OSL_SIDE_TEXT = (
+ "- 2026-10-14 (board 0113, from board 0086): the texts said that the prosthesis is the left leg; MyoSuite's model\n"
+ " puts the Open-Source Leg on the right (its joints are `osl_knee_angle_r` and `osl_ankle_angle_r`, under `femur_r`),\n"
+ " and the texts now say so. Text only: the instance, the tooling and the pass rule are the same.\n"
+)
+BAODING_END_TEXT = (
+ "- 2026-10-14 (board 0113, from board 0086): an episode that has not succeeded after 1000 control steps (25 s) ends\n"
+ " there as a failure, and the task text says so. MyoSuite's baoding targets are a precomputed trajectory of 1000\n"
+ " points; with MyoSuite's step budget lifted, the step after it raised an IndexError (the service answered 500 and\n"
+ " the episode ended as a simulator error; a privileged replay that long stopped with an error). Nothing before that\n"
+ " step changes: the instance, the tooling's physics and the pass rule are the same, and every verdict reached before\n"
+ " is the same now.\n"
+)
+
+
def change_log(fam: str, f: dict, std: bool) -> str:
"""The README's Change Log: the standard service's history, plus what changed for this family alone."""
out = CHANGE_LOG if std else ""
@@ -582,6 +936,7 @@ def change_log(fam: str, f: dict, std: bool) -> str:
if fam == "chase_tag":
out += CHASE_TAG_TEXT
out += R3_STD if std else R3_PRIV
+ out += {"osl_run": OSL_SIDE_TEXT, "baoding": BAODING_END_TEXT}.get(fam, "")
if out and not std:
out = CHANGE_LOG_HEAD + out
return out
@@ -725,7 +1080,7 @@ Generated by `scripts/myosuite/generate.py` from `scripts/myosuite/instances/{fa
{change_log(fam, f, std)}"""
-def make(fam: str, out: Path) -> tuple[Path, Path]:
+def make(fam: str, out: Path) -> tuple[Path, ...]:
f = json.loads((HERE / "instances" / f"{fam}.json").read_text())
inst = HERE / "instances" / f"{fam}.npz"
oracle = HERE / "oracles" / f"{fam}.npy"
@@ -780,7 +1135,34 @@ def make(fam: str, out: Path) -> tuple[Path, Path]:
put(s / "solution/oracle.npy", src=oracle)
normalize(p)
normalize(s)
- return p, s
+ if fam not in REALTIME:
+ return p, s
+
+ r = out / f"myosuite-{dash(fam)}-i00-standard-realtime"
+ reset_dir(r)
+ put(r / "task.toml", rt_task_toml(fam, f))
+ put(r / "instruction.md", rt_instruction(fam, f))
+ put(r / "README.md", rt_readme(fam, f))
+ put(r / "environment/Dockerfile", RT_ENV_DOCKERFILE)
+ put(r / "environment/docker-compose.yaml", rt_compose(fam))
+ put(
+ r / "environment/client-requirements.txt",
+ src=LAYER / _ENV.get("CLIENT_REQUIREMENTS", "client-requirements.txt").strip(),
+ )
+ put(r / "environment/rcb_client.py", src=RT / "rcb_client.py")
+ put(r / "environment/rcb1.py", src=LAYER / "rcb1.py")
+ for name in ("rcb_realtime.py", "rcb_service.py", "server.py"):
+ put(r / "environment/sim" / name, src=RT / name)
+ put(r / "environment/sim/rcb1.py", src=LAYER / "rcb1.py")
+ put(r / "environment/sim/rcb_myosuite.py", src=SHARED / "rcb_myosuite.py")
+ put(r / "environment/sim/instance.npz", src=inst)
+ put(r / "environment/sim/task.json", json.dumps({"task_text": f["goal"], "default_cameras": [default_camera(fam, f)],
+ "reference": f.get("reference")}, indent=1) + "\n")
+ put(r / "tests/Dockerfile", STD_TESTS_DOCKERFILE)
+ put(r / "tests/test.sh", src=SHARED / "standard/test.sh", mode=0o755)
+ put(r / "tests/grade.py", src=RT / "grade.py")
+ normalize(r)
+ return p, s, r
def render_all(out: Path) -> list[str]:
diff --git a/scripts/myosuite/instances/baoding.json b/scripts/myosuite/instances/baoding.json
index 39e396b3fd..6f43c141f8 100644
--- a/scripts/myosuite/instances/baoding.json
+++ b/scripts/myosuite/instances/baoding.json
@@ -8,7 +8,7 @@
"hold_steps": 200,
"hold_s": 5.0,
"instance_sha": "04882d9bf16a24f6",
- "goal": "Baoding balls: rotate the two balls around each other in the palm so that each follows its moving target (the targets circle the palm once every 5 s). Success: both balls within 0.015 m of their targets for 200 consecutive control steps (5.0 s: one full orbit of the targets). Failure (the episode ends): a ball's height (world z) drops below 1.25 m, i.e. it falls out of the palm.",
+ "goal": "Baoding balls: rotate the two balls around each other in the palm so that each follows its moving target (the targets circle the palm once every 5 s). Success: both balls within 0.015 m of their targets for 200 consecutive control steps (5.0 s: one full orbit of the targets). Failure (the episode ends): a ball's height (world z) drops below 1.25 m, i.e. it falls out of the palm. MyoSuite's target trajectory has 1000 points: an episode that has not succeeded after 1000 control steps (25 s) ends there, as a failure.",
"rule": "hold",
"set3": false,
"actuators": [],
diff --git a/scripts/myosuite/instances/osl_run.json b/scripts/myosuite/instances/osl_run.json
index 84749edf39..d97dfe6055 100644
--- a/scripts/myosuite/instances/osl_run.json
+++ b/scripts/myosuite/instances/osl_run.json
@@ -8,7 +8,7 @@
"hold_steps": 1,
"hold_s": 0.005,
"instance_sha": "c150e8ac3f15d79c",
- "goal": "Run the track: move forward 29 m along it to the finish (MyoChallenge 2024 OSL Run, its own win condition), staying within 1 m of the track's centre line and on your feet; the left leg is the prosthesis, driven by its own controller, so you command only the 54 muscles. Failure (the episode ends): falling, leaving the track sideways, or moving 2 m backwards. There is no time limit.",
+ "goal": "Run the track: move forward 29 m along it to the finish (MyoChallenge 2024 OSL Run, its own win condition), staying within 1 m of the track's centre line and on your feet; the right leg is the prosthesis (`osl_knee_angle_r`, `osl_ankle_angle_r`), driven by its own controller, so you command only the 54 muscles. Failure (the episode ends): falling, leaving the track sideways, or moving 2 m backwards. There is no time limit.",
"rule": "event",
"set3": true,
"actuators": [
@@ -39,7 +39,7 @@
"free_root": true,
"track_steps": null,
"reference": null,
- "scene": "A musculoskeletal body whose left leg is an Open-Source Leg prosthesis (knee and ankle motors driven by its own controller) stands at the start of a straight, flat running track.",
+ "scene": "A musculoskeletal body whose right leg is an Open-Source Leg prosthesis (knee and ankle motors driven by its own controller) stands at the start of a straight, flat running track.",
"difficulty": "hard",
"hours": 5.0,
"reward_path": "myosuite/envs/myo/myochallenge/run_track_v0.py",
diff --git a/scripts/myosuite/make_facts.py b/scripts/myosuite/make_facts.py
index ec1274a927..11b6dfd625 100644
--- a/scripts/myosuite/make_facts.py
+++ b/scripts/myosuite/make_facts.py
@@ -83,7 +83,8 @@ FAMILIES: dict[str, dict] = {
"Baoding balls: rotate the two balls around each other in the palm so that each follows its moving target "
"(the targets circle the palm once every {period} s). Success: both balls within {proximity_th} m of their "
"targets for {hold} consecutive control steps ({hold_s} s: one full orbit of the targets). Failure (the episode ends): a ball's height (world z) "
- "drops below {drop_th} m, i.e. it falls out of the palm."
+ "drops below {drop_th} m, i.e. it falls out of the palm. MyoSuite's target trajectory has 1000 points: an episode "
+ "that has not succeeded after 1000 control steps (25 s) ends there, as a failure."
),
scene="Two balls rest in the palm of the upturned hand; two small markers show where each ball should be now.",
difficulty="hard",
@@ -152,7 +153,7 @@ FAMILIES: dict[str, dict] = {
# the frozen instance drew the opponent policy `stationary` (MyoChallenge draws it once per environment): it stands still
"chase_tag": dict(goal="", scene="A musculoskeletal body with two legs stands in a square arena; an opponent stands somewhere in it (in this frozen instance it does not move).",
difficulty="hard", hours=5.0),
- "osl_run": dict(goal="", scene="A musculoskeletal body whose left leg is an Open-Source Leg prosthesis (knee and ankle motors driven by its own controller) stands at the start of a straight, flat running track.",
+ "osl_run": dict(goal="", scene="A musculoskeletal body whose right leg is an Open-Source Leg prosthesis (knee and ankle motors driven by its own controller) stands at the start of a straight, flat running track.",
difficulty="hard", hours=5.0),
"dex_apple_lift": dict(goal="", scene="An apple lies on the table; the MyoHand (with a 6-DoF positioned wrist) is at its pre-grasp pose next to it.", difficulty="hard", hours=3.0,
skill="lift", obj="apple"),
@@ -227,7 +228,7 @@ def goal_text_set3(fam: str, s) -> str:
if fam == "osl_run":
return (f"Run the track: move forward {b.start_pos - b.end_pos:g} m along it to the finish (MyoChallenge 2024 OSL "
f"Run, its own win condition), staying within {b.real_width:g} m of the track's centre line and on your feet; "
- "the left leg is the prosthesis, driven by its own controller, so you command only the 54 muscles. Failure "
+ "the right leg is the prosthesis (`osl_knee_angle_r`, `osl_ankle_angle_r`), driven by its own controller, so you command only the 54 muscles. Failure "
"(the episode ends): falling, leaving the track sideways, or moving 2 m backwards. There is no time limit.")
if fam.startswith("dex_"):
f = FAMILIES[fam]
diff --git a/scripts/myosuite/shared/rcb_myosuite.py b/scripts/myosuite/shared/rcb_myosuite.py
index 0f2fca5acd..930c0c73ba 100644
--- a/scripts/myosuite/shared/rcb_myosuite.py
+++ b/scripts/myosuite/shared/rcb_myosuite.py
@@ -465,6 +465,12 @@ class Session:
self.done, self.success, self.reason = True, True, "success"
elif failed:
self.done, self.reason = True, "failure: " + _FAILURE.get(self.task, "task terminated")
+ # Baoding's targets come from a precomputed trajectory of 1000 points (baoding_v1.create_goal_trajectory,
+ # "assumes that it's greater than env horizon"): with MyoSuite's step budget lifted, step 1001 would index past
+ # it (IndexError, board 0086). The episode ends there instead, as a failure unless it succeeded on that step.
+ n_goal = len(getattr(self.base, "goal", ())) if self.task == "baoding" else 0
+ if not self.done and n_goal and self.steps >= n_goal:
+ self.done, self.reason = True, f"failure: the targets' trajectory ends after {n_goal} control steps"
return not self.done
def _bimanual_contacts_ok(self) -> bool:
diff --git a/scripts/myosuite/shared/realtime/grade.py b/scripts/myosuite/shared/realtime/grade.py
new file mode 100644
index 0000000000..b5766dbbea
--- /dev/null
+++ b/scripts/myosuite/shared/realtime/grade.py
@@ -0,0 +1,97 @@
+"""Standard-mode verifier (eai-standard/2.1): grades what the sim service -- which the agent could not touch -- wrote.
+
+/data/final.json holds the live run's outcome and the verdict of a fresh replay of the graded episode:
+success = live_success AND replay_success AND deterministic (the replay ends in the live run's final state).
+Writes /logs/verifier/reward.txt (1/0) and a numeric-only reward.json, and copies the evidence (replay video, last
+frame, trajectory, event log) next to them for humans; the evidence never decides the verdict.
+infra_error = 1: no verdict on the agent (the service crashed past its restarts, never finished, or failed to finalize).
+Every branch writes success, deterministic, infra_error and grader_error (PROTOCOL §2.6); sim_error = 1 when the graded
+episode ended on a simulator error (reason "error").
+A standard-realtime service's final.json (eai-standard/2.3) also says how well its clock was kept; reward.json repeats
+it (REALTIME below; PROTOCOL §4b), and none of it decides the verdict.
+"""
+import json
+import math
+import os
+import shutil
+import traceback
+
+DATA = os.environ.get("RCB_DATA", "/data")
+LOG = os.environ.get("RCB_VERIFIER_LOG_DIR", "/logs/verifier")
+os.makedirs(LOG, exist_ok=True)
+
+
+def write(ok: bool, out: dict) -> None:
+ with open(os.path.join(LOG, "reward.txt"), "w") as f:
+ f.write("1\n" if ok else "0\n")
+ with open(os.path.join(LOG, "reward.json"), "w") as f:
+ json.dump(out, f)
+ print(json.dumps(out))
+
+
+def num(x, default: float = 0.0) -> float:
+ """A finite number, or the default (reward.json takes numbers only)."""
+ try:
+ x = float(x)
+ except (TypeError, ValueError):
+ return default
+ return x if math.isfinite(x) else default
+
+
+# Standard-realtime only: slowdown S; clock_s, the wall seconds the clock ran; max_lag_s, the furthest a control step ran
+# behind its time; stalls, the times the service dropped more than 0.5 s of debt instead of fast-forwarding the robot;
+# stalled_s, the wall time dropped that way (the robot stood still while its clock should have run); stalled_share =
+# stalled_s / clock_s; held_steps and late_steps; requests_refused (answered 4xx) and request_busy_s (simulator-thread
+# time spent on requests). A field final.json does not have is 0.
+REALTIME = {"slowdown": float, "clock_s": float, "max_lag_s": float, "stalls": int, "stalled_s": float,
+ "stalled_share": float, "held_steps": int, "late_steps": int, "requests_refused": int,
+ "request_busy_s": float}
+
+
+def realtime_fields(d: dict) -> dict:
+ if not (d.get("realtime") or str(d.get("protocol")) == "eai-standard/2.3"):
+ return {}
+ return {"realtime": 1, **{k: kind(num(d.get(k))) for k, kind in REALTIME.items()}}
+
+
+def main() -> None:
+ path = os.path.join(DATA, "final.json")
+ if not os.path.exists(path):
+ write(False, {"success": 0, "deterministic": 0, "missing_log": 1, "infra_error": 1, "sim_error": 0,
+ "grader_error": 0})
+ return
+ try:
+ with open(path) as f:
+ d = json.load(f)
+ if not isinstance(d, dict):
+ raise ValueError("final.json is not a JSON object")
+ live, replay, det = (bool(d.get(k)) for k in ("live_success", "replay_success", "deterministic"))
+ failed_finalize = bool(d.get("finalize_error"))
+ unfinished = not d.get("finalized")
+ infra = bool(d.get("infra_error")) or failed_finalize or unfinished
+ ok = live and replay and det and not infra
+ out = {"success": int(ok), "live_success": int(live), "replay_success": int(replay), "deterministic": int(det),
+ "finalized": int(not unfinished), "missing_log": 0, "infra_error": int(infra),
+ "grader_error": int(failed_finalize), "sim_error": int(d.get("reason") == "error"),
+ "sim_crashed_on_call": int(d.get("reason") == "sim_crashed_on_call"),
+ "n_actions": int(d.get("n_actions") or 0), "sim_steps": int(d.get("control_steps_total") or 0),
+ "calls": int(d.get("calls") or 0), "rows": int(d.get("rows") or 0), "resets": int(d.get("resets") or 0),
+ "reset_budget": int(d.get("reset_budget") or 0),
+ "sim_startups": int(d.get("sim_startups") or 0), "elapsed_s": float(d.get("elapsed_s") or 0.0),
+ "video_rendered": int(bool(d.get("video"))), **realtime_fields(d)}
+ except Exception: # noqa: BLE001 -- an unreadable final.json is the grader's failure, not the agent's
+ traceback.print_exc()
+ write(False, {"success": 0, "deterministic": 0, "missing_log": 0, "infra_error": 1, "sim_error": 0,
+ "grader_error": 1})
+ return
+ for name in ("replay.mp4", "replay.last.png", "trajectory.npz", "episode.jsonl"):
+ try:
+ if os.path.exists(os.path.join(DATA, name)):
+ shutil.copy(os.path.join(DATA, name), os.path.join(LOG, name))
+ except OSError:
+ traceback.print_exc()
+ write(ok, out)
+
+
+if __name__ == "__main__":
+ main()
diff --git a/scripts/myosuite/shared/realtime/rcb_client.py b/scripts/myosuite/shared/realtime/rcb_client.py
new file mode 100644
index 0000000000..f2b0559bf9
--- /dev/null
+++ b/scripts/myosuite/shared/realtime/rcb_client.py
@@ -0,0 +1,209 @@
+"""eai-standard/2.3 client (standard-realtime mode): your whole interface to the robot (installed at /opt/tools/rcb_client.py, on
+PYTHONPATH).
+
+ from rcb_client import Robot, EpisodeOver
+ r = Robot() # finds the service (SIM_URL, http://sim:8000, http://localhost:8000), waits until it is ready
+
+TIME DOES NOT STOP (spec()["timing"]). An episode starts paused; its clock starts at its first submit() or
+get_obs(seconds > 0) (again after every reset()) and from then on runs one control step every spec()["control_dt"] x
+spec()["timing"]["slowdown"] seconds of wall time, whether or not you call, until the episode ends. The service keeps a
+plan, one servo-target row per coming control step: submit() writes into it and answers at once; when the plan runs out
+the last target is held. Whatever you do in between -- think, print, run a model -- the robot keeps executing.
+
+Calls (each is one HTTP request; every answer that carries data is an RCB1 frame, decoded into a Result):
+ r.spec() GET /spec, fetched when connecting: the robot's facts. "action_types"
+ ("joint_position": dim, layout (actuator order), lower/upper, range_rad,
+ units, semantics), "control_dt" (0.02 s), "max_control_steps" (the
+ episode), "timing" (realtime, slowdown, max_ahead_steps), "cameras" {name:
+ rgb, depth, hw, fovy_deg, pose, mounted}, "proprio_layout" and "proprio"
+ (joints, axes, positions), "imu", "links" (masses, centres of mass, the
+ feet's boxes), "depth", "budget" {"reset", "wall_clock_s"}, "episode"
+ r.status() GET /status: calls, rows, control_steps, done, reason, elapsed_s, running,
+ planned_steps, lag_s (and episode, resets_left with a reset budget)
+ r.get_obs(cameras=None, seconds=0.0, video_fps=10, include_depth=False, video="frames")
+ POST /observe. seconds=0: the robot now, at once. seconds>0 (at most
+ spec()["max_watch_s"]): waits while that much robot time runs (starting the
+ clock if it is not running) and returns what the cameras saw; it never
+ changes the plan
+ r.submit("joint_position", actions, dt=None, interpolate="linear", at=None, cameras=None, video_fps=10,
+ video="none", include_depth=False, note="")
+ POST /submit: answers at once with the robot as it is now (the rows
+ have not run yet). actions (T, dim): rows dt apart (a multiple of
+ control_dt, the default; T x dt at most spec()["max_chunk_s"]); each step's
+ target is the row (interpolate="hold") or a straight line from the previous
+ row's ("linear", starting from the target in force just before `at`). They
+ become the plan from control step `at` (an episode step number; None = the
+ next step), replacing whatever was planned from `at` on and keeping what is
+ planned before it; at most spec()["timing"]["max_ahead_steps"] ahead. Rows
+ for steps already gone when the request arrives are dropped
+ (meta["late_steps"]). The first submit of an episode starts its clock
+ r.reset(cameras=None, include_depth=False)
+ POST /reset, only when spec()["budget"]["reset"] > 0: abandons the
+ running episode and starts a new one, paused, from the same starting state;
+ only the last episode is graded, and a success ends the trial
+
+cameras=None means every camera in spec()["cameras"]; cameras=[] means no images at all, only proprioception (the
+fastest call: what a controller loop wants). video: "mp4" (one H.264 file per camera), "frames" (the images as arrays),
+"both" or "none" (only the last frame); a submit returns no video (no time passes inside it). include_depth adds depth
+for the cameras that have it.
+
+Result (what get_obs / submit / reset return):
+ res.meta dict: calls, rows, control_steps (steps run in this episode: the next one is this number),
+ sim_time_s, time_s, time_left_s, proprio {"layout", "now"}, done, reason ("success" |
+ "failed" | "timeout" | "error"; None while the episode runs), running, planned_steps,
+ lag_s; per submit: start_step, late_steps, plan_end_step, clipped [{row, index, reason}]
+ (values outside [-1, 1] are clipped, not refused); with a reset budget also episode,
+ resets_left
+ res.last[cam] uint8 H x W x 3: the last frame (unless cameras=[])
+ res.frames[cam] uint8 N x H x W x 3 (get_obs, video="frames"/"both") res.video[cam] mp4 bytes
+ res.depth[cam] float16 N x H x W, metres, 0 = nothing in range (include_depth)
+ res.proprio float32 rows of spec()["proprio_layout"]: one at the call's start, one every 1/video_fps s
+ of a watch, one at its end
+ res.done the episode is over
+ res.save(dir) meta.json, <cam>.mp4, last/<cam>.jpg, frames/<cam>/NNN.jpg, depth_<cam>.npy, proprio.npy
+
+Errors: ValueError -- refused, nothing changed (400: wrong type or shape, NaN, a bad dt or at, too many frames, an
+unknown camera; or a request over 64 MB, refused before it is sent); EpisodeOver -- the episode, or the trial, is over
+(409); RuntimeError -- the service failed (5xx: see spec()["on_service_error"]) or the endpoint does not exist in this
+run (404).
+"""
+from __future__ import annotations
+
+import json
+import os
+import pathlib
+import time
+
+import numpy as np
+import requests
+
+import rcb1
+
+PROTOCOLS = ("eai-standard/2.", "rcb-limited/2.") # any 2.x service, under either name (rcb-limited is the old one)
+MAX_BODY = 64 * 1024 * 1024 # the protocol's request limit: a larger request is refused
+
+
+class EpisodeOver(RuntimeError):
+ pass
+
+
+class Result:
+ """A decoded RCB1 answer: meta, video, frames, last, depth, proprio; save(dir) writes them all out."""
+
+ def __init__(self, meta: dict, parts: dict):
+ self.meta, self.raw = meta, parts
+ self.video, self.frames, self.last, self.depth, self.proprio = {}, {}, {}, {}, None
+ for name, p in parts.items():
+ kind, data = p["kind"], p["data"]
+ if name.startswith("video/") and kind == "mp4":
+ self.video[name[6:]] = data
+ elif name.startswith("frames/"):
+ self.frames[name[7:]] = data
+ elif name.startswith("last/"):
+ self.last[name[5:]] = data
+ elif name.startswith("depth/"):
+ self.depth[name[6:]] = data
+ elif name == "proprio":
+ self.proprio = data
+
+ @property
+ def done(self) -> bool:
+ return bool(self.meta.get("done"))
+
+ def save(self, out_dir: str) -> str:
+ from PIL import Image
+ out = pathlib.Path(out_dir)
+ out.mkdir(parents=True, exist_ok=True)
+ (out / "meta.json").write_text(json.dumps(self.meta, indent=1))
+ for cam, b in self.video.items():
+ (out / f"{cam}.mp4").write_bytes(b)
+ for cam, img in self.last.items():
+ (out / "last").mkdir(exist_ok=True)
+ Image.fromarray(np.ascontiguousarray(img)).save(out / "last" / f"{cam}.jpg", quality=90)
+ for cam, arr in self.frames.items():
+ d = out / "frames" / cam
+ d.mkdir(parents=True, exist_ok=True)
+ for i, f in enumerate(arr):
+ Image.fromarray(np.ascontiguousarray(f)).save(d / f"{i:03d}.jpg", quality=85)
+ for cam, arr in self.depth.items():
+ np.save(out / f"depth_{cam}.npy", arr)
+ if self.proprio is not None:
+ np.save(out / "proprio.npy", self.proprio)
+ return str(out)
+
+
+class Robot:
+ def __init__(self, base: str | None = None, timeout: float = 600.0):
+ candidates = [c for c in (base or os.environ.get("SIM_URL"), "http://sim:8000", "http://localhost:8000") if c]
+ t0, last = time.time(), None
+ while time.time() - t0 < timeout:
+ for b in candidates:
+ try:
+ r = requests.get(b + "/spec", timeout=(5, 60))
+ if r.status_code == 503:
+ last = "warming up"
+ continue
+ r.raise_for_status()
+ self.base, self._spec = b, r.json()
+ if not str(self._spec.get("protocol", "")).startswith(PROTOCOLS):
+ raise RuntimeError(f"not an eai-standard/2.x service: {self._spec.get('protocol')}")
+ return
+ except (requests.exceptions.ConnectionError, requests.exceptions.Timeout) as e:
+ last = e # not up yet (the scene takes about a minute), or busy
+ time.sleep(2)
+ raise RuntimeError(f"sim service unavailable: {last}")
+
+ def spec(self) -> dict:
+ return self._spec
+
+ def status(self) -> dict:
+ r = requests.get(self.base + "/status", timeout=(5, 60))
+ r.raise_for_status()
+ return r.json()
+
+ def get_obs(self, cameras=None, seconds: float = 0.0, video_fps: int = 10, include_depth: bool = False,
+ video: str = "frames") -> Result:
+ meta = {"cameras": _names(cameras), "seconds": float(seconds), "video_fps": int(video_fps),
+ "include_depth": bool(include_depth), "video": video}
+ return self._post("/observe", meta, [], timeout=900)
+
+ def submit(self, action_type: str, actions, dt: float | None = None, interpolate: str = "linear",
+ at: int | None = None, cameras=None, video_fps: int = 10, video: str = "none",
+ include_depth: bool = False, note: str = "") -> Result:
+ a = np.asarray(actions, dtype=np.float64)
+ if a.ndim == 1:
+ a = a[None, :]
+ if not np.all(np.isfinite(a)):
+ raise ValueError("actions contain NaN or inf")
+ meta = {"type": action_type, "dt": None if dt is None else float(dt), "interpolate": interpolate,
+ "at": None if at is None else int(at), "cameras": _names(cameras), "video_fps": int(video_fps),
+ "video": video,
+ "include_depth": bool(include_depth), "note": note}
+ return self._post("/submit", meta, [rcb1.array_part("actions", a)], timeout=1800)
+
+ def reset(self, cameras=None, include_depth: bool = False) -> Result:
+ return self._post("/reset", {"cameras": _names(cameras), "include_depth": bool(include_depth)}, [], timeout=900)
+
+ def _post(self, path: str, meta: dict, parts: list, timeout: float) -> Result:
+ body = rcb1.encode(meta, parts)
+ if len(body) > MAX_BODY: # the service would close the connection on it unread
+ raise ValueError(f"{path}: {len(body)} bytes, over the protocol's {MAX_BODY}: send fewer rows")
+ r = requests.post(self.base + path, data=body, headers={"Content-Type": "application/x-rcb1"},
+ timeout=(5, timeout))
+ if r.headers.get("Content-Type", "").startswith("application/x-rcb1") and r.status_code == 200:
+ return Result(*rcb1.decode(r.content))
+ try:
+ err = r.json()
+ except ValueError:
+ err = {"error": r.text[:300]}
+ msg = f"{r.status_code} {path}: {err.get('error')}" + (f" ({err['detail']})" if err.get("detail") else "")
+ if r.status_code == 409:
+ raise EpisodeOver(msg + (f"; reason {err['reason']}" if err.get("reason") else "")
+ + (f"; resets_left {err['resets_left']}" if "resets_left" in err else ""))
+ if r.status_code in (400, 413):
+ raise ValueError(msg)
+ raise RuntimeError(msg)
+
+
+def _names(cameras) -> list[str] | None:
+ return None if cameras is None else [cameras] if isinstance(cameras, str) else list(cameras)
diff --git a/scripts/myosuite/shared/realtime/rcb_realtime.py b/scripts/myosuite/shared/realtime/rcb_realtime.py
new file mode 100644
index 0000000000..cf4204bb26
--- /dev/null
+++ b/scripts/myosuite/shared/realtime/rcb_realtime.py
@@ -0,0 +1,926 @@
+"""eai-standard/2.3, the standard-realtime mode's protocol: the sim service of HumanoidBench's `-standard-realtime` tasks
+(docs/PROTOCOL.md §4b, docs/STANDARD_MODE.md "Standard-realtime mode"; docs/benchmarks/humanoidbench.md). Entry
+point of the sidecar: `python /srv/rcb_realtime.py` (the adapter is server.py's, unchanged). 2.3 is 2.2 plus the realtime
+capability, every 2.2 addition optional: this service is rcb_service's 2.1 service with the capability, and reports
+eai-standard/2.3.
+
+It extends rcb_service.Service: the same endpoints, RCB1 frames, budgets (reset, wall clock), evidence, crash recovery
+and finalize. Only the time model differs. In standard mode time moves only inside calls; here the episode has a clock:
+
+- An episode starts paused, at the frozen start with the servos holding their starting targets. Its clock starts at its
+ first POST /submit or POST /observe with seconds > 0 (again after every reset, and after a crash recovery). From then
+ on one control step runs every control_dt x slowdown seconds of wall time, whether or not the agent sends anything,
+ until the episode ends; the clock stops there.
+- The service keeps a plan: one target row per coming control step. /submit expands its chunk (rows dt apart, `linear`
+ or `hold`, exactly as 2.1 does) into control steps and writes it into the plan from control step `at` (default: the
+ next one), replacing whatever was planned from there on, and answers at once. A chunk that arrives after its `at`
+ loses the rows whose steps are gone (`late_steps`). When the plan runs out the last target is held (zero-order
+ hold): a slow agent's robot keeps its last servo targets.
+- A simulator may have several action types (RoboOlympics: `stroke`, `toss`, `mallet`, `stop`, ...), and a plan may
+ mix them. A row is only ever combined with a row of its own type: a chunk interpolates from the target of ITS type
+ in force before `at` (the simulator's current command of that type when the plan holds another type there), and a
+ gap before `at` is filled with held steps (the simulator's hold row, as when the plan has run out) unless the type
+ is an interpolable target that is itself in force there (then that target, held). A command that is not a target
+ (not interpolable: one row is one control step) is never interpolated and never repeated into a gap.
+- /observe with seconds = 0 answers at once; with seconds > 0 it watches the running episode that long (robot time)
+ and returns what the cameras saw. It never changes the plan.
+- `cameras = []` asks for no images at all (in 2.1 an empty list means every camera): a controller's proprio-only call.
+
+Grading is unchanged: every executed control step's sim row is journaled (trajectory_live.f64) and the graded episode's
+rows are replayed in a fresh simulator at the end. What is graded is the rows the simulator applied, not what was
+submitted, so the replay equals the live run however the timing went. Evidence added (never visible to the agent):
+ticks.f64 (per journal row: wall time, and the call whose chunk gave the row, -1 for a held one), calls.jsonl (one line
+per call; no calls/NNNNNN/ directories: a controller calls 50 times a second) and submitted.f64 (every row as sent:
+call id, then the row). final.json adds realtime, slowdown, held_steps, late_steps, max_queue_wait_s (the longest a
+request waited for the simulator thread) and how well the clock was kept: clock_s (wall seconds the clock ran, over the
+whole trial), max_lag_s (the furthest a control step ran behind its time), stalls, stalled_s (the wall time the stalls
+dropped: the robot stood still that long while the clock should have run) and stalled_share = stalled_s / clock_s; and
+what the requests cost: requests_refused (answered 4xx), request_mb (request bodies read) and request_busy_s (simulator
+thread time spent on requests). The graders copy the clock's figures into reward.json; none of them decides a verdict.
+
+Threads: the simulator and its GL context belong to the main thread, which runs the clock and, between control steps,
+the part of a request that needs the simulator. Everything else about a request happens on its HTTP thread, off the
+clock: reading the body, decoding it, every check that does not depend on the episode's state, and encoding the answer.
+A request that is refused there (404, 413, a malformed body, a wrong type / shape / dt / camera ...) never reaches the
+simulator thread. What a request can cost the clock is bounded:
+- JSON text is parsed under this process's interpreter lock (which the simulator thread needs too) only up to
+ INLINE_JSON bytes; a body with more of it is parsed by a child process, so a flood of large bodies, refused or not,
+ takes no time from the clock. Bodies over LANE_BODY bytes are read, decoded and checked one at a time (one of them
+ in memory at a time) and paced (LANE_PACE); while serving, the interpreter switches threads every SWITCH_INTERVAL_S.
+- A control step that is due runs before any waiting request. Behind the clock (a long render, a simulator step slower
+ than control_dt, a busy host) the missed steps run back to back, skipping none; a waiting request is then served only
+ once the clock has caught up, or -- so that a simulator that cannot keep up at all still answers -- after
+ BEHIND_PACE times the time the last request of its kind took. Requests therefore queue behind each other, not into
+ the clock: an /observe may take longer to answer, the robot's time does not slow down. A request that renders and
+ alone takes more than MAX_LAG_S (a stall by itself: CPU rendering) keeps the next rendering request waiting
+ BEHIND_PACE times as long, clock on time or not; requests without images (cameras = [], /status) are not held
+ back by it, so a controller's loop goes on. Requests can so take at most 1 / (1 + BEHIND_PACE) of the clock's time.
+- Cameras are rendered (and the proprioception read) at most once per control step, however many requests and watches
+ ask for them: all of them get that step's frames.
+More than MAX_LAG_S behind, the service drops the debt instead of fast-forwarding the robot (a stall: counted, and its
+seconds added to stalled_s).
+
+A crash resumes the same episode from its last checkpoint (one per second of robot time while the clock runs), paused.
+A simulator crash is an infrastructure failure, never the agent's (PROTOCOL §4b, owner 2026-10-08). No call spans a
+crash here (submit answers at once), so a crash is never put down to the agent's call: past RCB_MAX_RESUMES restarts the
+trial ends `sim_crashed` (infra_error), a failed resume `resume_failed` (infra_error), and a simulator error that ends
+the graded episode (reason "error") sets infra_error in final.json too.
+"""
+from __future__ import annotations
+
+import collections
+import json
+import os
+import queue
+import struct
+import subprocess
+import sys
+import threading
+import time
+import traceback
+from http.server import BaseHTTPRequestHandler, ThreadingHTTPServer
+from urllib.parse import urlparse
+
+import numpy as np
+
+import rcb1
+import rcb_service
+from rcb_service import MAX_BODY, Recorder, Refused, Service, bad_request
+
+PROTOCOL = "eai-standard/2.3" # the standard-realtime mode's protocol (PROTOCOL §4b); rcb_service.py itself serves 2.1
+MAX_LAG_S = 0.5 # behind the clock by more than this: drop the debt (a stall) instead of catching up
+CHECKPOINT_S = 1.0 # robot time between checkpoints while the clock runs (the most a crash loses)
+HELD = (None, None, -1) # a plan entry for a step with nothing planned: the simulator's hold row, taken at that step
+# What a request may cost the clock (module docstring, "Threads"):
+INLINE_JSON = 1 << 20 # bytes of JSON text parsed in this process; more is parsed by a child process (off the lock)
+LANE_BODY = 64 << 10 # bodies above this are decoded and checked one at a time ...
+LANE_PACE = 3.0 # ... and the next one waits this many times the CPU time the last one took
+BEHIND_PACE = 4.0 # behind the clock: the next request waits this many times the time the last one took
+CHILD_TIMEOUT_S = 60.0 # the child process's time for one body
+CHILD_MEM = 1 << 30 # ... and its memory above what its interpreter needs (a body that needs more is refused)
+SWITCH_INTERVAL_S = 0.001 # the interpreter's thread switch interval while serving (default 0.005): HTTP threads
+ # hand the lock back to the simulator thread within a millisecond
+FIELDS = ("type", "dt", "interpolate", "at", "cameras", "video_fps", "video", "include_depth", "note", "seconds")
+
+# The child process that parses a body with more than INLINE_JSON bytes of JSON text (argv: the most elements a valid
+# `actions` can have, INLINE_JSON, CHILD_MEM; stdin: the body). It answers with the fields the service reads and the
+# actions as float64 -- or only their shape when no action type takes that many -- or the error a malformed body gets.
+_CHILD = r"""
+import json, struct, sys
+FIELDS = %r
+def conv(np, obj, lim):
+ try:
+ a = np.asarray(obj, dtype=np.float64)
+ except (TypeError, ValueError):
+ return {"kind": "bad"}, b""
+ if a.size > lim:
+ return {"kind": "shape", "shape": list(a.shape)}, b""
+ return {"kind": "array", "dtype": "<f8", "shape": list(a.shape)}, np.ascontiguousarray(a).astype("<f8").tobytes()
+def main():
+ lim, inline, mem = (int(x) for x in sys.argv[1:4])
+ import numpy as np
+ try:
+ import resource
+ with open("/proc/self/statm") as f:
+ vm = int(f.read().split()[0]) * resource.getpagesize()
+ resource.setrlimit(resource.RLIMIT_AS, (vm + mem, vm + mem))
+ except Exception:
+ pass
+ blob = b""
+ try:
+ buf = sys.stdin.buffer.read()
+ part = None
+ if buf[:4] == b"RCB1":
+ n = struct.unpack("<I", buf[4:8])[0]
+ header = json.loads(buf[8:8 + n].decode())
+ off = 8 + n
+ meta = header.get("meta", {})
+ for p in header["parts"]:
+ if p["name"] == "actions":
+ part = (p, off)
+ off += p["len"]
+ else:
+ meta = json.loads(buf or b"{}")
+ if not isinstance(meta, dict):
+ raise ValueError("the body must be a JSON object or an RCB1 frame")
+ desc = {"kind": "none"}
+ if part is not None:
+ p, off = part
+ raw = buf[off:off + p["len"]]
+ if p["kind"] == "array":
+ a = np.frombuffer(raw, dtype=np.dtype(p["dtype"])).reshape(p["shape"])
+ if a.size > lim:
+ desc = {"kind": "shape", "shape": list(a.shape)}
+ else:
+ desc, blob = {"kind": "array", "dtype": p["dtype"], "shape": list(a.shape)}, raw
+ elif p["kind"] == "json":
+ desc, blob = conv(np, json.loads(raw.decode()), lim)
+ else:
+ desc = {"kind": "bad"}
+ elif "actions" in meta:
+ desc, blob = conv(np, meta["actions"], lim)
+ small = {k: meta[k] for k in FIELDS if k in meta}
+ if "note" in small:
+ small["note"] = str(small["note"] or "")[:200]
+ head = json.dumps({"meta": small, "actions": desc}).encode()
+ if len(head) > inline:
+ raise ValueError("the request's fields other than the actions take more than %%d bytes" %% inline)
+ except MemoryError:
+ head, blob = json.dumps({"error": "the body takes too much memory to parse"}).encode(), b""
+ except Exception as exc:
+ head, blob = json.dumps({"error": str(exc)[:300] or type(exc).__name__}).encode(), b""
+ sys.stdout.buffer.write(struct.pack("<I", len(head)) + head + blob)
+main()
+""" % (FIELDS,)
+
+
+class _Job:
+ """One HTTP request, handed from its HTTP thread to the simulator's thread and back."""
+
+ def __init__(self, method: str, path: str, body: bytes = b"", content_length: int | None = None, req: tuple = ()):
+ self.method, self.path, self.body, self.content_length = method, path, body, content_length
+ self.req = req # (route, checked request) from RealtimeService.front(); () = not looked at yet
+ self.queued = time.monotonic()
+ self.done = threading.Event()
+ self.reply: tuple = () # ("raw", code, ctype, body) | ("rcb1", meta, parts) | ("rec", meta, Recorder)
+
+ def finish(self, *reply) -> None:
+ self.reply = reply
+ self.done.set()
+
+
+class _Watch:
+ """A pending /observe with seconds > 0: completes once its control steps have run (or the episode ended)."""
+
+ def __init__(self, job: _Job, rec: Recorder, start: int, steps: int, cameras: list[str], obs_id: int):
+ self.job, self.rec, self.cameras, self.obs_id = job, rec, cameras, obs_id
+ self.start, self.end = start, start + steps
+
+
+class _Views:
+ """The simulator as the Recorders see it: the cameras are rendered, and the proprioception read, at most once per
+ control step, however many requests and watches ask (each gets that step's arrays; nobody writes into them). The
+ state they show changes only with a control step or a reset, which is what the cache is keyed by."""
+
+ def __init__(self, svc: "RealtimeService"):
+ self._svc, self._sim = svc, svc.sim
+ self._key, self._rgb, self._dep, self._prop = None, {}, {}, None
+
+ def __getattr__(self, name):
+ return getattr(self._sim, name)
+
+ def _fresh(self) -> None:
+ key = (self._svc.journal_rows, self._svc.resets)
+ if key != self._key:
+ self._key, self._rgb, self._dep, self._prop = key, {}, {}, None
+
+ def images(self, cameras, depth):
+ self._fresh()
+ want = depth and set(self._sim.depth_cameras)
+ need = [c for c in cameras if c not in self._rgb or (want and c in want and c not in self._dep)]
+ if need:
+ rgb, dep = self._sim.images(list(dict.fromkeys(need)), depth)
+ self._rgb.update(rgb)
+ self._dep.update(dep)
+ return {c: self._rgb[c] for c in cameras}, ({c: self._dep[c] for c in cameras if c in self._dep} if depth else {})
+
+ def proprio(self):
+ self._fresh()
+ if self._prop is None:
+ self._prop = self._sim.proprio()
+ return self._prop
+
+
+class RealtimeService(Service):
+ """rcb_service.Service with the realtime clock (module docstring)."""
+
+ def __init__(self, sim, *, slowdown: float | None = None, **kw):
+ self.slowdown = float(os.environ.get("RCB_SLOWDOWN") or 1.0) if slowdown is None else float(slowdown)
+ if not self.slowdown > 0:
+ raise ValueError("RCB_SLOWDOWN must be > 0")
+ self.wall_dt = sim.control_dt * self.slowdown
+ self.max_ahead = int(round(float(kw.get("max_chunk_s", 10.0)) / sim.control_dt)) # how far ahead `at` may be
+ self.plan: collections.deque = collections.deque() # (type, row, call id) per coming control step, or HELD
+ self.running = False
+ self.anchor_t = self.anchor_n = 0 # the clock: control step anchor_n was due at anchor_t
+ self.watches: list[_Watch] = []
+ self.q: queue.Queue = queue.Queue()
+ self.max_lag = self.lag = 0.0
+ self.stalls = self.held_steps = self.late_steps = 0
+ self.stalled_s = 0.0 # wall time the stalls dropped
+ self.clock_s = 0.0 # wall time the clock ran, up to its last start
+ self._clock_t0 = 0.0 # ... and when that was
+ self.max_queue_wait = 0.0
+ self.requests_refused, self.request_bytes, self.request_busy_s = 0, 0, 0.0
+ self._ticks: list[tuple[float, float]] = [] # (wall time, call id) per journal row not yet on disk
+ super().__init__(sim, **kw)
+ self.ticks_path = os.path.join(self.data, "ticks.f64")
+ self.ckpt_every = max(1, int(round(CHECKPOINT_S / sim.control_dt)))
+ # the HTTP threads' side (front()): what they need of the simulator is read once, here
+ self.views = _Views(self)
+ self._interp = {t: bool(sim.interpolable(t)) for t in self.types}
+ self._max_rows = int((self.max_chunk_s + 1e-9) / sim.control_dt) # of any chunk that can be valid
+ self._max_elems = max([self._max_rows * int(v["dim"]) for v in self.types.values()] or [0])
+ self._spec_reply: tuple | None = None
+ self._t_seen: float | None = None # the first request's time, as its HTTP thread saw it
+ self._stat_lock, self._lane, self._child_lock = threading.Lock(), threading.Lock(), threading.Lock()
+ self._lane_free_at = 0.0
+ # the simulator thread's side (loop()): requests by kind -- True: it renders cameras, False: it does not
+ self._gate = {True: 0.0, False: 0.0} # behind the clock, none of the kind is served before this
+ self._render_gate = 0.0 # ... and no rendering request before this, behind or not
+ self._held: collections.deque = collections.deque() # requests taken from the queue before their time
+
+ # ------------------------------------------------------------------ spec, status, evidence
+ def _spec(self) -> dict:
+ spec = super()._spec()
+ spec["protocol"] = PROTOCOL
+ cdt = self.sim.control_dt
+ spec["timing"] = {
+ "realtime": True, "slowdown": self.slowdown, "control_dt": cdt,
+ "wall_s_per_control_step": cdt * self.slowdown,
+ "clock": "an episode starts paused; its clock starts at its first submit or observe with seconds > 0 (also "
+ "after a reset) and then runs one control step every wall_s_per_control_step seconds, whatever "
+ "you do, until the episode ends",
+ "submit": "answers at once: the chunk (rows dt apart, linear or hold) becomes the plan from control step "
+ "`at` (default: the next one), replacing what was planned from there; rows for steps already "
+ "gone are dropped (meta.late_steps); after the plan's last row that target is held",
+ "observe": "seconds = 0: now, at once; seconds > 0: watches the running episode that long and returns what "
+ "the cameras saw; never changes the plan",
+ "max_ahead_steps": self.max_ahead,
+ "no_images": "cameras = [] returns no images, only proprioception (the fastest call)"}
+ return spec
+
+ def _realtime_counters(self) -> dict:
+ return {"running": self.running, "planned_steps": len(self.plan), "lag_s": round(self.lag, 4)}
+
+ def status(self) -> dict:
+ return {**super().status(), **self._realtime_counters()}
+
+ def meta(self) -> dict:
+ return {**super().meta(), **self._realtime_counters()}
+
+ def _clock_seconds(self) -> float:
+ """Wall seconds the clock has run, over the whole trial."""
+ return self.clock_s + (time.monotonic() - self._clock_t0 if self.running else 0.0)
+
+ def _timing(self) -> dict:
+ """The counters a checkpoint keeps (and _resume restores)."""
+ return {"max_lag_s": self.max_lag, "stalls": self.stalls, "stalled_s": self.stalled_s,
+ "clock_s": self._clock_seconds(), "held_steps": self.held_steps, "late_steps": self.late_steps,
+ "max_queue_wait_s": self.max_queue_wait, "requests_refused": self.requests_refused,
+ "request_bytes": self.request_bytes, "request_busy_s": self.request_busy_s}
+
+ def _timing_record(self) -> dict:
+ clock = self._clock_seconds()
+ return {"protocol": PROTOCOL, "realtime": 1, "slowdown": self.slowdown, "max_lag_s": round(self.max_lag, 4),
+ "stalls": self.stalls, "stalled_s": round(self.stalled_s, 3), "clock_s": round(clock, 3),
+ "stalled_share": round(min(1.0, self.stalled_s / clock), 4) if clock > 0 else 0.0,
+ "held_steps": self.held_steps, "late_steps": self.late_steps,
+ "max_queue_wait_s": round(self.max_queue_wait, 4), "requests_refused": self.requests_refused,
+ "request_mb": round(self.request_bytes / 1e6, 3), "request_busy_s": round(self.request_busy_s, 3)}
+
+ def _write_final(self) -> None:
+ self.verdict.update(self._timing_record()) # final.json merges the verdict; none of these keys is a verdict's
+ if self.reason == "error": # a simulator error ended the graded episode: infrastructure (§4b)
+ self.verdict["infra_error"] = 1
+ else: # (an episode that erred and was reset away is not graded)
+ self.verdict.pop("infra_error", None)
+ super()._write_final()
+
+ def _checkpoint(self, ev: str, **kw) -> None:
+ if self._ticks: # the tick log first: a crash in between leaves it too long
+ with open(self.ticks_path, "ab") as f:
+ f.write(np.asarray(self._ticks, dtype=np.float64).tobytes())
+ self._ticks = []
+ super()._checkpoint(ev, **kw, timing=self._timing())
+
+ def _save_call(self, meta: dict, actions: np.ndarray | None) -> None:
+ keep = ("type", "n_rows", "dt", "interpolate", "at", "start_step", "late_steps", "planned_steps", "seconds",
+ "control_steps", "episode", "done", "reason", "n_clipped", "lag_s", "note")
+ line = {"call": self.calls, "ts": time.time(), **{k: meta[k] for k in keep if k in meta}}
+ with open(os.path.join(self.data, "calls.jsonl"), "a") as f:
+ f.write(json.dumps(line) + "\n")
+ if actions is not None:
+ rows = np.concatenate([np.full((len(actions), 1), float(self.calls)), actions], axis=1)
+ with open(os.path.join(self.data, "submitted.f64"), "ab") as f:
+ f.write(np.asarray(rows, dtype=np.float64).tobytes())
+
+ # ------------------------------------------------------------------ start, clock
+ def start(self) -> None:
+ super().start()
+ rows = np.fromfile(self.ticks_path, dtype=np.float64) if os.path.exists(self.ticks_path) else np.zeros(0)
+ n = self.journal_rows
+ if rows.size > 2 * n: # the steps a crash lost from the journal go from the tick log too
+ rows[: 2 * n].tofile(self.ticks_path)
+ elif rows.size < 2 * n: # (an older tick log, short of the journal: pad it as unknown)
+ np.concatenate([rows, np.full(2 * n - rows.size, np.nan)]).tofile(self.ticks_path)
+ self.running = False # a resumed episode waits, paused, for the agent's next call
+
+ def _resume(self, ev: list[dict]) -> None:
+ """Service._resume, with the realtime counters as of the checkpoint the episode resumes from (what the lost
+ steps had added goes with them): final.json's stalls, stalled_s, clock_s, held_steps, late_steps, max_lag_s,
+ max_queue_wait_s and the request counters cover the whole trial, not only the part after the last restart."""
+ cp = next((e for e in reversed(ev) if "state_sha" in e and "journal_rows" in e), {})
+ t = cp.get("timing") if isinstance(cp.get("timing"), dict) else {}
+ self.stalls, self.held_steps, self.late_steps, self.requests_refused, self.request_bytes = (
+ int(t.get(k) or 0) for k in ("stalls", "held_steps", "late_steps", "requests_refused", "request_bytes"))
+ self.max_lag, self.max_queue_wait, self.stalled_s, self.clock_s, self.request_busy_s = (
+ float(t.get(k) or 0.0) for k in ("max_lag_s", "max_queue_wait_s", "stalled_s", "clock_s", "request_busy_s"))
+ super()._resume(ev)
+
+ def _start_clock(self) -> None:
+ if not self.running and not self.done and not self.stopping:
+ self.running, self.anchor_t, self.anchor_n = True, time.monotonic(), len(self.episode)
+ self._clock_t0 = self.anchor_t
+ self.log("clock_start", control_steps=len(self.episode), episode=self.resets + 1)
+
+ def _stop_clock(self, why: str) -> None:
+ if self.running:
+ self.clock_s = self._clock_seconds()
+ self.running = False
+ self.log("clock_stop", why=why, control_steps=len(self.episode), max_lag_s=round(self.max_lag, 4),
+ stalls=self.stalls, stalled_s=round(self.stalled_s, 3), clock_s=round(self.clock_s, 3))
+ self.plan.clear()
+
+ def _end(self, reason: str) -> None:
+ super()._end(reason)
+ self._stop_clock(reason) # pending watches are completed by whoever ended it, after this
+
+ def due_in(self, now: float) -> float | None:
+ """Seconds until the next control step is due (<= 0: due now); None while the clock is stopped."""
+ if not self.running or self.done or self.stopping:
+ return None
+ return self.anchor_t + (len(self.episode) - self.anchor_n) * self.wall_dt - now
+
+ def tick(self) -> None:
+ """One control step of the clock: the plan's row for it, or the held target."""
+ now = time.monotonic()
+ lag = -(self.due_in(now) or 0.0)
+ self.max_lag = max(self.max_lag, lag)
+ if lag > MAX_LAG_S: # far behind: re-anchor rather than rush the robot through the debt
+ self.stalls += 1
+ self.stalled_s += lag # the robot stood still this long while its clock should have run
+ self.log("stall", behind_s=round(lag, 3), control_steps=len(self.episode))
+ self.anchor_t, self.anchor_n, lag = now, len(self.episode), 0.0
+ self.lag = max(lag, 0.0)
+ typ, row, src = self.plan.popleft() if self.plan else HELD
+ if typ is None: # nothing planned for this step (the plan ran out, or a gap in it)
+ typ, row = self.sim.hold()
+ self.held_steps += 1
+ before = self.journal_rows
+ self._ticks.append((time.time(), float(src)))
+ try:
+ self._step(typ, row)
+ except Exception as exc: # noqa: BLE001 -- it ran and failed: the episode ends, as on hardware
+ traceback.print_exc()
+ if self.journal_rows == before:
+ self._ticks.pop()
+ self.log("error", path="clock", error=repr(exc)[:500])
+ self._end("error")
+ self._release_watches()
+ return
+ for w in list(self.watches):
+ w.rec.step()
+ if self.done or len(self.episode) >= w.end:
+ self._complete_watch(w)
+ if not self.done and len(self.episode) % self.ckpt_every == 0:
+ self._checkpoint("tick")
+
+ # ------------------------------------------------------------------ requests, off the clock (HTTP threads)
+ def first_request(self) -> None:
+ if self.t0 is None:
+ self.t0 = self._t_seen or time.time() # when its HTTP thread saw it (front), logged by this thread
+ self.log("first_request", t0=self.t0)
+
+ def _count(self, refused: int = 0, read: int = 0) -> None:
+ with self._stat_lock:
+ self.requests_refused += refused
+ self.request_bytes += read
+
+ def _refusal(self, exc: Refused) -> tuple:
+ self._count(refused=1)
+ return ("reply", *self._reply_json(exc.code, exc.body))
+
+ def front(self, method: str, path: str, n: int = 0, read=None) -> tuple:
+ """Everything about a request that needs no simulator, on the caller's thread (an HTTP thread): the route, the
+ body (`read(n)`, called only for a body that will be decoded), its decoding, and every check that does not
+ depend on the episode's state. Returns ("reply", status, content type, body) for a request that ends here --
+ every refusal but the ones only the episode's state can decide (409, an `at` too far ahead, too many frames) --
+ or (route, checked request) for the simulator thread."""
+ if self._t_seen is None:
+ with self._stat_lock:
+ self._t_seen = self._t_seen or time.time()
+ path = urlparse(path).path
+ if method == "GET" and path == "/spec": # fixed for the service's life
+ if self._spec_reply is None:
+ self._spec_reply = self._reply_json(200, self.spec)
+ return ("reply", *self._spec_reply)
+ if method == "GET" and path == "/status":
+ return ("/status", None)
+ if method != "POST" or path not in ("/submit", "/observe") + (("/reset",) if self.reset_budget else ()):
+ return self._refusal(Refused(404, {"error": "no such endpoint"}))
+ if n < 0:
+ return self._refusal(Refused(400, {"error": "bad Content-Length"}))
+ if n > MAX_BODY:
+ return self._refusal(Refused(413, {"error": f"request body over {MAX_BODY} bytes"}))
+ try:
+ if n <= LANE_BODY:
+ body = read(n) if n else b""
+ self._count(read=len(body))
+ return (path, self._prepare(path, body))
+ with self._lane: # large bodies: one at a time (also in memory: the next one is
+ wait = self._lane_free_at - time.monotonic() # read when this one is done), paced by what the last took
+ if wait > 0:
+ time.sleep(wait)
+ body = read(n)
+ self._count(read=len(body))
+ t = time.thread_time()
+ try:
+ return (path, self._prepare(path, body))
+ finally:
+ self._lane_free_at = time.monotonic() + LANE_PACE * (time.thread_time() - t)
+ except Refused as exc:
+ # After a body that cannot be read at all, what the simulator thread answers before it looks at a
+ # request's fields (409) comes first here too.
+ return self._refusal(exc if getattr(exc, "malformed", False) else self._state_refusal(path) or exc)
+
+ def _state_refusal(self, path: str) -> Refused | None:
+ """The refusal a request to `path` gets whatever it says, as things stand (read without the simulator thread:
+ a request that races the episode's end may get either answer, as it always could)."""
+ if path == "/submit" and (self.done or self.stopping):
+ return self._conflict()
+ if path == "/reset":
+ try:
+ self._reset_view({})
+ except Refused as exc:
+ return exc
+ return None
+
+ def _prepare(self, path: str, body: bytes):
+ try:
+ meta, actions = self._decode_rt(body)
+ except Exception as exc: # noqa: BLE001 -- any malformed body is the caller's error
+ refused = Refused(400, {"error": f"malformed body: {exc}"[:300]})
+ refused.malformed = True
+ raise refused from None
+ if path == "/submit":
+ return self._check_submit(meta, actions)
+ if path == "/observe":
+ return self._check_observe(meta)
+ with bad_request(): # /reset: its view; the budget is the simulator thread's to check
+ self._view(meta, "frames")
+ return meta
+
+ def _decode_rt(self, body: bytes) -> tuple[dict, object]:
+ """Service._decode for this service: (meta, the actions as sent -- the RCB1 part `actions`, else meta's -- or
+ None). Only the part the service reads is decoded. More than INLINE_JSON bytes of JSON text go to the child."""
+ actions, part = None, None
+ if body[:4] == rcb1.MAGIC:
+ (n,) = struct.unpack("<I", body[4:8])
+ if n > INLINE_JSON:
+ return self._decode_child(body)
+ header = json.loads(body[8:8 + n].decode())
+ meta, off = header.get("meta", {}), 8 + n
+ for p in header["parts"]:
+ if p["name"] == "actions":
+ part = (p, off)
+ off += p["len"]
+ else:
+ if len(body) > INLINE_JSON:
+ return self._decode_child(body)
+ meta = json.loads(body or b"{}")
+ if not isinstance(meta, dict):
+ raise ValueError("the body must be a JSON object or an RCB1 frame")
+ if part is not None:
+ p, off = part
+ blob = body[off:off + p["len"]]
+ if p["kind"] == "array":
+ actions = np.frombuffer(blob, dtype=np.dtype(p["dtype"])).reshape(p["shape"])
+ elif p["kind"] == "json":
+ if len(blob) > INLINE_JSON:
+ return self._decode_child(body)
+ actions = json.loads(blob.decode())
+ else:
+ actions = blob
+ else:
+ actions = meta.get("actions")
+ return meta, actions
+
+ def _decode_child(self, body: bytes) -> tuple[dict, object]:
+ """_decode_rt by the child process (_CHILD): the same answer, without holding this process's interpreter lock
+ for the time a large JSON text takes to parse. One at a time."""
+ with self._child_lock:
+ try:
+ env = {**os.environ, "OMP_NUM_THREADS": "1", "OPENBLAS_NUM_THREADS": "1", "MKL_NUM_THREADS": "1"}
+ out = subprocess.run([sys.executable, "-c", _CHILD, str(self._max_elems), str(INLINE_JSON),
+ str(CHILD_MEM)], input=body, stdout=subprocess.PIPE, stderr=subprocess.DEVNULL,
+ timeout=CHILD_TIMEOUT_S, env=env).stdout
+ (n,) = struct.unpack("<I", out[:4])
+ head, blob = json.loads(out[4:4 + n].decode()), out[4 + n:]
+ except Exception as exc: # noqa: BLE001 -- no child, no answer: refused; never parsed on the clock's lock
+ raise ValueError(f"more than {INLINE_JSON} bytes of JSON text, which this service could not parse "
+ f"({type(exc).__name__}): send the rows as an RCB1 array part") from None
+ if "error" in head:
+ raise ValueError(head["error"])
+ d = head["actions"]
+ if d["kind"] == "array":
+ actions = np.frombuffer(blob, dtype=np.dtype(d["dtype"])).reshape(d["shape"])
+ elif d["kind"] == "shape": # more elements than any action type takes: its shape is all that
+ shape = tuple(int(x) for x in d["shape"]) # matters (it is refused for it), so none of them came back
+ actions = np.lib.stride_tricks.as_strided(np.zeros(1), shape=shape, strides=(0,) * len(shape),
+ writeable=False)
+ else:
+ actions = None if d["kind"] == "none" else b"?" # b"?": not numbers (refused as such)
+ return head["meta"], actions
+
+ def _check_submit(self, meta: dict, actions) -> dict:
+ """/submit's checks that do not depend on the episode's state; the chunk, ready for the simulator thread."""
+ with bad_request():
+ typ = meta.get("type")
+ if not isinstance(typ, str) or typ not in self.types:
+ raise ValueError(f"type must be one of {sorted(self.types)}")
+ dim = int(self.types[typ]["dim"])
+ cdt = self.sim.control_dt
+
+ def shape_ok(a) -> None: # before any pass over the rows: a refusal costs nothing
+ if a.ndim != 2 or a.shape[0] < 1 or a.shape[1] != dim:
+ raise ValueError(f"{typ} actions must be (T, {dim}) with T >= 1, got {list(a.shape)}")
+ if a.shape[0] > self._max_rows:
+ raise ValueError(f"the chunk covers {a.shape[0] * cdt:g} s: "
+ f"more than max_chunk_s = {self.max_chunk_s:g}")
+
+ if isinstance(actions, np.ndarray):
+ shape_ok(actions)
+ elif isinstance(actions, list) and len(actions) > self._max_rows:
+ raise ValueError(f"the chunk covers {len(actions) * cdt:g} s: "
+ f"more than max_chunk_s = {self.max_chunk_s:g}")
+ try:
+ A = np.asarray(actions, dtype=np.float64)
+ except (TypeError, ValueError, KeyError):
+ raise ValueError("send the actions as an RCB1 array part `actions` or a JSON list of rows") from None
+ shape_ok(A)
+ if not np.all(np.isfinite(A)):
+ raise ValueError("actions contain NaN or inf")
+ try:
+ dt = cdt if meta.get("dt") is None else float(meta["dt"])
+ except (TypeError, ValueError):
+ raise ValueError("dt must be a number of seconds") from None
+ k = round(dt / cdt) if np.isfinite(dt / cdt) else 0
+ if k < 1 or abs(k * cdt - dt) > 1e-9:
+ raise ValueError(f"dt must be a positive whole multiple of control_dt = {cdt} s (never rounded)")
+ if k > 1 and not self._interp[typ]:
+ raise ValueError(f"{typ}: one row is one control step, so dt must be control_dt = {cdt} s")
+ if len(A) * k * cdt > self.max_chunk_s + 1e-9:
+ raise ValueError(f"the chunk covers {len(A) * k * cdt:g} s: "
+ f"more than max_chunk_s = {self.max_chunk_s:g}")
+ interp = meta.get("interpolate", "linear")
+ if interp not in ("linear", "hold"):
+ raise ValueError("interpolate must be linear or hold")
+ at = meta.get("at")
+ if at is not None:
+ if isinstance(at, bool) or not isinstance(at, (int, float)) or at != int(at) or at < 0:
+ raise ValueError("at must be a control step of this episode (an integer >= 0), "
+ "or null for the next one")
+ at = int(at)
+ return {"type": typ, "actions": A, "dt": dt, "k": k, "interpolate": interp, "at": at,
+ "view": self._view_rt(meta, "none"), "note": str(meta.get("note") or "")[:200]}
+
+ def _check_observe(self, meta: dict) -> dict:
+ with bad_request():
+ try:
+ seconds = float(meta.get("seconds") or 0.0)
+ except (TypeError, ValueError):
+ raise ValueError("seconds must be a number") from None
+ if not 0.0 <= seconds <= self.max_watch_s:
+ raise ValueError(f"seconds must be in [0, {self.max_watch_s:g}]")
+ return {"seconds": seconds, "view": self._view_rt(meta, "frames")}
+
+ # ------------------------------------------------------------------ requests (simulator thread)
+ def dispatch(self, job: _Job) -> None:
+ """Serve one queued request; a watch finishes later, from tick(). A job that front() has not seen (handed over
+ whole: tests, tools) is taken through it here."""
+ self.max_queue_wait = max(self.max_queue_wait, time.monotonic() - job.queued)
+ try:
+ if not job.req:
+ n = len(job.body) if job.content_length is None else job.content_length
+ job.req = self.front(job.method, job.path, n, lambda _n: job.body)
+ self.first_request()
+ route = job.req[0]
+ if route == "reply":
+ return job.finish("raw", *job.req[1:])
+ if route == "/status":
+ return job.finish("raw", *self._reply_json(200, self.status()))
+ try:
+ {"/submit": self._rt_submit, "/observe": self._rt_observe,
+ "/reset": self._rt_reset}[route](job.req[1], job)
+ except Refused as exc: # 409, or a 400 only the episode's state decides: nothing changed
+ self._count(refused=1)
+ job.finish("raw", *self._reply_json(exc.code, exc.body))
+ except Exception as exc: # noqa: BLE001 -- it ran and failed: the episode ends, as on hardware
+ traceback.print_exc()
+ self.log("error", path=route, error=repr(exc)[:500])
+ self._end("error")
+ self._release_watches()
+ if not job.done.is_set():
+ job.finish("raw", *self._reply_json(500, {"error": "simulator error", "detail": repr(exc)[:200]}))
+ except Exception as exc: # noqa: BLE001 -- never leave an HTTP thread waiting forever
+ traceback.print_exc()
+ if not job.done.is_set():
+ job.finish("raw", *self._reply_json(500, {"error": "service error", "detail": repr(exc)[:200]}))
+
+ def _view_rt(self, meta: dict, default_video: str) -> tuple[list[str], int, str, bool]:
+ """2.1's _view, plus cameras = [] for no images."""
+ none = isinstance(meta.get("cameras"), list) and not meta["cameras"]
+ cams, fps, video, depth = self._view(meta, default_video)
+ return ([], fps, "none", False) if none else (cams, fps, video, depth)
+
+ def _in_force(self, step: int, nxt: int, typ: str) -> np.ndarray | None:
+ """The `typ` row in force at control step `step` (>= nxt - 1) as things stand: the plan's row for it, else the
+ plan's last row before it (held); with nothing planned up to there, or a step already run, the simulator's
+ current command of that type. None when the row in force there is of ANOTHER action type: rows of different
+ types have different layouts and are never combined, and what that row leaves in force for `typ` is not
+ known before it runs."""
+ for i in range(min(step - nxt, len(self.plan) - 1), -1, -1):
+ t, row, _ = self.plan[i]
+ if t is not None: # (a held step leaves in force what was before it)
+ return np.asarray(row, dtype=np.float64) if t == typ else None
+ return np.asarray(self.sim.command(typ), dtype=np.float64)
+
+ def _rt_submit(self, req: dict, job: _Job) -> None:
+ """A checked chunk (_check_submit) into the plan: what depends on the episode's state, and the planning."""
+ if self.done or self.stopping:
+ raise self._conflict()
+ nxt = len(self.episode)
+ typ, A, dt, k, interp, at = (req[x] for x in ("type", "actions", "dt", "k", "interpolate", "at"))
+ cams, fps, video, depth = req["view"]
+ if at is not None and at - nxt > self.max_ahead:
+ raise Refused(400, {"error": f"at = {at} is more than max_ahead_steps = {self.max_ahead} control steps "
+ f"ahead of the next one ({nxt})"})
+ at = nxt if at is None else at
+ rows, clipped = self.sim.clip(typ, A)
+ late = max(0, nxt - at)
+ keep = at - nxt if at > nxt else 0
+ try:
+ with bad_request(): # planning runs nothing: whatever fails here changed nothing (400)
+ rows = np.asarray(rows, dtype=np.float64)
+ target = self._interp[typ] # a target can be interpolated and held; a command cannot
+ held = self._in_force(at - 1, nxt, typ) if target else None
+ if held is not None and held.shape != rows.shape[1:]:
+ held = None # (an adapter whose command() has not the type's layout)
+ if target and interp == "linear": # linear starts from the target in force just before `at`
+ prev = np.asarray(self.sim.command(typ), dtype=np.float64) if held is None else held
+ if rows.ndim != 2 or prev.shape != rows.shape[1:]:
+ raise ValueError(f"{typ}: cannot interpolate rows of shape {list(rows.shape[1:])} from the "
+ f"command in force, of shape {list(prev.shape)}; send interpolate='hold'")
+ steps = []
+ for row in rows:
+ steps += [prev + (row - prev) * (j / k) for j in range(1, k + 1)]
+ prev = row
+ else: # hold; and a command stands for itself, one row one control step
+ steps = [row for row in rows for _ in range(k)]
+ head = list(self.plan)[:keep]
+ if len(head) < keep:
+ # A gap before `at`. A target of this type in force there is held through it (rows of call -1).
+ # Anything else -- a command (repeating it would issue it again at every step of the gap), or a
+ # plan that holds another action type there -- leaves the gap to the simulator's hold row, as
+ # when the plan has run out.
+ head += [HELD if held is None else (typ, held, -1)] * (keep - len(head))
+ except Refused as exc: # the request itself passed its checks: keep it on the record
+ self.log("plan_refused", type=typ, at=at, error=str(exc)[:300])
+ raise
+ self.calls += 1
+ self.rows += len(A)
+ self.plan = collections.deque(head + [(typ, s, self.calls) for s in steps[late:]])
+ self.late_steps += min(late, len(steps))
+ self._start_clock()
+ rec = Recorder(self.views, cams, fps, depth, "none")
+ rec.start()
+ rec.finish()
+ out = {**self.meta(), "chunk_id": self.calls, "type": typ, "n_rows": int(len(A)), "dt": dt,
+ "interpolate": interp, "at": at, "start_step": max(at, nxt), "late_steps": late,
+ "plan_end_step": nxt + len(self.plan), "clipped": clipped[:200], "n_clipped": len(clipped), "ok": True}
+ self._save_call({**out, "note": req["note"]}, A)
+ job.finish("rec", out, rec)
+
+ def _rt_observe(self, req: dict, job: _Job) -> None:
+ with bad_request():
+ seconds, (cams, fps, video, depth) = req["seconds"], req["view"]
+ n = int(round(seconds / self.sim.control_dt)) if not (self.done or self.stopping) else 0
+ rec = Recorder(self.views, cams, fps, depth, video if n or video != "mp4" else "frames") # no time: no video
+ self._frame_budget(rec, n)
+ self.calls += 1
+ rec.start()
+ if not n:
+ rec.finish()
+ out = {**self.meta(), "obs_id": self.calls, "seconds": 0.0, "control_steps_watch": 0, "ok": True}
+ self._save_call(out, None)
+ return job.finish("rec", out, rec)
+ self._start_clock()
+ self.watches.append(_Watch(job, rec, len(self.episode), n, cams, self.calls))
+
+ def _complete_watch(self, w: _Watch) -> None:
+ self.watches.remove(w)
+ w.rec.finish()
+ ran = max(0, min(len(self.episode), w.end) - w.start)
+ out = {**self.meta(), "obs_id": w.obs_id, "seconds": round(ran * self.sim.control_dt, 4),
+ "control_steps_watch": ran, "ok": True}
+ self._save_call(out, None)
+ w.job.finish("rec", out, w.rec)
+
+ def _release_watches(self) -> None:
+ for w in list(self.watches):
+ self._complete_watch(w)
+
+ def _rt_reset(self, meta: dict, job: _Job) -> None:
+ self._reset_view(meta) # a refused reset (409, 400) changes nothing: the watches go on
+ self._release_watches()
+ out, prts = self._reset(meta, {}) # 2.1's reset: abandons the episode, a new one (paused)
+ self._stop_clock("reset")
+ job.finish("rcb1", {**out, **self._realtime_counters()}, prts)
+
+ # ------------------------------------------------------------------ the loop (simulator thread) and HTTP
+ def poll(self) -> None:
+ """After SIGUSR1: end the episode, answer the pending watches, finalize (Service.poll, plus the watches)."""
+ if self.stopping and not self.finalized:
+ if not self.done:
+ self._end("timeout")
+ self._release_watches()
+ self._finalize()
+
+ @staticmethod
+ def _renders(job: _Job) -> bool:
+ """Does serving this request render cameras? (What a request without images costs is bounded by its checks.)"""
+ route, req = (job.req + (None, None))[:2]
+ if route in ("/submit", "/observe"):
+ return bool(req["view"][0])
+ return route == "/reset" and not (isinstance(req.get("cameras"), list) and not req["cameras"])
+
+ def _may_serve(self, job: _Job, behind: bool, now: float) -> bool:
+ if not self.running or self.done or self.stopping:
+ return True # no clock to keep
+ kind = self._renders(job)
+ return not (kind and now < self._render_gate) and not (behind and now < self._gate[kind])
+
+ def _take(self, behind: bool, timeout: float) -> _Job | None:
+ """The next request to serve now, waiting up to `timeout` for one: the oldest held one whose time has come,
+ else the queue's next -- which is held instead, if its time has not come."""
+ now = time.monotonic()
+ for job in self._held:
+ if self._may_serve(job, behind, now):
+ self._held.remove(job)
+ return job
+ if self._held and self.running: # do not sleep past a held request's time
+ gates = [self._gate[k] for k in (True, False)] + [self._render_gate]
+ timeout = min([timeout] + [g - now for g in gates if g > now])
+ try:
+ job = self.q.get(timeout=timeout) if timeout > 0 else self.q.get_nowait()
+ except queue.Empty:
+ return None
+ if job.req and not self._may_serve(job, behind, time.monotonic()):
+ self._held.append(job)
+ return None
+ return job
+
+ def loop(self, until: threading.Event | None = None) -> None:
+ """The simulator thread: control steps when due, requests in between, the wrap-up after SIGUSR1."""
+ while until is None or not until.is_set():
+ self.poll()
+ if self.t0 is None and self._t_seen is not None:
+ self.first_request() # a request that ended on its HTTP thread started the trial's time
+ wait = self.due_in(time.monotonic())
+ if wait is not None and wait <= 0:
+ self.tick()
+ # Behind the clock still (the next step is due at once): the steps come first. A waiting request is
+ # served once the clock has caught up -- so requests queue behind each other, not into the clock --
+ # or, for a simulator that cannot keep up by itself, BEHIND_PACE times the last request's time after
+ # it (a cheap request: at once, one per step; a long render: a fifth of the time at most).
+ wait = self.due_in(time.monotonic())
+ job = self._take(wait is not None and wait <= 0, 0.0)
+ else:
+ job = self._take(False, 0.25 if wait is None else min(wait, 0.25))
+ if job is None:
+ continue
+ kind, t = bool(job.req) and self._renders(job), time.monotonic()
+ self.dispatch(job)
+ now = time.monotonic()
+ self.request_busy_s += now - t
+ self._gate[kind] = now + BEHIND_PACE * (now - t)
+ if kind and now - t > MAX_LAG_S: # a stall by itself: the next render waits, clock on time or not
+ self._render_gate = self._gate[kind]
+
+ def submit_job(self, job: _Job) -> tuple[int, str, bytes]:
+ """From an HTTP thread: queue the request, wait for the simulator thread, encode the answer here."""
+ self.q.put(job)
+ job.done.wait()
+ kind, *rest = job.reply
+ if kind == "raw":
+ return tuple(rest)
+ meta, payload = rest # a Recorder's parts are encoded here, off the simulator's thread
+ return self._reply_rcb1(meta, payload.parts() if kind == "rec" else payload)
+
+ def http_server(self, address: tuple[str, int]) -> ThreadingHTTPServer:
+ svc = self
+
+ class Handler(BaseHTTPRequestHandler):
+ timeout = 120
+
+ def log_message(self, *a):
+ pass
+
+ def _send(self, code: int, ctype: str, body: bytes) -> None:
+ self.send_response(code)
+ self.send_header("Content-Type", ctype)
+ self.send_header("Content-Length", str(len(body)))
+ self.end_headers()
+ self.wfile.write(body)
+
+ def do_GET(self):
+ self._serve("GET", 0)
+
+ def do_POST(self):
+ try:
+ n = int(self.headers.get("Content-Length") or 0)
+ except ValueError:
+ n = -1
+ if not 0 <= n <= MAX_BODY: # refused unread: drop the connection with it
+ self.close_connection = True
+ self._serve("POST", n)
+
+ def _serve(self, method: str, n: int) -> None:
+ try:
+ req = svc.front(method, self.path, n, self.rfile.read) # all that needs no simulator: here
+ except OSError: # the client went away mid-body
+ raise
+ except Exception as exc: # noqa: BLE001 -- never an HTTP thread's traceback instead of an answer
+ traceback.print_exc()
+ req = ("reply", *svc._reply_json(500, {"error": "service error", "detail": repr(exc)[:200]}))
+ self._send(*(req[1:] if req[0] == "reply" else svc.submit_job(_Job(method, self.path, req=req))))
+
+ class Server(ThreadingHTTPServer):
+ daemon_threads = True
+
+ return Server(address, Handler)
+
+ def serve(self, port: int = 8000) -> None:
+ import signal
+ signal.signal(signal.SIGUSR1, self.on_sigusr1)
+ srv = self.bind(port)
+ sys.setswitchinterval(SWITCH_INTERVAL_S) # the HTTP threads give the simulator thread its turn quickly
+ threading.Thread(target=srv.serve_forever, name="http", daemon=True).start()
+ print(f"{rcb_service.PROTOCOL} sim service (realtime, slowdown {self.slowdown:g}) ready on :{port}", flush=True)
+ self.loop()
+
+
+def main(build, port: int = 8000) -> None:
+ """rcb_service.main -- the start count, back-off, startup watchdog, resume -- with this service in place of
+ Service (RCB_SLOWDOWN: 1 = real time), and eai-standard/2.3 in everything it logs and writes."""
+ rcb_service.Service = RealtimeService
+ rcb_service.PROTOCOL = PROTOCOL
+ rcb_service.main(build, port)
+
+
+if __name__ == "__main__":
+ sys.path.insert(0, os.path.dirname(os.path.abspath(__file__)))
+ import server # noqa: E402 -- the HumanoidBench adapter; picks the GL backend and restores the reward on import
+
+ main(server.build)
diff --git a/scripts/myosuite/shared/realtime/rcb_service.py b/scripts/myosuite/shared/realtime/rcb_service.py
new file mode 100644
index 0000000000..5b31ed7428
--- /dev/null
+++ b/scripts/myosuite/shared/realtime/rcb_service.py
@@ -0,0 +1,835 @@
+"""eai-standard/2.1 sim service: the simulator-agnostic half of a standard-mode sidecar (docs/PROTOCOL.md §4,
+docs/STANDARD_MODE.md).
+
+The robot as a service: the agent reaches a simulator it cannot open over HTTP. GET /spec, GET /status, POST /observe
+and POST /submit; POST /reset exists only when the run grants resets (budget.reset > 0); everything else is 404.
+Responses that carry data are RCB1 frames (rcb1.py), never base64. Everything simulator-specific comes in through an
+adapter (the `Sim` contract below), so this file imports no simulator: a sidecar is its benchmark's adapter plus this.
+
+One thread does all of it -- the HTTP requests, the SIGUSR1 wrap-up, the finalize -- because a simulator's GL context
+belongs to the thread that made it (MuJoCo renders black frames from any other).
+
+The episode starts when the service is ready. It ends on success, by the task's own rule ("failed": the robot fell,
+the episode's control steps ran out), on a simulator error ("error"), or at SIGUSR1 ("timeout": the collect hook, once
+the agent's container is gone). With a reset budget, POST /reset abandons the running episode and starts another from
+the frozen instance; only the last episode is graded, and a success ends the trial.
+
+A request is checked in full before anything runs: whatever is wrong with it is a 400 and nothing ran. Once it runs,
+any failure is the simulator's: a 500, and the episode ends ("error").
+
+Evidence under DATA, never visible to the agent:
+ episode.jsonl one line per event; every state the service may resume from is a checkpoint (it carries
+ state_sha and journal_rows)
+ trajectory_live.f64 the journal: every executed control step in the simulator's own action space, float64
+ calls/NNNNNN/ each call's response meta and the actions as sent
+ startups.jsonl one line per process start
+ final.json kept current (written atomically); "finalized": true once the verdict is in -- before the
+ video is rendered ("video_pending": true while it is), so the video never decides anything
+ trajectory.npz, replay.mp4, replay.last.png the graded episode and its fresh replay
+
+Crash recovery: compose restarts a crashed service (restart: on-failure). It replays the journal up to the last
+checkpoint, checks the state hash, and continues the SAME episode -- never a reset. A call in flight is lost as a
+whole (its caller got an error, not a result, so nothing was learned). More than RCB_MAX_RESUMES restarts since the
+episode began (every process start counts, also one that died building the scene or replaying the journal) end the
+trial: sim_crashed_on_call if every crash happened while serving a call (the agent's failure), else sim_crashed
+(infrastructure, infra_error = 1). A verdict already on disk is never overwritten; a crash, a hang or a slow render
+of the video loses only the video.
+"""
+from __future__ import annotations
+
+import contextlib
+import errno
+import faulthandler
+import json
+import os
+import signal
+import tempfile
+import threading
+import time
+import traceback
+from http.server import BaseHTTPRequestHandler, HTTPServer
+from typing import Callable
+from urllib.parse import urlparse
+
+import numpy as np
+
+import rcb1
+
+PROTOCOL = "eai-standard/2.1"
+MAX_BODY = 64 * 1024 * 1024 # request bodies above this are refused (413)
+MAX_RESPONSE = 256 * 1024 * 1024 # responses above this drop frames/ and depth/ (meta.degraded says so)
+MAX_VIDEO_FPS = 30
+MAX_FRAMES_PER_CALL = 3000 # images in one response (cameras x frames, depth counted again): bounds memory
+VIDEOS = ("mp4", "frames", "both", "none")
+ON_SERVICE_ERROR = ("if a call fails with a connection error or a 5xx, the service may have crashed and resumed the "
+ "same episode (never a reset): the state is the one after the last call that completed on the "
+ "service, which you may not have received; check GET /status before continuing")
+
+
+class Sim:
+ """The adapter contract, duck-typed (this class only documents it). One object = one frozen scene.
+
+ Rows come in three forms: what the agent sends (an action of one of spec()["action_types"]), the same after
+ clip(), and a "sim row" (to_sim()), the action in the simulator's own space, which step() executes and the
+ journal and the graded trajectory record (raw_dim floats)."""
+ control_dt: float # seconds per control step; a chunk's dt is a whole multiple of it
+ max_control_steps: int # the task's episode length: the episode ends there
+ cameras: tuple[str, ...] # every camera the agent may ask for
+ depth_cameras: tuple[str, ...] # the ones that also give depth
+ proprio_layout: list[str] # the proprio row's columns
+ raw_dim: int # a sim row's width
+
+ def spec(self) -> dict: ... # simulator, robot, task_text, episode, action_types, cameras, robot facts
+ def interpolable(self, type: str) -> bool: ... # may a row of this type span several control steps?
+ def images(self, cameras: list[str], depth: bool) -> tuple[dict, dict]: ... # ({cam: uint8 HxWx3}, {cam: float32 HxW m})
+ def proprio(self) -> np.ndarray: ... # float32, one value per proprio_layout column
+ def clip(self, type: str, rows: np.ndarray) -> tuple[np.ndarray, list]: ... # into the limits; [{row, index, reason}]
+ def to_sim(self, type: str, row: np.ndarray) -> np.ndarray: ... # an agent row -> a sim row
+ def command(self, type: str) -> np.ndarray: ... # the last row executed, exactly as clip() returned it: a chunk
+ # interpolates from it, so a chunk equals its rows sent one by one
+ def hold(self) -> tuple[str, np.ndarray]: ... # the row that holds the last command (/observe with seconds > 0)
+ def step(self, row: np.ndarray) -> tuple[bool, bool, dict]: ... # one control step -> (done, success, info)
+ def reset(self) -> None: ... # back to the frozen instance
+ def state_sha(self) -> str: ...
+ def verify(self, rows: np.ndarray) -> dict: ... # replay in a fresh simulator -> {"success", "state_sha"}
+ def save_trajectory(self, rows: np.ndarray, path: str, meta: dict) -> None: ...
+ def render_video(self, rows: np.ndarray, path: str) -> None: ... # path and path.last.png
+
+
+class GiveUp(Exception):
+ """The service will not run this trial's episode again; final.json says why."""
+
+
+class Refused(Exception):
+ """A request the service turns down with `code` and the JSON `body`; nothing was executed."""
+
+ def __init__(self, code: int, body: dict):
+ super().__init__(body.get("error"))
+ self.code, self.body = code, body
+
+
+def bad_request():
+ """Checking a request, before anything runs: whatever goes wrong here is the request's fault (400)."""
+ try:
+ yield
+ except Refused:
+ raise
+ except Exception as exc: # noqa: BLE001 -- a wrong type, value or shape of any field
+ raise Refused(400, {"error": str(exc)[:300] or type(exc).__name__}) from None
+
+
+def count_start(data: str) -> int:
+ """Record one process start in startups.jsonl; returns how many there have been."""
+ path = os.path.join(data, "startups.jsonl")
+ with open(path, "a") as f:
+ f.write(json.dumps({"ts": time.time(), "pid": os.getpid()}) + "\n")
+ with open(path) as f:
+ return sum(1 for _ in f)
+
+
+def write_json_atomic(path: str, obj: dict) -> None:
+ """Write then rename: the collect hook polling the file never reads half of it."""
+ fd, tmp = tempfile.mkstemp(dir=os.path.dirname(path), suffix=".tmp")
+ with os.fdopen(fd, "w") as f:
+ json.dump(obj, f)
+ os.replace(tmp, path)
+
+
+def encode_mp4(frames: np.ndarray, fps: float) -> bytes:
+ import imageio.v2 as imageio
+ fd, path = tempfile.mkstemp(suffix=".mp4")
+ os.close(fd)
+ try:
+ with imageio.get_writer(path, format="FFMPEG", mode="I", fps=fps, codec="libx264", macro_block_size=1,
+ output_params=["-movflags", "+faststart", "-pix_fmt", "yuv420p"]) as w:
+ for f in frames:
+ w.append_data(np.ascontiguousarray(f))
+ with open(path, "rb") as f:
+ return f.read()
+ finally:
+ os.unlink(path)
+
+
+class Recorder:
+ """What one call returns: a proprio row at its start, every `every` control steps (the requested video_fps, rounded to
+ whole control steps) and at its end; with them the cameras' frames when a video or the frames are asked for, else
+ only the last frame (rendering is most of a call's cost)."""
+
+ def __init__(self, sim, cameras: list[str], fps: int, depth: bool, video: str):
+ self.sim, self.cams, self.depth, self.video = sim, cameras, depth, video
+ self.every = max(1, round(1.0 / (sim.control_dt * fps)))
+ self.fps = 1.0 / (self.every * sim.control_dt)
+ self.images = video != "none"
+ self.frames = {c: [] for c in cameras}
+ self.dep = {c: [] for c in cameras}
+ self.prop: list[np.ndarray] = []
+ self.k, self.grabbed_at = 0, None
+
+ def images_per_call(self, steps: int) -> int:
+ per_frame = len(self.cams) + (len(set(self.cams) & set(self.sim.depth_cameras)) if self.depth else 0)
+ return (steps // self.every + 2 if self.images else 1) * per_frame
+
+ def start(self) -> None:
+ self._grab(self.images)
+
+ def step(self) -> None:
+ self.k += 1
+ if self.k % self.every == 0:
+ self._grab(self.images)
+
+ def finish(self) -> None:
+ if self.grabbed_at != self.k:
+ self._grab(self.images)
+ if not self.images:
+ self._grab(True, proprio=False) # the one frame a call without video returns: the last
+
+ def _grab(self, images: bool, proprio: bool = True) -> None:
+ if images:
+ rgb, dep = self.sim.images(self.cams, self.depth)
+ for c in self.cams:
+ self.frames[c].append(rgb[c])
+ if c in dep:
+ self.dep[c].append(dep[c])
+ if proprio:
+ self.prop.append(np.asarray(self.sim.proprio(), np.float32))
+ self.grabbed_at = self.k
+
+ def parts(self) -> list:
+ out = []
+ for c in self.cams:
+ fr = np.stack(self.frames[c])
+ if self.video in ("mp4", "both"):
+ out.append((f"video/{c}", "mp4", encode_mp4(fr, self.fps), {"fps": round(self.fps, 3), "n": int(len(fr))}))
+ if self.video in ("frames", "both"):
+ out.append(rcb1.array_part(f"frames/{c}", fr))
+ out.append(rcb1.array_part(f"last/{c}", fr[-1]))
+ if self.dep[c]:
+ out.append(rcb1.array_part(f"depth/{c}", np.stack(self.dep[c]).astype(np.float16)))
+ out.append(rcb1.array_part("proprio", np.stack(self.prop)))
+ return out
+
+
+class Service:
+ """Episode bookkeeping, evidence and the HTTP endpoints around one `Sim`."""
+
+ def __init__(self, sim, *, data: str = "/data", reset_budget: int = 0, wall_clock_s: float = 3600.0,
+ max_chunk_s: float = 10.0, max_watch_s: float = 30.0, max_frames_per_call: int = MAX_FRAMES_PER_CALL,
+ max_resumes: int = 2):
+ self.sim, self.data = sim, data
+ self.reset_budget, self.wall_clock_s = int(reset_budget), float(wall_clock_s)
+ self.max_chunk_s, self.max_watch_s = float(max_chunk_s), float(max_watch_s)
+ self.max_frames, self.max_resumes = int(max_frames_per_call), int(max_resumes)
+ os.makedirs(os.path.join(data, "calls"), exist_ok=True)
+ self.journal_path = os.path.join(data, "trajectory_live.f64")
+ self.spec = self._spec()
+ self.types = {k: v for k, v in self.spec["action_types"].items() if isinstance(v, dict)}
+ self.calls = self.rows = self.resets = 0
+ self.journal_rows = 0 # control steps executed over the whole trial (the journal's length)
+ self.ep_start = 0 # where the current episode begins in the journal
+ self.episode: list[np.ndarray] = [] # the current episode's sim rows: what finalize grades
+ self.done, self.reason, self.success = False, None, False
+ self.stopping = False # SIGUSR1 seen: the agent is gone
+ self.finalized, self.verdict = False, {}
+ self.t0 = self.t_end = None # the clock starts at the agent's first request
+ self.starts = 1
+ self._unwritten: list[np.ndarray] = [] # sim rows executed since the last checkpoint
+
+ # ------------------------------------------------------------------ spec and status
+ def _spec(self) -> dict:
+ own = self.sim.spec()
+ core = {"protocol": PROTOCOL, "control_dt": self.sim.control_dt, "max_control_steps": self.sim.max_control_steps,
+ "max_chunk_s": self.max_chunk_s, "max_watch_s": self.max_watch_s, "max_video_fps": MAX_VIDEO_FPS,
+ "max_frames_per_call": self.max_frames, "proprio_layout": list(self.sim.proprio_layout),
+ "budget": {"reset": self.reset_budget, "wall_clock_s": self.wall_clock_s},
+ "clock": "time_s and elapsed_s count from your first request to the service; time_left_s is "
+ "budget.wall_clock_s minus time_s",
+ "on_service_error": ON_SERVICE_ERROR}
+ if clash := sorted(set(core) & set(own)):
+ raise ValueError(f"the adapter's spec() must not set {clash}")
+ cams = {c: {"rgb": True, "depth": c in self.sim.depth_cameras, **own["cameras"].get(c, {})}
+ for c in self.sim.cameras}
+ spec = {**core, **own, "cameras": cams}
+ if self.reset_budget:
+ spec["reset"] = ("POST /reset abandons the running episode and starts a new one from the same starting "
+ "state, budget.reset times in all; only the last episode is graded, and a success ends "
+ "the trial")
+ return spec
+
+ def clock(self) -> float:
+ return 0.0 if self.t0 is None else time.time() - self.t0
+
+ def _counters(self) -> dict:
+ out = {"calls": self.calls, "rows": self.rows, "control_steps": len(self.episode), "done": self.done,
+ "reason": self.reason}
+ if self.reset_budget:
+ out.update(episode=self.resets + 1, control_steps_total=self.journal_rows,
+ resets_left=0 if self.success or self.stopping else self.reset_budget - self.resets)
+ return out
+
+ def status(self) -> dict:
+ return {**self._counters(), "elapsed_s": round(self.clock(), 1)}
+
+ def meta(self) -> dict:
+ t = self.clock()
+ return {**self._counters(), "sim_time_s": round(len(self.episode) * self.sim.control_dt, 4),
+ "time_s": round(t, 2), "time_left_s": max(0.0, round(self.wall_clock_s - t, 1)),
+ "proprio": {"layout": list(self.sim.proprio_layout),
+ "now": np.asarray(self.sim.proprio(), np.float32).tolist()}}
+
+ # ------------------------------------------------------------------ evidence
+ def log(self, ev: str, **kw) -> None:
+ with open(os.path.join(self.data, "episode.jsonl"), "a") as f:
+ f.write(json.dumps({"ts": time.time(), "ev": ev, **kw}) + "\n")
+
+ def _events(self) -> list[dict]:
+ try:
+ with open(os.path.join(self.data, "episode.jsonl")) as f:
+ return [json.loads(line) for line in f if line.strip()]
+ except OSError:
+ return []
+
+ def _checkpoint(self, ev: str, **kw) -> None:
+ """Log a state this service may resume from, after appending everything executed since the last one to the
+ journal. Rows of a call that never reaches its checkpoint are lost with it, as the call is."""
+ if self._unwritten:
+ with open(self.journal_path, "ab") as f:
+ f.write(np.asarray(self._unwritten, dtype=np.float64).tobytes())
+ self._unwritten = []
+ self.log(ev, **kw, calls=self.calls, rows=self.rows, resets=self.resets, ep_start=self.ep_start,
+ journal_rows=self.journal_rows, done=self.done, reason=self.reason, success=self.success,
+ state_sha=self.sim.state_sha())
+ self._write_final()
+
+ def _write_final(self) -> None:
+ end = self.t_end or time.time()
+ d = {"protocol": PROTOCOL, "success": False, "live_success": self.success, "done": self.done,
+ "reason": self.reason, "calls": self.calls, "rows": self.rows, "control_steps": len(self.episode),
+ "control_steps_total": self.journal_rows, "episodes": self.resets + 1, "resets": self.resets,
+ "reset_budget": self.reset_budget, "elapsed_s": round(end - self.t0, 1) if self.t0 else 0.0,
+ "sim_startups": self.starts, "finalized": self.finalized, **self.verdict}
+ if "deterministic" in self.verdict:
+ d["success"] = bool(self.success and self.verdict["replay_success"] and self.verdict["deterministic"])
+ write_json_atomic(os.path.join(self.data, "final.json"), d)
+
+ def _save_call(self, meta: dict, actions: np.ndarray | None) -> None:
+ d = os.path.join(self.data, "calls", f"{self.calls:06d}")
+ os.makedirs(d, exist_ok=True)
+ with open(os.path.join(d, "meta.json"), "w") as f:
+ json.dump(meta, f)
+ if actions is not None:
+ np.save(os.path.join(d, "actions.npy"), actions)
+
+ # ------------------------------------------------------------------ start, resume, give up
+ def start(self) -> None:
+ """A fresh episode, or -- after a crash -- the same one again. Raises GiveUp when it must not run again."""
+ try:
+ with open(os.path.join(self.data, "startups.jsonl")) as f:
+ self.starts = max(1, sum(1 for _ in f))
+ except OSError:
+ self.starts = 1
+ ev = self._events()
+ if not any(e["ev"] == "start" for e in ev):
+ open(self.journal_path, "wb").close()
+ self._checkpoint("start", protocol=PROTOCOL, spec=self.spec, startup=self.starts)
+ else:
+ self._resume(ev)
+ with open(os.path.join(self.data, "server.pid"), "w") as f:
+ f.write(str(os.getpid()))
+
+ def _resume(self, ev: list[dict]) -> None:
+ path = os.path.join(self.data, "final.json")
+ try:
+ with open(path) as f:
+ fin = json.load(f)
+ except (OSError, ValueError):
+ fin = {}
+ if fin.get("finalized") or "deterministic" in fin:
+ # Graded already. The process died rendering the video (video_pending; before 2026-10, a verdict not yet
+ # finalized): all it lost is the video.
+ if not fin.get("finalized") or fin.get("video_pending"):
+ write_json_atomic(path, {**fin, "finalized": True, "video": False, "video_pending": False})
+ self._give_up("restarted after finalize", infra=False)
+ restarts = self.starts - int(next(e for e in ev if e["ev"] == "start")["startup"])
+ last = max(i for i, e in enumerate(ev) if e["ev"] in ("start", "resuming"))
+ in_call = False # did the process that died do so while serving a call?
+ for e in ev[last:]:
+ in_call = {"call_begin": True, "call_end": False}.get(e["ev"], in_call)
+ in_calls = sum(bool(e["crashed_in_call"]) for e in ev if e["ev"] == "resuming") + int(in_call)
+ if restarts > self.max_resumes:
+ agent = in_calls == restarts # every crash happened inside a call: the agent's commands did it
+ self._give_up("sim_crashed_on_call" if agent else "sim_crashed", infra=not agent)
+ self.log("resuming", crashed_in_call=in_call, restart=restarts)
+ cp = [e for e in ev if "state_sha" in e and "journal_rows" in e][-1]
+ n, dim = int(cp["journal_rows"]), int(self.sim.raw_dim)
+ rows = np.fromfile(self.journal_path, dtype=np.float64) if os.path.exists(self.journal_path) else np.zeros(0)
+ if rows.size < n * dim:
+ self._give_up("resume_failed: journal shorter than the log", infra=True)
+ rows = rows[: n * dim].reshape(n, dim)
+ for row in rows[cp["ep_start"]:]:
+ self.sim.step(row)
+ if self.sim.state_sha() != cp["state_sha"]:
+ self._give_up("resume_failed: replayed state differs from the logged one", infra=True)
+ rows.tofile(self.journal_path) # drop the steps of the call that was in flight
+ self.calls, self.rows, self.resets = int(cp["calls"]), int(cp["rows"]), int(cp["resets"])
+ self.ep_start, self.journal_rows = int(cp["ep_start"]), n
+ self.episode = [r.copy() for r in rows[self.ep_start:]]
+ self.done, self.reason, self.success = bool(cp["done"]), cp["reason"], bool(cp["success"])
+ self.t0 = next((e["t0"] for e in ev if e["ev"] == "first_request"), None)
+ if self.done:
+ self.t_end = next((e["ts"] for e in reversed(ev) if e["ev"] == "end"), cp["ts"])
+ self._checkpoint("resumed", downtime_s=round(time.time() - float(ev[-1]["ts"]), 1))
+
+ def _give_up(self, reason: str, infra: bool) -> None:
+ """End the trial without running its episode again: mark it failed (a finalized final.json is kept as it is),
+ leave the PID file for the collect hook, and stop."""
+ self.log("restart_refused", reason=reason)
+ path = os.path.join(self.data, "final.json")
+ try:
+ with open(path) as f:
+ fin = json.load(f)
+ except (OSError, ValueError):
+ fin = {}
+ if not fin.get("finalized"):
+ fin.update({"protocol": PROTOCOL, "success": False, "done": True, "reason": reason, "finalized": True,
+ **({"infra_error": 1} if infra else {})})
+ write_json_atomic(path, fin)
+ with open(os.path.join(self.data, "server.pid"), "w") as f:
+ f.write(str(os.getpid()))
+ raise GiveUp(reason)
+
+ # ------------------------------------------------------------------ requests
+ def first_request(self) -> None:
+ if self.t0 is None:
+ self.t0 = time.time()
+ self.log("first_request", t0=self.t0)
+
+ def handle(self, method: str, path: str, body: bytes = b"", content_length: int | None = None) -> tuple[int, str, bytes]:
+ """One request -> (status, content type, body): the whole HTTP surface, without the HTTP."""
+ self.first_request()
+ path = urlparse(path).path
+ if method == "GET" and path in ("/spec", "/status"):
+ return self._reply_json(200, self.spec if path == "/spec" else self.status())
+ routes = {"/submit": self._submit, "/observe": self._observe}
+ if self.reset_budget:
+ routes["/reset"] = self._reset
+ if method != "POST" or path not in routes:
+ return self._reply_json(404, {"error": "no such endpoint"})
+ if (len(body) if content_length is None else content_length) > MAX_BODY:
+ return self._reply_json(413, {"error": f"request body over {MAX_BODY} bytes"})
+ try:
+ meta, parts = self._decode(body)
+ except Exception as exc: # noqa: BLE001 -- any malformed body is the caller's error
+ return self._reply_json(400, {"error": f"malformed body: {exc}"[:300]})
+ self.log("call_begin", path=path, call=self.calls + 1)
+ try:
+ out, blobs = routes[path](meta, parts)
+ except Refused as exc: # 400 (bad_request), 409: nothing ran
+ return self._reply_json(exc.code, exc.body)
+ except Exception as exc: # noqa: BLE001 -- it ran and failed: the episode ends, as on hardware
+ traceback.print_exc()
+ self.log("error", path=path, error=repr(exc)[:500])
+ self._end("error")
+ return self._reply_json(500, {"error": "simulator error", "detail": repr(exc)[:200]})
+ finally:
+ self.log("call_end", path=path)
+ return self._reply_rcb1(out, blobs)
+
+ @staticmethod
+ def _decode(body: bytes) -> tuple[dict, dict]:
+ if body[:4] == rcb1.MAGIC:
+ meta, parts = rcb1.decode(body)
+ else:
+ meta, parts = json.loads(body or b"{}"), {}
+ if not isinstance(meta, dict):
+ raise ValueError("the body must be a JSON object or an RCB1 frame")
+ return meta, parts
+
+ @staticmethod
+ def _reply_json(code: int, obj: dict) -> tuple[int, str, bytes]:
+ return code, "application/json", json.dumps(obj).encode()
+
+ @staticmethod
+ def _reply_rcb1(meta: dict, parts: list) -> tuple[int, str, bytes]:
+ body = rcb1.encode(meta, parts)
+ if len(body) > MAX_RESPONSE:
+ parts = [p for p in parts if not p[0].startswith(("frames/", "depth/"))]
+ meta = {**meta, "degraded": f"response over {MAX_RESPONSE} bytes: frames/ and depth/ dropped"}
+ body = rcb1.encode(meta, parts)
+ return 200, "application/x-rcb1", body
+
+ def _view(self, meta: dict, default_video: str) -> tuple[list[str], int, str, bool]:
+ cams = meta.get("cameras") or list(self.sim.cameras)
+ if isinstance(cams, str):
+ cams = [c for c in cams.split(",") if c]
+ if not isinstance(cams, list) or (unknown := sorted(set(map(str, cams)) - set(self.sim.cameras))):
+ raise ValueError(f"unknown camera(s) {unknown if isinstance(cams, list) else cams}; "
+ f"the cameras are {list(self.sim.cameras)}")
+ # Each camera once, in the order asked for. A name repeated N times would be rendered into one list N times and
+ # returned N times (N * N images), which the frame budget counted as N: a small request could exhaust the memory.
+ cams = list(dict.fromkeys(map(str, cams)))
+ try:
+ fps = int(meta.get("video_fps", 10))
+ except (TypeError, ValueError, OverflowError):
+ raise ValueError("video_fps must be an integer") from None
+ if not 1 <= fps <= MAX_VIDEO_FPS:
+ raise ValueError(f"video_fps must be in [1, {MAX_VIDEO_FPS}]")
+ video = meta.get("video", default_video)
+ if video not in VIDEOS:
+ raise ValueError(f"video must be one of {list(VIDEOS)}")
+ depth = bool(meta.get("include_depth", False))
+ if depth and not set(cams) & set(self.sim.depth_cameras):
+ raise ValueError(f"include_depth: none of {cams} has depth; spec()['cameras'][name]['depth'] says which do")
+ return cams, fps, video, depth
+
+ def _frame_budget(self, rec: Recorder, steps: int) -> None:
+ n = rec.images_per_call(steps)
+ if n > self.max_frames:
+ raise ValueError(f"this call would return {n} images (cameras x frames, depth counted again); the limit is "
+ f"{self.max_frames}: ask for fewer cameras, a lower video_fps or a shorter chunk")
+
+ def _conflict(self) -> Refused:
+ if self.stopping:
+ return Refused(409, {"error": "trial finished: your time is up", "reason": self.reason})
+ body = {"error": "episode finished", "reason": self.reason}
+ if self.reset_budget:
+ body["resets_left"] = 0 if self.success else self.reset_budget - self.resets
+ return Refused(409, body)
+
+ def _step(self, typ: str, row: np.ndarray) -> dict:
+ """One control step: execute, journal, and end the episode if it is over."""
+ sim_row = np.asarray(self.sim.to_sim(typ, row), dtype=np.float64)
+ done, success, info = self.sim.step(sim_row)
+ self._unwritten.append(sim_row)
+ self.journal_rows += 1
+ self.episode.append(sim_row)
+ if success:
+ self._end("success")
+ elif done or len(self.episode) >= self.sim.max_control_steps:
+ self._end("failed")
+ return info
+
+ def _end(self, reason: str) -> None:
+ if self.done:
+ return
+ self.done, self.reason, self.success = True, reason, reason == "success"
+ self.t_end = time.time()
+ try:
+ self._checkpoint("end")
+ except Exception as exc: # noqa: BLE001 -- a broken simulator may not even hash: the end is still logged
+ self.log("end", reason=reason, checkpoint_error=repr(exc)[:300])
+ self._write_final()
+
+ def _submit(self, meta: dict, parts: dict) -> tuple[dict, list]:
+ if self.done or self.stopping:
+ raise self._conflict()
+ with bad_request():
+ typ = meta.get("type")
+ if not isinstance(typ, str) or typ not in self.types:
+ raise ValueError(f"type must be one of {sorted(self.types)}")
+ try:
+ A = np.asarray(parts["actions"]["data"] if "actions" in parts else meta.get("actions"), dtype=np.float64)
+ except (TypeError, ValueError, KeyError):
+ raise ValueError("send the actions as an RCB1 array part `actions` or a JSON list of rows") from None
+ dim = int(self.types[typ]["dim"])
+ if A.ndim != 2 or A.shape[0] < 1 or A.shape[1] != dim:
+ raise ValueError(f"{typ} actions must be (T, {dim}) with T >= 1, got {list(A.shape)}")
+ if not np.all(np.isfinite(A)):
+ raise ValueError("actions contain NaN or inf")
+ cdt = self.sim.control_dt
+ try:
+ dt = cdt if meta.get("dt") is None else float(meta["dt"])
+ except (TypeError, ValueError):
+ raise ValueError("dt must be a number of seconds") from None
+ k = round(dt / cdt) if np.isfinite(dt / cdt) else 0
+ if k < 1 or abs(k * cdt - dt) > 1e-9:
+ raise ValueError(f"dt must be a positive whole multiple of control_dt = {cdt} s (never rounded)")
+ if k > 1 and not self.sim.interpolable(typ):
+ raise ValueError(f"{typ}: one row is one control step, so dt must be control_dt = {cdt} s")
+ if len(A) * k * cdt > self.max_chunk_s + 1e-9:
+ raise ValueError(f"the chunk covers {len(A) * k * cdt:g} s: more than max_chunk_s = {self.max_chunk_s:g}")
+ interp = meta.get("interpolate", "linear")
+ if interp not in ("linear", "hold"):
+ raise ValueError("interpolate must be linear or hold")
+ cams, fps, video, depth = self._view(meta, "mp4")
+ rec = Recorder(self.sim, cams, fps, depth, video)
+ self._frame_budget(rec, len(A) * k)
+ rows, clipped = self.sim.clip(typ, A)
+ before, info, executed, stop = len(self.episode), {}, 0, False
+ prev = np.asarray(self.sim.command(typ), dtype=np.float64)
+ rec.start()
+ for i, row in enumerate(rows):
+ for j in range(1, k + 1):
+ info = self._step(typ, row if interp == "hold" else prev + (row - prev) * (j / k))
+ rec.step()
+ if self.done or self.stopping or info.get("blocked"):
+ stop = True
+ break
+ executed, prev = i + 1, row
+ if stop:
+ break
+ rec.finish()
+ self.calls += 1
+ self.rows += executed
+ out = {**self.meta(), "chunk_id": self.calls, "type": typ, "executed_rows": executed,
+ "control_steps_chunk": len(self.episode) - before, "clipped": clipped[:200], "n_clipped": len(clipped),
+ **info, "ok": True}
+ self._checkpoint("submit", type=typ, n_rows=len(A), dt=dt, interpolate=interp, cameras=cams,
+ note=str(meta.get("note") or "")[:500], executed_rows=executed, n_clipped=len(clipped), **info)
+ self._save_call(out, A)
+ return out, rec.parts()
+
+ def _observe(self, meta: dict, parts: dict) -> tuple[dict, list]:
+ with bad_request():
+ try:
+ seconds = float(meta.get("seconds") or 0.0)
+ except (TypeError, ValueError):
+ raise ValueError("seconds must be a number") from None
+ if not 0.0 <= seconds <= self.max_watch_s:
+ raise ValueError(f"seconds must be in [0, {self.max_watch_s:g}]")
+ cams, fps, video, depth = self._view(meta, "frames")
+ n = int(round(seconds / self.sim.control_dt)) if not (self.done or self.stopping) else 0
+ rec = Recorder(self.sim, cams, fps, depth, video if n or video != "mp4" else "frames") # no time: no video
+ self._frame_budget(rec, n)
+ before = len(self.episode)
+ rec.start()
+ if n:
+ typ, row = self.sim.hold()
+ for _ in range(n):
+ self._step(typ, row)
+ rec.step()
+ if self.done or self.stopping:
+ break
+ rec.finish()
+ self.calls += 1
+ ran = len(self.episode) - before
+ out = {**self.meta(), "obs_id": self.calls, "seconds": round(ran * self.sim.control_dt, 4),
+ "control_steps_watch": ran, "ok": True}
+ self._checkpoint("observe", seconds=out["seconds"], cameras=cams)
+ self._save_call(out, None)
+ return out, rec.parts()
+
+ def _reset_view(self, meta: dict) -> tuple[list[str], int, str, bool]:
+ """A reset's checks, all of them, before anything changes: Refused (409, 400) or the view it returns."""
+ if self.success or self.stopping:
+ raise Refused(409, {"error": "trial finished: " + ("an episode succeeded" if self.success else "your time is up"),
+ "reason": self.reason})
+ if self.resets >= self.reset_budget:
+ raise Refused(409, {"error": "reset budget exhausted", "budget": self.reset_budget})
+ with bad_request():
+ return self._view(meta, "frames")
+
+ def _reset(self, meta: dict, parts: dict) -> tuple[dict, list]:
+ cams, fps, video, depth = self._reset_view(meta)
+ if not self.done:
+ self._end("reset") # the running episode is abandoned
+ try:
+ self.sim.reset()
+ except Exception:
+ # The simulator failed to start the new episode: the trial ends on a simulator error, not on the episode
+ # the agent had every right to abandon (handle() answers 500; its _end("error") finds the episode ended).
+ self.done, self.reason, self.success = True, "error", False
+ try:
+ self._checkpoint("end", reset_failed=True)
+ except Exception as exc: # noqa: BLE001 -- a broken simulator may not even hash
+ self.log("end", reason="error", reset_failed=True, checkpoint_error=repr(exc)[:300])
+ self._write_final()
+ raise
+ self.resets += 1
+ self.ep_start, self.episode = self.journal_rows, []
+ self.done, self.reason, self.success, self.t_end = False, None, False, None
+ rec = Recorder(self.sim, cams, fps, depth, "frames" if video == "mp4" else video) # no time: no video
+ rec.start()
+ rec.finish()
+ self.calls += 1
+ out = {**self.meta(), "ok": True}
+ self._checkpoint("reset", cameras=cams)
+ self._save_call(out, None)
+ return out, rec.parts()
+
+ # ------------------------------------------------------------------ the end of the trial
+ def on_sigusr1(self, *_) -> None:
+ """The agent is gone (the [[verifier.collect]] hook). Only a flag: the handler may fire in the middle of a
+ request; a chunk stops at its next control step and poll() wraps up between requests."""
+ self.stopping = True
+
+ def poll(self) -> None:
+ """Between requests (the HTTP loop calls it every 0.5 s): after SIGUSR1, end the episode and finalize."""
+ if self.stopping and not self.finalized:
+ if not self.done:
+ self._end("timeout")
+ self._finalize()
+
+ def _finalize(self) -> None:
+ """The graded episode (every control step since the last reset) as a trajectory and its replay in a fresh
+ simulator (success, and bit-exact agreement with the live run): the verdict, final from the moment it is
+ written ("finalized": true). Then a video for humans, which can no longer change or void it: while it renders
+ final.json says "video_pending": true (the collect hook waits for it as long as its time allows)."""
+ rows = None
+ try:
+ rows = np.asarray(self.episode, dtype=np.float64).reshape(-1, self.sim.raw_dim)
+ live_sha = self.sim.state_sha()
+ self.sim.save_trajectory(rows, os.path.join(self.data, "trajectory.npz"),
+ {"source": "standard-mode episode", "protocol": PROTOCOL, "reason": self.reason,
+ "episode": self.resets + 1})
+ rep = self.sim.verify(rows)
+ self.verdict = {"n_actions": int(len(rows)), "replay_success": bool(rep["success"]),
+ "deterministic": rep["state_sha"] == live_sha, "state_sha": live_sha, "video": False}
+ self.log("replay", **self.verdict)
+ except Exception as exc: # noqa: BLE001
+ traceback.print_exc()
+ self.log("finalize_error", error=repr(exc)[:500])
+ self.verdict = {"finalize_error": repr(exc)[:300]}
+ graded = "finalize_error" not in self.verdict
+ self.verdict["video_pending"] = graded
+ self.finalized = True
+ self._write_final() # the verdict, finalized, on disk before any video is rendered
+ if not graded:
+ return
+ try:
+ self.sim.render_video(rows, os.path.join(self.data, "replay.mp4"))
+ self.verdict["video"] = True
+ except Exception as exc: # noqa: BLE001 -- the video is for humans; it never decides the verdict
+ traceback.print_exc()
+ self.log("video_error", error=repr(exc)[:500])
+ self.verdict["video_pending"] = False
+ try:
+ self._write_final()
+ except Exception: # noqa: BLE001 -- the finalized verdict is on disk already; only the video's flag is lost
+ traceback.print_exc()
+
+ # ------------------------------------------------------------------ HTTP
+ def http_server(self, address: tuple[str, int]) -> HTTPServer:
+ svc = self
+
+ class Handler(BaseHTTPRequestHandler):
+ timeout = 120 # a client that stalls mid-request is dropped: one thread serves everyone
+
+ def log_message(self, *a):
+ pass
+
+ def _send(self, code: int, ctype: str, body: bytes) -> None:
+ self.send_response(code)
+ self.send_header("Content-Type", ctype)
+ self.send_header("Content-Length", str(len(body)))
+ self.end_headers()
+ self.wfile.write(body)
+
+ def do_GET(self):
+ self._send(*svc.handle("GET", self.path))
+
+ def do_POST(self):
+ try:
+ n = int(self.headers.get("Content-Length") or 0)
+ except ValueError:
+ n = -1
+ if n < 0: # (read(-1) would read to the end of the stream, without a limit)
+ self.close_connection = True
+ return self._send(*svc._reply_json(400, {"error": "bad Content-Length"}))
+ if n > MAX_BODY: # refused unread: drop the connection with it
+ self.close_connection = True
+ return self._send(*svc.handle("POST", self.path, b"", content_length=n))
+ self._send(*svc.handle("POST", self.path, self.rfile.read(n)))
+
+ class Server(HTTPServer):
+ def service_actions(self):
+ svc.poll()
+
+ return Server(address, Handler)
+
+ def bind(self, port: int) -> HTTPServer:
+ """The HTTP server on `port`. Something else may hold the port -- in the network namespace it shares with the
+ agent's container, an agent's process can take it while this service (re)starts -- so wait, not die (a death
+ here would count as a crash); the trial can still end meanwhile (SIGUSR1)."""
+ busy = False
+ while True:
+ try:
+ return self.http_server(("0.0.0.0", port))
+ except OSError as exc:
+ if exc.errno != errno.EADDRINUSE:
+ raise
+ if not busy:
+ busy = True
+ self.log("port_busy", port=port)
+ self.poll()
+ time.sleep(1.0)
+
+ def serve(self, port: int = 8000) -> None:
+ signal.signal(signal.SIGUSR1, self.on_sigusr1)
+ srv = self.bind(port)
+ print(f"{PROTOCOL} sim service ready on :{port}", flush=True)
+ srv.serve_forever()
+
+
+def refuse_forever(reason: str, port: int = 8000) -> None:
+ """After a give-up: answer every request 409 with the reason until the container is removed. Before, nothing
+ listened once the service had given up, and a client waited out its whole connection timeout for a simulator that
+ would never answer. The verdict is on disk already: nothing here reads or changes it."""
+ body = json.dumps({"error": "the sim service gave up on this episode and will not serve it", "reason": reason,
+ "gave_up": True}).encode()
+
+ class Refusal(BaseHTTPRequestHandler):
+ timeout = 30 # a client that stops sending must not hold the only thread
+
+ def _refuse(self):
+ try:
+ n = int(self.headers.get("Content-Length") or 0)
+ if 0 < n <= MAX_BODY: # take the request first: a client still sending would see a reset
+ self.rfile.read(n)
+ self.send_response(409)
+ self.send_header("Content-Type", "application/json")
+ self.send_header("Content-Length", str(len(body)))
+ self.end_headers()
+ self.wfile.write(body)
+ except (OSError, ValueError):
+ pass
+ self.close_connection = True
+
+ do_GET = do_POST = _refuse
+
+ def log_message(self, *args):
+ pass
+
+ def answer():
+ while True:
+ try:
+ HTTPServer(("0.0.0.0", port), Refusal).serve_forever()
+ except Exception: # noqa: BLE001 -- a busy port, or anything else: the give-up stands without the answer
+ time.sleep(5.0)
+
+ threading.Thread(target=answer, daemon=True).start()
+ while True:
+ time.sleep(3600)
+
+
+def main(build: Callable[[], Sim], port: int = 8000) -> None:
+ """A sidecar's life: count the start, back off after a failed one, build the scene under a startup watchdog, resume
+ or start the episode, serve. Settings: RCB_DATA (/data), RCB_RESET_BUDGET (0), RCB_WALL_CLOCK_S (3600),
+ RCB_STARTUP_TIMEOUT_S (300), RCB_MAX_RESUMES (2)."""
+ env = os.environ.get
+ data = env("RCB_DATA", "/data")
+ os.makedirs(data, exist_ok=True)
+ starts = count_start(data)
+ # Compose restarts a failed start at once; a scene that failed to build (e.g. the GPU full of other trials) would
+ # meet the same conditions every time, so wait before rebuilding -- only before the episode exists.
+ if starts > 1 and not os.path.exists(os.path.join(data, "episode.jsonl")):
+ time.sleep(min(120, 20 * (starts - 1)))
+ pending: list[int] = []
+ signal.signal(signal.SIGUSR1, lambda *_: pending.append(1)) # its default action would kill the process
+ # A scene build that hangs (no crash, so compose never restarts it) must end the process. Neither a Python thread
+ # (a native hang may hold the GIL) nor SIGALRM's default action (PID 1 ignores it) can; faulthandler's C watchdog
+ # dumps every thread's traceback and exits. Before the scene exists no episode exists: the retry is no reset.
+ faulthandler.dump_traceback_later(float(env("RCB_STARTUP_TIMEOUT_S", "300")), exit=True)
+ sim = build()
+ faulthandler.cancel_dump_traceback_later() # a resume replay may take minutes
+ svc = Service(sim, data=data, reset_budget=int(env("RCB_RESET_BUDGET") or 0),
+ wall_clock_s=float(env("RCB_WALL_CLOCK_S") or 3600), max_resumes=int(env("RCB_MAX_RESUMES") or 2))
+ try:
+ svc.start()
+ except GiveUp as exc:
+ print(f"restart refused: {exc}", flush=True)
+ signal.signal(signal.SIGUSR1, lambda *_: None)
+ refuse_forever(str(exc), port)
+ if pending:
+ svc.on_sigusr1()
+ svc.serve(port)
diff --git a/scripts/myosuite/shared/realtime/server.py b/scripts/myosuite/shared/realtime/server.py
new file mode 100644
index 0000000000..2895899a20
--- /dev/null
+++ b/scripts/myosuite/shared/realtime/server.py
@@ -0,0 +1,333 @@
+"""eai-standard/2.3 sim service for MyoSuite's standard-realtime tasks -- the adapter behind the shared realtime layer
+(`python /srv/rcb_realtime.py`, which imports this module and calls build(); docs/PROTOCOL.md §4b,
+docs/benchmarks/myosuite.md).
+
+The simulator-specific half only: `MyoSuite` implements rcb_service.Sim (the contract documented in rcb_service.py) for
+one frozen instance, on the benchmark's own tooling module rcb_myosuite.Session -- the same scene, action, success rule
+and replay as the standard twin's eai-standard/2.2 service (environment/sim/server.py of `-standard`). The protocol, the
+clock, the plan, the episode, crash recovery and the finalize are rcb_realtime.py's and rcb_service.py's, unchanged
+copies of the HumanoidBench sources (ci_checks/check-tooling-identity.sh).
+
+What the agent gets is what the standard twin gives, minus the 2.2 additions the realtime layer does not serve
+(`/kinematics`, per-frame camera poses, depth modes): the cameras (RGB, and depth on request from every camera that
+rides on the world or the body), the body's own proprioception (joint positions and velocities, muscle activations and
+lengths, the actuators' generalized force on each joint, and a free-standing body's root pose and velocity) and the
+actuator commands. Nothing about objects, balls, opponents or goals leaves this process; a camera that rides on a body
+of the scene gives images only (its pose would place that body).
+
+Rendering: MuJoCo on the GPU (EGL) when the container can reach the NVIDIA driver, else OSMesa on the CPU. In
+table_tennis an OSMesa frame takes 1.3 s and an EGL frame 5 ms; on a clock that does not stop, a render is time the
+simulator thread cannot step, so the realtime twins reserve a GPU. The NVIDIA container toolkit mounts libEGL_nvidia
+under NVIDIA_DRIVER_CAPABILITIES=graphics but, on some hosts, not its ICD manifest; this module writes one to /tmp and
+points libglvnd at it (inert where the library is absent: libglvnd falls back to Mesa), then tries EGL in a throwaway
+process, because MuJoCo binds its backend once per process at `import mujoco`.
+
+Settings (environment/docker-compose.yaml): RCB_INSTANCE (the frozen instance), RCB_TASK_INFO (task.json: task text,
+default camera), RCB_IMAGE_HW (the cameras' "H,W", default "480,640", the standard twin's), RCB_DATA, RCB_RESET_BUDGET,
+RCB_WALL_CLOCK_S, RCB_SLOWDOWN (rcb_realtime / rcb_service), RCB_GL (auto: EGL when it works, else OSMesa).
+"""
+from __future__ import annotations
+
+import json
+import math
+import os
+import subprocess
+import sys
+
+import numpy as np
+
+HERE = os.path.dirname(os.path.abspath(__file__))
+sys.path.insert(0, HERE)
+
+
+def _resolve_gl_backend() -> str:
+ """MuJoCo's offscreen backend: RCB_GL (egl | osmesa) if set, else EGL if a probe process can open a context on the
+ NVIDIA driver, else OSMesa. The image's own MUJOCO_GL (osmesa, for the CPU-only standard twin) is not a choice."""
+ if os.environ.get("RCB_GL", "auto") in ("egl", "osmesa"):
+ return os.environ["RCB_GL"]
+ if not os.path.exists("/dev/nvidiactl"):
+ return "osmesa"
+ vendors = [p for d in ("/usr/share/glvnd/egl_vendor.d", "/etc/glvnd/egl_vendor.d") if os.path.isdir(d)
+ for p in sorted(os.listdir(d))]
+ if not any("nvidia" in v for v in vendors) and not os.environ.get("__EGL_VENDOR_LIBRARY_FILENAMES"):
+ os.makedirs("/tmp/rcb_egl", exist_ok=True)
+ manifest = "/tmp/rcb_egl/10_nvidia.json"
+ with open(manifest, "w") as f:
+ f.write('{"file_format_version": "1.0.0", "ICD": {"library_path": "libEGL_nvidia.so.0"}}\n')
+ mesa = "/usr/share/glvnd/egl_vendor.d/50_mesa.json"
+ os.environ["__EGL_VENDOR_LIBRARY_FILENAMES"] = manifest + (f":{mesa}" if os.path.exists(mesa) else "")
+ probe = ("import os; os.environ['MUJOCO_GL'] = os.environ['PYOPENGL_PLATFORM'] = 'egl'; import mujoco; "
+ "c = mujoco.GLContext(8, 8); c.make_current(); import OpenGL.GL as g; "
+ "r = g.glGetString(g.GL_RENDERER).decode(); c.free(); print(r)")
+ try:
+ out = subprocess.run([sys.executable, "-c", probe], capture_output=True, text=True, timeout=120)
+ ok = out.returncode == 0 and "llvmpipe" not in out.stdout.lower() # Mesa's EGL renders on the CPU as well
+ except (OSError, subprocess.SubprocessError):
+ ok = False
+ return "egl" if ok else "osmesa"
+
+
+GL_BACKEND = _resolve_gl_backend()
+os.environ["MUJOCO_GL"] = os.environ["PYOPENGL_PLATFORM"] = GL_BACKEND
+os.environ["RCB_SESSION_LOG"] = ""
+import rcb_myosuite as M # noqa: E402 (after the backend: it only sets a default)
+import mujoco # noqa: E402
+
+M.SESSION_LOG = "" # the sidecar's Session is not the agent's interaction log
+from rcb_service import main # noqa: E402
+
+ROBOT = {
+ "hand": ("MyoHand", "a musculoskeletal model of a human forearm and hand (MyoSuite)"),
+ "arm": ("MyoArm", "a musculoskeletal model of a human shoulder, arm and hand (MyoSuite)"),
+ "body": ("MyoArm with torso", "a musculoskeletal model of a human torso, right arm and hand on a pelvis that slides "
+ "on the floor (MyoSuite)"),
+ "bimanual": ("MyoArm + MPL", "a musculoskeletal right arm and hand (MyoSuite's MyoArm) next to a motorized "
+ "prosthetic left arm (the Modular Prosthetic Limb)"),
+ "legs": ("MyoLegs", "a free-standing musculoskeletal model of a human torso and both legs (MyoSuite)"),
+ "osl": ("MyoLeg + OSL", "a free-standing musculoskeletal model of a human with the right leg's muscles and an "
+ "Open-Source Leg knee-ankle prosthesis on the left (MyoSuite)"),
+}
+MUSCLE_TEXT = (
+ "muscles: one real number a per muscle; excitation u = 1 / (1 + exp(-5 (a - 0.5))), a not clipped (a = -1: u ~ "
+ "0.0006, 0.5: 0.5, 1: 0.92, 2: 0.9994); activations follow excitations with a first-order delay (~10 ms up, ~40 ms "
+ "down); muscles only pull, one can cross several joints, and its force depends on its length and velocity")
+GL2CV = np.diag([1.0, -1.0, -1.0, 1.0])
+CAM_MODES = {0: "fixed", 1: "track", 2: "trackcom", 3: "targetbody", 4: "targetbodycom"}
+NO_POSE = ("not published: this camera rides on a body of the scene that is not yours, so its pose would be that "
+ "body's position; it gives images only")
+
+
+class MyoSuite:
+ """One frozen MyoSuite instance behind the rcb_service.Sim contract (see the module docstring)."""
+
+ def __init__(self, instance: str, task_info: dict, image_hw: tuple[int, int] = (480, 640)):
+ self.instance = instance
+ self.task = str(np.load(instance, allow_pickle=False)["task"])
+ self.info = task_info
+ self.s = M.Session(self.task, instance=instance)
+ m = self.m = self.s.model
+ self.H, self.W = image_hw
+ self.control_dt = float(self.s.dt)
+ self.max_control_steps = int(self.s.max_steps) # a runaway guard: the task's own rule ends an episode
+ self.nu = self.raw_dim = int(self.s.nu)
+ acts = self.s.actuators()
+ self.layout = [a["name"] for a in acts]
+ self.mus_idx = [a["index"] for a in acts if a["kind"] == "muscle"]
+ self.muscles = [acts[i]["name"] for i in self.mus_idx]
+ self.other = [a for a in acts if a["kind"] != "muscle"]
+ self.atype = "muscle" if not self.other else "actuators"
+ self.act_adr = np.array([int(m.actuator_actadr[i]) for i in self.mus_idx], dtype=int)
+ # the body's own joints (the key's hinge and the OBJ* joints are object state, never sent; free joints of balls,
+ # objects and paddles are not in S.joints(); a free-standing body's root is reported on its own)
+ limb = [j for j in self.s.joints() if j["name"] != "keyjoint" and not j["name"].startswith("OBJ")]
+ self.j_names = [j["name"] for j in limb]
+ self.j_q = np.array([j["qpos_adr"] for j in limb], dtype=int)
+ self.j_v = np.array([j["dof_adr"] for j in limb], dtype=int)
+ self.j_id = [m.joint(n).id for n in self.j_names]
+ root = mujoco.mj_name2id(m, mujoco.mjtObj.mjOBJ_JOINT, "root")
+ self.root = root if root >= 0 and int(m.jnt_type[root]) == 0 else None
+ self.own_joints = set(self.j_id) | ({self.root} if self.root is not None else set())
+ self.cameras = tuple(self.s.cameras())
+ self.scene_cams = {c for c in self.cameras if self._scene_camera(c)}
+ self.depth_cameras = tuple(c for c in self.cameras if c not in self.scene_cams)
+ self.proprio_layout = ([f"q_{n}" for n in self.j_names] + [f"qd_{n}" for n in self.j_names]
+ + [f"act_{n}" for n in self.muscles] + [f"len_{n}" for n in self.muscles]
+ + [f"effort_{n}" for n in self.j_names]
+ + (["root_x", "root_y", "root_z", "root_qw", "root_qx", "root_qy", "root_qz",
+ "root_vx", "root_vy", "root_vz", "root_wx", "root_wy", "root_wz"]
+ if self.root is not None else []))
+ self._depth_on = False
+ self.render_frame(self.cameras[0]) # warms up the GL context
+ self.cmd = self.s.rest_action()
+
+ # ------------------------------------------------------------------ what the agent may know
+ def _own_body(self, b: int) -> bool:
+ """A body fixed in the world or moved by the body's own joints alone; a mocap body (the opponent, goal markers)
+ or one behind any other joint (objects, balls, the paddle) is scene state."""
+ m = self.m
+ while b > 0:
+ if int(m.body_mocapid[b]) >= 0:
+ return False
+ first = int(m.body_jntadr[b])
+ if any(j not in self.own_joints for j in range(first, first + int(m.body_jntnum[b]))):
+ return False
+ b = int(m.body_parentid[b])
+ return True
+
+ def _scene_camera(self, cam: str) -> bool:
+ m, i = self.m, self.m.camera(cam).id
+ bodies = [int(m.cam_bodyid[i])] + ([int(m.cam_targetbodyid[i])] if int(m.cam_mode[i]) in (3, 4) else [])
+ return not all(self._own_body(b) for b in bodies if b >= 0)
+
+ def _intrinsic(self, cam: str) -> list:
+ f = 0.5 * self.H / math.tan(math.radians(float(self.m.cam_fovy[self.m.camera(cam).id])) / 2)
+ return [[f, 0.0, self.W / 2], [0.0, f, self.H / 2], [0.0, 0.0, 1.0]]
+
+ def _cam_T_gl(self, cam: str) -> np.ndarray:
+ i, d = self.m.camera(cam).id, self.s.data
+ T = np.eye(4)
+ T[:3, :3] = d.cam_xmat[i].reshape(3, 3) # MuJoCo cameras: OpenGL axes (x right, y up, z back)
+ T[:3, 3] = d.cam_xpos[i]
+ return T
+
+ def _camera_fact(self, cam: str) -> dict:
+ m, i = self.m, self.m.camera(cam).id
+ out = {"hw": [self.H, self.W], "intrinsic_cv": self._intrinsic(cam),
+ "fovy_deg": round(float(m.cam_fovy[i]), 2)}
+ if cam in self.scene_cams:
+ return {**out, "mounted": "on a body of the scene that is not yours (it moves with it): images only",
+ "pose": NO_POSE}
+ body, mode = m.body(int(m.cam_bodyid[i])).name, int(m.cam_mode[i])
+ if mode != 0:
+ return {**out, "mounted": f"tracking ({CAM_MODES.get(mode, mode)}) the body {body}: it follows that body's "
+ "position (or its subtree's centre of mass) without turning with it"}
+ if body == "world":
+ ext = np.linalg.inv(self._cam_T_gl(cam) @ GL2CV)[:3, :]
+ return {**out, "mounted": "static", "extrinsic_cv": ext.round(6).tolist(),
+ "pose_note": "extrinsic_cv: world -> camera, OpenCV axes (x right, y down, z forward)"}
+ b = m.body(body).id
+ Tb = np.eye(4)
+ Tb[:3, :3] = self.s.data.xmat[b].reshape(3, 3)
+ Tb[:3, 3] = self.s.data.xpos[b]
+ link_T_cam = np.linalg.inv(Tb) @ self._cam_T_gl(cam) @ GL2CV
+ return {**out, "mounted": f"fixed to the body {body}: it moves with it",
+ "link": body, "link_T_cam_cv": link_T_cam.round(6).tolist()}
+
+ # ------------------------------------------------------------------ the rcb_service.Sim contract
+ def spec(self) -> dict:
+ m, robot = self.m, ROBOT[M.BODY[self.task]]
+ lo = [float(m.jnt_range[j, 0]) if m.jnt_limited[j] else None for j in self.j_id]
+ hi = [float(m.jnt_range[j, 1]) if m.jnt_limited[j] else None for j in self.j_id]
+ action = {"dim": self.nu, "layout": self.layout, "lower": None, "upper": None,
+ "units": MUSCLE_TEXT + ("; the other actuators: see `actuators`" if self.other else ""),
+ "semantics": "one row per control step, every actuator in `layout` order; rows dt apart (a whole "
+ "multiple of control_dt), `linear` (default: a straight line from the command in force) "
+ "or `hold` between rows; nothing is clipped by the service (MuJoCo clips the non-muscle "
+ "controls to their ranges)"}
+ if self.other:
+ action["actuators"] = [{k: a[k] for k in ("index", "name", "kind", "command")} for a in self.other]
+ return {
+ "simulator": "MyoSuite 2.12.2 on MuJoCo 3.6.0 (CPU physics)",
+ "robot": f"{robot[0]}: {robot[1]}, {len(self.muscles)} muscles"
+ + (f" and {len(self.other)} other actuators" if self.other else ""),
+ "task_text": self.info["task_text"],
+ "episode": "an episode ends at the task's own success event or failure (see task_text), when your time is "
+ "up, or on a simulator error",
+ "action_types": {self.atype: action},
+ "cameras": {c: self._camera_fact(c) for c in self.cameras},
+ "default_camera": self.info["default_cameras"][0],
+ "proprio": {
+ "groups": {"q_*": "joint positions of the body (rad; m for sliding joints)",
+ "qd_*": "joint velocities (rad/s; m/s)",
+ "act_*": "muscle activations in [0, 1] (they follow the excitations with a first-order delay)",
+ "len_*": "muscle-tendon lengths (m)",
+ "effort_*": "generalized force of all actuators on each joint (N m; N)",
+ **({"root_*": "the free-standing root (the pelvis) in the world frame: position (m), wxyz "
+ "quaternion, linear (m/s, world axes) and angular velocity (rad/s, its own axes)"}
+ if self.root is not None else {})}},
+ "static": {"muscles": self.muscles},
+ "world": {"frame": "MyoSuite's world frame, z up (no positions of the scene are given)", "up": "+z",
+ "units": {"length": "m", "angle": "rad", "time": "s"}, "quaternion_order": "wxyz"},
+ "robot_spec": {
+ "name": robot[0],
+ "joints": [{"name": n, "type": "prismatic" if int(m.jnt_type[j]) == 2 else "revolute",
+ "position_limits": [lo[k], hi[k]] if lo[k] is not None else None}
+ for k, (n, j) in enumerate(zip(self.j_names, self.j_id))],
+ "controller": "muscle actuators (MuJoCo muscle model); no joint-level controller"
+ + ("; plus the non-muscle actuators in action_types" if self.other else ""),
+ "physics_hz": round(1.0 / self.control_dt * int(self.s.base.frame_skip), 3),
+ },
+ "depth": {"cameras": list(self.depth_cameras), "units": "z-depth in metres, pixel for pixel with the "
+ "camera's image; noise-free"} if self.depth_cameras else None,
+ "gl_backend": GL_BACKEND,
+ }
+
+ def interpolable(self, typ: str) -> bool:
+ return True
+
+ def render_frame(self, cam: str) -> np.ndarray:
+ if self._depth_on:
+ self.s._renderer.disable_depth_rendering()
+ self._depth_on = False
+ return self.s.render_frame(cam, (self.H, self.W))
+
+ def images(self, cameras: list[str], depth: bool) -> tuple[dict, dict]:
+ rgb = {c: self.render_frame(c) for c in cameras}
+ dep = {}
+ if depth:
+ r = self.s._renderer # the same renderer (a second context reads garbage
+ for c in cameras: # under OSMesa)
+ if c in self.depth_cameras:
+ r.enable_depth_rendering()
+ self._depth_on = True
+ r.update_scene(self.s.data, camera=c)
+ z = np.asarray(r.render(), dtype=np.float32).copy()
+ dep[c] = np.where(np.isfinite(z) & (z > 0), z, 0.0).astype(np.float32)
+ if self._depth_on:
+ r.disable_depth_rendering()
+ self._depth_on = False
+ return rgb, dep
+
+ def proprio(self) -> np.ndarray:
+ d, phys = self.s.data, self.s.base.robot.mj_data
+ # joint_effort from the physics, not the mirror: MyoSuite steps its robot wrapper's own MjData and refreshes
+ # S.data from qpos / qvel / act only (the standard twin's server.py, 2026-10-09)
+ parts = [d.qpos[self.j_q], d.qvel[self.j_v], d.act[self.act_adr], d.actuator_length[self.mus_idx],
+ phys.qfrc_actuator[self.j_v]]
+ if self.root is not None:
+ qa, va = int(self.m.jnt_qposadr[self.root]), int(self.m.jnt_dofadr[self.root])
+ parts += [d.qpos[qa:qa + 7], d.qvel[va:va + 6]]
+ return np.concatenate(parts).astype(np.float32)
+
+ def clip(self, typ: str, rows: np.ndarray) -> tuple[np.ndarray, list]:
+ return np.asarray(rows, dtype=np.float64), [] # as the standard twin: nothing clipped here
+
+ def to_sim(self, typ: str, row: np.ndarray) -> np.ndarray:
+ return np.asarray(row, dtype=np.float64)
+
+ def command(self, typ: str) -> np.ndarray:
+ return self.cmd
+
+ def hold(self) -> tuple[str, np.ndarray]:
+ return self.atype, self.cmd
+
+ def step(self, row: np.ndarray) -> tuple[bool, bool, dict]:
+ running = self.s.step(row)
+ self.cmd = np.asarray(row, dtype=np.float64).copy()
+ return (not running), bool(self.s.success), ({} if running else {"why": self.s.reason})
+
+ def reset(self) -> None:
+ self.s.reset()
+ self.cmd = self.s.rest_action()
+
+ def state_sha(self) -> str:
+ return self.s.state_sha()
+
+ def verify(self, rows: np.ndarray) -> dict:
+ """The recorded episode from the frozen instance in a fresh MyoSuite session (physics only)."""
+ fresh = M.Session(self.task, instance=self.instance)
+ try:
+ st = fresh.replay(rows)
+ return {"success": bool(st["success"]), "state_sha": st["state_sha"]}
+ finally:
+ fresh.close()
+
+ def save_trajectory(self, rows: np.ndarray, path: str, meta: dict) -> None:
+ self.s.save_trajectory(rows, path, meta)
+
+ def render_video(self, rows: np.ndarray, path: str) -> None:
+ import imageio.v2 as imageio
+ cam = self.info["default_cameras"][0]
+ self.s.render_video(rows, path, cameras=[cam], every=2, fps=25, size=(240, 320))
+ imageio.imwrite(os.path.splitext(path)[0] + ".last.png", self.s.render_frame(cam, (self.H, self.W)))
+
+
+def build() -> MyoSuite:
+ env = os.environ.get
+ hw = [int(v) for v in env("RCB_IMAGE_HW", "480,640").split(",")]
+ info = json.load(open(env("RCB_TASK_INFO", os.path.join(HERE, "task.json"))))
+ return MyoSuite(env("RCB_INSTANCE", os.path.join(HERE, "instance.npz")), info, (hw[0], hw[1]))
+
+
+if __name__ == "__main__":
+ main(build)
diff --git a/scripts/robocasa-gr1/generate.py b/scripts/robocasa-gr1/generate.py
index 3ecce01a07..0a783a6bf3 100644
--- a/scripts/robocasa-gr1/generate.py
+++ b/scripts/robocasa-gr1/generate.py
@@ -135,6 +135,12 @@ CHANGES_STANDARD = [
"mode results from before this date used the offset calibration**: a point located from the images and the "
"reported calibration came out about 3.4 cm to the side (agents missed a microwave button by 3.5 cm), so they are "
"not strictly comparable with later ones.",
+ "2026-10-14 (board 0113, from board 0086): the instruction describes the hands: each hand command is the "
+ "position target (rad) of its finger's joints, 0 open and larger closing, within each joint's range; the grip site "
+ "is the origin of the hand's mounting frame at the wrist end of the hand; the pink spheres beside each hand are "
+ "visual markers of the robot model, not objects (board 0086: most standard runs took them for objects). Text only: "
+ "the scene, the frozen instance, the reference solution, the success check and the verifier are unchanged, so the "
+ "task keeps its name.",
]
# Upstream success checks (robocasa-gr1-tabletop-tasks @ 4840e67, robocasa/environments/tabletop/): the
@@ -263,6 +269,22 @@ def action_text(dim: int) -> str:
)
+# The GR1's hands as the model defines them (board 0086: standard runs spent their first minutes, and objects, finding
+# this out, and took the pink marker spheres for objects): each hand command is the position target of its finger's
+# joints (robosuite's Fourier hand copies command k to the finger's actuators, position servos, ctrl = target in rad;
+# finger joints range from 0 to 1.22-1.74 rad); the grip site is the origin of `gripper0_<side>_eef`, at the wrist end
+# of the hand; `robot0_<side>_pinch_spheres_0..3` are visible sites without collision, 13.5-13.8 cm from it.
+HANDS = (
+ "- The hands: each hand command is the position target, in radians, of that finger's joints (`thumb_flexion` drives\n"
+ " two thumb joints, `thumb_rotation` one): 0 is the straight, open finger, larger values bend it toward the palm\n"
+ " (the hand closes), and every finger joint stops at the end of its range (1.22 to 1.74 rad), so values below 0 or\n"
+ " above about 1.7 change nothing more. The grip site that the pose targets place is the origin of the hand's mounting\n"
+ " frame (`gripper0_right_grip_site`, `gripper0_left_grip_site`), at the wrist end of the hand, not a point between\n"
+ " the fingertips. The four small pink spheres beside each hand are part of the robot model (visual markers\n"
+ " `robot0_<side>_pinch_spheres_0..3`, about 13.5 cm from its grip site, with no collision): they are not objects.\n"
+)
+
+
def hold_note(rec: dict) -> str:
if rec["success_after_hold"]:
return ""
@@ -374,7 +396,7 @@ res.video[cam]; res.frames[cam]; res.last[cam]; res.proprio; res.save("/app/runs
channel is an absolute target, so the grip sites stay within about a millimetre, but the IK's posture term can still
move the waist and elbows by a few hundredths of a radian while you watch.
- The cameras are mounted on the robot and move with it: use the calibration that comes with each response.
-- World frame: metres, z up. In the proprio, `eef_left_*` / `eef_right_*` are the grip sites' positions in the world
+{HANDS if dim == 24 else ""}- World frame: metres, z up. In the proprio, `eef_left_*` / `eef_right_*` are the grip sites' positions in the world
frame; the action's pose targets and the `cmd_*` entries (the last command) are in the robot's base frame.
{frame_text(rec)}
"""
diff --git a/scripts/robocasa/envs.py b/scripts/robocasa/envs.py
index ffe290f77a..828cb89417 100644
--- a/scripts/robocasa/envs.py
+++ b/scripts/robocasa/envs.py
@@ -157,8 +157,15 @@ _MICRO_DOOR = (
"the microwave only registers a button while its door is closed (at most 0.5% open), and opening the door "
"switches it off"
)
+# The sink's faucet handle has two hinges at one pivot (RoboCasa's sink models, read from the frozen scenes): the water
+# on / off hinge `<sink>_handle_joint` (range 0-0.52 rad) and the temperature hinge `<sink>_handle_temp_joint`
+# (about +-0.79 rad); their axes differ between sink models (board 0086: runs moved the wrong hinge, or pressed down).
+_HANDLE = ("the handle turns on two hinges at one pivot: the on / off hinge, whose range is 0 to 0.52 rad (so 0.40 rad "
+ "is near the end of its travel), and a temperature hinge (about ±0.79 rad) that moves the handle in another "
+ "direction and never turns the water on or off; which way each hinge turns depends on the sink model")
_WATER_ON = (
- "the water is on: the faucet handle's angle (mod 360°) is between 0.40 rad (about 23°) and π from its off position"
+ "the water is on: the faucet handle's angle on its on / off hinge (mod 360°) is between 0.40 rad (about 23°) and π "
+ "from its off position (" + _HANDLE + ")"
)
_KNOB = "0.35 rad (about 20°)"
@@ -235,8 +242,9 @@ SUCCESS = {
"the gripper's grip site is more than 15 cm from the stop button: move the hand away after pressing",
],
"TurnOffSinkFaucet": [
- "the water is off: the faucet handle's angle (mod 360°) is outside 0.40 rad (about 23°) to π from its off "
- "position, i.e. lowered back to below 0.40 rad"
+ "the water is off: the faucet handle's angle on its on / off hinge (mod 360°) is outside 0.40 rad (about 23°) "
+ "to π from its off position, i.e. turned back toward 0 to below 0.40 rad (it starts at about 0.47 rad; "
+ + _HANDLE + "; what counts is that hinge's angle, not how high the handle is)"
],
"TurnOffStove": ["the front right burner's knob is within " + _KNOB + " of its off position"],
"TurnOnMicrowave": [
diff --git a/scripts/robocasa/generate.py b/scripts/robocasa/generate.py
index c0db8ec3a5..5e806b50e5 100644
--- a/scripts/robocasa/generate.py
+++ b/scripts/robocasa/generate.py
@@ -80,6 +80,15 @@ ACTION_TYPE = {12: "ee_delta", 13: "joint_position", 24: "ee_pose"}
# (scripts/task_cutoffs.toml) keeps it safe; look at it when the cutoff is raised (docs/CUTOFFS.md).
COLLECT_TIMEOUT_S = 3600
COLLECT_WAIT_S = 3450
+# README Change Log entries of single tasks (by environment), after those of every task; both modes.
+_FAUCET_CHANGE = (
+ '2026-10-14 (board 0113, from board 0086): the success text now says that the faucet handle has two hinges at one'
+ ' pivot: the water on / off hinge (range 0 to 0.52 rad) and a temperature hinge (about ±0.79 rad) that never '
+ 'turns the water on or off, and that "off" means turned back toward 0 on that hinge, not lowered (board 0086: '
+ 'runs moved the temperature hinge or pressed the handle down). Text only: the scene, the frozen instance, the '
+ "reference solution, the environment's success check and the verifier are unchanged, so the task keeps its name."
+)
+CHANGES_TASK = {env: [_FAUCET_CHANGE] for env in ['TurnOnSinkFaucet', 'TurnOffSinkFaucet', 'PreSoakPan']}
# README Change Log of the standard tasks (PROTOCOL §2.8: published tasks are fixed in place), oldest first.
CHANGES_STANDARD = [
"2026-10-10: the sim service writes its verdict together with `finalized` in one atomic step and renders the "
@@ -344,9 +353,11 @@ def privileged_readme(rec: dict) -> str:
("Deliverable", f"`(T, {rec['action_dim']})` native controller commands, 1 ≤ T ≤ {MAX_STEPS}, 20 Hz"),
("Success", "the environment's own check (`_check_success`) on the state after the last action"),
]
- return readme(
+ text = readme(
readme_head(rec, "privileged"), task, DIFFICULTY_PRIVILEGED.replace("\n", " "), reference, verification, details
)
+ log = CHANGES_TASK.get(rec["env"], [])
+ return text + ("\n### Change Log\n" + "".join(f"- {c}\n" for c in log) if log else "")
def standard_readme(rec: dict) -> str:
@@ -374,7 +385,7 @@ def standard_readme(rec: dict) -> str:
text = readme(
readme_head(rec, "standard"), task, DIFFICULTY_STANDARD.replace("\n", " "), reference, verification, details
)
- return text + "\n### Change Log\n" + "".join(f"- {c}\n" for c in CHANGES_STANDARD)
+ return text + "\n### Change Log\n" + "".join(f"- {c}\n" for c in CHANGES_STANDARD + CHANGES_TASK.get(rec["env"], []))
def privileged_toml(rec: dict) -> str:
diff --git a/scripts/robocasa365/envs.py b/scripts/robocasa365/envs.py
index fe6cb5fea0..9e542f9094 100644
--- a/scripts/robocasa365/envs.py
+++ b/scripts/robocasa365/envs.py
@@ -129,6 +129,14 @@ EXCLUDED = {
# What each environment's `_check_success` (robocasa @ 4f8a298) requires, in words, for the instructions: conditions
# the task's sentence leaves out (letting go and moving the gripper away, thresholds, timers). The terms "on / in",
# "touches", "inside" and "the gripper ends more than d m from" are defined once in generate.py (SUCCESS_TERMS).
+# The sink's faucet handle has two hinges at one pivot (RoboCasa's sink model, read from the frozen scenes of
+# turn-on-sink-faucet, thaw-in-sink and drain-veggies): `<sink>_handle_joint`, the water on / off, range 0-0.5 rad about
+# the sink's front-to-back axis, and `<sink>_handle_temp_joint`, the temperature, +-0.785 rad about its left-right axis
+# (board 0086: every standard run moved the temperature hinge and never got water).
+_HANDLE = ("The handle turns on two hinges at one pivot: the on / off hinge, whose range is 0 to 0.5 rad (so the water "
+ "is on only over the last fifth of its travel), turns about the sink's front-to-back axis, so the handle "
+ "swings sideways; the temperature hinge (about ±0.79 rad) tilts the handle toward or away from you and never "
+ "turns the water on or off.")
SUCCESS = {
"PickPlaceSinkToCounter": "the egg is on the plate, the plate touches the counter next to the sink, and the "
"gripper ends more than 0.25 m from the egg.",
@@ -147,12 +155,14 @@ SUCCESS = {
"ends more than 0.15 m from the stop button.",
"PackDessert": "the hot dog and the marshmallow are both in the tupperware, and the gripper ends more than 0.25 m "
"from the marshmallow.",
- "TurnOnSinkFaucet": "the faucet handle is in its water-on range (handle joint between 0.40 rad and π rad).",
+ "TurnOnSinkFaucet": "the faucet handle is in its water-on range (its on / off hinge between 0.40 rad and π rad). "
+ + _HANDLE,
"CloseStandMixerHead": "the mixer head is closed: less than 1% open (normalised: 0 closed, 1 fully open).",
"RetrieveMeat": "the fish is on the cutting board, and the gripper ends more than 0.25 m from the fish.",
"DrainVeggies": "the lettuce has been washed: inside the sink with the water on for more than 10 consecutive "
"checks (the check runs twice per control step: at least 6 control steps; once washed it stays washed); at the "
- "end the lettuce is in the pot and the water is off.",
+ "end the lettuce is in the pot and the water is off (the water is on while the faucet handle's on / off hinge is "
+ "between 0.40 rad and π rad). " + _HANDLE,
"SanitizePrepCuttingBoard": "the spray bottle and the dish brush each touch a counter with their centres at most "
"0.35 m (horizontally) from the cutting board's, the board touches a counter, and the gripper ends more than "
"0.25 m from the spray bottle and from the brush.",
@@ -212,7 +222,8 @@ SUCCESS = {
"that burner's knob is on (angle from off, in absolute value, between 0.35 rad and 2π − 0.35 rad), the drumstick "
"is in the pan (touching it, centres within 0.07 m horizontally; on the stove beside the pan does not count), and "
"the gripper ends more than 0.25 m from the drumstick.",
- "ThawInSink": "the orange is inside the sink and the water is on (faucet handle joint between 0.40 rad and π rad).",
+ "ThawInSink": "the orange is inside the sink and the water is on (the faucet handle's on / off hinge between 0.40 "
+ "rad and π rad). " + _HANDLE,
"ScrubBowl": "the sponge has scrubbed the bowl for at least 15 control steps in all (steps on which it is on the "
"bowl and moved more than 1 mm), and the gripper ends more than 0.25 m from the bowl.",
"PlateSteakMeal": "the steak and the asparagus are on the plate, the chili pepper is still on the tray, and the "
diff --git a/scripts/robocasa365/generate.py b/scripts/robocasa365/generate.py
index 8cc446dbbf..26cf3d0d21 100644
--- a/scripts/robocasa365/generate.py
+++ b/scripts/robocasa365/generate.py
@@ -85,6 +85,16 @@ ACTION_TYPE = {12: "ee_delta", 13: "joint_position", 24: "ee_pose"}
# (scripts/task_cutoffs.toml) keeps it safe; look at it when the cutoff is raised (docs/CUTOFFS.md).
COLLECT_TIMEOUT_S = 3600
COLLECT_WAIT_S = 3450
+# README Change Log entries of single tasks (by environment), after those of every task; both modes.
+_FAUCET_CHANGE = (
+ '2026-10-14 (board 0113, from board 0086): the success text now says that the faucet handle has two hinges at one'
+ " pivot: the water on / off hinge (range 0 to 0.5 rad, about the sink's front-to-back axis, so the handle swings "
+ 'sideways) and a temperature hinge (about ±0.79 rad) that tilts it toward or away from you and never turns the '
+ 'water on; before, it named only "the handle joint" (board 0086: every standard run of `turn-on-sink-faucet` and '
+ '`thaw-in-sink` moved the temperature hinge). Text only: the scene, the frozen instance, the reference solution, '
+ "the environment's success check and the verifier are unchanged, so the task keeps its name."
+)
+CHANGES_TASK = {env: [_FAUCET_CHANGE] for env in ['TurnOnSinkFaucet', 'DrainVeggies', 'ThawInSink']}
# README Change Log of the standard tasks (PROTOCOL §2.8: published tasks are fixed in place), oldest first.
CHANGES_STANDARD = [
"2026-10-10: the sim service writes its verdict together with `finalized` in one atomic step and renders the "
@@ -404,9 +414,11 @@ def privileged_readme(rec: dict) -> str:
("Success", f"the environment's own check (`_check_success`) on the state after the last action: "
f"{SUCCESS[env]}"),
]
- return readme(
+ text = readme(
readme_head(rec, "privileged"), task, DIFFICULTY_PRIVILEGED.replace("\n", " "), reference, verification, details
)
+ log = CHANGES_TASK.get(rec["env"], [])
+ return text + ("\n### Change Log\n" + "".join(f"- {c}\n" for c in log) if log else "")
def standard_readme(rec: dict) -> str:
@@ -434,7 +446,7 @@ def standard_readme(rec: dict) -> str:
text = readme(
readme_head(rec, "standard"), task, DIFFICULTY_STANDARD.replace("\n", " "), reference, verification, details
)
- return text + "\n### Change Log\n" + "".join(f"- {c}\n" for c in CHANGES_STANDARD)
+ return text + "\n### Change Log\n" + "".join(f"- {c}\n" for c in CHANGES_STANDARD + CHANGES_TASK.get(rec["env"], []))
def privileged_toml(rec: dict) -> str:
diff --git a/scripts/robodojo/descr/hang_mugs.json b/scripts/robodojo/descr/hang_mugs.json
index f8f9b85cf4..c7d55d5438 100644
--- a/scripts/robodojo/descr/hang_mugs.json
+++ b/scripts/robodojo/descr/hang_mugs.json
@@ -8,8 +8,8 @@
"For each mug m in {mug0, mug1, mug2} (all three, simultaneously):",
" - lifted: m's current origin z is more than 0.045 m above its origin z at episode start (is_lift, z_threshold=0.045; current height, not 'ever lifted').",
" - m's 'handle' active functional point has world z > 0.9 m (any handle point; is_A_functional_point_higher_than_z, z=0.9).",
- " - m's 'handle' active functional-point frame z-axis is within 60 deg of the z-axis of the cup_holder's 'handle_support_point' passive functional-point frame.",
- " - m's handle frame x-axis is within 60 deg of the cup_holder handle_support_point frame's +x OR -x axis (either sign accepted).",
+ " - m's 'handle' active functional-point frame z-axis is within 60 deg of the z-axis of the cup_holder's 'handle_support_point' passive functional-point frame. Both frames are the FIRST functional point of that tag (RoboDojo's is_axis_aligned takes element [0] of each list): the mug's first 'handle' point and the rack's first 'handle_support_point', whichever hook the mug actually hangs on.",
+ " - m's handle frame x-axis is within 60 deg of the cup_holder handle_support_point frame's +x OR -x axis (either sign accepted; again the first handle point and the rack's first support point).",
" - m's 'handle' active functional point is within 0.045 m (3-D Euclidean) of a cup_holder 'handle_support_point' passive functional point (is_A_functional_point_close_to_B_functional_point, threshold=0.045, no orientation-alignment option).",
"all_robot_back_to_origin(): every agent-controlled (target) arm's end-effector must be within 0.15 m on EACH of x, y, z (per-axis) and within 20 degrees of its pose at the start of the episode, at the same action boundary as all other conditions. Grippers are NOT required to be open by the success test (only by partial-credit scoring)."
],
diff --git a/scripts/robodojo/descr/hang_mugs_random.json b/scripts/robodojo/descr/hang_mugs_random.json
index 792999bd9d..b5d1c2b381 100644
--- a/scripts/robodojo/descr/hang_mugs_random.json
+++ b/scripts/robodojo/descr/hang_mugs_random.json
@@ -7,8 +7,8 @@
"success": [
"For each of mug0, mug1, mug2: root z is more than 0.045 m above the mug's initial root z (is_lift z_threshold=0.045, strict), i.e. the mug has been lifted at least 4.5 cm from where it started.",
"For each mug: at least one of its active functional points tagged 'handle' has world z > 0.9 m (is_A_functional_point_higher_than_z, absolute z in the simulation world frame; the exact reference frame/table height is not determinable statically, so this is an absolute height).",
- "For each mug: the handle point's local +z axis is within 60 degrees (strict <60) of the rack's passive functional point 'handle_support_point' local +z axis (is_axis_aligned with functional_point_A='handle' (active), functional_point_B='handle_support_point' (passive)).",
- "For each mug: the handle point's local +x axis is within 60 degrees of either the support point's local +x or its local -x axis (parallel or anti-parallel).",
+ "For each mug: the handle point's local +z axis is within 60 degrees (strict <60) of the rack's passive functional point 'handle_support_point' local +z axis (is_axis_aligned with functional_point_A='handle' (active), functional_point_B='handle_support_point' (passive)). Both frames are the FIRST point of that tag (element [0] of each list): the mug's first 'handle' point and the rack's first 'handle_support_point', whichever hook the mug actually hangs on.",
+ "For each mug: the handle point's local +x axis is within 60 degrees of either the support point's local +x or its local -x axis (parallel or anti-parallel; again the first handle point and the rack's first support point).",
"For each mug: the handle point (active) is within 0.045 m (3D Euclidean distance, strict <) of at least one 'handle_support_point' (passive) of cup_holder (is_A_functional_point_close_to_B_functional_point threshold=0.045, no orientation check on that one). Each mug is tested against any support point of the rack independently; the code does not require different mugs to use different hooks.",
"All target robots back at start end-effector pose: |dx|,|dy|,|dz| each <= 0.15 m and rotation <= 20 degrees (all_robot_back_to_origin defaults)."
],
diff --git a/scripts/robodojo/descr/sort_nesting_dolls_by_size.json b/scripts/robodojo/descr/sort_nesting_dolls_by_size.json
index 9a42799f66..5970365a9d 100644
--- a/scripts/robodojo/descr/sort_nesting_dolls_by_size.json
+++ b/scripts/robodojo/descr/sort_nesting_dolls_by_size.json
@@ -1,13 +1,13 @@
{
"task": "sort_nesting_dolls_by_size",
- "goal": "Arrange the five nesting dolls upright in a straight row, ordered by size, from smallest to largest (left to right as the instruction says), and return the arms to their start pose.",
+ "goal": "Arrange the five nesting dolls upright in a straight row, ordered by size from the smallest at the smallest world x to the largest at the largest world x, and return the arms to their start pose.",
"instruction_templates": [
"Arrange the five nesting dolls in a row from left to right, from smallest to largest."
],
"success": [
- "Define the order of the dolls by their model index mod 5 (0 = smallest, 4 = largest; derived from the yml comments in the config: index 5 smallest half-extent 0.0152 m, index 9 largest 0.0335 m). The code lists the dolls sorted by (index mod 5) DESCENDING: L0 = largest, L1, L2, L3, L4 = smallest.",
+ "Define the order of the dolls by their model index mod 5, as the code does: 0 = the LARGEST doll, 4 = the SMALLEST (measured on the dolls of the frozen layout: their bounding boxes shrink as index mod 5 grows; the comments in RoboDojo's config, index 5 smallest and index 9 largest, have it the wrong way round). The code lists the dolls sorted by (index mod 5) DESCENDING: L0 = the smallest doll, L1, L2, L3, L4 = the largest.",
"Row straightness: the maximum minus the minimum world Y of the five dolls' root positions is <= 0.035 m.",
- "Order along X: for each consecutive pair (L_k, L_{k+1}), k=0..3: x(L_{k+1}) - x(L_k) > 0.03 m (strict), i.e. the largest doll has the smallest world x, and x increases by more than 3 cm with each smaller doll; the smallest doll has the largest world x. (The instruction says 'left to right, smallest to largest' while the code compares world/table x with the largest at the smallest x; presumably 'left' is as seen from the front camera, but this is not verifiable statically - use the numeric x rule.)",
+ "Order along X: for each consecutive pair (L_k, L_{k+1}), k=0..3: x(L_{k+1}) - x(L_k) > 0.03 m (strict), i.e. the SMALLEST doll has the smallest world x, and x increases by more than 3 cm with each larger doll; the largest doll has the largest world x. (This matches the instruction's 'left to right, from smallest to largest' with left = smaller world x.)",
"Upright: each of doll0..doll4 has its local +Z axis within 15 deg (default threshold, strict) of world +Z (standing, not lying).",
"Both target arms are back at origin: all target arms (left_arm and right_arm; the support arm is not counted) have their end-effector pose within 0.15 m on EACH axis (|dx|,|dy|,|dz| all <= 0.15 m) and within 20 degrees (quaternion angle) of the pose they had at the start of the episode (recorded after the scene settled).",
"All conditions hold simultaneously after one action."
@@ -21,5 +21,5 @@
"needs_memory_or_language": "none (the order is by size; sizes can be inferred visually, the instruction is generic).",
"end_requirements": "Both arms within 0.15 m per axis and 20 deg of their start pose in the same action as the dolls are correctly sorted, upright, and in a row.",
"physics_kind": "rigid (free-standing dolls)",
- "notes": "The required X gap between consecutive dolls is only > 0.03 m between root positions (roots, not edges), while the row's Y spread must be <= 3.5 cm for all five, so the dolls need to be tightly aligned in y but only loosely spaced in x. The yml comments suggest the intended target positions are x = -0.24 (largest, index mod 5 = 4), -0.12, 0.0, 0.12, 0.24 (smallest) in the table frame, but any positions satisfying the stated rules succeed. The mapping index->size is inferred from yml comments, not verified from assets. Dolls must stay upright (within 15 deg) - a doll left lying fails."
+ "notes": "The required X gap between consecutive dolls is only > 0.03 m between root positions (roots, not edges), while the row's Y spread must be <= 3.5 cm for all five, so the dolls need to be tightly aligned in y but only loosely spaced in x. The yml comments place the targets at table-frame x = -0.24 (index mod 5 = 4), -0.12, 0.0, 0.12, 0.24 (index mod 5 = 0) and call index mod 5 = 4 the largest; the dolls' measured sizes say the opposite (index mod 5 = 4 is the smallest), so the smallest doll goes at the smallest x. Any positions satisfying the stated rules succeed. Dolls must stay upright (within 15 deg) - a doll left lying fails."
}
\ No newline at end of file
diff --git a/scripts/robodojo/descr/sort_nesting_dolls_by_size_random.json b/scripts/robodojo/descr/sort_nesting_dolls_by_size_random.json
index 4044f17254..aeec6e12d1 100644
--- a/scripts/robodojo/descr/sort_nesting_dolls_by_size_random.json
+++ b/scripts/robodojo/descr/sort_nesting_dolls_by_size_random.json
@@ -1,13 +1,13 @@
{
"task": "sort_nesting_dolls_by_size_random",
- "goal": "Arrange the five nesting dolls upright in a straight row, ordered by size, from smallest to largest (left to right as the instruction says), and return the arms to their start pose.",
+ "goal": "Arrange the five nesting dolls upright in a straight row, ordered by size from the smallest at the smallest world x to the largest at the largest world x, and return the arms to their start pose.",
"instruction_templates": [
"Arrange the five nesting dolls in a row from left to right, from smallest to largest."
],
"success": [
- "Define the order of the dolls by their model index mod 5 (0 = smallest, 4 = largest; derived from the yml comments in the config: index 5 smallest half-extent 0.0152 m, index 9 largest 0.0335 m). The code lists the dolls sorted by (index mod 5) DESCENDING: L0 = largest, L1, L2, L3, L4 = smallest.",
+ "Define the order of the dolls by their model index mod 5, as the code does: 0 = the LARGEST doll, 4 = the SMALLEST (measured on the dolls of the frozen layout: their bounding boxes shrink as index mod 5 grows; the comments in RoboDojo's config, index 5 smallest and index 9 largest, have it the wrong way round). The code lists the dolls sorted by (index mod 5) DESCENDING: L0 = the smallest doll, L1, L2, L3, L4 = the largest.",
"Row straightness: the maximum minus the minimum world Y of the five dolls' root positions is <= 0.035 m.",
- "Order along X: for each consecutive pair (L_k, L_{k+1}), k=0..3: x(L_{k+1}) - x(L_k) > 0.03 m (strict), i.e. the largest doll has the smallest world x, and x increases by more than 3 cm with each smaller doll; the smallest doll has the largest world x. (The instruction says 'left to right, smallest to largest' while the code compares world/table x with the largest at the smallest x; presumably 'left' is as seen from the front camera, but this is not verifiable statically - use the numeric x rule.)",
+ "Order along X: for each consecutive pair (L_k, L_{k+1}), k=0..3: x(L_{k+1}) - x(L_k) > 0.03 m (strict), i.e. the SMALLEST doll has the smallest world x, and x increases by more than 3 cm with each larger doll; the largest doll has the largest world x. (This matches the instruction's 'left to right, from smallest to largest' with left = smaller world x.)",
"Upright: each of doll0..doll4 has its local +Z axis within 15 deg (default threshold, strict) of world +Z (standing, not lying).",
"Both target arms are back at origin: all target arms (left_arm and right_arm; the support arm is not counted) have their end-effector pose within 0.15 m on EACH axis (|dx|,|dy|,|dz| all <= 0.15 m) and within 20 degrees (quaternion angle) of the pose they had at the start of the episode (recorded after the scene settled).",
"All conditions hold simultaneously after one action."
@@ -22,5 +22,5 @@
"needs_memory_or_language": "none (the order is by size; sizes can be inferred visually, the instruction is generic).",
"end_requirements": "Both arms within 0.15 m per axis and 20 deg of their start pose in the same action as the dolls are correctly sorted, upright, and in a row.",
"physics_kind": "rigid (free-standing dolls)",
- "notes": "The required X gap between consecutive dolls is only > 0.03 m between root positions (roots, not edges), while the row's Y spread must be <= 3.5 cm for all five, so the dolls need to be tightly aligned in y but only loosely spaced in x. The yml comments suggest the intended target positions are x = -0.24 (largest, index mod 5 = 4), -0.12, 0.0, 0.12, 0.24 (smallest) in the table frame, but any positions satisfying the stated rules succeed. The mapping index->size is inferred from yml comments, not verified from assets. Dolls must stay upright (within 15 deg) - a doll left lying fails. The random variant has different texture models and clutter distractors; reward identical."
+ "notes": "The required X gap between consecutive dolls is only > 0.03 m between root positions (roots, not edges), while the row's Y spread must be <= 3.5 cm for all five, so the dolls need to be tightly aligned in y but only loosely spaced in x. The yml comments place the targets at table-frame x = -0.24 (index mod 5 = 4), -0.12, 0.0, 0.12, 0.24 (index mod 5 = 0) and call index mod 5 = 4 the largest; the dolls' measured sizes say the opposite (index mod 5 = 4 is the smallest), so the smallest doll goes at the smallest x. Any positions satisfying the stated rules succeed. Dolls must stay upright (within 15 deg) - a doll left lying fails. The random variant has different texture models and clutter distractors; reward identical."
}
\ No newline at end of file
diff --git a/scripts/robodojo/descr/swap_blocks.json b/scripts/robodojo/descr/swap_blocks.json
index 9996a1d021..2ee11f3f6b 100644
--- a/scripts/robodojo/descr/swap_blocks.json
+++ b/scripts/robodojo/descr/swap_blocks.json
@@ -26,9 +26,9 @@
"target0, target1: two cubes (category 'cube', index 3), 2 total; each starts at the centre of one of the mats mat0/mat1/mat2 (relative_plane chosen at random per object from the three; the third mat is the empty mat E), random yaw up to +-30 deg",
"button0: articulated spring button (SpringButton index 2), 1 total, at Table-relative x=0.0, y=-0.3 (in front of the mat row); springs back after release"
],
- "scene_dynamics": "The spring button returns to its rest position by itself after being released (spring). Otherwise none.",
+ "scene_dynamics": "The spring button pushes back up by itself after being released, but in this scene it comes to rest at about 0.79 of its travel ratio (measured in two trials on this frozen layout), below the 0.9 that a 'released' stage needs and the 0.95 that re-arms the press counter: after each press the button cap has to be pulled back up (above 0.95) before the next stage or press can count. Otherwise none.",
"needs_memory_or_language": "none (fixed instruction), but the solver must track which mat is empty and how many moves/presses have happened (state across the three-move swap sequence).",
"end_requirements": "Final stage requires all target arms back within 0.15 m per axis and 20 degrees of their start end-effector pose, simultaneously with button0 ratio > 0.9 (released). Nothing about gripper opening.",
"physics_kind": "mixed: rigid cubes, static mats, articulated spring button (joint with spring return)",
- "notes": "Many predicates use OR inside the nested brackets (a bracket list inside a stage is evaluated with OR semantics in check_once; only flat items of a stage are ANDed), so individual stages are looser than the intent (e.g. 'either block lifted'), but the ordered sequence plus the final stage 10 (both blocks swapped) and the exact-3-presses query effectively force the 3-move swap: move one block to the empty mat, press, move the other block into the first block's old mat, press, move the first block into the second block's old mat, press. Which block goes first is free (either order passes since the checks use OR). The button must be pressed fully (<0.5 of travel) and released (>0.9 for stage checks, >0.95 for a new press to be counted by the query) after every move, and extra presses (e.g. a double tap that re-arms above 0.95) are fatal. Pressing the button before the first lift/move counts toward the 3-press total. 'Lifted' is relative to the z recorded at the start or at the last re-record (stages 2 and 6). Block release is not checked directly, but final positions must be within 0.03 m (root to root) of the mat centres. Success is evaluated on object root poses/points, not on gripper contact. Object geometry sizes/asset dimensions are not in the source and cannot be determined statically."
+ "notes": "Many predicates use OR inside the nested brackets (a bracket list inside a stage is evaluated with OR semantics in check_once; only flat items of a stage are ANDed), so individual stages are looser than the intent (e.g. 'either block lifted'), but the ordered sequence plus the final stage 10 (both blocks swapped) and the exact-3-presses query effectively force the 3-move swap: move one block to the empty mat, press, move the other block into the first block's old mat, press, move the first block into the second block's old mat, press. Which block goes first is free (either order passes since the checks use OR). The button must be pressed fully (<0.5 of travel) and released (>0.9 for stage checks, >0.95 for a new press to be counted by the query) after every move; its spring alone returns it only to about 0.79, so lift its cap back up after each press, and extra presses (e.g. a double tap that re-arms above 0.95) are fatal. Pressing the button before the first lift/move counts toward the 3-press total. 'Lifted' is relative to the z recorded at the start or at the last re-record (stages 2 and 6). Block release is not checked directly, but final positions must be within 0.03 m (root to root) of the mat centres. Success is evaluated on object root poses/points, not on gripper contact. Object geometry sizes/asset dimensions are not in the source and cannot be determined statically."
}
\ No newline at end of file
diff --git a/scripts/robodojo/instance_overrides.json b/scripts/robodojo/instance_overrides.json
index e47d4d49bf..4103bb07f6 100644
--- a/scripts/robodojo/instance_overrides.json
+++ b/scripts/robodojo/instance_overrides.json
@@ -19,5 +19,10 @@
"change_log": ["2026-10-10: determinism class `state` -> `verdict` (`instance.json`: `\"determinism\": \"verdict\"`; the scene, the layout, the recorded reset hash and RoboDojo's success check are unchanged). The task could not be passed before: every trial that folded the garment was scored 0 as \"not deterministic\", because GPU PhysX cloth does not replay bit for bit (the determinism test that classed it `state` moved the arms without handling the cloth). Both replays must succeed now; `state_identical` is still reported. Standard mode: the service resumes `approx`. Trials graded before this date with `deterministic` 0 are re-scored from their recorded fields (`all_success` / `replay_success`), not rerun.", "2026-10-11: standard only, graded by the live episode. The physics does not replay bit for bit, so an open-loop trajectory cannot be graded reproducibly: the privileged twin is removed, and this task scores the live episode the trusted service ran (`success` = `live_success`); the collect hook still replays it twice in fresh processes and reports `replay_success`, `state_identical` and `deterministic`, which no longer decide. The scene, the layout and RoboDojo's success check are unchanged. A trial recorded before this date can be re-scored from its `live_success`."]},
"fold_clothes_random": {"determinism": "verdict",
"determinism_why": "the garment is simulated as GPU PhysX cloth, which is not bit-reproducible across processes once the arms handle it (a live episode and its two fresh-process replays, all started from the exact frozen state, ended in three different states, measured 2026-10-10)",
- "change_log": ["2026-10-10: determinism class `state` -> `verdict` (`instance.json`: `\"determinism\": \"verdict\"`; the scene, the layout, the recorded reset hash and RoboDojo's success check are unchanged). The task could not be passed before: every trial that folded the garment was scored 0 as \"not deterministic\", because GPU PhysX cloth does not replay bit for bit (the determinism test that classed it `state` moved the arms without handling the cloth). Both replays must succeed now; `state_identical` is still reported. Standard mode: the service resumes `approx`. Trials graded before this date with `deterministic` 0 are re-scored from their recorded fields (`all_success` / `replay_success`), not rerun.", "2026-10-11: standard only, graded by the live episode. The physics does not replay bit for bit, so an open-loop trajectory cannot be graded reproducibly: the privileged twin is removed, and this task scores the live episode the trusted service ran (`success` = `live_success`); the collect hook still replays it twice in fresh processes and reports `replay_success`, `state_identical` and `deterministic`, which no longer decide. The scene, the layout and RoboDojo's success check are unchanged. A trial recorded before this date can be re-scored from its `live_success`."]}
+ "change_log": ["2026-10-10: determinism class `state` -> `verdict` (`instance.json`: `\"determinism\": \"verdict\"`; the scene, the layout, the recorded reset hash and RoboDojo's success check are unchanged). The task could not be passed before: every trial that folded the garment was scored 0 as \"not deterministic\", because GPU PhysX cloth does not replay bit for bit (the determinism test that classed it `state` moved the arms without handling the cloth). Both replays must succeed now; `state_identical` is still reported. Standard mode: the service resumes `approx`. Trials graded before this date with `deterministic` 0 are re-scored from their recorded fields (`all_success` / `replay_success`), not rerun.", "2026-10-11: standard only, graded by the live episode. The physics does not replay bit for bit, so an open-loop trajectory cannot be graded reproducibly: the privileged twin is removed, and this task scores the live episode the trusted service ran (`success` = `live_success`); the collect hook still replays it twice in fresh processes and reports `replay_success`, `state_identical` and `deterministic`, which no longer decide. The scene, the layout and RoboDojo's success check are unchanged. A trial recorded before this date can be re-scored from its `live_success`."]},
+ "sort_nesting_dolls_by_size": {"change_log": ["2026-10-14 (board 0113, from board 0086): the instruction's description of RoboDojo's check (`instruction.md`, \"What RoboDojo checks\") had the size order the wrong way round. It said the LARGEST doll must have the smallest world x, following the comments in RoboDojo's config; the check actually wants the SMALLEST doll at the smallest world x and x growing with size (index mod 5 = 0 is the largest doll: the dolls' measured bounding boxes in the frozen layout shrink as index mod 5 grows), which is what RoboDojo's own sentence \"from left to right, from smallest to largest\" says. Text only: the scene, the frozen layout, RoboDojo's success check and the verifier are unchanged, so the task keeps its name. Trials that followed the old text and arranged the row as it said failed for that reason (board 0086: models that then mirrored the row passed)."]},
+ "sort_nesting_dolls_by_size_random": {"change_log": ["2026-10-14 (board 0113, from board 0086): the instruction's description of RoboDojo's check (`instruction.md`, \"What RoboDojo checks\") had the size order the wrong way round. It said the LARGEST doll must have the smallest world x, following the comments in RoboDojo's config; the check actually wants the SMALLEST doll at the smallest world x and x growing with size (index mod 5 = 0 is the largest doll: the dolls' measured bounding boxes in the frozen layout shrink as index mod 5 grows), which is what RoboDojo's own sentence \"from left to right, from smallest to largest\" says. Text only: the scene, the frozen layout, RoboDojo's success check and the verifier are unchanged, so the task keeps its name. Trials that followed the old text and arranged the row as it said failed for that reason (board 0086: models that then mirrored the row passed)."]},
+ "swap_blocks": {"change_log": ["2026-10-14 (board 0113, from board 0086): the instruction said the spring button returns to its rest position by itself after release. In this scene it comes back only to about 0.79 of its travel ratio (measured in two trials on this frozen layout), below the 0.9 that the 'released' stages need and the 0.95 that re-arms RoboDojo's press counter; the instruction now says so, and that the cap has to be lifted back up after each press. Text only: the scene, the frozen layout, RoboDojo's success check and the verifier are unchanged, so the task keeps its name."]},
+ "hang_mugs": {"change_log": ["2026-10-14 (board 0113, from board 0086): the instruction now says that the two orientation conditions compare the mug's FIRST 'handle' functional point with the rack's FIRST 'handle_support_point' (RoboDojo's `is_axis_aligned` takes element [0] of each list), not with the hook the mug hangs on; the position condition is against any hook. Text only: the scene, the frozen layout, RoboDojo's success check and the verifier are unchanged, so the task keeps its name."]},
+ "hang_mugs_random": {"change_log": ["2026-10-14 (board 0113, from board 0086): the instruction now says that the two orientation conditions compare the mug's FIRST 'handle' functional point with the rack's FIRST 'handle_support_point' (RoboDojo's `is_axis_aligned` takes element [0] of each list), not with the hook the mug hangs on; the position condition is against any hook. Text only: the scene, the frozen layout, RoboDojo's success check and the verifier are unchanged, so the task keeps its name."]}
}
diff --git a/scripts/robodojo/mk_tasks.py b/scripts/robodojo/mk_tasks.py
index 6c09ae00dc..049693043c 100644
--- a/scripts/robodojo/mk_tasks.py
+++ b/scripts/robodojo/mk_tasks.py
@@ -243,7 +243,8 @@ def plan_instance(task: str, i: int, D: dict, dims: dict, conv: dict, version: s
common["CUTOFF"] = fill(CUTOFF[mode], {"STARTUP": conv["STARTUP"]})
common["AGENT_TIMEOUT_S"] = str(agent_timeout(mode))
every = [ONE_REPLAY_LIVE if (live and x is ONE_REPLAY["standard"]) else x for x in CHANGE_LOG_ALL[mode]]
- common["CHANGE_LOG"] = "\n### Change Log\n" + "".join(f"- {x}\n" for x in log + every)
+ # oldest first: a per-task entry dated after the entries of every task (e.g. 2026-10-14) goes after them
+ common["CHANGE_LOG"] = "\n### Change Log\n" + "".join(f"- {x}\n" for x in sorted(log + every, key=lambda x: x[:10]))
files: dict[str, tuple[bytes, int]] = {}
tdir = HERE / "template" / mode
for p in sorted(tdir.rglob("*")):
diff --git a/scripts/robolab/families.json b/scripts/robolab/families.json
index a0fb3bc628..9aba129990 100644
--- a/scripts/robolab/families.json
+++ b/scripts/robolab/families.json
@@ -1,5 +1,5 @@
{
- "_doc": "RoboLab families for scripts/robolab/mk_tasks.py (one entry per RoboLab benchmark task that has tasks here; tasks = true: generated into tasks/ in both modes). Frozen instances: instances/<family>.json; oracle inputs: oracles/<family>/.",
+ "_doc": "RoboLab families for scripts/robolab/mk_tasks.py (one entry per RoboLab benchmark task that has tasks here; tasks = true: generated into tasks/ in both modes). Frozen instances: instances/<family>.json; oracle inputs: oracles/<family>/. change_log (optional): README Change Log entries of the family's standard task alone, after those of every task.",
"families": {
"apple_and_yogurt_in_bowl": {
"tasks": true,
@@ -28,7 +28,7 @@
"terms": "class Terminations:\n time_out = DoneTerm(func=mdp.time_out, time_out=True)\n success = DoneTerm(\n func=object_in_container,\n params={\n \"object\": [\"apple_01\", \"yogurt_cup\"],\n \"container\": \"wooden_bowl\",\n \"logical\": \"all\",\n \"require_gripper_detached\": True\n },\n )",
"note": "",
"difficulty": "RoboLab labels this task \"moderate\" (2 subtasks; attributes: affordance); its success term (`object_in_container`) asks for 2 objects into wooden_bowl. The tooling gives only thin primitives (move_ee on Isaac Lab's differential IK, joint moves, gripper, hold), so every grasp, transport and placement is the agent's to build. RoboLab ships no expert solution.",
- "standard_goal": "Put the apple and yogurt in the bowl. Success: The red apple and the small yogurt cup must end inside the wooden bowl and the gripper must touch neither. \"Inside\" means the centre of the object's convex hull lies within the container's convex hull, extended without limit along the container's own up axis (the open side; straight up for a container standing normally).",
+ "standard_goal": "Put the apple and yogurt in the bowl. Success: The red apple and the small yogurt cup must end inside the wooden bowl and, at that moment, the gripper must touch neither (let go of both at the end). \"Inside\" means the centre of the object's convex hull lies within the container's convex hull, extended without limit along the container's own up axis (the open side; straight up for a container standing normally).",
"standard_objects": "Objects here: a ceramic mug, a pair of eyeglasses, a keyboard, a whiteboard marker, a remote control, a smartphone, a wooden bowl, a metal spoon, a computer mouse, a yogurt cup, a pitcher, a plastic pail (trash bin) and a red apple. The Robotiq opens to 8.5 cm. The episode ends by itself the moment the success condition above holds.\n",
"oracle": "none",
"criterion_group": "object_in_container",
@@ -46,6 +46,9 @@
"site_log": "https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/apple_and_yogurt_in_bowl/unlimited/",
"hf_replay": "https://huggingface.co/datasets/eai-v2-internal/agent-runs/resolve/main/robolab/codex-gpt6_luna-xhigh/apple_and_yogurt_in_bowl/unlimited/replay.mp4"
}
+ ],
+ "change_log": [
+ "2026-10-14 (board 0113, from board 0086): the success text says that the gripper must not be touching the objects at the moment of success (\"at that moment\"); the old wording \"the gripper must touch none of them\" was read by an agent as \"never touch them\" (board 0086). Text only: the scene, the frozen instance, RoboLab's success term and the verifier are unchanged, so the task keeps its name."
]
},
"banana_in_bowl": {
@@ -110,7 +113,7 @@
"terms": "class Terminations:\n \"\"\"Termination configuration for banana task.\"\"\"\n time_out = DoneTerm(func=mdp.time_out, time_out=True)\n\n success = DoneTerm(\n func=objects_placed_in_container_in_order,\n params={\n \"objects\": [\"banana\", \"rubiks_cube\"],\n \"container\": \"bowl\",\n \"tolerance\": 0.0,\n \"require_contact_with\": True,\n \"require_gripper_detached\": True,\n },\n )",
"note": "`objects_placed_in_container_in_order` is stateful: it latches the first control step at which each object is placed, and a later removal does not erase it, so the order of the first placements decides.",
"difficulty": "RoboLab labels this task \"simple\" (2 subtasks; attributes: conjunction); its success term (`objects_placed_in_container_in_order`) asks for 2 objects into bowl in a given order. The tooling gives only thin primitives (move_ee on Isaac Lab's differential IK, joint moves, gripper, hold), so every grasp, transport and placement is the agent's to build. RoboLab ships no expert solution.",
- "standard_goal": "Put the banana then the cube in the bowl. Success: The yellow banana and then the Rubik's cube must end inside the bowl, both touching it, and the gripper must touch neither. Order counts: the service remembers the first moment each object is placed and a later removal does not erase it, so if the Rubik's cube is placed first the episode can no longer succeed. \"Inside\" means the centre of the object's convex hull lies within the container's convex hull, extended without limit along the container's own up axis (the open side; straight up for a container standing normally).",
+ "standard_goal": "Put the banana then the cube in the bowl. Success: The yellow banana and then the Rubik's cube must end inside the bowl, both touching it, and, at that moment, the gripper must touch neither (let go of both at the end). Order counts: the service remembers the first moment each object is placed and a later removal does not erase it, so if the Rubik's cube is placed first the episode can no longer succeed. \"Inside\" means the centre of the object's convex hull lies within the container's convex hull, extended without limit along the container's own up axis (the open side; straight up for a container standing normally).",
"standard_objects": "Objects here: a bowl, a Rubik's cube and a banana. The Robotiq opens to 8.5 cm. The episode ends by itself the moment the success condition above holds.\n",
"oracle": "none",
"criterion_group": "objects_placed_in_container_in_order",
@@ -140,6 +143,9 @@
"site_log": "https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/banana_then_rubiks_cube/limited/",
"hf_replay": "https://huggingface.co/datasets/eai-v2-internal/agent-runs/resolve/main/robolab/codex-gpt6_luna-xhigh/banana_then_rubiks_cube/limited/replay.mp4"
}
+ ],
+ "change_log": [
+ "2026-10-14 (board 0113, from board 0086): the success text says that the gripper must not be touching the objects at the moment of success (\"at that moment\"); the old wording \"the gripper must touch none of them\" was read by an agent as \"never touch them\" (board 0086). Text only: the scene, the frozen instance, RoboLab's success term and the verifier are unchanged, so the task keeps its name."
]
},
"bananas_in_bin_one_more": {
@@ -162,7 +168,7 @@
"terms": "class ChooseOneBananaInBin:\n time_out = DoneTerm(func=mdp.time_out, time_out=True)\n success = DoneTerm(\n func=object_in_container,\n params={\"object\": [\"banana\", \"banana_01\", \"banana_03\"], \"container\": \"grey_bin\", \"logical\": \"choose\", \"K\": 1, \"require_gripper_detached\": True, 'require_contact_with': False})",
"note": "`logical=\"choose\"` with `K=1` means exactly 1 of the listed objects (`evaluate_logicals` in `/opt/robolab/robolab/core/task/predicate_logic.py`): more than 1 fails as well.",
"difficulty": "RoboLab labels this task \"moderate\" (1 subtasks; attributes: semantics, counting); its success term (`object_in_container`) asks for exactly one of the three bananas outside the grey bin moved into it (no more). The tooling gives only thin primitives (move_ee on Isaac Lab's differential IK, joint moves, gripper, hold), so every grasp, transport and placement is the agent's to build. RoboLab ships no expert solution.",
- "standard_goal": "Put one (1) more bananas in the grey bin. Success: Exactly 1 of the three bananas that start outside the grey bin (no more, no fewer) must end inside the grey bin and the gripper must touch none of them. \"Inside\" means the centre of the object's convex hull lies within the container's convex hull, extended without limit along the container's own up axis (the open side; straight up for a container standing normally).",
+ "standard_goal": "Put one (1) more bananas in the grey bin. Success: Exactly 1 of the three bananas that start outside the grey bin (no more, no fewer) must end inside the grey bin and, at that moment, the gripper must touch none of them (moving them with the gripper is how the task is done; let go of them at the end). \"Inside\" means the centre of the object's convex hull lies within the container's convex hull, extended without limit along the container's own up axis (the open side; straight up for a container standing normally).",
"standard_objects": "Objects here: five bananas and a grey bin. Two of the five bananas already lie in the grey bin, which stands on the robot's right (-y); they do not count, whatever happens to them. The Robotiq opens to 8.5 cm. The episode ends by itself the moment the success condition above holds.\n",
"oracle": "none",
"criterion_group": "object_in_container",
@@ -180,6 +186,9 @@
"site_log": "https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/bananas_in_bin_one_more/unlimited/",
"hf_replay": "https://huggingface.co/datasets/eai-v2-internal/agent-runs/resolve/main/robolab/codex-gpt6_luna-xhigh/bananas_in_bin_one_more/unlimited/replay.mp4"
}
+ ],
+ "change_log": [
+ "2026-10-14 (board 0113, from board 0086): the success text says that the gripper must not be touching the objects at the moment of success (\"at that moment\"); the old wording \"the gripper must touch none of them\" was read by an agent as \"never touch them\" (board 0086). Text only: the scene, the frozen instance, RoboLab's success term and the verifier are unchanged, so the task keeps its name."
]
},
"bananas_in_bin_three_total": {
@@ -202,7 +211,7 @@
"terms": "class BananasInBinTerminations:\n time_out = DoneTerm(func=mdp.time_out, time_out=True)\n success = DoneTerm(\n func=object_in_container,\n params={\"object\": [\"banana\", \"banana_01\", \"banana_03\"], \"container\": \"grey_bin\", \"logical\": \"choose\", \"K\": 1, \"require_gripper_detached\": True, 'require_contact_with': False})",
"note": "`logical=\"choose\"` with `K=1` means exactly 1 of the listed objects (`evaluate_logicals` in `/opt/robolab/robolab/core/task/predicate_logic.py`): more than 1 fails as well.",
"difficulty": "RoboLab labels this task \"moderate\" (1 subtasks; attributes: semantics, counting); its success term (`object_in_container`) asks for exactly one of the three bananas outside the grey bin moved into it (no more). The tooling gives only thin primitives (move_ee on Isaac Lab's differential IK, joint moves, gripper, hold), so every grasp, transport and placement is the agent's to build. RoboLab ships no expert solution.",
- "standard_goal": "Make sure there are 3 (three) bananas in the grey bin. Success: Exactly 1 of the three bananas that start outside the grey bin (no more, no fewer) must end inside the grey bin and the gripper must touch none of them. \"Inside\" means the centre of the object's convex hull lies within the container's convex hull, extended without limit along the container's own up axis (the open side; straight up for a container standing normally).",
+ "standard_goal": "Make sure there are 3 (three) bananas in the grey bin. Success: Exactly 1 of the three bananas that start outside the grey bin (no more, no fewer) must end inside the grey bin and, at that moment, the gripper must touch none of them (moving them with the gripper is how the task is done; let go of them at the end). \"Inside\" means the centre of the object's convex hull lies within the container's convex hull, extended without limit along the container's own up axis (the open side; straight up for a container standing normally).",
"standard_objects": "Objects here: five bananas and a grey bin. Two of the five bananas already lie in the grey bin, which stands on the robot's right (-y); they do not count, whatever happens to them. The Robotiq opens to 8.5 cm. The episode ends by itself the moment the success condition above holds.\n",
"oracle": "none",
"criterion_group": "object_in_container",
@@ -220,6 +229,9 @@
"site_log": "https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/bananas_in_bin_three_total/unlimited/",
"hf_replay": "https://huggingface.co/datasets/eai-v2-internal/agent-runs/resolve/main/robolab/codex-gpt6_luna-xhigh/bananas_in_bin_three_total/unlimited/replay.mp4"
}
+ ],
+ "change_log": [
+ "2026-10-14 (board 0113, from board 0086): the success text says that the gripper must not be touching the objects at the moment of success (\"at that moment\"); the old wording \"the gripper must touch none of them\" was read by an agent as \"never touch them\" (board 0086). Text only: the scene, the frozen instance, RoboLab's success term and the verifier are unchanged, so the task keeps its name."
]
},
"bananas_out_of_bin": {
@@ -242,7 +254,7 @@
"terms": "class BananasOutOfBinTerminations:\n time_out = DoneTerm(func=mdp.time_out, time_out=True)\n success = DoneTerm(\n func=object_outside_of,\n params={\"object\": [\"banana_02\", \"banana_04\"], \"container\": \"grey_bin\", \"logical\": \"all\", \"tolerance\": 0.00, \"require_gripper_detached\": True}\n )",
"note": "",
"difficulty": "RoboLab labels this task \"moderate\" (2 subtasks; attributes: semantics, spatial); its success term (`object_outside_of`) asks for 2 objects out of grey_bin. The tooling gives only thin primitives (move_ee on Isaac Lab's differential IK, joint moves, gripper, hold), so every grasp, transport and placement is the agent's to build. RoboLab ships no expert solution.",
- "standard_goal": "Take the bananas out. Success: Both bananas that start inside the grey bin must end outside the grey bin, and the gripper must touch neither. \"Outside\" means the centre of the object's convex hull is not within the container's convex hull extended along the container's own up axis.",
+ "standard_goal": "Take the bananas out. Success: Both bananas that start inside the grey bin must end outside the grey bin, and, at that moment, the gripper must touch neither (let go of both at the end). \"Outside\" means the centre of the object's convex hull is not within the container's convex hull extended along the container's own up axis.",
"standard_objects": "Objects here: five bananas and a grey bin. The Robotiq opens to 8.5 cm. The episode ends by itself the moment the success condition above holds.\n",
"oracle": "none",
"criterion_group": "object_outside_of",
@@ -272,6 +284,9 @@
"site_log": "https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/bananas_out_of_bin/limited/",
"hf_replay": "https://huggingface.co/datasets/eai-v2-internal/agent-runs/resolve/main/robolab/codex-gpt6_luna-xhigh/bananas_out_of_bin/limited/replay.mp4"
}
+ ],
+ "change_log": [
+ "2026-10-14 (board 0113, from board 0086): the success text says that the gripper must not be touching the objects at the moment of success (\"at that moment\"); the old wording \"the gripper must touch none of them\" was read by an agent as \"never touch them\" (board 0086). Text only: the scene, the frozen instance, RoboLab's success term and the verifier are unchanged, so the task keeps its name."
]
},
"bbq_sauce_in_bin": {
@@ -300,7 +315,7 @@
"terms": "class BBQSauceInBinTerminations:\n \"\"\"Termination configuration for BBQ sauce in bin task.\"\"\"\n time_out = DoneTerm(func=mdp.time_out, time_out=True)\n success = DoneTerm(\n func=object_in_container,\n params={\n \"object\": [\"bbq_sauce_bottle\", \"bbq_sauce_bottle_01\"],\n \"container\": \"grey_bin\",\n \"logical\": \"all\",\n \"require_gripper_detached\": True\n },\n )",
"note": "",
"difficulty": "RoboLab labels this task \"simple\" (2 subtasks; attributes: color, semantics); its success term (`object_in_container`) asks for 2 objects into grey_bin. The tooling gives only thin primitives (move_ee on Isaac Lab's differential IK, joint moves, gripper, hold), so every grasp, transport and placement is the agent's to build. RoboLab ships no expert solution.",
- "standard_goal": "Put the red BBQ sauce bottles in the grey bin. Success: Both red BBQ sauce bottles must end inside the grey bin and the gripper must touch neither. \"Inside\" means the centre of the object's convex hull lies within the container's convex hull, extended without limit along the container's own up axis (the open side; straight up for a container standing normally).",
+ "standard_goal": "Put the red BBQ sauce bottles in the grey bin. Success: Both red BBQ sauce bottles must end inside the grey bin and, at that moment, the gripper must touch neither (let go of both at the end). \"Inside\" means the centre of the object's convex hull lies within the container's convex hull, extended without limit along the container's own up axis (the open side; straight up for a container standing normally).",
"standard_objects": "Objects here: a grey bin, a mug, a mustard bottle, a bowl, a ranch dressing bottle, two red BBQ sauce bottles, a box of oatmeal raisin cookies, a tuna can, a Soft Scrub bottle, a wooden block and a coffee pot. The Robotiq opens to 8.5 cm. The episode ends by itself the moment the success condition above holds.\n",
"oracle": "none",
"criterion_group": "object_in_container",
@@ -318,6 +333,9 @@
"site_log": "https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/bbq_sauce_in_bin/unlimited/",
"hf_replay": "https://huggingface.co/datasets/eai-v2-internal/agent-runs/resolve/main/robolab/codex-gpt6_luna-xhigh/bbq_sauce_in_bin/unlimited/replay.mp4"
}
+ ],
+ "change_log": [
+ "2026-10-14 (board 0113, from board 0086): the success text says that the gripper must not be touching the objects at the moment of success (\"at that moment\"); the old wording \"the gripper must touch none of them\" was read by an agent as \"never touch them\" (board 0086). Text only: the scene, the frozen instance, RoboLab's success term and the verifier are unchanged, so the task keeps its name."
]
},
"block_stacking_order_agnostic": {
@@ -445,7 +463,7 @@
"terms": "class Terminations:\n time_out = DoneTerm(func=mdp.time_out, time_out=True)\n success = DoneTerm(\n func=object_in_container,\n params={\n \"object\": [\"yellow_block\", \"red_block\", \"green_block\", \"blue_block\"],\n \"container\": \"grey_bin\",\n \"logical\": \"all\",\n \"require_gripper_detached\": True\n },\n )",
"note": "",
"difficulty": "RoboLab labels this task \"complex\" (4 subtasks; attributes: sorting); its success term (`object_in_container`) asks for 4 objects into grey_bin. The tooling gives only thin primitives (move_ee on Isaac Lab's differential IK, joint moves, gripper, hold), so every grasp, transport and placement is the agent's to build. RoboLab ships no expert solution.",
- "standard_goal": "Sort all colored blocks into the bin. Success: The four coloured blocks (yellow, red, green, blue) must end inside the grey bin and the gripper must touch none of them. \"Inside\" means the centre of the object's convex hull lies within the container's convex hull, extended without limit along the container's own up axis (the open side; straight up for a container standing normally).",
+ "standard_goal": "Sort all colored blocks into the bin. Success: The four coloured blocks (yellow, red, green, blue) must end inside the grey bin and, at that moment, the gripper must touch none of them (moving them with the gripper is how the task is done; let go of them at the end). \"Inside\" means the centre of the object's convex hull lies within the container's convex hull, extended without limit along the container's own up axis (the open side; straight up for a container standing normally).",
"standard_objects": "Objects here: three Rubik's cubes, a grey bin, two green lizard figurines, a birdhouse, a yellow block, a red block, a green block and a blue block. The Robotiq opens to 8.5 cm. The episode ends by itself the moment the success condition above holds.\n",
"oracle": "none",
"criterion_group": "object_in_container",
@@ -475,6 +493,9 @@
"site_log": "https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/blocks_in_bin/limited/",
"hf_replay": "https://huggingface.co/datasets/eai-v2-internal/agent-runs/resolve/main/robolab/codex-gpt6_luna-xhigh/blocks_in_bin/limited/replay.mp4"
}
+ ],
+ "change_log": [
+ "2026-10-14 (board 0113, from board 0086): the success text says that the gripper must not be touching the objects at the moment of success (\"at that moment\"); the old wording \"the gripper must touch none of them\" was read by an agent as \"never touch them\" (board 0086). Text only: the scene, the frozen instance, RoboLab's success term and the verifier are unchanged, so the task keeps its name."
]
},
"butter_above_raisin": {
@@ -544,7 +565,7 @@
"terms": "class ClutterPlasticTerminations:\n time_out = DoneTerm(func=mdp.time_out, time_out=True)\n success = DoneTerm(func=object_in_container, params={\"object\": [ \"whitepackerbottle_a01\", \"milkjug_a01\",\"utilityjug_a03\"], \"container\": \"right_bin\", \"logical\": \"all\", \"require_gripper_detached\": True})",
"note": "",
"difficulty": "RoboLab labels this task \"moderate\" (3 subtasks; attributes: semantics); its success term (`object_in_container`) asks for 3 objects into right_bin. The tooling gives only thin primitives (move_ee on Isaac Lab's differential IK, joint moves, gripper, hold), so every grasp, transport and placement is the agent's to build. RoboLab ships no expert solution.",
- "standard_goal": "Put all plastic bottles away in the bin. Success: The three plastic bottles (the small white plastic bottle, the milk jug and the utility jug) must end inside the blue bin and the gripper must touch none of them. \"Inside\" means the centre of the object's convex hull lies within the container's convex hull, extended without limit along the container's own up axis (the open side; straight up for a container standing normally).",
+ "standard_goal": "Put all plastic bottles away in the bin. Success: The three plastic bottles (the small white plastic bottle, the milk jug and the utility jug) must end inside the blue bin and, at that moment, the gripper must touch none of them (moving them with the gripper is how the task is done; let go of them at the end). \"Inside\" means the centre of the object's convex hull lies within the container's convex hull, extended without limit along the container's own up axis (the open side; straight up for a container standing normally).",
"standard_objects": "Objects here: two lemons, two limes, two oranges, a pomegranate, a large pumpkin, a small pumpkin, a small white plastic bottle, an avocado, a crab-shaped pen holder, a milk jug, a serving bowl, a utility jug, a blue bin and a red onion. The Robotiq opens to 8.5 cm. The episode ends by itself the moment the success condition above holds.\n",
"oracle": "none",
"criterion_group": "object_in_container",
@@ -562,6 +583,9 @@
"site_log": "https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/clutter_plastic/unlimited/",
"hf_replay": "https://huggingface.co/datasets/eai-v2-internal/agent-runs/resolve/main/robolab/codex-gpt6_luna-xhigh/clutter_plastic/unlimited/replay.mp4"
}
+ ],
+ "change_log": [
+ "2026-10-14 (board 0113, from board 0086): the success text says that the gripper must not be touching the objects at the moment of success (\"at that moment\"); the old wording \"the gripper must touch none of them\" was read by an agent as \"never touch them\" (board 0086). Text only: the scene, the frozen instance, RoboLab's success term and the verifier are unchanged, so the task keeps its name."
]
},
"clutter_pumpkin": {
@@ -595,7 +619,7 @@
"terms": "class ClutterPumpkinTerminations:\n time_out = DoneTerm(func=mdp.time_out, time_out=True)\n success = DoneTerm(func=object_in_container, params={\"object\": [\"pumpkinlarge\", \"pumpkinsmall\"], \"container\": \"right_bin\", \"logical\": \"all\", \"require_gripper_detached\": True})",
"note": "",
"difficulty": "RoboLab labels this task \"simple\" (2 subtasks; attributes: semantics); its success term (`object_in_container`) asks for 2 objects into right_bin. The tooling gives only thin primitives (move_ee on Isaac Lab's differential IK, joint moves, gripper, hold), so every grasp, transport and placement is the agent's to build. RoboLab ships no expert solution.",
- "standard_goal": "Put all the pumpkins away in the bin. Success: The large and the small pumpkin must end inside the blue bin and the gripper must touch neither. \"Inside\" means the centre of the object's convex hull lies within the container's convex hull, extended without limit along the container's own up axis (the open side; straight up for a container standing normally).",
+ "standard_goal": "Put all the pumpkins away in the bin. Success: The large and the small pumpkin must end inside the blue bin and, at that moment, the gripper must touch neither (let go of both at the end). \"Inside\" means the centre of the object's convex hull lies within the container's convex hull, extended without limit along the container's own up axis (the open side; straight up for a container standing normally).",
"standard_objects": "Objects here: two lemons, two limes, two oranges, a pomegranate, a large pumpkin, a small pumpkin, a small white plastic bottle, an avocado, a crab-shaped pen holder, a milk jug, a serving bowl, a utility jug, a blue bin and a red onion. The Robotiq opens to 8.5 cm. The episode ends by itself the moment the success condition above holds.\n",
"oracle": "none",
"criterion_group": "object_in_container",
@@ -613,6 +637,9 @@
"site_log": "https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/clutter_pumpkin/unlimited/",
"hf_replay": "https://huggingface.co/datasets/eai-v2-internal/agent-runs/resolve/main/robolab/codex-gpt6_luna-xhigh/clutter_pumpkin/unlimited/replay.mp4"
}
+ ],
+ "change_log": [
+ "2026-10-14 (board 0113, from board 0086): the success text says that the gripper must not be touching the objects at the moment of success (\"at that moment\"); the old wording \"the gripper must touch none of them\" was read by an agent as \"never touch them\" (board 0086). Text only: the scene, the frozen instance, RoboLab's success term and the verifier are unchanged, so the task keeps its name."
]
},
"condiments_in_bin": {
@@ -641,7 +668,7 @@
"terms": "class CondimentsInBinTerminations:\n time_out = DoneTerm(func=mdp.time_out, time_out=True)\n success = DoneTerm(\n func=object_in_container,\n params={\n \"object\": [\"mustard\", \"ranch_dressing\", \"bbq_sauce_bottle\", \"bbq_sauce_bottle_01\"],\n \"container\": \"grey_bin\",\n \"logical\": \"all\",\n \"require_gripper_detached\": True\n },\n )",
"note": "",
"difficulty": "RoboLab labels this task \"complex\" (4 subtasks; attributes: semantics, sorting); its success term (`object_in_container`) asks for 4 objects into grey_bin. The tooling gives only thin primitives (move_ee on Isaac Lab's differential IK, joint moves, gripper, hold), so every grasp, transport and placement is the agent's to build. RoboLab ships no expert solution.",
- "standard_goal": "Sort the sauce condiments into the grey bin. Success: The four sauce bottles (the mustard, the ranch dressing and both red BBQ sauce bottles) must end inside the grey bin and the gripper must touch none of them. \"Inside\" means the centre of the object's convex hull lies within the container's convex hull, extended without limit along the container's own up axis (the open side; straight up for a container standing normally).",
+ "standard_goal": "Sort the sauce condiments into the grey bin. Success: The four sauce bottles (the mustard, the ranch dressing and both red BBQ sauce bottles) must end inside the grey bin and, at that moment, the gripper must touch none of them (moving them with the gripper is how the task is done; let go of them at the end). \"Inside\" means the centre of the object's convex hull lies within the container's convex hull, extended without limit along the container's own up axis (the open side; straight up for a container standing normally).",
"standard_objects": "Objects here: a grey bin, a mug, a mustard bottle, a bowl, a ranch dressing bottle, two red BBQ sauce bottles, a box of oatmeal raisin cookies, a tuna can, a Soft Scrub bottle, a wooden block and a coffee pot. The Robotiq opens to 8.5 cm. The episode ends by itself the moment the success condition above holds.\n",
"oracle": "none",
"criterion_group": "object_in_container",
@@ -659,6 +686,9 @@
"site_log": "https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/condiments_in_bin/unlimited/",
"hf_replay": "https://huggingface.co/datasets/eai-v2-internal/agent-runs/resolve/main/robolab/codex-gpt6_luna-xhigh/condiments_in_bin/unlimited/replay.mp4"
}
+ ],
+ "change_log": [
+ "2026-10-14 (board 0113, from board 0086): the success text says that the gripper must not be touching the objects at the moment of success (\"at that moment\"); the old wording \"the gripper must touch none of them\" was read by an agent as \"never touch them\" (board 0086). Text only: the scene, the frozen instance, RoboLab's success term and the verifier are unchanged, so the task keeps its name."
]
},
"cooking_clear_plate": {
@@ -690,7 +720,7 @@
"terms": "class CookingClearPlateTerminations:\n time_out = DoneTerm(func=mdp.time_out, time_out=True)\n success = DoneTerm(\n func=object_outside_of,\n params={\"object\": [\"spoon_1\", \"measuring_cups_1\"], \"container\": \"clay_plates\", \"logical\": \"all\", \"tolerance\": 0.0, \"require_gripper_detached\": True}\n )",
"note": "",
"difficulty": "RoboLab labels this task \"moderate\" (2 subtasks; attributes: color, sorting); its success term (`object_outside_of`) asks for 2 objects out of clay_plates. The tooling gives only thin primitives (move_ee on Isaac Lab's differential IK, joint moves, gripper, hold), so every grasp, transport and placement is the agent's to build. RoboLab ships no expert solution.",
- "standard_goal": "Put the two measuring cups outside of the plate. Success: The orange measuring cup and the small blue measuring scoop, which start on the clay plate, must end outside the clay plate, and the gripper must touch neither. \"Outside\" means the centre of the object's convex hull is not within the container's convex hull extended along the container's own up axis.",
+ "standard_goal": "Put the two measuring cups outside of the plate. Success: The orange measuring cup and the small blue measuring scoop, which start on the clay plate, must end outside the clay plate, and, at that moment, the gripper must touch neither (let go of both at the end). \"Outside\" means the centre of the object's convex hull is not within the container's convex hull extended along the container's own up axis.",
"standard_objects": "Objects here: a red onion, a serving bowl, a flat clay plate, a wooden spatula, a spatula, two utensil holders, a tomato sauce can, an orange measuring cup, a pink spaghetti spoon, a small blue measuring scoop, a green serving spoon, a ladle, a wooden bowl and a potato masher. The Robotiq opens to 8.5 cm. The episode ends by itself the moment the success condition above holds.\n",
"oracle": "none",
"criterion_group": "object_outside_of",
@@ -720,6 +750,9 @@
"site_log": null,
"hf_replay": null
}
+ ],
+ "change_log": [
+ "2026-10-14 (board 0113, from board 0086): the success text says that the gripper must not be touching the objects at the moment of success (\"at that moment\"); the old wording \"the gripper must touch none of them\" was read by an agent as \"never touch them\" (board 0086). Text only: the scene, the frozen instance, RoboLab's success term and the verifier are unchanged, so the task keeps its name."
]
},
"cubes_and_blocks_in_bin": {
@@ -747,7 +780,7 @@
"terms": "class Terminations:\n time_out = DoneTerm(func=mdp.time_out, time_out=True)\n success = DoneTerm(\n func=object_in_container,\n params={\n \"object\": [\n \"rubiks_cube\", \"rubiks_cube_1\", \"rubiks_cube_2\",\n \"yellow_block\", \"red_block\", \"green_block\", \"blue_block\"\n ],\n \"container\": \"grey_bin\",\n \"logical\": \"all\",\n \"require_gripper_detached\": True\n },\n )",
"note": "",
"difficulty": "RoboLab labels this task \"complex\" (7 subtasks; attributes: sorting, conjunction); its success term (`object_in_container`) asks for 7 objects into grey_bin. The tooling gives only thin primitives (move_ee on Isaac Lab's differential IK, joint moves, gripper, hold), so every grasp, transport and placement is the agent's to build. RoboLab ships no expert solution.",
- "standard_goal": "Put all the cubes and blocks in the bin. Success: The seven cubes (the three Rubik's cubes and the four coloured blocks (yellow, red, green, blue)) must end inside the grey bin and the gripper must touch none of them. \"Inside\" means the centre of the object's convex hull lies within the container's convex hull, extended without limit along the container's own up axis (the open side; straight up for a container standing normally).",
+ "standard_goal": "Put all the cubes and blocks in the bin. Success: The seven cubes (the three Rubik's cubes and the four coloured blocks (yellow, red, green, blue)) must end inside the grey bin and, at that moment, the gripper must touch none of them (moving them with the gripper is how the task is done; let go of them at the end). \"Inside\" means the centre of the object's convex hull lies within the container's convex hull, extended without limit along the container's own up axis (the open side; straight up for a container standing normally).",
"standard_objects": "Objects here: three Rubik's cubes, a grey bin, two green lizard figurines, a birdhouse, a yellow block, a red block, a green block and a blue block. The Robotiq opens to 8.5 cm. The episode ends by itself the moment the success condition above holds.\n",
"oracle": "none",
"criterion_group": "object_in_container",
@@ -777,6 +810,9 @@
"site_log": "https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/cubes_and_blocks_in_bin/limited/",
"hf_replay": "https://huggingface.co/datasets/eai-v2-internal/agent-runs/resolve/main/robolab/codex-gpt6_luna-xhigh/cubes_and_blocks_in_bin/limited/replay.mp4"
}
+ ],
+ "change_log": [
+ "2026-10-14 (board 0113, from board 0086): the success text says that the gripper must not be touching the objects at the moment of success (\"at that moment\"); the old wording \"the gripper must touch none of them\" was read by an agent as \"never touch them\" (board 0086). Text only: the scene, the frozen instance, RoboLab's success term and the verifier are unchanged, so the task keeps its name."
]
},
"food_packing2_cans": {
@@ -799,7 +835,7 @@
"terms": "class FoodPacking2CansTerminations:\n time_out = DoneTerm(func=mdp.time_out, time_out=True)\n success = DoneTerm(func=object_in_container, params={\"object\": [ \"tomato_soup_can\", \"tuna_can\"], \"container\": \"bin_a06\", \"logical\": \"all\", \"require_gripper_detached\": True})",
"note": "",
"difficulty": "RoboLab labels this task \"simple\" (2 subtasks; attributes: semantics); its success term (`object_in_container`) asks for 2 objects into bin_a06. The tooling gives only thin primitives (move_ee on Isaac Lab's differential IK, joint moves, gripper, hold), so every grasp, transport and placement is the agent's to build. RoboLab ships no expert solution.",
- "standard_goal": "Pack canned foods into the bin. Success: Both cans (the tomato soup can and the tuna can) must end inside the dark grey bin and the gripper must touch neither. \"Inside\" means the centre of the object's convex hull lies within the container's convex hull, extended without limit along the container's own up axis (the open side; straight up for a container standing normally).",
+ "standard_goal": "Pack canned foods into the bin. Success: Both cans (the tomato soup can and the tuna can) must end inside the dark grey bin and, at that moment, the gripper must touch neither (let go of both at the end). \"Inside\" means the centre of the object's convex hull lies within the container's convex hull, extended without limit along the container's own up axis (the open side; straight up for a container standing normally).",
"standard_objects": "Objects here: a dark grey bin, a Cheez-It box, a mustard bottle, a sugar box, a tomato soup can and a tuna can. The Robotiq opens to 8.5 cm. The episode ends by itself the moment the success condition above holds.\n",
"oracle": "none",
"criterion_group": "object_in_container",
@@ -817,6 +853,9 @@
"site_log": "https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/food_packing2_cans/unlimited/",
"hf_replay": "https://huggingface.co/datasets/eai-v2-internal/agent-runs/resolve/main/robolab/codex-gpt6_luna-xhigh/food_packing2_cans/unlimited/replay.mp4"
}
+ ],
+ "change_log": [
+ "2026-10-14 (board 0113, from board 0086): the success text says that the gripper must not be touching the objects at the moment of success (\"at that moment\"); the old wording \"the gripper must touch none of them\" was read by an agent as \"never touch them\" (board 0086). Text only: the scene, the frozen instance, RoboLab's success term and the verifier are unchanged, so the task keeps its name."
]
},
"food_packing3_cans": {
@@ -841,7 +880,7 @@
"terms": "class FoodPacking3CansTerminations:\n time_out = DoneTerm(func=mdp.time_out, time_out=True)\n success = DoneTerm(func=object_in_container, params={\"object\": [ \"spam_can\", \"tomato_soup_can\", \"tuna_can\"], \"container\": \"bin_a06\", \"logical\": \"all\", \"require_gripper_detached\": True})",
"note": "",
"difficulty": "RoboLab labels this task \"moderate\" (3 subtasks; attributes: semantics); its success term (`object_in_container`) asks for 3 objects into bin_a06. The tooling gives only thin primitives (move_ee on Isaac Lab's differential IK, joint moves, gripper, hold), so every grasp, transport and placement is the agent's to build. RoboLab ships no expert solution.",
- "standard_goal": "Pack canned foods into the bin. Success: The three cans (the Spam can, the tomato soup can and the tuna can) must end inside the dark grey bin and the gripper must touch none of them. \"Inside\" means the centre of the object's convex hull lies within the container's convex hull, extended without limit along the container's own up axis (the open side; straight up for a container standing normally).",
+ "standard_goal": "Pack canned foods into the bin. Success: The three cans (the Spam can, the tomato soup can and the tuna can) must end inside the dark grey bin and, at that moment, the gripper must touch none of them (moving them with the gripper is how the task is done; let go of them at the end). \"Inside\" means the centre of the object's convex hull lies within the container's convex hull, extended without limit along the container's own up axis (the open side; straight up for a container standing normally).",
"standard_objects": "Objects here: a dark grey bin, a Cheez-It box, a chocolate pudding box, a mustard bottle, a Spam can, a sugar box, a tomato soup can and a tuna can. The Robotiq opens to 8.5 cm. The episode ends by itself the moment the success condition above holds.\n",
"oracle": "none",
"criterion_group": "object_in_container",
@@ -859,6 +898,9 @@
"site_log": "https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/food_packing3_cans/unlimited/",
"hf_replay": "https://huggingface.co/datasets/eai-v2-internal/agent-runs/resolve/main/robolab/codex-gpt6_luna-xhigh/food_packing3_cans/unlimited/replay.mp4"
}
+ ],
+ "change_log": [
+ "2026-10-14 (board 0113, from board 0086): the success text says that the gripper must not be touching the objects at the moment of success (\"at that moment\"); the old wording \"the gripper must touch none of them\" was read by an agent as \"never touch them\" (board 0086). Text only: the scene, the frozen instance, RoboLab's success term and the verifier are unchanged, so the task keeps its name."
]
},
"fruits_green_limes_on_plate": {
@@ -890,7 +932,7 @@
"terms": "class FruitsGreenLimesOnPlateTerminations:\n time_out = DoneTerm(func=mdp.time_out, time_out=True)\n success = DoneTerm(\n func=object_on_top,\n params={\"object\": [\"lime01\", \"lime01_01\"], \"reference_object\": \"clay_plates\", \"logical\": \"all\", \"require_gripper_detached\": True}\n )",
"note": "",
"difficulty": "RoboLab labels this task \"simple\" (2 subtasks; attributes: color); its success term (`object_on_top`) asks for 2 objects onto clay_plates. The tooling gives only thin primitives (move_ee on Isaac Lab's differential IK, joint moves, gripper, hold), so every grasp, transport and placement is the agent's to build. RoboLab ships no expert solution.",
- "standard_goal": "Put all the green fruit on the plate. Success: Both green limes must end on the flat beige plate, and the gripper must touch neither. \"On\" means the surface supports the object (the contact force it exerts on the object points upwards, within 45 deg of vertical) and the object's centre lies over the surface's bounding-box footprint, +/- 1 cm.",
+ "standard_goal": "Put all the green fruit on the plate. Success: Both green limes must end on the flat beige plate, and, at that moment, the gripper must touch neither (let go of both at the end). \"On\" means the surface supports the object (the contact force it exerts on the object points upwards, within 45 deg of vertical) and the object's centre lies over the surface's bounding-box footprint, +/- 1 cm.",
"standard_objects": "Objects here: two lemons, two limes, two oranges, a pomegranate, a large pumpkin, a small pumpkin, a red onion, a serving bowl, a flat clay plate, a wooden spatula, a spatula and a utensil holder. The Robotiq opens to 8.5 cm. The episode ends by itself the moment the success condition above holds.\n",
"oracle": "none",
"criterion_group": "object_on_top",
@@ -908,6 +950,9 @@
"site_log": "https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/fruits_green_limes_on_plate/limited/",
"hf_replay": "https://huggingface.co/datasets/eai-v2-internal/agent-runs/resolve/main/robolab/codex-gpt6_luna-xhigh/fruits_green_limes_on_plate/limited/replay.mp4"
}
+ ],
+ "change_log": [
+ "2026-10-14 (board 0113, from board 0086): the success text says that the gripper must not be touching the objects at the moment of success (\"at that moment\"); the old wording \"the gripper must touch none of them\" was read by an agent as \"never touch them\" (board 0086). Text only: the scene, the frozen instance, RoboLab's success term and the verifier are unchanged, so the task keeps its name."
]
},
"fruits_on_plate3": {
@@ -939,7 +984,7 @@
"terms": "class FruitsOnPlate3Terminations:\n \"\"\"Termination configuration for banana task.\"\"\"\n time_out = DoneTerm(func=mdp.time_out, time_out=True)\n success = DoneTerm(\n func=object_on_top,\n params={\"object\": [\"lemon_01\", \"lemon_02\", \"lime01\", \"lime01_01\", \"orange_01\", \"orange_02\", \"pomegranate01\"], \"reference_object\": \"clay_plates\", \"logical\": \"choose\", \"K\": 3, \"require_gripper_detached\": True}\n )",
"note": "`logical=\"choose\"` with `K=3` means exactly 3 of the listed objects (`evaluate_logicals` in `/opt/robolab/robolab/core/task/predicate_logic.py`): more than 3 fails as well.",
"difficulty": "RoboLab labels this task \"complex\" (3 subtasks; attributes: semantics, vague, counting); its success term (`object_on_top`) asks for exactly 3 of 7 objects onto clay_plates. The tooling gives only thin primitives (move_ee on Isaac Lab's differential IK, joint moves, gripper, hold), so every grasp, transport and placement is the agent's to build. RoboLab ships no expert solution.",
- "standard_goal": "Put three (3) fruits on the plate. Success: Exactly 3 of the seven fruits (both lemons, both limes, both oranges, the pomegranate) (no more, no fewer) must end on the flat beige plate, and the gripper must touch none of them. \"On\" means the surface supports the object (the contact force it exerts on the object points upwards, within 45 deg of vertical) and the object's centre lies over the surface's bounding-box footprint, +/- 1 cm.",
+ "standard_goal": "Put three (3) fruits on the plate. Success: Exactly 3 of the seven fruits (both lemons, both limes, both oranges, the pomegranate) (no more, no fewer) must end on the flat beige plate, and, at that moment, the gripper must touch none of them (moving them with the gripper is how the task is done; let go of them at the end). \"On\" means the surface supports the object (the contact force it exerts on the object points upwards, within 45 deg of vertical) and the object's centre lies over the surface's bounding-box footprint, +/- 1 cm.",
"standard_objects": "Objects here: two lemons, two limes, two oranges, a pomegranate, a large pumpkin, a small pumpkin, a red onion, a serving bowl, a flat clay plate, a wooden spatula, a spatula and a utensil holder. One orange already lies on the plate. The Robotiq opens to 8.5 cm. The episode ends by itself the moment the success condition above holds.\n",
"oracle": "none",
"criterion_group": "object_on_top",
@@ -969,6 +1014,9 @@
"site_log": "https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/fruits_on_plate3/limited/",
"hf_replay": "https://huggingface.co/datasets/eai-v2-internal/agent-runs/resolve/main/robolab/codex-gpt6_luna-xhigh/fruits_on_plate3/limited/replay.mp4"
}
+ ],
+ "change_log": [
+ "2026-10-14 (board 0113, from board 0086): the success text says that the gripper must not be touching the objects at the moment of success (\"at that moment\"); the old wording \"the gripper must touch none of them\" was read by an agent as \"never touch them\" (board 0086). Text only: the scene, the frozen instance, RoboLab's success term and the verifier are unchanged, so the task keeps its name."
]
},
"keyboard_out_of_bin": {
@@ -1416,7 +1464,7 @@
"terms": "class Terminations:\n time_out = DoneTerm(func=mdp.time_out, time_out=True)\n success = DoneTerm(\n func=object_in_container,\n params={\n \"object\": [\"whitepackerbottle_a01\", \"whitepackerbottle_a02\", \"whitepackerbottle_a03\"],\n \"container\": \"squarepail_a01\",\n \"logical\": \"all\",\n \"require_gripper_detached\": True\n },\n )",
"note": "",
"difficulty": "RoboLab labels this task \"complex\" (3 subtasks; attributes: semantics, sorting, size); its success term (`object_in_container`) asks for 3 objects into squarepail_a01. The tooling gives only thin primitives (move_ee on Isaac Lab's differential IK, joint moves, gripper, hold), so every grasp, transport and placement is the agent's to build. RoboLab ships no expert solution.",
- "standard_goal": "Put all the small plastic bottles in the square pail. Success: The three small white plastic bottles must end inside the square pail and the gripper must touch none of them. \"Inside\" means the centre of the object's convex hull lies within the container's convex hull, extended without limit along the container's own up axis (the open side; straight up for a container standing normally).",
+ "standard_goal": "Put all the small plastic bottles in the square pail. Success: The three small white plastic bottles must end inside the square pail and, at that moment, the gripper must touch none of them (moving them with the gripper is how the task is done; let go of them at the end). \"Inside\" means the centre of the object's convex hull lies within the container's convex hull, extended without limit along the container's own up axis (the open side; straight up for a container standing normally).",
"standard_objects": "Objects here: three small white plastic bottles, two utility jugs, a square pail and a plastic pail (trash bin). The Robotiq opens to 8.5 cm. The episode ends by itself the moment the success condition above holds.\n",
"oracle": "none",
"criterion_group": "object_in_container",
@@ -1434,6 +1482,9 @@
"site_log": "https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/plastic_bottles_in_square_pail/unlimited/",
"hf_replay": "https://huggingface.co/datasets/eai-v2-internal/agent-runs/resolve/main/robolab/codex-gpt6_luna-xhigh/plastic_bottles_in_square_pail/unlimited/replay.mp4"
}
+ ],
+ "change_log": [
+ "2026-10-14 (board 0113, from board 0086): the success text says that the gripper must not be touching the objects at the moment of success (\"at that moment\"); the old wording \"the gripper must touch none of them\" was read by an agent as \"never touch them\" (board 0086). Text only: the scene, the frozen instance, RoboLab's success term and the verifier are unchanged, so the task keeps its name."
]
},
"put_two_mugs_on_shelf": {
@@ -1458,7 +1509,7 @@
"terms": "class Terminations:\n time_out = DoneTerm(func=mdp.time_out, time_out=True)\n success = DoneTerm(func=object_in_container, params={\"object\": [\"ceramic_mug\", \"mug\", \"mug_01\"], \"container\": \"wireshelving_a01\", \"require_contact_with\": True, \"require_gripper_detached\": True, \"logical\": \"choose\", \"K\": 2})",
"note": "`logical=\"choose\"` with `K=2` means exactly 2 of the listed objects (`evaluate_logicals` in `/opt/robolab/robolab/core/task/predicate_logic.py`): more than 2 fails as well.",
"difficulty": "RoboLab labels this task \"moderate\" (2 subtasks; attributes: affordance, spatial, counting); its success term (`object_in_container`) asks for exactly 2 of 3 objects into wireshelving_a01. The tooling gives only thin primitives (move_ee on Isaac Lab's differential IK, joint moves, gripper, hold), so every grasp, transport and placement is the agent's to build. RoboLab ships no expert solution.",
- "standard_goal": "Put two (2) mugs on the wire shelf. Success: Exactly 2 of the three mugs (no more, no fewer) must end on the wire shelf, touching it, and the gripper must touch none of them. \"Inside\" means the centre of the object's convex hull lies within the container's convex hull, extended without limit along the container's own up axis (the open side; straight up for a container standing normally). (RoboLab checks \"on\" the wire shelf this way.)",
+ "standard_goal": "Put two (2) mugs on the wire shelf. Success: Exactly 2 of the three mugs (no more, no fewer) must end on the wire shelf, touching it, and, at that moment, the gripper must touch none of them (moving them with the gripper is how the task is done; let go of them at the end). \"Inside\" means the centre of the object's convex hull lies within the container's convex hull, extended without limit along the container's own up axis (the open side; straight up for a container standing normally). (RoboLab checks \"on\" the wire shelf this way.)",
"standard_objects": "Objects here: a wire shelf, a spatula, a small plate, a large fork, a small fork, a ceramic mug and two mugs. The Robotiq opens to 8.5 cm. The episode ends by itself the moment the success condition above holds.\n",
"oracle": "none",
"criterion_group": "object_in_container",
@@ -1476,6 +1527,9 @@
"site_log": "https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/put_two_mugs_on_shelf/unlimited/",
"hf_replay": "https://huggingface.co/datasets/eai-v2-internal/agent-runs/resolve/main/robolab/codex-gpt6_luna-xhigh/put_two_mugs_on_shelf/unlimited/replay.mp4"
}
+ ],
+ "change_log": [
+ "2026-10-14 (board 0113, from board 0086): the success text says that the gripper must not be touching the objects at the moment of success (\"at that moment\"); the old wording \"the gripper must touch none of them\" was read by an agent as \"never touch them\" (board 0086). Text only: the scene, the frozen instance, RoboLab's success term and the verifier are unchanged, so the task keeps its name."
]
},
"recycle_carton": {
@@ -1500,7 +1554,7 @@
"terms": "class RecycleCartonTerminations:\n time_out = DoneTerm(func=mdp.time_out, time_out=True)\n success = DoneTerm(func=object_in_container, params={\"object\": [\"milk_carton\", \"orange_juice_carton\"], \"container\": \"container_a01\", \"logical\": \"all\", \"require_gripper_detached\": True})",
"note": "",
"difficulty": "RoboLab labels this task \"simple\" (2 subtasks; attributes: semantics, vague); its success term (`object_in_container`) asks for 2 objects into container_a01. The tooling gives only thin primitives (move_ee on Isaac Lab's differential IK, joint moves, gripper, hold), so every grasp, transport and placement is the agent's to build. RoboLab ships no expert solution.",
- "standard_goal": "Put the recyclable cartons in the grey bin. Success: The milk carton and the orange juice carton must end inside the grey bin in the centre of the table and the gripper must touch neither. \"Inside\" means the centre of the object's convex hull lies within the container's convex hull, extended without limit along the container's own up axis (the open side; straight up for a container standing normally).",
+ "standard_goal": "Put the recyclable cartons in the grey bin. Success: The milk carton and the orange juice carton must end inside the grey bin in the centre of the table and, at that moment, the gripper must touch neither (let go of both at the end). \"Inside\" means the centre of the object's convex hull lies within the container's convex hull, extended without limit along the container's own up axis (the open side; straight up for a container standing normally).",
"standard_objects": "Objects here: an alphabet soup can, a milk carton, an orange juice carton, a smartphone, a mug, a container, a mayonnaise bottle and a ketchup bottle. The Robotiq opens to 8.5 cm. The episode ends by itself the moment the success condition above holds.\n",
"oracle": "none",
"criterion_group": "object_in_container",
@@ -1518,6 +1572,9 @@
"site_log": "https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/recycle_carton/unlimited/",
"hf_replay": "https://huggingface.co/datasets/eai-v2-internal/agent-runs/resolve/main/robolab/codex-gpt6_luna-xhigh/recycle_carton/unlimited/replay.mp4"
}
+ ],
+ "change_log": [
+ "2026-10-14 (board 0113, from board 0086): the success text says that the gripper must not be touching the objects at the moment of success (\"at that moment\"); the old wording \"the gripper must touch none of them\" was read by an agent as \"never touch them\" (board 0086). Text only: the scene, the frozen instance, RoboLab's success term and the verifier are unchanged, so the task keeps its name."
]
},
"recycle_cartons_vertical_crate": {
@@ -1542,7 +1599,7 @@
"terms": "class RecycleCartonsVerticalCrateTerminations:\n time_out = DoneTerm(func=mdp.time_out, time_out=True)\n success = DoneTerm(func=object_inside, params={\"object\": [\"milk_carton\", \"orange_juice_carton\"], \"container\": \"container_a01\", \"logical\": \"all\", \"require_gripper_detached\": True})",
"note": "",
"difficulty": "RoboLab labels this task \"moderate\" (2 subtasks; attributes: semantics, spatial); its success term (`object_inside`) asks for 2 objects into container_a01. The tooling gives only thin primitives (move_ee on Isaac Lab's differential IK, joint moves, gripper, hold), so every grasp, transport and placement is the agent's to build. RoboLab ships no expert solution.",
- "standard_goal": "Put the cartons that can be recycled in the vertical crate. Success: The milk carton and the orange juice carton must end inside the vertical crate in the centre of the table and the gripper must touch neither. \"Inside\" means the centre of the object's convex hull lies within the container's convex hull, extended without limit along the container's own up axis (the open side; straight up for a container standing normally).",
+ "standard_goal": "Put the cartons that can be recycled in the vertical crate. Success: The milk carton and the orange juice carton must end inside the vertical crate in the centre of the table and, at that moment, the gripper must touch neither (let go of both at the end). \"Inside\" means the centre of the object's convex hull lies within the container's convex hull, extended without limit along the container's own up axis (the open side; straight up for a container standing normally).",
"standard_objects": "Objects here: an alphabet soup can, a milk carton, an orange juice carton, a smartphone, a mug, a container, a mayonnaise bottle and a ketchup bottle. The Robotiq opens to 8.5 cm. The episode ends by itself the moment the success condition above holds.\n",
"oracle": "none",
"criterion_group": "object_inside",
@@ -1560,6 +1617,9 @@
"site_log": "https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/recycle_cartons_vertical_crate/unlimited/",
"hf_replay": "https://huggingface.co/datasets/eai-v2-internal/agent-runs/resolve/main/robolab/codex-gpt6_luna-xhigh/recycle_cartons_vertical_crate/unlimited/replay.mp4"
}
+ ],
+ "change_log": [
+ "2026-10-14 (board 0113, from board 0086): the success text says that the gripper must not be touching the objects at the moment of success (\"at that moment\"); the old wording \"the gripper must touch none of them\" was read by an agent as \"never touch them\" (board 0086). Text only: the scene, the frozen instance, RoboLab's success term and the verifier are unchanged, so the task keeps its name."
]
},
"red_dishes_in_bin": {
@@ -1581,7 +1641,7 @@
"terms": "class Terminations:\n time_out = DoneTerm(func=mdp.time_out, time_out=True)\n success = DoneTerm(func=object_in_container, params={\"object\": [\"mug\", \"bowl\"], \"container\": \"grey_bin\", \"logical\": \"all\", \"require_gripper_detached\": True})",
"note": "",
"difficulty": "RoboLab labels this task \"simple\" (2 subtasks; attributes: color, semantics); its success term (`object_in_container`) asks for 2 objects into grey_bin. The tooling gives only thin primitives (move_ee on Isaac Lab's differential IK, joint moves, gripper, hold), so every grasp, transport and placement is the agent's to build. RoboLab ships no expert solution.",
- "standard_goal": "Put the red dishware in the grey bin. Success: The red mug and the red bowl must end inside the grey bin and the gripper must touch neither. \"Inside\" means the centre of the object's convex hull lies within the container's convex hull, extended without limit along the container's own up axis (the open side; straight up for a container standing normally).",
+ "standard_goal": "Put the red dishware in the grey bin. Success: The red mug and the red bowl must end inside the grey bin and, at that moment, the gripper must touch neither (let go of both at the end). \"Inside\" means the centre of the object's convex hull lies within the container's convex hull, extended without limit along the container's own up axis (the open side; straight up for a container standing normally).",
"standard_objects": "Objects here: a bowl, a Rubik's cube, a banana, a grey bin and a mug. The Robotiq opens to 8.5 cm. The episode ends by itself the moment the success condition above holds.\n",
"oracle": "none",
"criterion_group": "object_in_container",
@@ -1599,6 +1659,9 @@
"site_log": "https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/red_dishes_in_bin/unlimited/",
"hf_replay": "https://huggingface.co/datasets/eai-v2-internal/agent-runs/resolve/main/robolab/codex-gpt6_luna-xhigh/red_dishes_in_bin/unlimited/replay.mp4"
}
+ ],
+ "change_log": [
+ "2026-10-14 (board 0113, from board 0086): the success text says that the gripper must not be touching the objects at the moment of success (\"at that moment\"); the old wording \"the gripper must touch none of them\" was read by an agent as \"never touch them\" (board 0086). Text only: the scene, the frozen instance, RoboLab's success term and the verifier are unchanged, so the task keeps its name."
]
},
"reorient_red_mug": {
@@ -1671,7 +1734,7 @@
"terms": "class RubiksCubeAndBananaTerminations:\n \"\"\"Termination configuration for banana task.\"\"\"\n time_out = DoneTerm(func=mdp.time_out, time_out=True)\n success = DoneTerm(\n func=object_in_container,\n params={\"object\": [\"rubiks_cube\", \"banana\"], \"container\": \"bowl\", \"logical\": \"all\", \"tolerance\": 0.0, \"require_contact_with\": True, \"require_gripper_detached\": True}\n )",
"note": "",
"difficulty": "RoboLab labels this task \"simple\" (2 subtasks; attributes: conjunction); its success term (`object_in_container`) asks for 2 objects into bowl. The tooling gives only thin primitives (move_ee on Isaac Lab's differential IK, joint moves, gripper, hold), so every grasp, transport and placement is the agent's to build. RoboLab ships no expert solution.",
- "standard_goal": "Put the cube and the banana in the bowl. Success: The Rubik's cube and the yellow banana must end inside the bowl, touching it, and the gripper must touch neither. \"Inside\" means the centre of the object's convex hull lies within the container's convex hull, extended without limit along the container's own up axis (the open side; straight up for a container standing normally).",
+ "standard_goal": "Put the cube and the banana in the bowl. Success: The Rubik's cube and the yellow banana must end inside the bowl, touching it, and, at that moment, the gripper must touch neither (let go of both at the end). \"Inside\" means the centre of the object's convex hull lies within the container's convex hull, extended without limit along the container's own up axis (the open side; straight up for a container standing normally).",
"standard_objects": "Objects here: a bowl, a Rubik's cube and a banana. The Robotiq opens to 8.5 cm. The episode ends by itself the moment the success condition above holds.\n",
"oracle": "none",
"criterion_group": "object_in_container",
@@ -1689,6 +1752,9 @@
"site_log": "https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/rubiks_cube_and_banana/limited/",
"hf_replay": "https://huggingface.co/datasets/eai-v2-internal/agent-runs/resolve/main/robolab/codex-gpt6_luna-xhigh/rubiks_cube_and_banana/limited/replay.mp4"
}
+ ],
+ "change_log": [
+ "2026-10-14 (board 0113, from board 0086): the success text says that the gripper must not be touching the objects at the moment of success (\"at that moment\"); the old wording \"the gripper must touch none of them\" was read by an agent as \"never touch them\" (board 0086). Text only: the scene, the frozen instance, RoboLab's success term and the verifier are unchanged, so the task keeps its name."
]
},
"rubiks_cube_behind_bowl": {
@@ -1914,7 +1980,7 @@
"terms": "class Terminations:\n \"\"\"Termination configuration for banana task.\"\"\"\n time_out = DoneTerm(func=mdp.time_out, time_out=True)\n\n success = DoneTerm(\n func=objects_placed_in_container_in_order,\n params={\n \"objects\": [\"rubiks_cube\", \"banana\"],\n \"container\": \"bowl\",\n \"tolerance\": 0.0,\n \"require_contact_with\": True,\n \"require_gripper_detached\": True,\n },\n )",
"note": "`objects_placed_in_container_in_order` is stateful: it latches the first control step at which each object is placed, and a later removal does not erase it, so the order of the first placements decides.",
"difficulty": "RoboLab labels this task \"simple\" (2 subtasks; attributes: conjunction); its success term (`objects_placed_in_container_in_order`) asks for 2 objects into bowl in a given order. The tooling gives only thin primitives (move_ee on Isaac Lab's differential IK, joint moves, gripper, hold), so every grasp, transport and placement is the agent's to build. RoboLab ships no expert solution.",
- "standard_goal": "Put the cube then the banana in the bowl. Success: The Rubik's cube and then the yellow banana must end inside the bowl, both touching it, and the gripper must touch neither. Order counts: the service remembers the first moment each object is placed and a later removal does not erase it, so if the yellow banana is placed first the episode can no longer succeed. \"Inside\" means the centre of the object's convex hull lies within the container's convex hull, extended without limit along the container's own up axis (the open side; straight up for a container standing normally).",
+ "standard_goal": "Put the cube then the banana in the bowl. Success: The Rubik's cube and then the yellow banana must end inside the bowl, both touching it, and, at that moment, the gripper must touch neither (let go of both at the end). Order counts: the service remembers the first moment each object is placed and a later removal does not erase it, so if the yellow banana is placed first the episode can no longer succeed. \"Inside\" means the centre of the object's convex hull lies within the container's convex hull, extended without limit along the container's own up axis (the open side; straight up for a container standing normally).",
"standard_objects": "Objects here: a bowl, a Rubik's cube and a banana. The Robotiq opens to 8.5 cm. The episode ends by itself the moment the success condition above holds.\n",
"oracle": "none",
"criterion_group": "objects_placed_in_container_in_order",
@@ -1932,6 +1998,9 @@
"site_log": "https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/rubiks_cube_then_banana/unlimited/",
"hf_replay": "https://huggingface.co/datasets/eai-v2-internal/agent-runs/resolve/main/robolab/codex-gpt6_luna-xhigh/rubiks_cube_then_banana/unlimited/replay.mp4"
}
+ ],
+ "change_log": [
+ "2026-10-14 (board 0113, from board 0086): the success text says that the gripper must not be touching the objects at the moment of success (\"at that moment\"); the old wording \"the gripper must touch none of them\" was read by an agent as \"never touch them\" (board 0086). Text only: the scene, the frozen instance, RoboLab's success term and the verifier are unchanged, so the task keeps its name."
]
},
"rubiks_cubes_in_bin": {
@@ -1959,7 +2028,7 @@
"terms": "class Terminations:\n time_out = DoneTerm(func=mdp.time_out, time_out=True)\n success = DoneTerm(\n func=object_in_container,\n params={\n \"object\": [\"rubiks_cube\", \"rubiks_cube_1\", \"rubiks_cube_2\"],\n \"container\": \"grey_bin\",\n \"logical\": \"all\",\n \"require_gripper_detached\": True\n },\n )",
"note": "",
"difficulty": "RoboLab labels this task \"complex\" (3 subtasks; attributes: sorting); its success term (`object_in_container`) asks for 3 objects into grey_bin. The tooling gives only thin primitives (move_ee on Isaac Lab's differential IK, joint moves, gripper, hold), so every grasp, transport and placement is the agent's to build. RoboLab ships no expert solution.",
- "standard_goal": "Sort all rubiks cubes into the bin. Success: The three Rubik's cubes must end inside the grey bin and the gripper must touch none of them. \"Inside\" means the centre of the object's convex hull lies within the container's convex hull, extended without limit along the container's own up axis (the open side; straight up for a container standing normally).",
+ "standard_goal": "Sort all rubiks cubes into the bin. Success: The three Rubik's cubes must end inside the grey bin and, at that moment, the gripper must touch none of them (moving them with the gripper is how the task is done; let go of them at the end). \"Inside\" means the centre of the object's convex hull lies within the container's convex hull, extended without limit along the container's own up axis (the open side; straight up for a container standing normally).",
"standard_objects": "Objects here: three Rubik's cubes, a grey bin, two green lizard figurines, a birdhouse, a yellow block, a red block, a green block and a blue block. The Robotiq opens to 8.5 cm. The episode ends by itself the moment the success condition above holds.\n",
"oracle": "none",
"criterion_group": "object_in_container",
@@ -1989,6 +2058,9 @@
"site_log": "https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/rubiks_cubes_in_bin/limited/",
"hf_replay": "https://huggingface.co/datasets/eai-v2-internal/agent-runs/resolve/main/robolab/codex-gpt6_luna-xhigh/rubiks_cubes_in_bin/limited/replay.mp4"
}
+ ],
+ "change_log": [
+ "2026-10-14 (board 0113, from board 0086): the success text says that the gripper must not be touching the objects at the moment of success (\"at that moment\"); the old wording \"the gripper must touch none of them\" was read by an agent as \"never touch them\" (board 0086). Text only: the scene, the frozen instance, RoboLab's success term and the verifier are unchanged, so the task keeps its name."
]
},
"spoon_in_mug": {
@@ -2216,7 +2288,7 @@
"terms": "class Terminations:\n time_out = DoneTerm(func=mdp.time_out, time_out=True)\n success = DoneTerm(func=object_outside_of_and_on_surface, params={\"object\": [\"ceramic_mug\", \"mug\"], \"container\": \"rack_l04\", \"surface\": \"table\", \"logical\": \"all\", \"require_gripper_detached\": True})",
"note": "",
"difficulty": "RoboLab labels this task \"moderate\" (2 subtasks; attributes: semantics, affordance); its success term (`object_outside_of_and_on_surface`) asks for 2 objects out of rack_l04. The tooling gives only thin primitives (move_ee on Isaac Lab's differential IK, joint moves, gripper, hold), so every grasp, transport and placement is the agent's to build. RoboLab ships no expert solution.",
- "standard_goal": "Take the mugs off the shelf. Success: Both mugs on the shelf must end outside the shelf and standing on the table, and the gripper must touch neither. \"Outside\" means the centre of the object's convex hull is not within the container's convex hull extended along the container's own up axis. \"On the table\" means the table supports it (its contact force points upwards, within 45 deg of vertical).",
+ "standard_goal": "Take the mugs off the shelf. Success: Both mugs on the shelf must end outside the shelf and standing on the table, and, at that moment, the gripper must touch neither (let go of both at the end). \"Outside\" means the centre of the object's convex hull is not within the container's convex hull extended along the container's own up axis. \"On the table\" means the table supports it (its contact force points upwards, within 45 deg of vertical).",
"standard_objects": "Objects here: a ceramic mug, a mug, a shelf, a serving bowl and a utility jug. The Robotiq opens to 8.5 cm. The episode ends by itself the moment the success condition above holds.\n",
"oracle": "none",
"criterion_group": "object_outside_of_and_on_surface",
@@ -2234,6 +2306,9 @@
"site_log": "https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/take_mugs_off_of_shelf/unlimited/",
"hf_replay": "https://huggingface.co/datasets/eai-v2-internal/agent-runs/resolve/main/robolab/codex-gpt6_luna-xhigh/take_mugs_off_of_shelf/unlimited/replay.mp4"
}
+ ],
+ "change_log": [
+ "2026-10-14 (board 0113, from board 0086): the success text says that the gripper must not be touching the objects at the moment of success (\"at that moment\"); the old wording \"the gripper must touch none of them\" was read by an agent as \"never touch them\" (board 0086). Text only: the scene, the frozen instance, RoboLab's success term and the verifier are unchanged, so the task keeps its name."
]
},
"take_spatula_off_shelf": {
@@ -2317,7 +2392,7 @@
"terms": "class Terminations:\n time_out = DoneTerm(func=mdp.time_out, time_out=True)\n success = DoneTerm(\n func=object_in_container,\n params={\n \"object\": [\"apple_01\", \"yogurt_cup\"],\n \"container\": \"plasticpail_a02\",\n \"logical\": \"all\",\n \"require_gripper_detached\": True\n },\n )",
"note": "",
"difficulty": "RoboLab labels this task \"simple\" (2 subtasks; attributes: semantics); its success term (`object_in_container`) asks for 2 objects into plasticpail_a02. The tooling gives only thin primitives (move_ee on Isaac Lab's differential IK, joint moves, gripper, hold), so every grasp, transport and placement is the agent's to build. RoboLab ships no expert solution.",
- "standard_goal": "Put away the snacks in the bin. Success: The red apple and the yogurt cup must end inside the plastic pail (the trash bin) and the gripper must touch neither. \"Inside\" means the centre of the object's convex hull lies within the container's convex hull, extended without limit along the container's own up axis (the open side; straight up for a container standing normally).",
+ "standard_goal": "Put away the snacks in the bin. Success: The red apple and the yogurt cup must end inside the plastic pail (the trash bin) and, at that moment, the gripper must touch neither (let go of both at the end). \"Inside\" means the centre of the object's convex hull lies within the container's convex hull, extended without limit along the container's own up axis (the open side; straight up for a container standing normally).",
"standard_objects": "Objects here: a ceramic mug, a pair of eyeglasses, a keyboard, a whiteboard marker, a remote control, a smartphone, a wooden bowl, a metal spoon, a computer mouse, a yogurt cup, a pitcher, a plastic pail (trash bin) and a red apple. The Robotiq opens to 8.5 cm. The episode ends by itself the moment the success condition above holds.\n",
"oracle": "none",
"criterion_group": "object_in_container",
@@ -2347,6 +2422,9 @@
"site_log": "https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/throw_away_snacks/limited/",
"hf_replay": "https://huggingface.co/datasets/eai-v2-internal/agent-runs/resolve/main/robolab/codex-gpt6_luna-xhigh/throw_away_snacks/limited/replay.mp4"
}
+ ],
+ "change_log": [
+ "2026-10-14 (board 0113, from board 0086): the success text says that the gripper must not be touching the objects at the moment of success (\"at that moment\"); the old wording \"the gripper must touch none of them\" was read by an agent as \"never touch them\" (board 0086). Text only: the scene, the frozen instance, RoboLab's success term and the verifier are unchanged, so the task keeps its name."
]
},
"tools_picking_drill": {
@@ -2374,7 +2452,7 @@
"terms": "class ToolsPickingDrillTerminations:\n \"\"\"Termination configuration for drill picking task.\"\"\"\n time_out = DoneTerm(func=mdp.time_out, time_out=True)\n success = DoneTerm(\n func=object_outside_of_and_on_surface,\n params={\"object\": [\"cordless_drill\"], \"container\": \"left_bin\", \"surface\": \"table\", \"logical\": \"all\", \"tolerance\": 0.0, \"require_gripper_detached\": True}\n )",
"note": "",
"difficulty": "RoboLab labels this task \"simple\" (1 subtasks; attributes: spatial, semantics); its success term (`object_outside_of_and_on_surface`) asks for the cordless drill, which starts in the right bin, set on the table (the success term: outside the left bin and on the table). The tooling gives only thin primitives (move_ee on Isaac Lab's differential IK, joint moves, gripper, hold), so every grasp, transport and placement is the agent's to build. RoboLab ships no expert solution.",
- "standard_goal": "Select the cordless drill and put it on the table. Success: The cordless drill must end outside the left bin and resting on the table, and the gripper must no longer touch it. \"Outside\" means the centre of the object's convex hull is not within the container's convex hull extended along the container's own up axis. \"On the table\" means the table supports it (its contact force points upwards, within 45 deg of vertical).",
+ "standard_goal": "Select the cordless drill and put it on the table. Success: The cordless drill must end outside the left bin and resting on the table, and the gripper must no longer touch it. (RoboLab's term names the left bin although the drill starts in the right bin: the drill has to come out of its bin and rest on the table itself.) \"Outside\" means the centre of the object's convex hull is not within the container's convex hull extended along the container's own up axis. \"On the table\" means the table supports it (its contact force points upwards, within 45 deg of vertical).",
"standard_objects": "Objects here: two clamps, a cordless drill, a spring clamp, three bins, a black hammer, a wooden hammer, a red hammer and a blue hammer. Three dark grey bins stand side by side at the far edge of the table: the left one (on the robot's left, +y) holds the four hammers, the middle one the clamps, the right one (on the robot's right, -y) the cordless drill. The Robotiq opens to 8.5 cm. The episode ends by itself the moment the success condition above holds.\n",
"oracle": "none",
"criterion_group": "object_outside_of_and_on_surface",
@@ -2392,6 +2470,9 @@
"site_log": null,
"hf_replay": null
}
+ ],
+ "change_log": [
+ "2026-10-14 (board 0113, from board 0086): the success text explains that RoboLab's term names the left bin although the drill starts in the right bin, and that the drill has to rest on the table itself. Text only: the scene, the frozen instance, RoboLab's success term and the verifier are unchanged, so the task keeps its name."
]
},
"unstack_rubiks_cube": {
@@ -2416,7 +2497,7 @@
"terms": "class UnstackRubiksCubeTerminations:\n time_out = DoneTerm(func=mdp.time_out, time_out=True)\n success = DoneTerm(func=object_outside_of_and_on_surface, params={\"object\": [\"rubiks_cube_middle\", \"rubiks_cube_top\", \"rubiks_cube_bottom\"], \"container\": \"grey_bin\", \"surface\": \"table\", \"logical\": \"all\", \"require_gripper_detached\": True})",
"note": "",
"difficulty": "RoboLab labels this task \"moderate\" (2 subtasks; attributes: stacking); its success term (`object_outside_of_and_on_surface`) asks for 3 objects out of grey_bin. The tooling gives only thin primitives (move_ee on Isaac Lab's differential IK, joint moves, gripper, hold), so every grasp, transport and placement is the agent's to build. RoboLab ships no expert solution.",
- "standard_goal": "Unstack the rubiks cube tower. Success: The three Rubik's cubes of the tower must end outside the grey bin and standing on the table, and the gripper must touch none of them. \"Outside\" means the centre of the object's convex hull is not within the container's convex hull extended along the container's own up axis. \"On the table\" means the table supports it (its contact force points upwards, within 45 deg of vertical).",
+ "standard_goal": "Unstack the rubiks cube tower. Success: The three Rubik's cubes of the tower must end outside the grey bin and standing on the table, and, at that moment, the gripper must touch none of them (moving them with the gripper is how the task is done; let go of them at the end). \"Outside\" means the centre of the object's convex hull is not within the container's convex hull extended along the container's own up axis. \"On the table\" means the table supports it (its contact force points upwards, within 45 deg of vertical).",
"standard_objects": "Objects here: a bowl, a ketchup bottle, a grey bin, a mug, three Rubik's cubes and a banana. The three Rubik's cubes start as one tower standing on the table. The Robotiq opens to 8.5 cm. The episode ends by itself the moment the success condition above holds.\n",
"oracle": "none",
"criterion_group": "object_outside_of_and_on_surface",
@@ -2446,6 +2527,9 @@
"site_log": "https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/unstack_rubiks_cube/limited/",
"hf_replay": "https://huggingface.co/datasets/eai-v2-internal/agent-runs/resolve/main/robolab/codex-gpt6_luna-xhigh/unstack_rubiks_cube/limited/replay.mp4"
}
+ ],
+ "change_log": [
+ "2026-10-14 (board 0113, from board 0086): the success text says that the gripper must not be touching the objects at the moment of success (\"at that moment\"); the old wording \"the gripper must touch none of them\" was read by an agent as \"never touch them\" (board 0086). Text only: the scene, the frozen instance, RoboLab's success term and the verifier are unchanged, so the task keeps its name."
]
},
"utensils_in_mug": {
@@ -2481,7 +2565,7 @@
"terms": "class Terminations:\n time_out = DoneTerm(func=mdp.time_out, time_out=True)\n success = DoneTerm(\n func=object_in_container,\n params={\n \"object\": [\"fork_big\", \"spoon_big\"],\n \"container\": \"ceramic_mug\",\n \"logical\": \"all\",\n \"require_gripper_detached\": True\n },\n )",
"note": "",
"difficulty": "RoboLab labels this task \"moderate\" (2 subtasks; attributes: semantics, affordance); its success term (`object_in_container`) asks for 2 objects into ceramic_mug. The tooling gives only thin primitives (move_ee on Isaac Lab's differential IK, joint moves, gripper, hold), so every grasp, transport and placement is the agent's to build. RoboLab ships no expert solution.",
- "standard_goal": "Put the fork and spoon in the ceramicmug. Success: The fork on the bagel plate and the spoon in the red bowl must end inside the ceramic mug and the gripper must touch neither. \"Inside\" means the centre of the object's convex hull lies within the container's convex hull, extended without limit along the container's own up axis (the open side; straight up for a container standing normally).",
+ "standard_goal": "Put the fork and spoon in the ceramicmug. Success: The fork on the bagel plate and the spoon in the red bowl must end inside the ceramic mug and, at that moment, the gripper must touch neither (let go of both at the end). \"Inside\" means the centre of the object's convex hull lies within the container's convex hull, extended without limit along the container's own up axis (the open side; straight up for a container standing normally).",
"standard_objects": "Objects here: a bowl, two bananas, three bagels, a coffee can, a yogurt cup, a coffee pot, a ceramic mug, a pitcher, a large fork, a metal spoon, a red apple, an orange, a milk carton, an orange juice carton, a small plate and a large plate. The Robotiq opens to 8.5 cm. The episode ends by itself the moment the success condition above holds.\n",
"oracle": "none",
"criterion_group": "object_in_container",
@@ -2499,6 +2583,9 @@
"site_log": "https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/utensils_in_mug/unlimited/",
"hf_replay": "https://huggingface.co/datasets/eai-v2-internal/agent-runs/resolve/main/robolab/codex-gpt6_luna-xhigh/utensils_in_mug/unlimited/replay.mp4"
}
+ ],
+ "change_log": [
+ "2026-10-14 (board 0113, from board 0086): the success text says that the gripper must not be touching the objects at the moment of success (\"at that moment\"); the old wording \"the gripper must touch none of them\" was read by an agent as \"never touch them\" (board 0086). Text only: the scene, the frozen instance, RoboLab's success term and the verifier are unchanged, so the task keeps its name."
]
},
"white_mug_in_center_of_table": {
diff --git a/scripts/robolab/mk_tasks.py b/scripts/robolab/mk_tasks.py
index 8e66681bb8..57db1bb495 100755
--- a/scripts/robolab/mk_tasks.py
+++ b/scripts/robolab/mk_tasks.py
@@ -159,7 +159,8 @@ def plan_family(fam: str, F: dict, version: str) -> dict[str, dict[str, tuple[by
priv = {"GOAL": F["goal"], "GOAL_TOML": toml_str(F["goal"]), "ROBOLAB_TASK": F["robolab_task"], "README_GOAL": F.get("readme_goal", F["goal"]),
"DIFFICULTY": F["difficulty"], "OBJECTS_BLOCK": objects_block(F), **per_mode("privileged"),
"REFERENCE_SOLUTION": reference("privileged", F.get("reference_solution", ""))}
- std = {"INSTANCE_TEXT": toml_str(I["text"]), "STANDARD_GOAL": F["standard_goal"], "STANDARD_OBJECTS": F["standard_objects"].rstrip() + "\n",
+ std = {"INSTANCE_TEXT": toml_str(I["text"]), "STANDARD_GOAL": F["standard_goal"],
+ "CHANGE_LOG_EXTRA": "".join(f"- {x}\n" for x in F.get("change_log", [])), "STANDARD_OBJECTS": F["standard_objects"].rstrip() + "\n",
**per_mode("standard"),
"REFERENCE_SOLUTION": reference("standard", "`solution/oracle_replay.py` streams the privileged oracle's joint targets "
"(`oracle_qpos.npy`, one row per control step, `hold`) as one-second `joint_position` chunks "
diff --git a/scripts/robolab/template/standard/README.md.tmpl b/scripts/robolab/template/standard/README.md.tmpl
index e5ffcaadb0..5c9ed1e989 100644
--- a/scripts/robolab/template/standard/README.md.tmpl
+++ b/scripts/robolab/template/standard/README.md.tmpl
@@ -56,3 +56,4 @@ replays.
- 2026-10-10 (review round 3, board 0067): the shared service layer and the client only (`environment/sim/rcb_limited.py`, `environment/rcb_client.py`). A camera named more than once in `cameras` is rendered and returned once (N copies of one name made the service build N x N images within the frame budget, which could exhaust its memory). A service that has given up on the episode (it could not start it, or could not resume it after its crashes) answers every request with 409 and the reason, and `Robot()` / `wait_ready()` raise `EpisodeOver` at once; before, nothing answered and they waited out their timeout (up to an hour). The adapter, the collect hook, the grader, the scene and the pass rule are unchanged.
- 2026-10-10 (owner decision, board 0072): the agent's network egress is restricted to the model APIs (`[agent] network_mode = "allowlist"`, the hosts of `scripts/set_agent_egress.py`; since 2026-10-11 these include the package registries, PyPI and the Debian / Ubuntu archives, so that `pip install` and `apt-get install` work); it was open before (Harbor's default), and agents could download upstream sources, reference solutions and robot models from their own shell. The simulator image is `eai-robolab:0.1.2` (was 0.1.1): it holds the MDL materials and textures that RoboLab's asset library fetched from NVIDIA's Omniverse content bucket whenever a scene was built, so that the cameras show under the allowlist what they showed with open egress (on 0.1.1 without that download the table renders red). The scene, the frozen instance, the physics, the tooling and the pass rule are unchanged, so the task keeps its name; results from before this date were obtained with open egress.
- 2026-10-13 (owner, boards 0085 and 0087): the sim service's clock (`environment/sim/rcb_limited.py`, `environment/sim/server.py`, the docstring of `environment/rcb_client.py`). `meta["time_s"]` and `meta["time_left_s"]` count from the agent's first request to the service (any endpoint), and a restarted service keeps that moment; before, they counted from the service's own start of the episode (when its scene was ready). Where the service is ready before the agent's phase begins (the harness is installed in between: a median of 79 s over the recorded standard trials), that clock ran ahead of the agent's and `time_left_s` reached 0 while the agent still had that time. `/spec` says what the clock counts from (`budget.clock`), and the instruction's `## Time limit` tells the agent to plan by `time_left_s`. The service never ended an episode on this clock: the scene, the frozen instance, the pass rule and the grader are unchanged, so the task keeps its name and earlier results stand.
+@@CHANGE_LOG_EXTRA@@
\ No newline at end of file
diff --git a/scripts/robowits/families.py b/scripts/robowits/families.py
index df39c7ef71..a57c21caa5 100644
--- a/scripts/robowits/families.py
+++ b/scripts/robowits/families.py
@@ -9,6 +9,9 @@ Every family here has tasks. Eleven further families without a valid positive ar
`criterion` (`mk_tasks.py --instance <its instance.json> <family>`).
DROPPED lists the families that are not coming back, with the reason.
+change_log (optional): {"priv": [...], "std": [...]}, README Change Log entries of this family alone, after those of every
+family (mk_tasks.CHANGE_LOG).
+
host_cpu_vendor (optional): the family's frozen instance is rebuilt within tolerance only on CPUs of this vendor (the
`vendor_id` of /proc/cpuinfo; docs/benchmarks/robowits.md, Hosts). mk_tasks.py writes it to `[metadata].host_cpu_vendor`
of both tasks and to their READMEs, for a scheduler to read; on another vendor's host the trial ends as infra_error = 1.
@@ -62,6 +65,9 @@ FAMILIES = {
"Neither the ball nor the bottle has fallen more than 5 cm below the table top."],
unchecked="how the ball got there (the blue container and the funnel are not referenced), whether a gripper holds anything, and any speed",
text_vs_code="\"Into the red bottle\" means deep inside it: seen from above inside the bottle's outline, with the ball's top more than 11 cm below the bottle's top and the ball off the bottle's base. \"From the blue container\" is not checked."),
+ priv_note="The red bottle's base is thin in the simulator (its collision shape is about 3.5 mm thick, and the bottle is fixed to the table): a ball that falls fast down the bottle can pass through the base onto the table, which fails the task. In one measured case about half of the pours did (16 replays of one pouring trajectory with tiny noise, 9 balls through the base): let the ball in slowly.",
+ std_note="The red bottle's base is thin in the simulator (its collision shape is about 3.5 mm thick, and the bottle is fixed to the table): a ball that falls fast down the bottle can pass through the base onto the table, which fails the task. In one measured case about half of the pours did (16 replays of one pouring trajectory with tiny noise, 9 balls through the base): let the ball in slowly.",
+ change_log={"priv": ["2026-10-14 (board 0113, from board 0086): the instruction says that the bottle's base is thin in the simulator and a fast-falling ball can pass through it (board 0086: a privileged trial measured 9 of 16 slightly perturbed replays of one pour losing the ball through the base; every standard run failed). Text only: the scene, the frozen instance, the tooling, the predicate and the verifier are unchanged, so the task keeps its name. A thicker base would change the frozen scene (not done)."], "std": ["2026-10-14 (board 0113, from board 0086): the instruction says that the bottle's base is thin in the simulator and a fast-falling ball can pass through it (board 0086: a privileged trial measured 9 of 16 slightly perturbed replays of one pour losing the ball through the base; every standard run failed). Text only: the scene, the frozen instance, the tooling, the predicate and the verifier are unchanged, so the task keeps its name. A thicker base would change the frozen scene (not done)."]},
),
"ball_onto_tower": dict(
status="pr", rid="17-ball-onto-tower", eval_index=0, blenderkit=True, backend="cpu", oracle="none",
diff --git a/scripts/robowits/mk_tasks.py b/scripts/robowits/mk_tasks.py
index 984d16a79a..fb76ae242e 100755
--- a/scripts/robowits/mk_tasks.py
+++ b/scripts/robowits/mk_tasks.py
@@ -450,12 +450,13 @@ def reference_text(fam: str, f: dict, mode: str) -> str:
"published on the results site).")
-def readme(title: str, rows: list[tuple[str, str]], task: str, difficulty: str, reference: str, verification: str, mode: str) -> str:
+def readme(title: str, rows: list[tuple[str, str]], task: str, difficulty: str, reference: str, verification: str, mode: str,
+ extra: tuple = ()) -> str:
"""README content in the standard layout (PROTOCOL §2.1). normalize() (scripts/task_conventions.py) then writes the key
table (benchmark, mode, twin, simulator / robot, family / instance, oracle, budget, base image) from task.toml and the
Dockerfiles, and moves these extra rows and the Change Log (PROTOCOL §2.8) into ## Details."""
table = "\n".join(["| | |", "|---|---|"] + [f"| {k} | {v} |" for k, v in rows])
- log = "\n".join(f"- {e}" for e in CHANGE_LOG[mode])
+ log = "\n".join(f"- {e}" for e in [*CHANGE_LOG[mode], *extra]) # extra: the family's own entries (families.py change_log)
return (f"# {title}\n\n{table}\n\n## Task\n{task}\n\n## Difficulty\n{difficulty}\n\n## Reference solution\n{reference}\n\n"
f"## Verification\n{verification}\n\n## Change Log\n{log}\n")
@@ -508,7 +509,8 @@ def priv_readme(fam: str, f: dict) -> str:
task = wrap(f"{f_text(fam)} One replayable trajectory from the frozen instance (evaluation scene {f['eval_index']} of RoboWits task "
f"{int(num(f))}), with the privileged tooling (`rcb_robowits.py`: the scene's state, resets, thin primitives, replay); the "
"success criterion as graded is in `instruction.md`.")
- return readme(priv_dir(fam), rows, task, wrap(f["difficulty"]), reference_text(fam, f, "priv"), ver, "priv")
+ return readme(priv_dir(fam), rows, task, wrap(f["difficulty"]), reference_text(fam, f, "priv"), ver, "priv",
+ f.get("change_log", {}).get("priv", ()))
def std_readme(fam: str, f: dict) -> str:
@@ -536,7 +538,8 @@ def std_readme(fam: str, f: dict) -> str:
diff = wrap(f"That of `{priv_dir(fam)}` (its Difficulty), without resets, privileged state or the tooling's primitives: the agent "
"perceives the scene through calibrated RGB-D cameras and its own proprioception, and every action it sends is part of "
"the one graded episode.")
- return readme(std_dir(fam), rows, task, diff, reference_text(fam, f, "std"), ver, "std")
+ return readme(std_dir(fam), rows, task, diff, reference_text(fam, f, "std"), ver, "std",
+ f.get("change_log", {}).get("std", ()))
def toml_str(s: str) -> str:
diff --git a/scripts/task_conventions.py b/scripts/task_conventions.py
index eb3809db70..282a6e6a8d 100644
--- a/scripts/task_conventions.py
+++ b/scripts/task_conventions.py
@@ -22,6 +22,8 @@ What it does:
memory cap from it, and the cap sits above the limit so that the monitor sees a violation before the OOM killer).
- `instruction.md` ends with `## External material` (§2.11: the same text in every task) and `## Resource limits`, rendered
from `[metadata.resources.limits]` (§2.10).
+- `[metadata] dynamics` (§2.14: `dynamic` when something in the scene other than the robot does work by a force other
+ than the robot's and gravity, else `static`) from `scripts/task_dynamics.toml`, the one table.
- the agent's time cutoff (§2.12) from `scripts/task_cutoffs.toml`, the one table: `[agent] timeout_sec`, the standard
modes' `RCB_WALL_CLOCK_S` in `environment/docker-compose.yaml`, and the last section of `instruction.md`, `## Time
limit` ("You have N seconds to complete this task. Do not cheat ...", then what the cutoff means in the mode).
@@ -75,23 +77,67 @@ BENCHMARKS = [ # (directory prefix, label)
("robofunolympics-", "RoboFunOlympics"),
("pandadyn-", "Panda Dynamic Scenes"),
]
-# Documented exceptions: tasks without a privileged twin (the same lists: ci_checks/check-instance-copies.sh NO_TWIN and
-# STANDARD_ONLY). RoboOlympics (PROTOCOL §4b, owner 2026-10-08): standard-realtime only. RoboDojo's six tasks whose physics
-# does not replay bit for bit (PROTOCOL §4.8 exception, owner 2026-10-11): standard only, graded by the live episode.
-NO_TWIN = {"roboolympics-": "— (none: RoboOlympics is standard-realtime only, PROTOCOL §4b)",
+# Modes (PROTOCOL §2.15, owner 2026-10-14): a benchmark ships any subset of the three modes. A standard or
+# standard-realtime task without a privileged twin is expected when its benchmark has no privileged mode (no privileged
+# number in its row of scripts/task_cutoffs.toml); in a benchmark that has one, the task is listed here with the reason
+# its README's Twin row gives (ci_checks/check-instance-copies.sh fails a missing twin that is neither). RoboDojo's six
+# tasks whose physics does not replay bit for bit (PROTOCOL §4.8, owner 2026-10-11): standard only, graded by the live
+# episode. The three RoboOlympics tasks keep the wording their READMEs were written with (2026-10-08).
+_RO_NO_TWIN = "— (none: RoboOlympics is standard-realtime only, PROTOCOL §4b)"
+NO_TWIN = {"roboolympics-airhockey-match-i00-standard-realtime": _RO_NO_TWIN,
+ "roboolympics-badminton-match-i00-standard-realtime": _RO_NO_TWIN,
+ "roboolympics-tabletennis-match-i00-standard-realtime": _RO_NO_TWIN,
"robodojo-fold-clothes-i00-standard": "— (none: standard only, graded by the live episode; its physics does not replay bit for bit, PROTOCOL §4.8)",
"robodojo-fold-clothes-random-i00-standard": "— (none: standard only, graded by the live episode; its physics does not replay bit for bit, PROTOCOL §4.8)",
"robodojo-make-toast-random-i00-standard": "— (none: standard only, graded by the live episode; its physics does not replay bit for bit, PROTOCOL §4.8)",
"robodojo-pour-by-language-i00-standard": "— (none: standard only, graded by the live episode; its physics does not replay bit for bit, PROTOCOL §4.8)",
"robodojo-pour-liquid-into-cup-i00-standard": "— (none: standard only, graded by the live episode; its physics does not replay bit for bit, PROTOCOL §4.8)",
"robodojo-pour-liquid-into-cup-random-i00-standard": "— (none: standard only, graded by the live episode; its physics does not replay bit for bit, PROTOCOL §4.8)"}
+# PROTOCOL §2.16 (owner, 2026-10-14): THE registry of modes. A mode's name is the whole suffix of its task directories
+# (`<base>-<mode>`) and its `[metadata] mode`; everything that parses or branches on a mode reads this table (mode_of,
+# base_of, the cutoff table's columns, the limits' scope, the checks under ci_checks/, which get it once per run through
+# `python3 scripts/task_conventions.py modes`). A new mode is one entry here plus its protocol section, by a mode
+# pull request (PROTOCOL §2.16 lists what it must contain). Fields:
+# short the column label in docs/CUTOFFS.md
+# section the PROTOCOL section that defines the mode
+# sidecar the simulator runs in a trusted `sim` sidecar the agent's container talks to (robot as a service):
+# the instance copy is environment/sim/, the agent's image holds no simulator, the service is told the
+# cutoff (RCB_WALL_CLOCK_S), the verifier collects its evidence from the sidecar
+# limit_scope whom the resource limits bind (§2.10): "trial" (every container) or "agent" (the agent's container)
+# protocols the `[metadata] protocol` values a task of the mode may declare (empty: the field is not set)
+# twin the mode whose task of the same base holds the pair's frozen instance and seeds (None: itself)
+# same_cutoff_as the mode whose cutoff this mode's must equal wherever a benchmark, family or task has both (None: free)
+MODE_REGISTRY = {
+ "privileged": {"short": "P", "section": "§3", "sidecar": False, "limit_scope": "trial", "protocols": (),
+ "twin": None, "same_cutoff_as": None},
+ "standard": {"short": "S", "section": "§4", "sidecar": True, "limit_scope": "agent",
+ "protocols": ("eai-standard/2.1", "eai-standard/2.2", "eai-standard/2.4"),
+ "twin": "privileged", "same_cutoff_as": "privileged"}, # owner, 2026-10-12: one cutoff for both modes
+ "standard-realtime": {"short": "RT", "section": "§4b", "sidecar": True, "limit_scope": "agent",
+ "protocols": ("eai-standard/2.3",), "twin": "privileged", "same_cutoff_as": None},
+}
+MODES = tuple(MODE_REGISTRY)
+SIDECAR_MODES = tuple(m for m, d in MODE_REGISTRY.items() if d["sidecar"]) # robot as a service (§4, §4b)
+for _m, _d in MODE_REGISTRY.items(): # a registry that cannot parse names is refused at import (every check imports it)
+ assert re.fullmatch(r"[a-z]+(?:-[a-z]+)*", _m) and not any(_m.endswith("-" + o) for o in MODE_REGISTRY if o != _m), \
+ f"mode {_m!r}: lower-case words joined by '-', not ending in another mode's suffix (PROTOCOL §2.16)"
+ assert _d["limit_scope"] in ("trial", "agent") and _d["twin"] in (None, *MODE_REGISTRY) \
+ and _d["same_cutoff_as"] in (None, *MODE_REGISTRY), f"mode {_m!r}: a bad registry field (PROTOCOL §2.16)"
HEADER_KEYS = {"Benchmark", "Mode", "Twin", "Simulator / robot", "Family / instance", "Oracle", "Agent time budget",
- "Base image"}
+ "Real time", "Base image"}
DROP_ROWS = {"simulator / robot", "family / instance", "base image", "agent budget", "task type / tier", "tier / mode",
"task type / mode"}
CORE = ["Task", "Difficulty", "Reference solution", "Verification"]
# PROTOCOL §2.9: the minimum compute a task needs, `[metadata.resources]` in task.toml (same list: ci_checks/check-resources.py)
RESOURCES_FILE = Path(__file__).resolve().parent / "task_resources.toml"
+# PROTOCOL §2.14: scene dynamics, `[metadata] dynamics` in task.toml (same list: ci_checks/check-dynamics.py)
+DYNAMICS_FILE = Path(__file__).resolve().parent / "task_dynamics.toml"
+DYNAMICS = ("static", "dynamic")
+# PROTOCOL §4b: a standard-realtime task states its slowdown S, `[metadata] slowdown` in task.toml, written from its sim
+# service's `RCB_SLOWDOWN` default (environment/docker-compose.yaml); its reset budget is the `RCB_RESET_BUDGET` default
+# there (same rules: ci_checks/check-realtime.py)
+SLOWDOWN_RE = re.compile(r'RCB_SLOWDOWN:\s*"?\$\{RCB_SLOWDOWN:-([0-9.]+)\}"?')
+RESET_RE = re.compile(r'RCB_RESET_BUDGET:\s*"?\$\{RCB_RESET_BUDGET:-([0-9]+)\}"?')
RESOURCE_KEYS = { # key -> type; recommended_gpu is optional when gpu_required is false
"min_cpus": float, "min_memory_gb": float, "min_gpu_memory_gb": float, "gpu_required": bool, "gpu_class": str,
"recommended_gpu": str, "verifier_min_cpus": float, "verifier_min_memory_gb": float,
@@ -106,7 +152,7 @@ LIMIT_KEYS = { # key -> type (sizes in GiB)
CAP_KEYS = {"cpus": float, "memory_gb": float, "pids": int} # [metadata.resources.caps] per compose service (§2.10)
ENFORCE_ENV = "RCB_LIMITS_ENFORCE" # §2.10: run-time switch (compose default 0 = record only; 1 = a violation fails)
EMERGENCY_GPU_FACTOR = 1.5 # §2.10: host protection in enforced runs (the same value: rcb_monitor/monitor.py)
-LIMIT_SCOPES = {"privileged": "trial", "standard": "agent", "standard-realtime": "agent"} # §2.10: who the limits bind
+LIMIT_SCOPES = {m: d["limit_scope"] for m, d in MODE_REGISTRY.items()} # §2.10: who the limits bind
LIMITS_HEADING = "## Resource limits"
# PROTOCOL §2.11 (owner, 2026-10-11, board 0072): an honour rule, the same words in every task's instruction.md, just
# before `## Resource limits` (generator-owned: edit it here, never in a task; ci_checks/check-resources.py compares).
@@ -160,6 +206,7 @@ CUTOFF_NOTE = {
"spent before that request. Plan by it." + _CUTOFF_CLOCK + _CUTOFF_CLOSE),
}
CUTOFF_NOTE["standard-realtime"] = CUTOFF_NOTE["standard"]
+assert set(CUTOFF_NOTE) == set(MODE_REGISTRY), "every mode of MODE_REGISTRY needs its CUTOFF_NOTE (PROTOCOL §2.12)"
# A benchmark whose service does not count from the agent's first request keeps the earlier, weaker sentence, and its own
# text says what its service counts (PROTOCOL §2.12). None today: RoboARC's service counts from the first request
# since its MuJoCo release (2026-10-13).
@@ -184,23 +231,36 @@ def replace_words(s: str) -> str:
return s
-# PROTOCOL D12: the three modes, each the whole suffix of its task directories. A mode is recognised by its complete
-# suffix only: `-standard-realtime` is never `-standard` (it does not end in it), and `-standard` is never standard-realtime.
-MODES = ("privileged", "standard", "standard-realtime")
-SIDECAR_MODES = ("standard", "standard-realtime") # robot as a service: the simulator runs in the `sim` sidecar (§4, §4b)
-
-
+# PROTOCOL D12 / §2.16: each mode of MODE_REGISTRY is the whole suffix of its task directories. A mode is recognised by
+# its complete suffix only, longest first: `-standard-realtime` is never `-standard` (it does not end in it).
def mode_of(task: Path) -> str:
for mode in sorted(MODES, key=len, reverse=True): # longest suffix first
if task.name.endswith(f"-{mode}"):
return mode
- raise ValueError(f"{task.name}: task directories end in -privileged, -standard or -standard-realtime "
- "(docs/PROTOCOL.md §2.2)")
+ raise ValueError(f"{task.name}: task directories end in " + ", ".join(f"-{m}" for m in MODES)
+ + " (docs/PROTOCOL.md §2.2, the modes of §2.16)")
def base_of(task: Path) -> str:
- """The name without its mode suffix (`-privileged`, `-standard`, `-standard-realtime`; PROTOCOL §2.2)."""
- return re.sub(r"-(standard-realtime|privileged|standard)$", "", task.name)
+ """The name without its mode suffix (`-privileged`, `-standard`, `-standard-realtime`, ...; PROTOCOL §2.2)."""
+ for mode in sorted(MODES, key=len, reverse=True):
+ if task.name.endswith(f"-{mode}"):
+ return task.name[: -len(mode) - 1]
+ return task.name
+
+
+def twin_of(task: Path) -> Path | None:
+ """The task of the same base in the mode MODE_REGISTRY names as this mode's twin (the one that holds the pair's frozen
+ instance and seeds: `<base>-privileged` for the standard modes), whether it exists or not; None for a mode without one."""
+ twin = MODE_REGISTRY[mode_of(task)]["twin"]
+ return task.with_name(f"{base_of(task)}-{twin}") if twin else None
+
+
+def modes_table() -> str:
+ """One line per mode, longest name first, for the shell checks (`python3 scripts/task_conventions.py modes`):
+ `<mode> <sidecar 0|1> <twin mode or ->`."""
+ return "\n".join(f"{m} {int(MODE_REGISTRY[m]['sidecar'])} {MODE_REGISTRY[m]['twin'] or '-'}"
+ for m in sorted(MODES, key=len, reverse=True))
def bench_of(task: Path) -> str:
@@ -577,6 +637,23 @@ def cutoff_for(task: Path) -> int | None:
return table.get(bench, {}).get(mode)
+def no_twin_reason(task: Path) -> str | None:
+ """Why a standard or standard-realtime task has no privileged twin `<base>-privileged` (PROTOCOL §2.15), as its
+ README's Twin row says it; None for a privileged task and for one that must have its twin. Whether the twin exists
+ is not looked at: the answer is the same before and after a generator has rendered it."""
+ twin = MODE_REGISTRY[mode_of(task)]["twin"]
+ if twin is None:
+ return None
+ for prefix, why in NO_TWIN.items():
+ if task.name.startswith(prefix):
+ return why
+ bench = bench_of(task)
+ row = cutoff_table().get(bench)
+ if row and twin not in row:
+ return f"— (none: {bench} has no {twin} mode, PROTOCOL §2.15)"
+ return None
+
+
def cutoff_sec(bench: str, mode: str) -> int:
"""The benchmark's number for a mode (`[cutoff_sec]`): what a generator that still writes the number (task.toml,
the sim service's wall clock) takes it from, so that no generator keeps a copy of its own. A family's or a task's
@@ -840,17 +917,23 @@ def render_readme(task: Path) -> str:
ev = md.get("oracle_evidence", "")
head = [("Benchmark", bench_of(task) or "—"),
("Mode", mode + (f" (`{md['protocol']}`)" if mode in SIDECAR_MODES and md.get("protocol") else ""))]
- # a standard or realtime task always has its privileged twin (PROTOCOL §2.2), even when a generator renders it first
- # (an exempted name whose twin still exists, e.g. before a removal lands, keeps the twin)
- no_twin = None if task.with_name(twin).is_dir() else next((v for p, v in NO_TWIN.items() if task.name.startswith(p)), None)
+ # a standard or realtime task has its privileged twin unless no_twin_reason() says why not (PROTOCOL §2.15), even
+ # when a generator renders it first (an exempted name whose twin still exists, e.g. before a removal lands, keeps it)
+ no_twin = None if task.with_name(twin).is_dir() else no_twin_reason(task)
head.append(("Twin", no_twin or (f"`{twin}`" if mode in SIDECAR_MODES or task.with_name(twin).is_dir()
else "— (privileged only)")))
head += [("Simulator / robot", f"`{md.get('simulator', '')}` / `{md.get('robot', '')}`"),
("Family / instance", f"`{md.get('family', '')}`" + (f" / seed {seed}" if seed is not None else "")
+ (f" (allowed seeds: {seeds_text(sd['allowed'])}; §2.13)" if sd["allowed"] != [sd["default"]] else "")),
("Oracle", f"`{oracle}`" + (f" — {ev}" if ev else "") if oracle else "—"),
- ("Agent time budget", f"{int(float(ag.get('timeout_sec', 0)))} s; tokens: protocol default (PROTOCOL §5)"),
- ("Base image", f"`{image_of(task)}`")]
+ ("Agent time budget", f"{int(float(ag.get('timeout_sec', 0)))} s; tokens: protocol default (PROTOCOL §5)")]
+ rt = realtime_of(task)
+ if rt is not None:
+ head.append(("Real time", f"slowdown S = {slowdown_text(rt[0])} (the world runs at 1/S of real speed: one control "
+ f"step every S x control_dt s of wall time; `RCB_SLOWDOWN`'s default)"
+ + (f"; reset budget {rt[1]} (`RCB_RESET_BUDGET`'s default)" if rt[1] is not None else "")
+ + " (PROTOCOL §4b)"))
+ head.append(("Base image", f"`{image_of(task)}`"))
used = set()
task_text = intro or meta.get("task", {}).get("description", "")
difficulty = secd.get("Difficulty") or default_difficulty(task, rowd)
@@ -959,12 +1042,84 @@ def normalize_seeds(task: Path) -> None:
write_if_changed(p, s)
+def dynamics_table() -> tuple[dict, dict]:
+ """scripts/task_dynamics.toml: ({benchmark: entry}, {benchmark: {family: entry}}), entry = {dynamics, why}."""
+ t = tomllib.loads(DYNAMICS_FILE.read_text()) if DYNAMICS_FILE.is_file() else {}
+ return t.get("benchmark", {}), t.get("family", {})
+
+
+def dynamics_for(task: Path) -> str | None:
+ """The task's `dynamics` (PROTOCOL §2.14): its family's entry in scripts/task_dynamics.toml, else its benchmark's.
+ None: the table has no row for the benchmark (ci_checks/check-dynamics.py fails such a task)."""
+ rows, fams = dynamics_table()
+ bench = bench_of(task)
+ e = fams.get(bench, {}).get(family_of(task)) or rows.get(bench)
+ if e is None:
+ return None
+ if e.get("dynamics") not in DYNAMICS:
+ raise RuntimeError(f"{task.name}: scripts/task_dynamics.toml gives dynamics {e.get('dynamics')!r}, not one of {DYNAMICS}")
+ return e["dynamics"]
+
+
+def normalize_dynamics(task: Path) -> None:
+ """Writes `dynamics = "..."` (PROTOCOL §2.14) right after `mode` in `[metadata]`, replacing an earlier line, when
+ scripts/task_dynamics.toml has a row for the task's benchmark."""
+ d = dynamics_for(task)
+ if d is None:
+ return
+ p = task / "task.toml"
+ s = re.sub(r'(?m)^dynamics = .*\n', "", p.read_text())
+ line = f'dynamics = "{d}"'
+ line += " " * max(1, 34 - len(line)) + "# scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml\n"
+ s, n = re.subn(r'(?m)^(mode = "[^"]*".*\n)', lambda m: m.group(1) + line, s, count=1)
+ if n != 1:
+ raise RuntimeError(f"{task.name}: task.toml has no `mode` line in [metadata]")
+ write_if_changed(p, s)
+
+
+def realtime_of(task: Path) -> tuple[float, int | None] | None:
+ """A standard-realtime task's (slowdown S, reset budget) as its sim service starts by default: the `RCB_SLOWDOWN`
+ and `RCB_RESET_BUDGET` defaults in environment/docker-compose.yaml (PROTOCOL §4b). None: not a standard-realtime task,
+ or a service without `RCB_SLOWDOWN` (ci_checks/check-realtime.py fails a 2.3 one)."""
+ if mode_of(task) != "standard-realtime":
+ return None
+ c = read_text(task / "environment/docker-compose.yaml") or ""
+ m, r = SLOWDOWN_RE.search(c), RESET_RE.search(c)
+ if not m:
+ return None
+ return float(m.group(1)), (int(r.group(1)) if r else None)
+
+
+def slowdown_text(S: float) -> str:
+ return f"{S:g}"
+
+
+def normalize_slowdown(task: Path) -> None:
+ """Writes `slowdown = S` (PROTOCOL §4b) right after `dynamics` (else `mode`) in `[metadata]` of a standard-realtime
+ task, from its sim service's `RCB_SLOWDOWN` default, replacing an earlier line."""
+ rt = realtime_of(task)
+ p = task / "task.toml"
+ s = re.sub(r'(?m)^slowdown = .*\n', "", p.read_text())
+ if rt is not None:
+ line = f"slowdown = {slowdown_text(rt[0])}"
+ line += " " * max(1, 34 - len(line)) + ("# standard-realtime: the world runs at 1/S of real speed (docs/PROTOCOL.md "
+ "§4b); RCB_SLOWDOWN's default\n")
+ s, n = re.subn(r'(?m)^(dynamics = "[^"]*".*\n)', lambda m: m.group(1) + line, s, count=1)
+ if n != 1:
+ s, n = re.subn(r'(?m)^(mode = "[^"]*".*\n)', lambda m: m.group(1) + line, s, count=1)
+ if n != 1:
+ raise RuntimeError(f"{task.name}: task.toml has no `mode` line in [metadata]")
+ write_if_changed(p, s)
+
+
def normalize(task: Path) -> None:
task = Path(task)
mode_of(task)
if (task / "results").exists():
shutil.rmtree(task / "results")
normalize_toml(task)
+ normalize_dynamics(task)
+ normalize_slowdown(task)
normalize_cutoff(task)
normalize_resources(task)
normalize_seeds(task)
@@ -982,6 +1137,9 @@ def snapshot(task: Path) -> dict[str, bytes]:
def main(argv: list[str]) -> int:
+ if argv[:1] == ["modes"]: # the mode registry for the shell checks (PROTOCOL §2.16)
+ print(modes_table())
+ return 0
check = "--check" in argv
dirs = [Path(a) for a in argv if a != "--check" and (Path(a) / "task.toml").is_file()]
bad = 0
diff --git a/scripts/task_cutoffs.py b/scripts/task_cutoffs.py
index dac27bff86..1b28136750 100644
--- a/scripts/task_cutoffs.py
+++ b/scripts/task_cutoffs.py
@@ -33,7 +33,7 @@ import task_conventions as tc # noqa: E402
DOC = ROOT / "docs/CUTOFFS.md"
BASELINE = ROOT / "scripts/task_cutoffs.baseline.toml" # the table on the day it was created: never edited
MANIFEST = ROOT / "tasks/dataset.toml"
-SHORT = {"privileged": "P", "standard": "S", "standard-realtime": "RT"}
+SHORT = {m: d["short"] for m, d in tc.MODE_REGISTRY.items()} # the mode registry (PROTOCOL §2.16)
def all_tasks() -> list[Path]:
@@ -46,11 +46,12 @@ def _dur(sec: int) -> str:
def table_rows(tasks: list[Path]) -> list[str]:
n = Counter((tc.bench_of(t), tc.mode_of(t)) for t in tasks)
- rows = ["| Benchmark | privileged | standard | standard-realtime | Tasks (P / S / RT) |", "|---|---|---|---|---|"]
+ rows = ["| Benchmark | " + " | ".join(tc.MODES) + " | Tasks (" + " / ".join(SHORT[m] for m in tc.MODES) + ") |",
+ "|---|" + "---|" * (len(tc.MODES) + 1)]
for bench, row in sorted(tc.cutoff_table().items(), key=lambda kv: kv[0].lower()):
cells = [_dur(row[m]) if m in row else "—" for m in tc.MODES]
counts = " / ".join(str(n[(bench, m)]) if n[(bench, m)] else "—" for m in tc.MODES)
- rows.append(f"| {bench} | {cells[0]} | {cells[1]} | {cells[2]} | {counts} |")
+ rows.append(f"| {bench} | " + " | ".join(cells) + f" | {counts} |")
return rows
@@ -203,7 +204,7 @@ def render_doc(tasks: list[Path]) -> str:
"",
*table_rows(tasks),
"",
- f"{len(tasks)} task directories. P / S / RT: privileged, standard, standard-realtime. A cutoff is a whole number "
+ f"{len(tasks)} task directories. {' / '.join(SHORT[m] for m in tc.MODES)}: {', '.join(tc.MODES)}. A cutoff is a whole number "
f"of seconds, at most {tc.CUTOFF_CAP} (PROTOCOL D8"
+ "".join(f"; {b}'s exception: {c}" for b, c in sorted(tc.CUTOFF_CAP_EXCEPTIONS.items())) + ").",
"",
diff --git a/scripts/task_cutoffs.toml b/scripts/task_cutoffs.toml
index 432bdea80b..73906257ce 100644
--- a/scripts/task_cutoffs.toml
+++ b/scripts/task_cutoffs.toml
@@ -32,7 +32,7 @@
"MetaWorld+" = { privileged = 1800, standard = 1800 }
"MolmoSpaces" = { privileged = 3600, standard = 3600 }
"MuJoCo Playground" = { privileged = 1800, standard = 1800 }
-"MyoSuite" = { privileged = 3600, standard = 3600 }
+"MyoSuite" = { privileged = 3600, standard = 3600, standard-realtime = 3600 }
"Panda Dynamic Scenes" = { privileged = 3600, standard = 3600 }
"RoboCasa" = { privileged = 3600, standard = 3600 }
"RoboCasa-GR1" = { privileged = 3600, standard = 3600 }
@@ -258,3 +258,12 @@ from_sec = 0
to_sec = 10800
decided_by = "owner"
note = "new row: RoboARC, the MuJoCo release that replaces RoboTangle's Genesis tasks (prepared as RoboTangle-New, then RoboTangle; their entries, 0 -> 7200 on 2026-10-12 and 7200 -> 10800 on 2026-10-13, are folded into this one): 3 h wall clock for every family, within the default cap"
+
+[[history]]
+date = "2026-10-14"
+benchmark = "MyoSuite"
+modes = ["standard-realtime"]
+from_sec = 0
+to_sec = 3600
+decided_by = "owner (board 0112: a realtime twin's cutoff equals its standard twin's)"
+note = "the first standard-realtime task (table_tennis): the standard twin's number"
diff --git a/scripts/task_dynamics.toml b/scripts/task_dynamics.toml
new file mode 100644
index 0000000000..e96182019b
--- /dev/null
+++ b/scripts/task_dynamics.toml
@@ -0,0 +1,87 @@
+# Scene dynamics per task: THE table (docs/PROTOCOL.md §2.14; owner 2026-10-14, board 0106). One row per benchmark, and
+# per family an exception where the family differs from its benchmark's row; every entry gives its reason (`why`).
+#
+# dynamic something in the scene other than the robot can do work by a force other than the robot's own and gravity:
+# a body with a velocity at the start (thrown, served, rolling, sliding, swinging), an actuated or scripted
+# body (conveyor, motor-driven rail, scripted push, a timer that moves a part), an opponent or another agent.
+# static everything else: objects move only because the robot moves them or gravity pulls them (settling after
+# reset, cloth / liquid / granular material the robot moves, passive springs that only give back the
+# robot's energy, visual-only or state-only changes, episode horizons). A robot that must balance actively
+# (humanoid, legged, drone) does not make a task dynamic by itself.
+#
+# scripts/task_conventions.py (`normalize`) writes `[metadata] dynamics` into every task.toml from this file; never edit
+# it in a task, a template or a generator: ci_checks/check-dynamics.py fails a task that disagrees with this table, a
+# benchmark without a row, and an exception that matches no family. To change a value: edit it here (with its `why`),
+# then `python3 scripts/task_cutoffs.py apply` (normalizes every task and refreshes tasks/dataset.toml, ~3 min).
+# The benchmark names are the labels of scripts/task_conventions.py BENCHMARKS; a family is `[metadata] family`.
+# A dynamic task may get a `-standard-realtime` twin (PROTOCOL §4b); static tasks need none. Per-family evidence
+# (file:line): agentboard task 0106, dynamic_tasks.csv.
+
+[benchmark]
+"AutoBio" = { dynamics = "static", why = "Fixed-base arms on a laboratory bench; devices (centrifuge, thermal cycler, mixer, pipette) move only when the robot moves them." }
+"BEHAVIOR-1K" = { dynamics = "static", why = "R1Pro (wheeled base) in household scenes; objects move only by the robot or gravity; cooking / freezing / toggled states change no motion." }
+"BiGym 2.0" = { dynamics = "static", why = "G1 (legs held by the GR00T-WBC policy) at kitchen fixtures; objects and fixtures move only by the robot or gravity." }
+"DexJoCo" = { dynamics = "static", why = "Franka + Allegro hand(s) at a table; objects and devices move only by the robot or gravity." }
+"DexToolBench" = { dynamics = "static", why = "Arm + hand using a tool on objects at rest; nothing moves except by the robot or gravity." }
+"HumanoidBench" = { dynamics = "static", why = "H1 humanoid (floating base, must balance) in scenes whose objects move only by the robot or gravity; balancing is not scene dynamics (owner 2026-10-14)." }
+"KinDER" = { dynamics = "static", why = "TidyBot++ (wheeled base + arm) in KinDER Dynamic3D; objects move only when the robot pushes or carries them." }
+"LIBERO-MAX" = { dynamics = "static", why = "Fixed Franka (LIBERO); the mid-episode perturbation is triggered by the robot's proximity and applied once, nothing moves by itself afterwards." }
+"LabUtopia" = { dynamics = "static", why = "Fixed Franka on a lab bench; apparatus moves only when the robot moves it." }
+"ManiGuard" = { dynamics = "static", why = "Fixed Franka in BEHAVIOR-1K scenes; safety constraints on objects the robot moves, no scripted hazards." }
+"MetaWorld+" = { dynamics = "static", why = "Fixed Sawyer at a table; objects, buttons and drawers move only by the robot or gravity." }
+"MolmoSpaces" = { dynamics = "static", why = "Franka or RB-Y1 (holonomic base) in house scenes; objects and doors move only by the robot or gravity." }
+"MuJoCo Playground" = { dynamics = "static", why = "Arms or legged robots on flat / rough terrain; joystick commands are targets, not bodies; nothing in the scene moves by itself." }
+"MyoSuite" = { dynamics = "static", why = "Musculoskeletal arm / hand / legs; objects move only by the model's muscles or gravity." }
+"Panda Dynamic Scenes" = { dynamics = "dynamic", why = "One or two Pandas in scenes whose balls, pucks, carts, conveyors, rails and pendulums move by themselves (initial velocities, driven rails / conveyors, scripted impulses); the six static families are listed below." }
+"RoboARC" = { dynamics = "static", why = "Fixed-base arm(s) on MuJoCo puzzles, ropes and cloth that move only when the robot moves them (rope / cloth elasticity only returns the robot's energy)." }
+"RoboCasa" = { dynamics = "static", why = "PandaOmron in robosuite kitchens; fixtures change only visual sites or on/off states, objects move only by the robot or gravity." }
+"RoboCasa-GR1" = { dynamics = "static", why = "Fixed-base GR1 at kitchen counters; objects move only by the robot or gravity." }
+"RoboCasa365" = { dynamics = "static", why = "PandaOmron in RoboCasa365 kitchens; fixtures change only visual sites or on/off states, objects move only by the robot or gravity." }
+"RoboDojo" = { dynamics = "static", why = "Two fixed ARX X5 arms at a table; objects move only by the robot or gravity (cloth and liquid moved by the robot)." }
+"RoboFunOlympics" = { dynamics = "static", why = "A humanoid (floating base) at a venue whose ball / implement is at rest until the robot touches it; only the robot and gravity do work." }
+"RoboLab" = { dynamics = "static", why = "Fixed Franka + Robotiq at a table; objects move only by the robot or gravity." }
+"RoboOlympics" = { dynamics = "dynamic", why = "Real-time sports against a scripted opponent: an incoming ball / shuttle / puck and an opponent that moves by itself." }
+"RoboPaint" = { dynamics = "static", why = "Fixed Franka painting or writing; paint deposits only where the robot's brush touches." }
+"RoboTwin 2.0" = { dynamics = "static", why = "Dual-arm aloha-agilex at a table; objects move only by the robot or gravity (RoboTwin's envs set no object velocities)." }
+"RoboWits" = { dynamics = "static", why = "Fixed bimanual Marvin at a table; objects move only by the robot or gravity (a domino chain falls under gravity after the robot's push)." }
+"VLABench" = { dynamics = "static", why = "Fixed Franka at a table; objects move only by the robot or gravity." }
+"YAMLab" = { dynamics = "static", why = "Two fixed YAM arms at a workstation table; objects move only by the robot or gravity." }
+
+[family."MuJoCo Playground"]
+"mujoco_playground/BarkourJoystick" = { dynamics = "dynamic", why = "Scripted random pushes (kicks) apply an external xfrc force to the floating-base Barkour torso at frozen times (12 kicks in the instance), work from neither the robot nor gravity." }
+
+[family."MyoSuite"]
+"table_tennis" = { dynamics = "dynamic", why = "The ping-pong ball is served toward the player with an initial velocity set at reset (5.6 m/s, measured in the frozen instance), kinetic energy not from the robot." }
+
+[family."Panda Dynamic Scenes"]
+"d05_moving_ramp_relay" = { dynamics = "static", why = "Nothing moves by itself; the ball only rolls under gravity once the robot puts it in the channel." }
+"d08_pendulum_energy_relay" = { dynamics = "static", why = "Passive pendulum at rest at t = 0; only the arms put energy into it." }
+"d11_moving_fixture_unlock" = { dynamics = "static", why = "Free fixture at rest; its latch spring only returns energy the robot put in (spring-return)." }
+"d12_reciprocal_sliding_transfer" = { dynamics = "static", why = "Cube at rest on a slippery scoop; nothing moves by itself." }
+"d13_supported_cube_flip_relay" = { dynamics = "static", why = "Cube at rest on the tray; nothing moves by itself." }
+"s06_cube_flip_route_recover" = { dynamics = "static", why = "Cube at rest; it slides down the chute only under gravity after the robot's push." }
+
+[family."RoboCasa365"]
+"robocasa365/TurnOnToasterOven" = { dynamics = "dynamic", why = "The toaster oven's timer dial, once set by the robot, turns back by itself over time (set_joint_state each step: val - elapsed/3000)." }
+
+[family."RoboDojo"]
+"imitate_sorting_sequence" = { dynamics = "dynamic", why = "A second, non-target arm replays a recorded trajectory that sorts the demonstrator objects." }
+"make_kong" = { dynamics = "dynamic", why = "The support arm replays a recorded opponent 'discard' motion." }
+"match_and_pick_from_conveyor" = { dynamics = "dynamic", why = "A powered conveyor belt carries the objects during the episode." }
+"pick_from_conveyor_by_image" = { dynamics = "dynamic", why = "A powered conveyor belt carries the four objects during the episode." }
+"play_tic_tac_toe" = { dynamics = "dynamic", why = "An opponent arm (support_arm0) places its pieces, running in parallel with the agent's actions." }
+
+[family."RoboFunOlympics"]
+"catch" = { dynamics = "dynamic", why = "A launcher 40 m away fires the baseball at t = 1 s (21 m/s)." }
+"catch_and_shoot" = { dynamics = "dynamic", why = "A machine throws the chest pass at a fixed time." }
+"header" = { dynamics = "dynamic", why = "The cross is launched toward the robot at launch_time_s." }
+"juggle_three" = { dynamics = "dynamic", why = "The venue throws the first ball at launch_time_s." }
+"paddle_bounce" = { dynamics = "dynamic", why = "The venue releases the ball at drop_time_s (a clock-timed release; gravity then does the work)." }
+"penalty_kick" = { dynamics = "dynamic", why = "A goalkeeper dummy on a slide is driven by the referee toward the ball after the kick." }
+"table_tennis_return" = { dynamics = "dynamic", why = "The ball is served toward the robot at serve_time_s with spin." }
+"tennis_return" = { dynamics = "dynamic", why = "The ball is served toward the robot at serve_time_s with spin." }
+"trap_ball" = { dynamics = "dynamic", why = "The ball is launched toward the robot at launch_time_s." }
+"volley_shot" = { dynamics = "dynamic", why = "The cross is launched toward the robot at launch_time_s." }
+"volleyball_bump" = { dynamics = "dynamic", why = "The ball is launched toward the robot at launch_time_s." }
+"volleyball_set" = { dynamics = "dynamic", why = "The ball is launched toward the robot at launch_time_s." }
+"volleyball_spike" = { dynamics = "dynamic", why = "The set ball is launched at launch_time_s." }
diff --git a/scripts/task_resources.toml b/scripts/task_resources.toml
index f2efd66a7a..50ef0cd030 100644
--- a/scripts/task_resources.toml
+++ b/scripts/task_resources.toml
@@ -468,6 +468,20 @@ verifier_min_memory_gb = 3.0
verifier_min_gpu_memory_gb = 0.0
measured = "memory: privileged container median 6.2 GB, sim p90 1.4 GB, verifier 1.4 GB, mem_guard (docker memory per container, ws-1, 10-02..10); CPU-only; CPU author estimate"
+# myosuite, standard-realtime: 1 task(s)
+[[rule]]
+match = "^myosuite-.*-standard-realtime$"
+min_cpus = 2.0
+min_memory_gb = 4.0
+min_gpu_memory_gb = 1.0
+gpu_required = true
+gpu_class = "any"
+recommended_gpu = "any NVIDIA GPU with EGL (RTX 4090 / L4)"
+verifier_min_cpus = 2.0
+verifier_min_memory_gb = 3.0
+verifier_min_gpu_memory_gb = 1.0
+measured = "as the standard twin (sim 1.2-1.4 GB, docker stats, board 0112); the GPU only renders the cameras over EGL (5 ms a frame; board 0112, 2026-10-14, ws-1); the sim keeps the GPU through the collect hook (fresh replay and video)"
+
# robocasa, privileged: 29 task(s)
[[rule]]
match = "^robocasa-(?!gr1-).*-privileged$"
@@ -1472,6 +1486,21 @@ trial_max_gpu_memory_gb = 0.0
trial_max_memory_gb = 14.0
measured = "memory: no samples of the agent's container; floor 8 GiB (at most the 8 GiB cap); whole trial: + the sim service's 6 GiB (docker peak 1.4 GiB, 5 samples); processes 512 and disk 20 GiB: design defaults, not yet measured"
+[[limit]]
+match = "^myosuite-.*-standard-realtime$"
+scope = "agent"
+max_gpu_memory_gb = 0.0
+max_memory_gb = 8.0
+max_processes = 512
+max_disk_gb = 20.0
+disk_paths = ["/app", "/tmp", "/root"]
+disk_exclude = []
+sustain_sec = 60.0
+window_sec = 120.0
+trial_max_gpu_memory_gb = 4.0
+trial_max_memory_gb = 14.0
+measured = "GPU: EGL rendering only, under 1 GiB (board 0112) -> peak + 4; memory: no samples of the agent's container; floor 8 GiB (at most the 8 GiB cap); whole trial: + the sim service's 6 GiB (docker peak 1.4 GiB, 5 samples); processes 512 and disk 20 GiB: design defaults, not yet measured"
+
[[limit]]
match = "^robocasa-(?!gr1-).*-privileged$"
scope = "trial"
@@ -2061,6 +2090,11 @@ match = "^myosuite-.*-standard$"
main = { cpus = 4.0, memory_gb = 10.0, pids = 8192 }
sim = { cpus = 4.0, memory_gb = 8.0, pids = 8192 }
+[[cap]]
+match = "^myosuite-.*-standard-realtime$"
+main = { cpus = 4.0, memory_gb = 10.0, pids = 8192 }
+sim = { cpus = 4.0, memory_gb = 8.0, pids = 8192 }
+
[[cap]]
match = "^robocasa-(?!gr1-).*-privileged$"
main = { cpus = 4.0, memory_gb = 20.0, pids = 8192 }
diff --git a/scripts/task_seed.py b/scripts/task_seed.py
index 811081fd6e..5a4c7df869 100755
--- a/scripts/task_seed.py
+++ b/scripts/task_seed.py
@@ -34,11 +34,8 @@ def seeds(task: Path) -> dict:
def holder(task: Path) -> Path:
- if tc.mode_of(task) in tc.SIDECAR_MODES:
- twin = task.with_name(tc.base_of(task) + "-privileged")
- if twin.is_dir():
- return twin
- return task
+ twin = tc.twin_of(task) # the mode registry's twin mode (PROTOCOL §2.16)
+ return twin if twin is not None and twin.is_dir() else task
def places(task: Path) -> list[str]:
diff --git a/tasks/autobio-insert-centrifuge-5430-i00-privileged/README.md b/tasks/autobio-insert-centrifuge-5430-i00-privileged/README.md
index 4a958a13fc..45b8deda2e 100644
--- a/tasks/autobio-insert-centrifuge-5430-i00-privileged/README.md
+++ b/tasks/autobio-insert-centrifuge-5430-i00-privileged/README.md
@@ -12,7 +12,7 @@
| Base image | `ghcr.io/mll-lab-nu/eai-autobio:0.1.0` |
## Task
-AutoBio's `insert_centrifuge_5430` (Load centrifuge rotor), with a UR5e arm (6 joints) with a Robotiq 2F-85 gripper, fixed to the bench: Load a second 1.5 mL tube into the centrifuge's rotor, into the slot diametrically opposite the tube that is already in it (slot 8 of 30, `s.task.slot_id`; the tube in the rotor is in slot 23). The tube to load stands in the plate on the bench. Success at the first control step where the tube's body is within 5 mm of its seated position in that slot (`s.task.final_tar_tubepose`) and its height is between 0.955 and 0.961 m.
+AutoBio's `insert_centrifuge_5430` (Load centrifuge rotor), with a UR5e arm (6 joints) with a Robotiq 2F-85 gripper, fixed to the bench: Load a second 1.5 mL tube into the centrifuge's rotor, into the slot diametrically opposite the tube that is already in it (slot 8 of 30, `s.task.slot_id`; the tube in the rotor is in slot 23). The tube to load stands in the plate on the bench. Success at the first control step where the tube's body is within 5 mm of its seated position in that slot (`s.task.final_tar_tubepose`) and its height is between 0.955 and 0.961 m. That reference position is computed when the episode starts: turning the rotor moves the slot away from it. In this instance a tube that has been let go comes to rest 5.0 to 5.1 mm from it, just outside the 5 mm (AutoBio's own expert's tube included), so the check holds while the tube is still being lowered into the slot; the episode ends at that moment.
On the bench: a centrifuge (an Eppendorf 5430 model) with its lid open and a 30-slot rotor holding one 1.5 mL tube, and a well plate with a second 1.5 mL tube standing in it. The deliverable is one replayable trajectory of joint position targets (`/app/output/trajectory.npz`, `(T, 7)`, one row per 2 ms control step).
## Difficulty
@@ -52,3 +52,4 @@ Generated by `scripts/autobio/generate.py` from `scripts/autobio/instances/inser
### Change Log
- 2026-10-10: the grader's time limits and flags; the scene, the frozen instance and the pass rule are unchanged, so the task keeps its name. `tests/grade.py`: each verdict replay may take 2400 s and the video replay gets what is left of the verifier's 5400 s (every replay had 1700 s), so a reward is written before Harbor's timeout whatever the trajectory's length; a replay without a result also sets `infra_error` (with `grader_error`, as before), every `reward.json` carries `infra_error`, and an error inside the grader still writes a reward (`grader_error = 1`). A trajectory that was graded before gets the same verdict.
- 2026-10-10 (review round 3, board 0067): `tests/grade.py` only. A byte that is not valid UTF-8 in the session log the agent's tools write (`/app/output/rcb_session_log.jsonl`, read only for the self-reported `agent_*` cost fields) was graded `grader_error = 1` and `infra_error = 1` with reward 0, whatever the trajectory; the log is now read tolerantly and can no longer decide anything. A trial whose log is valid is graded as before; the pass rule is unchanged.
+- 2026-10-14 (board 0113, from board 0086): the instruction now states the whole of AutoBio's test: the reference position is computed when the episode starts (turning the rotor moves the slot away from it), and in this instance a released tube rests 5.0 to 5.1 mm from it, so the test holds while the tube is still being lowered (before, this was only in `docs/benchmarks/autobio.md`, Known limits). Text only: the scene, the frozen instance, the reference solution, AutoBio's check and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/autobio-insert-centrifuge-5430-i00-privileged/instruction.md b/tasks/autobio-insert-centrifuge-5430-i00-privileged/instruction.md
index 2129d8c6c1..ceddc7224f 100644
--- a/tasks/autobio-insert-centrifuge-5430-i00-privileged/instruction.md
+++ b/tasks/autobio-insert-centrifuge-5430-i00-privileged/instruction.md
@@ -2,7 +2,7 @@
You control a **UR5e** arm (6 joints) with a Robotiq 2F-85 gripper, fixed to the bench, in AutoBio's simulated biology laboratory (MuJoCo 3.3.0), via Python, through joint
position targets.
-**Task:** Load a second 1.5 mL tube into the centrifuge's rotor, into the slot diametrically opposite the tube that is already in it (slot 8 of 30, `s.task.slot_id`; the tube in the rotor is in slot 23). The tube to load stands in the plate on the bench. Success at the first control step where the tube's body is within 5 mm of its seated position in that slot (`s.task.final_tar_tubepose`) and its height is between 0.955 and 0.961 m.
+**Task:** Load a second 1.5 mL tube into the centrifuge's rotor, into the slot diametrically opposite the tube that is already in it (slot 8 of 30, `s.task.slot_id`; the tube in the rotor is in slot 23). The tube to load stands in the plate on the bench. Success at the first control step where the tube's body is within 5 mm of its seated position in that slot (`s.task.final_tar_tubepose`) and its height is between 0.955 and 0.961 m. That reference position is computed when the episode starts: turning the rotor moves the slot away from it. In this instance a tube that has been let go comes to rest 5.0 to 5.1 mm from it, just outside the 5 mm (AutoBio's own expert's tube included), so the check holds while the tube is still being lowered into the slot; the episode ends at that moment.
On the bench: a centrifuge (an Eppendorf 5430 model) with its lid open and a 30-slot rotor holding one 1.5 mL tube, and a well plate with a second 1.5 mL tube standing in it.
diff --git a/tasks/autobio-insert-centrifuge-5430-i00-privileged/task.toml b/tasks/autobio-insert-centrifuge-5430-i00-privileged/task.toml
index 7704f17433..d95e0360f6 100644
--- a/tasks/autobio-insert-centrifuge-5430-i00-privileged/task.toml
+++ b/tasks/autobio-insert-centrifuge-5430-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "autobio" # AutoBio @ 0850421 on MuJoCo 3.3.0 (CPU)
robot = "ur5e-robotiq-2f85"
family = "insert_centrifuge_5430"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
difficulty = "hard" # AutoBio's own level for the task
success_rule = "first"
diff --git a/tasks/autobio-insert-centrifuge-5430-i00-standard/README.md b/tasks/autobio-insert-centrifuge-5430-i00-standard/README.md
index 34823841e5..367a4f1e74 100644
--- a/tasks/autobio-insert-centrifuge-5430-i00-standard/README.md
+++ b/tasks/autobio-insert-centrifuge-5430-i00-standard/README.md
@@ -12,7 +12,7 @@
| Base image | `ghcr.io/mll-lab-nu/eai-autobio:0.1.0` |
## Task
-AutoBio's `insert_centrifuge_5430` (Load centrifuge rotor) on the same frozen instance as `autobio-insert-centrifuge-5430-i00-privileged`: Load a second 1.5 mL tube into the centrifuge's rotor, into the slot diametrically opposite the tube that is already in it (the rotor has 30 slots). The tube to load stands in the plate on the bench. Success the moment the tube is seated in that slot: within 5 mm of its seated position.
+AutoBio's `insert_centrifuge_5430` (Load centrifuge rotor) on the same frozen instance as `autobio-insert-centrifuge-5430-i00-privileged`: Load a second 1.5 mL tube into the centrifuge's rotor, into the slot diametrically opposite the tube that is already in it (the rotor has 30 slots). The tube to load stands in the plate on the bench. Success the moment the tube is seated in that slot: its body within 5 mm of its seated position there, and its body's height within a fixed 6 mm window (AutoBio's test; standard mode gives no world heights). That seated position is computed when the episode starts: turning the rotor moves the slot away from it, so leave the rotor where it is. In this instance a tube that has been let go comes to rest just outside the 5 mm (5.0 to 5.1 mm away, AutoBio's own expert's tube included), so the test holds while the tube is still being lowered into the slot, held by the gripper; the episode ends in success at that moment.
On the bench: a centrifuge (an Eppendorf 5430 model) with its lid open and a 30-slot rotor holding one 1.5 mL tube, and a well plate with a second 1.5 mL tube standing in it. Standard mode (eai-standard/2.2): the agent drives the robot through a client in one episode, with only its proprioception and the scene's cameras.
## Difficulty
@@ -53,3 +53,4 @@ Generated by `scripts/autobio/generate.py` from `scripts/autobio/instances/inser
- 2026-10-10 (review round 2, board 0067): the client only (`environment/rcb_client.py`). `Robot()` waits up to 3600 s for the service (was 600 s) and prints its progress to stderr once a minute, so the first `Robot()` outlasts the simulator's start (no effect on this benchmark, whose service is up within a minute); `cameras="name"` (a string) is one camera, no longer its characters. The service, the collect hook, the grader, the scene and the pass rule are unchanged.
- 2026-10-10 (review round 3, board 0067): the shared service layer and the client only (`environment/sim/rcb_limited.py`, `environment/rcb_client.py`). A camera named more than once in `cameras` is rendered and returned once (N copies of one name made the service build N x N images within the frame budget, which could exhaust its memory). A service that has given up on the episode (it could not start it, or could not resume it after its crashes) answers every request with 409 and the reason, and `Robot()` / `wait_ready()` raise `EpisodeOver` at once; before, nothing answered and they waited out their timeout (up to an hour). The adapter, the collect hook, the grader, the scene and the pass rule are unchanged.
- 2026-10-13 (owner, boards 0085 and 0087): the sim service's clock (`environment/sim/rcb_limited.py`, `environment/sim/server.py`, the docstring of `environment/rcb_client.py`). `meta["time_s"]` and `meta["time_left_s"]` count from the agent's first request to the service (any endpoint), and a restarted service keeps that moment; before, they counted from the service's own start of the episode (when its scene was ready). Where the service is ready before the agent's phase begins (the harness is installed in between: a median of 79 s over the recorded standard trials), that clock ran ahead of the agent's and `time_left_s` reached 0 while the agent still had that time. `/spec` says what the clock counts from (`budget.clock`), and the instruction's `## Time limit` tells the agent to plan by `time_left_s`. The service never ended an episode on this clock: the scene, the frozen instance, the pass rule and the grader are unchanged, so the task keeps its name and earlier results stand.
+- 2026-10-14 (board 0113, from board 0086): the instruction now states the whole of AutoBio's test (before, the standard text left out its 6 mm height window): the reference position is computed when the episode starts (turning the rotor moves the slot away from it), and in this instance a released tube rests 5.0 to 5.1 mm from it, so the test holds while the tube is still being lowered (before, this was only in `docs/benchmarks/autobio.md`, Known limits). Text only: the scene, the frozen instance, the reference solution, AutoBio's check and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/autobio-insert-centrifuge-5430-i00-standard/environment/sim/task.json b/tasks/autobio-insert-centrifuge-5430-i00-standard/environment/sim/task.json
index dfe1cc4654..5526aa2bf7 100644
--- a/tasks/autobio-insert-centrifuge-5430-i00-standard/environment/sim/task.json
+++ b/tasks/autobio-insert-centrifuge-5430-i00-standard/environment/sim/task.json
@@ -1,5 +1,5 @@
{
- "task_text": "Load a second 1.5 mL tube into the centrifuge's rotor, into the slot diametrically opposite the tube that is already in it (the rotor has 30 slots). The tube to load stands in the plate on the bench. Success the moment the tube is seated in that slot: within 5 mm of its seated position.",
+ "task_text": "Load a second 1.5 mL tube into the centrifuge's rotor, into the slot diametrically opposite the tube that is already in it (the rotor has 30 slots). The tube to load stands in the plate on the bench. Success the moment the tube is seated in that slot: its body within 5 mm of its seated position there, and its body's height within a fixed 6 mm window (AutoBio's test; standard mode gives no world heights). That seated position is computed when the episode starts: turning the rotor moves the slot away from it, so leave the rotor where it is. In this instance a tube that has been let go comes to rest just outside the 5 mm (5.0 to 5.1 mm away, AutoBio's own expert's tube included), so the test holds while the tube is still being lowered into the slot, held by the gripper; the episode ends in success at that moment.",
"default_cameras": [
"table_cam_front",
"wrist_cam"
diff --git a/tasks/autobio-insert-centrifuge-5430-i00-standard/instruction.md b/tasks/autobio-insert-centrifuge-5430-i00-standard/instruction.md
index 7a2497e3c7..0a82310fc2 100644
--- a/tasks/autobio-insert-centrifuge-5430-i00-standard/instruction.md
+++ b/tasks/autobio-insert-centrifuge-5430-i00-standard/instruction.md
@@ -2,7 +2,7 @@
You control a **UR5e** arm (6 joints) with a Robotiq 2F-85 gripper, fixed to the bench, **as if it were real**: the simulator (AutoBio's biology laboratory on MuJoCo 3.3.0) runs
in a service you cannot open. You get one client and one episode.
-**Task:** Load a second 1.5 mL tube into the centrifuge's rotor, into the slot diametrically opposite the tube that is already in it (the rotor has 30 slots). The tube to load stands in the plate on the bench. Success the moment the tube is seated in that slot: within 5 mm of its seated position.
+**Task:** Load a second 1.5 mL tube into the centrifuge's rotor, into the slot diametrically opposite the tube that is already in it (the rotor has 30 slots). The tube to load stands in the plate on the bench. Success the moment the tube is seated in that slot: its body within 5 mm of its seated position there, and its body's height within a fixed 6 mm window (AutoBio's test; standard mode gives no world heights). That seated position is computed when the episode starts: turning the rotor moves the slot away from it, so leave the rotor where it is. In this instance a tube that has been let go comes to rest just outside the 5 mm (5.0 to 5.1 mm away, AutoBio's own expert's tube included), so the test holds while the tube is still being lowered into the slot, held by the gripper; the episode ends in success at that moment.
On the bench: a centrifuge (an Eppendorf 5430 model) with its lid open and a 30-slot rotor holding one 1.5 mL tube, and a well plate with a second 1.5 mL tube standing in it.
diff --git a/tasks/autobio-insert-centrifuge-5430-i00-standard/task.toml b/tasks/autobio-insert-centrifuge-5430-i00-standard/task.toml
index 40172cf780..8fc696808c 100644
--- a/tasks/autobio-insert-centrifuge-5430-i00-standard/task.toml
+++ b/tasks/autobio-insert-centrifuge-5430-i00-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "autobio" # AutoBio @ 0850421 on MuJoCo 3.3.0 (CPU)
robot = "ur5e-robotiq-2f85"
family = "insert_centrifuge_5430"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "full"
difficulty = "hard" # AutoBio's own level for the task
diff --git a/tasks/autobio-insert-i00-privileged/task.toml b/tasks/autobio-insert-i00-privileged/task.toml
index 9601ceab05..07e622a5f7 100644
--- a/tasks/autobio-insert-i00-privileged/task.toml
+++ b/tasks/autobio-insert-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "autobio" # AutoBio @ 0850421 on MuJoCo 3.3.0 (CPU)
robot = "ur5e-robotiq-2f85"
family = "insert"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
difficulty = "medium" # AutoBio's own level for the task
success_rule = "first"
diff --git a/tasks/autobio-insert-i00-standard/task.toml b/tasks/autobio-insert-i00-standard/task.toml
index 131fb86ef2..05519d6311 100644
--- a/tasks/autobio-insert-i00-standard/task.toml
+++ b/tasks/autobio-insert-i00-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "autobio" # AutoBio @ 0850421 on MuJoCo 3.3.0 (CPU)
robot = "ur5e-robotiq-2f85"
family = "insert"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "full"
difficulty = "medium" # AutoBio's own level for the task
diff --git a/tasks/autobio-pickup-i00-privileged/task.toml b/tasks/autobio-pickup-i00-privileged/task.toml
index 33d69cbf15..ffcddb5f7e 100644
--- a/tasks/autobio-pickup-i00-privileged/task.toml
+++ b/tasks/autobio-pickup-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "autobio" # AutoBio @ 0850421 on MuJoCo 3.3.0 (CPU)
robot = "aloha-2-arm"
family = "pickup"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
difficulty = "easy" # AutoBio's own level for the task
success_rule = "first"
diff --git a/tasks/autobio-pickup-i00-standard/task.toml b/tasks/autobio-pickup-i00-standard/task.toml
index 35b290f78a..158414c60c 100644
--- a/tasks/autobio-pickup-i00-standard/task.toml
+++ b/tasks/autobio-pickup-i00-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "autobio" # AutoBio @ 0850421 on MuJoCo 3.3.0 (CPU)
robot = "aloha-2-arm"
family = "pickup"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "full"
difficulty = "easy" # AutoBio's own level for the task
diff --git a/tasks/autobio-pipette-i00-privileged/task.toml b/tasks/autobio-pipette-i00-privileged/task.toml
index 6f0db26687..8f54c18d04 100644
--- a/tasks/autobio-pipette-i00-privileged/task.toml
+++ b/tasks/autobio-pipette-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "autobio" # AutoBio @ 0850421 on MuJoCo 3.3.0 (CPU)
robot = "ur5e-dexhand-2f85-bimanual"
family = "pipette"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
difficulty = "medium" # AutoBio's own level for the task
success_rule = "first"
diff --git a/tasks/autobio-pipette-i00-standard/task.toml b/tasks/autobio-pipette-i00-standard/task.toml
index d519ffc773..4f3213d124 100644
--- a/tasks/autobio-pipette-i00-standard/task.toml
+++ b/tasks/autobio-pipette-i00-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "autobio" # AutoBio @ 0850421 on MuJoCo 3.3.0 (CPU)
robot = "ur5e-dexhand-2f85-bimanual"
family = "pipette"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "full"
difficulty = "medium" # AutoBio's own level for the task
diff --git a/tasks/autobio-screw-loose-i00-privileged/task.toml b/tasks/autobio-screw-loose-i00-privileged/task.toml
index ca52335c64..a6576850af 100644
--- a/tasks/autobio-screw-loose-i00-privileged/task.toml
+++ b/tasks/autobio-screw-loose-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "autobio" # AutoBio @ 0850421 on MuJoCo 3.3.0 (CPU)
robot = "aloha-2"
family = "screw_loose"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
difficulty = "medium" # AutoBio's own level for the task
success_rule = "first"
diff --git a/tasks/autobio-screw-loose-i00-standard/task.toml b/tasks/autobio-screw-loose-i00-standard/task.toml
index ecfc6d3d05..a503284fd5 100644
--- a/tasks/autobio-screw-loose-i00-standard/task.toml
+++ b/tasks/autobio-screw-loose-i00-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "autobio" # AutoBio @ 0850421 on MuJoCo 3.3.0 (CPU)
robot = "aloha-2"
family = "screw_loose"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "full"
difficulty = "medium" # AutoBio's own level for the task
diff --git a/tasks/autobio-screw-tighten-i00-privileged/task.toml b/tasks/autobio-screw-tighten-i00-privileged/task.toml
index 45109a405d..405b1cac1a 100644
--- a/tasks/autobio-screw-tighten-i00-privileged/task.toml
+++ b/tasks/autobio-screw-tighten-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "autobio" # AutoBio @ 0850421 on MuJoCo 3.3.0 (CPU)
robot = "aloha-2"
family = "screw_tighten"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
difficulty = "hard" # AutoBio's own level for the task
success_rule = "first"
diff --git a/tasks/autobio-screw-tighten-i00-standard/task.toml b/tasks/autobio-screw-tighten-i00-standard/task.toml
index cc0ac654d3..b31f9dceea 100644
--- a/tasks/autobio-screw-tighten-i00-standard/task.toml
+++ b/tasks/autobio-screw-tighten-i00-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "autobio" # AutoBio @ 0850421 on MuJoCo 3.3.0 (CPU)
robot = "aloha-2"
family = "screw_tighten"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "full"
difficulty = "hard" # AutoBio's own level for the task
diff --git a/tasks/autobio-thermal-cycler-close-i00-privileged/task.toml b/tasks/autobio-thermal-cycler-close-i00-privileged/task.toml
index e2381be596..850641e7ad 100644
--- a/tasks/autobio-thermal-cycler-close-i00-privileged/task.toml
+++ b/tasks/autobio-thermal-cycler-close-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "autobio" # AutoBio @ 0850421 on MuJoCo 3.3.0 (CPU)
robot = "ur5e-robotiq-2f85"
family = "thermal_cycler_close"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
difficulty = "easy" # AutoBio's own level for the task
success_rule = "first"
diff --git a/tasks/autobio-thermal-cycler-close-i00-standard/task.toml b/tasks/autobio-thermal-cycler-close-i00-standard/task.toml
index 6c38f21ec5..ee8c9a8d47 100644
--- a/tasks/autobio-thermal-cycler-close-i00-standard/task.toml
+++ b/tasks/autobio-thermal-cycler-close-i00-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "autobio" # AutoBio @ 0850421 on MuJoCo 3.3.0 (CPU)
robot = "ur5e-robotiq-2f85"
family = "thermal_cycler_close"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "full"
difficulty = "easy" # AutoBio's own level for the task
diff --git a/tasks/autobio-thermal-cycler-open-i00-privileged/task.toml b/tasks/autobio-thermal-cycler-open-i00-privileged/task.toml
index 168a8166b8..6fa908d99e 100644
--- a/tasks/autobio-thermal-cycler-open-i00-privileged/task.toml
+++ b/tasks/autobio-thermal-cycler-open-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "autobio" # AutoBio @ 0850421 on MuJoCo 3.3.0 (CPU)
robot = "ur5e-robotiq-2f85"
family = "thermal_cycler_open"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
difficulty = "easy" # AutoBio's own level for the task
success_rule = "first"
diff --git a/tasks/autobio-thermal-cycler-open-i00-standard/task.toml b/tasks/autobio-thermal-cycler-open-i00-standard/task.toml
index 72c4cd3571..9212db88ab 100644
--- a/tasks/autobio-thermal-cycler-open-i00-standard/task.toml
+++ b/tasks/autobio-thermal-cycler-open-i00-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "autobio" # AutoBio @ 0850421 on MuJoCo 3.3.0 (CPU)
robot = "ur5e-robotiq-2f85"
family = "thermal_cycler_open"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "full"
difficulty = "easy" # AutoBio's own level for the task
diff --git a/tasks/autobio-thermal-mixer-i00-privileged/task.toml b/tasks/autobio-thermal-mixer-i00-privileged/task.toml
index 776029afe3..b83cf8d1ce 100644
--- a/tasks/autobio-thermal-mixer-i00-privileged/task.toml
+++ b/tasks/autobio-thermal-mixer-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "autobio" # AutoBio @ 0850421 on MuJoCo 3.3.0 (CPU)
robot = "ur5e-robotiq-2f85"
family = "thermal_mixer"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
difficulty = "hard" # AutoBio's own level for the task
success_rule = "final"
diff --git a/tasks/autobio-thermal-mixer-i00-standard/task.toml b/tasks/autobio-thermal-mixer-i00-standard/task.toml
index 28055ecc90..431e211ecd 100644
--- a/tasks/autobio-thermal-mixer-i00-standard/task.toml
+++ b/tasks/autobio-thermal-mixer-i00-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "autobio" # AutoBio @ 0850421 on MuJoCo 3.3.0 (CPU)
robot = "ur5e-robotiq-2f85"
family = "thermal_mixer"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "full"
difficulty = "hard" # AutoBio's own level for the task
diff --git a/tasks/behavior-assembling-gift-baskets-i00-privileged/task.toml b/tasks/behavior-assembling-gift-baskets-i00-privileged/task.toml
index 99f95c0955..ab3dd5042b 100644
--- a/tasks/behavior-assembling-gift-baskets-i00-privileged/task.toml
+++ b/tasks/behavior-assembling-gift-baskets-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "behavior" # BEHAVIOR-1K v3.9.2 = OmniGibson 3.9.2 on Isa
robot = "r1pro"
family = "assembling_gift_baskets"
mode = "privileged" # privileged state, unlimited resets, thin primitives (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (869 s on average); demonstration video (869 s): https://player.vimeo.com/video/1114059476 (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-04); human review per family (protocol §2.7)."
instance_seed = 301 # scene = environment/instance.json (2026 challenge public_test instance 301 + its state hash); identity = dataset.toml digest
diff --git a/tasks/behavior-assembling-gift-baskets-i00-standard/task.toml b/tasks/behavior-assembling-gift-baskets-i00-standard/task.toml
index d4fb25be86..ab0082eca8 100644
--- a/tasks/behavior-assembling-gift-baskets-i00-standard/task.toml
+++ b/tasks/behavior-assembling-gift-baskets-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "behavior"
robot = "r1pro"
family = "assembling_gift_baskets"
mode = "standard" # robot as a service: RGB-D + proprioception, no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (869 s on average); demonstration video (869 s): https://player.vimeo.com/video/1114059476 (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-04); human review per family (protocol §2.7)."
diff --git a/tasks/behavior-attach-a-camera-to-a-tripod-i00-privileged/task.toml b/tasks/behavior-attach-a-camera-to-a-tripod-i00-privileged/task.toml
index 1ae929fd13..03bcb07659 100644
--- a/tasks/behavior-attach-a-camera-to-a-tripod-i00-privileged/task.toml
+++ b/tasks/behavior-attach-a-camera-to-a-tripod-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "behavior" # BEHAVIOR-1K v3.9.2 = OmniGibson 3.9.2 on Isa
robot = "r1pro"
family = "attach_a_camera_to_a_tripod"
mode = "privileged" # privileged state, unlimited resets, thin primitives (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (130 s on average); demonstration video (130 s): https://player.vimeo.com/video/1114060393 (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-07); human review in PR #75 (protocol §2.7)."
instance_seed = 301 # scene = environment/instance.json (2026 challenge public_test instance 301 + its state hash); identity = dataset.toml digest
diff --git a/tasks/behavior-attach-a-camera-to-a-tripod-i00-standard/task.toml b/tasks/behavior-attach-a-camera-to-a-tripod-i00-standard/task.toml
index c26a0ab264..682cf3f947 100644
--- a/tasks/behavior-attach-a-camera-to-a-tripod-i00-standard/task.toml
+++ b/tasks/behavior-attach-a-camera-to-a-tripod-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "behavior"
robot = "r1pro"
family = "attach_a_camera_to_a_tripod"
mode = "standard" # robot as a service: RGB-D + proprioception, no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (130 s on average); demonstration video (130 s): https://player.vimeo.com/video/1114060393 (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-07); human review in PR #75 (protocol §2.7)."
diff --git a/tasks/behavior-boxing-books-up-for-storage-i00-privileged/task.toml b/tasks/behavior-boxing-books-up-for-storage-i00-privileged/task.toml
index ecd42cb110..31e32a54d3 100644
--- a/tasks/behavior-boxing-books-up-for-storage-i00-privileged/task.toml
+++ b/tasks/behavior-boxing-books-up-for-storage-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "behavior" # BEHAVIOR-1K v3.9.2 = OmniGibson 3.9.2 on Isa
robot = "r1pro"
family = "boxing_books_up_for_storage"
mode = "privileged" # privileged state, unlimited resets, thin primitives (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (808 s on average); demonstration video (808 s): https://player.vimeo.com/video/1114059113 (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-04); human review per family (protocol §2.7)."
instance_seed = 301 # scene = environment/instance.json (2026 challenge public_test instance 301 + its state hash); identity = dataset.toml digest
diff --git a/tasks/behavior-boxing-books-up-for-storage-i00-standard/task.toml b/tasks/behavior-boxing-books-up-for-storage-i00-standard/task.toml
index 73c664f983..0d0686b086 100644
--- a/tasks/behavior-boxing-books-up-for-storage-i00-standard/task.toml
+++ b/tasks/behavior-boxing-books-up-for-storage-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "behavior"
robot = "r1pro"
family = "boxing_books_up_for_storage"
mode = "standard" # robot as a service: RGB-D + proprioception, no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (808 s on average); demonstration video (808 s): https://player.vimeo.com/video/1114059113 (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-04); human review per family (protocol §2.7)."
diff --git a/tasks/behavior-boxing-food-after-dinner-i00-privileged/task.toml b/tasks/behavior-boxing-food-after-dinner-i00-privileged/task.toml
index 6e8f5aaf34..d4389b4d5c 100644
--- a/tasks/behavior-boxing-food-after-dinner-i00-privileged/task.toml
+++ b/tasks/behavior-boxing-food-after-dinner-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "behavior" # BEHAVIOR-1K v3.9.2 = OmniGibson 3.9.2 on Isa
robot = "r1pro"
family = "boxing_food_after_dinner"
mode = "privileged" # privileged state, unlimited resets, thin primitives (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (223 s on average); demonstration video (240 s): https://www.youtube.com/embed/OmpbFiXPwMo (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-07); human review in PR #71 (protocol §2.7)."
instance_seed = 301 # scene = environment/instance.json (2026 challenge public_test instance 301 + its state hash); identity = dataset.toml digest
diff --git a/tasks/behavior-boxing-food-after-dinner-i00-standard/task.toml b/tasks/behavior-boxing-food-after-dinner-i00-standard/task.toml
index 68535cbbd8..38ddd94fcd 100644
--- a/tasks/behavior-boxing-food-after-dinner-i00-standard/task.toml
+++ b/tasks/behavior-boxing-food-after-dinner-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "behavior"
robot = "r1pro"
family = "boxing_food_after_dinner"
mode = "standard" # robot as a service: RGB-D + proprioception, no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (223 s on average); demonstration video (240 s): https://www.youtube.com/embed/OmpbFiXPwMo (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-07); human review in PR #71 (protocol §2.7)."
diff --git a/tasks/behavior-bringing-in-wood-i00-privileged/task.toml b/tasks/behavior-bringing-in-wood-i00-privileged/task.toml
index aa7f68d96d..599ed0f45c 100644
--- a/tasks/behavior-bringing-in-wood-i00-privileged/task.toml
+++ b/tasks/behavior-bringing-in-wood-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "behavior" # BEHAVIOR-1K v3.9.2 = OmniGibson 3.9.2 on Isa
robot = "r1pro"
family = "bringing_in_wood"
mode = "privileged" # privileged state, unlimited resets, thin primitives (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (451 s on average); demonstration video (451 s): https://player.vimeo.com/video/1114058068 (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-04); human review per family (protocol §2.7)."
instance_seed = 301 # scene = environment/instance.json (2026 challenge public_test instance 301 + its state hash); identity = dataset.toml digest
diff --git a/tasks/behavior-bringing-in-wood-i00-standard/task.toml b/tasks/behavior-bringing-in-wood-i00-standard/task.toml
index b7c7b0b9ba..a17daa5057 100644
--- a/tasks/behavior-bringing-in-wood-i00-standard/task.toml
+++ b/tasks/behavior-bringing-in-wood-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "behavior"
robot = "r1pro"
family = "bringing_in_wood"
mode = "standard" # robot as a service: RGB-D + proprioception, no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (451 s on average); demonstration video (451 s): https://player.vimeo.com/video/1114058068 (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-04); human review per family (protocol §2.7)."
diff --git a/tasks/behavior-bringing-paper-to-recycling-i00-privileged/task.toml b/tasks/behavior-bringing-paper-to-recycling-i00-privileged/task.toml
index c37e0d8192..a493c989ee 100644
--- a/tasks/behavior-bringing-paper-to-recycling-i00-privileged/task.toml
+++ b/tasks/behavior-bringing-paper-to-recycling-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "behavior" # BEHAVIOR-1K v3.9.2 = OmniGibson 3.9.2 on Isa
robot = "r1pro"
family = "bringing_paper_to_recycling"
mode = "privileged" # privileged state, unlimited resets, thin primitives (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (374 s on average); demonstration video (318 s): https://www.youtube.com/embed/5OU6Zq9bGG4 (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-04); human review per family (protocol §2.7)."
instance_seed = 301 # scene = environment/instance.json (2026 challenge public_test instance 301 + its state hash); identity = dataset.toml digest
diff --git a/tasks/behavior-bringing-paper-to-recycling-i00-standard/task.toml b/tasks/behavior-bringing-paper-to-recycling-i00-standard/task.toml
index 8ac6e94771..e671850e64 100644
--- a/tasks/behavior-bringing-paper-to-recycling-i00-standard/task.toml
+++ b/tasks/behavior-bringing-paper-to-recycling-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "behavior"
robot = "r1pro"
family = "bringing_paper_to_recycling"
mode = "standard" # robot as a service: RGB-D + proprioception, no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (374 s on average); demonstration video (318 s): https://www.youtube.com/embed/5OU6Zq9bGG4 (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-04); human review per family (protocol §2.7)."
diff --git a/tasks/behavior-bringing-water-i00-privileged/task.toml b/tasks/behavior-bringing-water-i00-privileged/task.toml
index 8197009e3d..ea91136738 100644
--- a/tasks/behavior-bringing-water-i00-privileged/task.toml
+++ b/tasks/behavior-bringing-water-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "behavior" # BEHAVIOR-1K v3.9.2 = OmniGibson 3.9.2 on Isa
robot = "r1pro"
family = "bringing_water"
mode = "privileged" # privileged state, unlimited resets, thin primitives (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (315 s on average); demonstration video (315 s): https://player.vimeo.com/video/1114058418 (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-07); human review in PR #71 (protocol §2.7)."
instance_seed = 301 # scene = environment/instance.json (2026 challenge public_test instance 301 + its state hash); identity = dataset.toml digest
diff --git a/tasks/behavior-bringing-water-i00-standard/task.toml b/tasks/behavior-bringing-water-i00-standard/task.toml
index c82a0bab1f..f9d522c597 100644
--- a/tasks/behavior-bringing-water-i00-standard/task.toml
+++ b/tasks/behavior-bringing-water-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "behavior"
robot = "r1pro"
family = "bringing_water"
mode = "standard" # robot as a service: RGB-D + proprioception, no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (315 s on average); demonstration video (315 s): https://player.vimeo.com/video/1114058418 (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-07); human review in PR #71 (protocol §2.7)."
diff --git a/tasks/behavior-can-meat-i00-privileged/task.toml b/tasks/behavior-can-meat-i00-privileged/task.toml
index 1c3db01316..5e784d2b28 100644
--- a/tasks/behavior-can-meat-i00-privileged/task.toml
+++ b/tasks/behavior-can-meat-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "behavior" # BEHAVIOR-1K v3.9.2 = OmniGibson 3.9.2 on Isa
robot = "r1pro"
family = "can_meat"
mode = "privileged" # privileged state, unlimited resets, thin primitives (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (395 s on average); demonstration video (395 s): https://player.vimeo.com/video/1114054618 (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-04); human review per family (protocol §2.7)."
instance_seed = 301 # scene = environment/instance.json (2026 challenge public_test instance 301 + its state hash); identity = dataset.toml digest
diff --git a/tasks/behavior-can-meat-i00-standard/task.toml b/tasks/behavior-can-meat-i00-standard/task.toml
index 13eecd347b..f0c5fc11cc 100644
--- a/tasks/behavior-can-meat-i00-standard/task.toml
+++ b/tasks/behavior-can-meat-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "behavior"
robot = "r1pro"
family = "can_meat"
mode = "standard" # robot as a service: RGB-D + proprioception, no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (395 s on average); demonstration video (395 s): https://player.vimeo.com/video/1114054618 (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-04); human review per family (protocol §2.7)."
diff --git a/tasks/behavior-canning-food-i00-privileged/task.toml b/tasks/behavior-canning-food-i00-privileged/task.toml
index 4549b1f95c..f8860c600a 100644
--- a/tasks/behavior-canning-food-i00-privileged/task.toml
+++ b/tasks/behavior-canning-food-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "behavior" # BEHAVIOR-1K v3.9.2 = OmniGibson 3.9.2 on Isa
robot = "r1pro"
family = "canning_food"
mode = "privileged" # privileged state, unlimited resets, thin primitives (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (766 s on average); demonstration video (766 s): https://player.vimeo.com/video/1114061821 (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-04); human review per family (protocol §2.7)."
instance_seed = 301 # scene = environment/instance.json (2026 challenge public_test instance 301 + its state hash); identity = dataset.toml digest
diff --git a/tasks/behavior-canning-food-i00-standard/task.toml b/tasks/behavior-canning-food-i00-standard/task.toml
index 3d98f92800..ea32d772e2 100644
--- a/tasks/behavior-canning-food-i00-standard/task.toml
+++ b/tasks/behavior-canning-food-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "behavior"
robot = "r1pro"
family = "canning_food"
mode = "standard" # robot as a service: RGB-D + proprioception, no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (766 s on average); demonstration video (766 s): https://player.vimeo.com/video/1114061821 (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-04); human review per family (protocol §2.7)."
diff --git a/tasks/behavior-carrying-in-groceries-i00-privileged/task.toml b/tasks/behavior-carrying-in-groceries-i00-privileged/task.toml
index 81afa08d98..8cdfd9202b 100644
--- a/tasks/behavior-carrying-in-groceries-i00-privileged/task.toml
+++ b/tasks/behavior-carrying-in-groceries-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "behavior" # BEHAVIOR-1K v3.9.2 = OmniGibson 3.9.2 on Isa
robot = "r1pro"
family = "carrying_in_groceries"
mode = "privileged" # privileged state, unlimited resets, thin primitives (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (476 s on average); demonstration video (476 s): https://player.vimeo.com/video/1114057919 (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-04); human review per family (protocol §2.7)."
instance_seed = 301 # scene = environment/instance.json (2026 challenge public_test instance 301 + its state hash); identity = dataset.toml digest
diff --git a/tasks/behavior-carrying-in-groceries-i00-standard/task.toml b/tasks/behavior-carrying-in-groceries-i00-standard/task.toml
index beb8b708b3..fe73d89afa 100644
--- a/tasks/behavior-carrying-in-groceries-i00-standard/task.toml
+++ b/tasks/behavior-carrying-in-groceries-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "behavior"
robot = "r1pro"
family = "carrying_in_groceries"
mode = "standard" # robot as a service: RGB-D + proprioception, no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (476 s on average); demonstration video (476 s): https://player.vimeo.com/video/1114057919 (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-04); human review per family (protocol §2.7)."
diff --git a/tasks/behavior-carrying-out-garden-furniture-i00-privileged/task.toml b/tasks/behavior-carrying-out-garden-furniture-i00-privileged/task.toml
index 08e75ad904..96b9024a7c 100644
--- a/tasks/behavior-carrying-out-garden-furniture-i00-privileged/task.toml
+++ b/tasks/behavior-carrying-out-garden-furniture-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "behavior" # BEHAVIOR-1K v3.9.2 = OmniGibson 3.9.2 on Isa
robot = "r1pro"
family = "carrying_out_garden_furniture"
mode = "privileged" # privileged state, unlimited resets, thin primitives (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (369 s on average); demonstration video (358 s): https://www.youtube.com/embed/NCQc0IIRmos (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-04); human review per family (protocol §2.7)."
instance_seed = 301 # scene = environment/instance.json (2026 challenge public_test instance 301 + its state hash); identity = dataset.toml digest
diff --git a/tasks/behavior-carrying-out-garden-furniture-i00-standard/task.toml b/tasks/behavior-carrying-out-garden-furniture-i00-standard/task.toml
index ee8717dfe1..af29280df9 100644
--- a/tasks/behavior-carrying-out-garden-furniture-i00-standard/task.toml
+++ b/tasks/behavior-carrying-out-garden-furniture-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "behavior"
robot = "r1pro"
family = "carrying_out_garden_furniture"
mode = "standard" # robot as a service: RGB-D + proprioception, no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (369 s on average); demonstration video (358 s): https://www.youtube.com/embed/NCQc0IIRmos (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-04); human review per family (protocol §2.7)."
diff --git a/tasks/behavior-chop-an-onion-i00-privileged/task.toml b/tasks/behavior-chop-an-onion-i00-privileged/task.toml
index e0122a298f..6517080896 100644
--- a/tasks/behavior-chop-an-onion-i00-privileged/task.toml
+++ b/tasks/behavior-chop-an-onion-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "behavior" # BEHAVIOR-1K v3.9.2 = OmniGibson 3.9.2 on Isa
robot = "r1pro"
family = "chop_an_onion"
mode = "privileged" # privileged state, unlimited resets, thin primitives (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (213 s on average); demonstration video (213 s): https://player.vimeo.com/video/1114061062 (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-07); human review in PR #75 (protocol §2.7)."
instance_seed = 301 # scene = environment/instance.json (2026 challenge public_test instance 301 + its state hash); identity = dataset.toml digest
diff --git a/tasks/behavior-chop-an-onion-i00-standard/task.toml b/tasks/behavior-chop-an-onion-i00-standard/task.toml
index df495432d1..ad03310098 100644
--- a/tasks/behavior-chop-an-onion-i00-standard/task.toml
+++ b/tasks/behavior-chop-an-onion-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "behavior"
robot = "r1pro"
family = "chop_an_onion"
mode = "standard" # robot as a service: RGB-D + proprioception, no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (213 s on average); demonstration video (213 s): https://player.vimeo.com/video/1114061062 (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-07); human review in PR #75 (protocol §2.7)."
diff --git a/tasks/behavior-chopping-wood-i00-privileged/task.toml b/tasks/behavior-chopping-wood-i00-privileged/task.toml
index ff33f92be8..5a9a0897fa 100644
--- a/tasks/behavior-chopping-wood-i00-privileged/task.toml
+++ b/tasks/behavior-chopping-wood-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "behavior" # BEHAVIOR-1K v3.9.2 = OmniGibson 3.9.2 on Isa
robot = "r1pro"
family = "chopping_wood"
mode = "privileged" # privileged state, unlimited resets, thin primitives (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (358 s on average); demonstration video (358 s): https://player.vimeo.com/video/1114061262 (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-04); human review per family (protocol §2.7)."
instance_seed = 301 # scene = environment/instance.json (2026 challenge public_test instance 301 + its state hash); identity = dataset.toml digest
diff --git a/tasks/behavior-chopping-wood-i00-standard/task.toml b/tasks/behavior-chopping-wood-i00-standard/task.toml
index 8310bfb6c9..3b2260e2b1 100644
--- a/tasks/behavior-chopping-wood-i00-standard/task.toml
+++ b/tasks/behavior-chopping-wood-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "behavior"
robot = "r1pro"
family = "chopping_wood"
mode = "standard" # robot as a service: RGB-D + proprioception, no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (358 s on average); demonstration video (358 s): https://player.vimeo.com/video/1114061262 (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-04); human review per family (protocol §2.7)."
diff --git a/tasks/behavior-clean-a-keyboard-i00-privileged/task.toml b/tasks/behavior-clean-a-keyboard-i00-privileged/task.toml
index d8be214e1c..67e3843bea 100644
--- a/tasks/behavior-clean-a-keyboard-i00-privileged/task.toml
+++ b/tasks/behavior-clean-a-keyboard-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "behavior" # BEHAVIOR-1K v3.9.2 = OmniGibson 3.9.2 on Isa
robot = "r1pro"
family = "clean_a_keyboard"
mode = "privileged" # privileged state, unlimited resets, thin primitives (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (129 s on average); demonstration video (142 s): https://www.youtube.com/embed/yEV3cFqyu-4 (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-07); human review in PR #75 (protocol §2.7)."
instance_seed = 301 # scene = environment/instance.json (2026 challenge public_test instance 301 + its state hash); identity = dataset.toml digest
diff --git a/tasks/behavior-clean-a-keyboard-i00-standard/task.toml b/tasks/behavior-clean-a-keyboard-i00-standard/task.toml
index c0d0567a79..b64b9a37e7 100644
--- a/tasks/behavior-clean-a-keyboard-i00-standard/task.toml
+++ b/tasks/behavior-clean-a-keyboard-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "behavior"
robot = "r1pro"
family = "clean_a_keyboard"
mode = "standard" # robot as a service: RGB-D + proprioception, no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (129 s on average); demonstration video (142 s): https://www.youtube.com/embed/yEV3cFqyu-4 (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-07); human review in PR #75 (protocol §2.7)."
diff --git a/tasks/behavior-clean-a-patio-i00-privileged/task.toml b/tasks/behavior-clean-a-patio-i00-privileged/task.toml
index ca802d7e55..19b2a74f1e 100644
--- a/tasks/behavior-clean-a-patio-i00-privileged/task.toml
+++ b/tasks/behavior-clean-a-patio-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "behavior" # BEHAVIOR-1K v3.9.2 = OmniGibson 3.9.2 on Isa
robot = "r1pro"
family = "clean_a_patio"
mode = "privileged" # privileged state, unlimited resets, thin primitives (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (402 s on average); demonstration video (402 s): https://player.vimeo.com/video/1114060492 (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-04); human review per family (protocol §2.7)."
instance_seed = 301 # scene = environment/instance.json (2026 challenge public_test instance 301 + its state hash); identity = dataset.toml digest
diff --git a/tasks/behavior-clean-a-patio-i00-standard/task.toml b/tasks/behavior-clean-a-patio-i00-standard/task.toml
index 70c30d1ebf..c1d08680c8 100644
--- a/tasks/behavior-clean-a-patio-i00-standard/task.toml
+++ b/tasks/behavior-clean-a-patio-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "behavior"
robot = "r1pro"
family = "clean_a_patio"
mode = "standard" # robot as a service: RGB-D + proprioception, no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (402 s on average); demonstration video (402 s): https://player.vimeo.com/video/1114060492 (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-04); human review per family (protocol §2.7)."
diff --git a/tasks/behavior-clean-a-trumpet-i00-privileged/task.toml b/tasks/behavior-clean-a-trumpet-i00-privileged/task.toml
index 7639de2831..93d9d63f84 100644
--- a/tasks/behavior-clean-a-trumpet-i00-privileged/task.toml
+++ b/tasks/behavior-clean-a-trumpet-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "behavior" # BEHAVIOR-1K v3.9.2 = OmniGibson 3.9.2 on Isa
robot = "r1pro"
family = "clean_a_trumpet"
mode = "privileged" # privileged state, unlimited resets, thin primitives (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (177 s on average); demonstration video (177 s): https://player.vimeo.com/video/1114060606 (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-07); human review in PR #75 (protocol §2.7)."
instance_seed = 301 # scene = environment/instance.json (2026 challenge public_test instance 301 + its state hash); identity = dataset.toml digest
diff --git a/tasks/behavior-clean-a-trumpet-i00-standard/task.toml b/tasks/behavior-clean-a-trumpet-i00-standard/task.toml
index 6011ac365b..9233178a32 100644
--- a/tasks/behavior-clean-a-trumpet-i00-standard/task.toml
+++ b/tasks/behavior-clean-a-trumpet-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "behavior"
robot = "r1pro"
family = "clean_a_trumpet"
mode = "standard" # robot as a service: RGB-D + proprioception, no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (177 s on average); demonstration video (177 s): https://player.vimeo.com/video/1114060606 (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-07); human review in PR #75 (protocol §2.7)."
diff --git a/tasks/behavior-clean-boxing-gloves-i00-privileged/task.toml b/tasks/behavior-clean-boxing-gloves-i00-privileged/task.toml
index d48551cad2..53e4fb34a2 100644
--- a/tasks/behavior-clean-boxing-gloves-i00-privileged/task.toml
+++ b/tasks/behavior-clean-boxing-gloves-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "behavior" # BEHAVIOR-1K v3.9.2 = OmniGibson 3.9.2 on Isa
robot = "r1pro"
family = "clean_boxing_gloves"
mode = "privileged" # privileged state, unlimited resets, thin primitives (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (275 s on average); demonstration video (275 s): https://player.vimeo.com/video/1114060008 (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-07); human review in PR #75 (protocol §2.7)."
instance_seed = 301 # scene = environment/instance.json (2026 challenge public_test instance 301 + its state hash); identity = dataset.toml digest
diff --git a/tasks/behavior-clean-boxing-gloves-i00-standard/task.toml b/tasks/behavior-clean-boxing-gloves-i00-standard/task.toml
index 508fca4d0c..7192b64014 100644
--- a/tasks/behavior-clean-boxing-gloves-i00-standard/task.toml
+++ b/tasks/behavior-clean-boxing-gloves-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "behavior"
robot = "r1pro"
family = "clean_boxing_gloves"
mode = "standard" # robot as a service: RGB-D + proprioception, no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (275 s on average); demonstration video (275 s): https://player.vimeo.com/video/1114060008 (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-07); human review in PR #75 (protocol §2.7)."
diff --git a/tasks/behavior-clean-up-broken-glass-i00-privileged/task.toml b/tasks/behavior-clean-up-broken-glass-i00-privileged/task.toml
index f37d1fc3b9..b0135fcab0 100644
--- a/tasks/behavior-clean-up-broken-glass-i00-privileged/task.toml
+++ b/tasks/behavior-clean-up-broken-glass-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "behavior" # BEHAVIOR-1K v3.9.2 = OmniGibson 3.9.2 on Isa
robot = "r1pro"
family = "clean_up_broken_glass"
mode = "privileged" # privileged state, unlimited resets, thin primitives (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance. Solvable: model trial behavior-clean-up-broken-glass-i__Pw85YoX (behavior-lunaxh-t8-0927, privileged) graded success=1, deterministic=1 by the separate verifier on a pre-v1.0 version of this task (dev/qineng 912567fa; grader verdict and frozen instance unchanged since); agent log https://embodied-agent-interface-v2-internal.github.io/runs/behavior-1k/codex-gpt6_luna-xhigh/clean_up_broken_glass/unlimited/, verifier replay https://huggingface.co/datasets/eai-v2-internal/agent-runs/resolve/main/behavior-1k/codex-gpt6_luna-xhigh/clean_up_broken_glass/unlimited/replay.mp4; not re-run on this version (static validation only). Human review per family (protocol §2.7)."
instance_seed = 301 # scene = environment/instance.json (2026 challenge public_test instance 301 + its state hash); identity = dataset.toml digest
diff --git a/tasks/behavior-clean-up-broken-glass-i00-standard/task.toml b/tasks/behavior-clean-up-broken-glass-i00-standard/task.toml
index 829a3f6bfb..47a8c7f5a2 100644
--- a/tasks/behavior-clean-up-broken-glass-i00-standard/task.toml
+++ b/tasks/behavior-clean-up-broken-glass-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "behavior"
robot = "r1pro"
family = "clean_up_broken_glass"
mode = "standard" # robot as a service: RGB-D + proprioception, no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance. Solvable: model trial behavior-clean-up-broken-glass-i__Pw85YoX (behavior-lunaxh-t8-0927, the privileged twin behavior-clean-up-broken-glass-i00-privileged on the same frozen instance and BDDL goal) graded success=1, deterministic=1 by the separate verifier on a pre-v1.0 version of this task (dev/qineng 912567fa; grader verdict and frozen instance unchanged since); agent log https://embodied-agent-interface-v2-internal.github.io/runs/behavior-1k/codex-gpt6_luna-xhigh/clean_up_broken_glass/unlimited/, verifier replay https://huggingface.co/datasets/eai-v2-internal/agent-runs/resolve/main/behavior-1k/codex-gpt6_luna-xhigh/clean_up_broken_glass/unlimited/replay.mp4; not re-run on this version (static validation only). Human review per family (protocol §2.7)."
diff --git a/tasks/behavior-clean-up-your-desk-i00-privileged/task.toml b/tasks/behavior-clean-up-your-desk-i00-privileged/task.toml
index 494a728cc1..709bda6b4f 100644
--- a/tasks/behavior-clean-up-your-desk-i00-privileged/task.toml
+++ b/tasks/behavior-clean-up-your-desk-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "behavior" # BEHAVIOR-1K v3.9.2 = OmniGibson 3.9.2 on Isa
robot = "r1pro"
family = "clean_up_your_desk"
mode = "privileged" # privileged state, unlimited resets, thin primitives (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (714 s on average); demonstration video (714 s): https://player.vimeo.com/video/1114059839 (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-04); human review per family (protocol §2.7)."
instance_seed = 301 # scene = environment/instance.json (2026 challenge public_test instance 301 + its state hash); identity = dataset.toml digest
diff --git a/tasks/behavior-clean-up-your-desk-i00-standard/task.toml b/tasks/behavior-clean-up-your-desk-i00-standard/task.toml
index eef6c2f461..7760b4b251 100644
--- a/tasks/behavior-clean-up-your-desk-i00-standard/task.toml
+++ b/tasks/behavior-clean-up-your-desk-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "behavior"
robot = "r1pro"
family = "clean_up_your_desk"
mode = "standard" # robot as a service: RGB-D + proprioception, no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (714 s on average); demonstration video (714 s): https://player.vimeo.com/video/1114059839 (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-04); human review per family (protocol §2.7)."
diff --git a/tasks/behavior-clean-your-rusty-garden-tools-i00-privileged/task.toml b/tasks/behavior-clean-your-rusty-garden-tools-i00-privileged/task.toml
index cb515fddc3..3722dd807e 100644
--- a/tasks/behavior-clean-your-rusty-garden-tools-i00-privileged/task.toml
+++ b/tasks/behavior-clean-your-rusty-garden-tools-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "behavior" # BEHAVIOR-1K v3.9.2 = OmniGibson 3.9.2 on Isa
robot = "r1pro"
family = "clean_your_rusty_garden_tools"
mode = "privileged" # privileged state, unlimited resets, thin primitives (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (505 s on average); demonstration video (484 s): https://www.youtube.com/embed/tHiesuHDgkc (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-04); human review per family (protocol §2.7)."
instance_seed = 301 # scene = environment/instance.json (2026 challenge public_test instance 301 + its state hash); identity = dataset.toml digest
diff --git a/tasks/behavior-clean-your-rusty-garden-tools-i00-standard/task.toml b/tasks/behavior-clean-your-rusty-garden-tools-i00-standard/task.toml
index e49bf7d1d3..6e47eb3590 100644
--- a/tasks/behavior-clean-your-rusty-garden-tools-i00-standard/task.toml
+++ b/tasks/behavior-clean-your-rusty-garden-tools-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "behavior"
robot = "r1pro"
family = "clean_your_rusty_garden_tools"
mode = "standard" # robot as a service: RGB-D + proprioception, no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (505 s on average); demonstration video (484 s): https://www.youtube.com/embed/tHiesuHDgkc (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-04); human review per family (protocol §2.7)."
diff --git a/tasks/behavior-cleaning-up-branches-and-twigs-i00-privileged/task.toml b/tasks/behavior-cleaning-up-branches-and-twigs-i00-privileged/task.toml
index 02b5b94d71..31c53334ee 100644
--- a/tasks/behavior-cleaning-up-branches-and-twigs-i00-privileged/task.toml
+++ b/tasks/behavior-cleaning-up-branches-and-twigs-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "behavior" # BEHAVIOR-1K v3.9.2 = OmniGibson 3.9.2 on Isa
robot = "r1pro"
family = "cleaning_up_branches_and_twigs"
mode = "privileged" # privileged state, unlimited resets, thin primitives (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (418 s on average); demonstration video (434 s): https://www.youtube.com/embed/iztsNYYMpv0 (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-04); human review per family (protocol §2.7)."
instance_seed = 301 # scene = environment/instance.json (2026 challenge public_test instance 301 + its state hash); identity = dataset.toml digest
diff --git a/tasks/behavior-cleaning-up-branches-and-twigs-i00-standard/task.toml b/tasks/behavior-cleaning-up-branches-and-twigs-i00-standard/task.toml
index 6a28c4ec81..41a7e65d34 100644
--- a/tasks/behavior-cleaning-up-branches-and-twigs-i00-standard/task.toml
+++ b/tasks/behavior-cleaning-up-branches-and-twigs-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "behavior"
robot = "r1pro"
family = "cleaning_up_branches_and_twigs"
mode = "standard" # robot as a service: RGB-D + proprioception, no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (418 s on average); demonstration video (434 s): https://www.youtube.com/embed/iztsNYYMpv0 (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-04); human review per family (protocol §2.7)."
diff --git a/tasks/behavior-cleaning-up-plates-and-food-i00-privileged/task.toml b/tasks/behavior-cleaning-up-plates-and-food-i00-privileged/task.toml
index 678c298580..f8c3b72c29 100644
--- a/tasks/behavior-cleaning-up-plates-and-food-i00-privileged/task.toml
+++ b/tasks/behavior-cleaning-up-plates-and-food-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "behavior" # BEHAVIOR-1K v3.9.2 = OmniGibson 3.9.2 on Isa
robot = "r1pro"
family = "cleaning_up_plates_and_food"
mode = "privileged" # privileged state, unlimited resets, thin primitives (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (457 s on average); demonstration video (457 s): https://player.vimeo.com/video/1114054554 (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-04); human review per family (protocol §2.7)."
instance_seed = 301 # scene = environment/instance.json (2026 challenge public_test instance 301 + its state hash); identity = dataset.toml digest
diff --git a/tasks/behavior-cleaning-up-plates-and-food-i00-standard/task.toml b/tasks/behavior-cleaning-up-plates-and-food-i00-standard/task.toml
index 171139c707..5d03f49f8a 100644
--- a/tasks/behavior-cleaning-up-plates-and-food-i00-standard/task.toml
+++ b/tasks/behavior-cleaning-up-plates-and-food-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "behavior"
robot = "r1pro"
family = "cleaning_up_plates_and_food"
mode = "standard" # robot as a service: RGB-D + proprioception, no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (457 s on average); demonstration video (457 s): https://player.vimeo.com/video/1114054554 (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-04); human review per family (protocol §2.7)."
diff --git a/tasks/behavior-clearing-food-from-table-into-fridge-i00-privileged/task.toml b/tasks/behavior-clearing-food-from-table-into-fridge-i00-privileged/task.toml
index ac7678c4e5..984355814c 100644
--- a/tasks/behavior-clearing-food-from-table-into-fridge-i00-privileged/task.toml
+++ b/tasks/behavior-clearing-food-from-table-into-fridge-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "behavior" # BEHAVIOR-1K v3.9.2 = OmniGibson 3.9.2 on Isa
robot = "r1pro"
family = "clearing_food_from_table_into_fridge"
mode = "privileged" # privileged state, unlimited resets, thin primitives (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (436 s on average); demonstration video (436 s): https://player.vimeo.com/video/1114059394 (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-04); human review per family (protocol §2.7)."
instance_seed = 301 # scene = environment/instance.json (2026 challenge public_test instance 301 + its state hash); identity = dataset.toml digest
diff --git a/tasks/behavior-clearing-food-from-table-into-fridge-i00-standard/task.toml b/tasks/behavior-clearing-food-from-table-into-fridge-i00-standard/task.toml
index 7910fae561..d5cf6690a0 100644
--- a/tasks/behavior-clearing-food-from-table-into-fridge-i00-standard/task.toml
+++ b/tasks/behavior-clearing-food-from-table-into-fridge-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "behavior"
robot = "r1pro"
family = "clearing_food_from_table_into_fridge"
mode = "standard" # robot as a service: RGB-D + proprioception, no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (436 s on average); demonstration video (436 s): https://player.vimeo.com/video/1114059394 (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-04); human review per family (protocol §2.7)."
diff --git a/tasks/behavior-collecting-aluminum-cans-i00-privileged/task.toml b/tasks/behavior-collecting-aluminum-cans-i00-privileged/task.toml
index ad8d260b62..56887d604c 100644
--- a/tasks/behavior-collecting-aluminum-cans-i00-privileged/task.toml
+++ b/tasks/behavior-collecting-aluminum-cans-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "behavior" # BEHAVIOR-1K v3.9.2 = OmniGibson 3.9.2 on Isa
robot = "r1pro"
family = "collecting_aluminum_cans"
mode = "privileged" # privileged state, unlimited resets, thin primitives (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance. Solvable: model trial behavior-collecting-aluminum-can__VRvxwLv (behavior-lunaxh-t8-0927, privileged) graded success=1, deterministic=1 by the separate verifier on a pre-v1.0 version of this task (dev/qineng 20a382a8; grader verdict and frozen instance unchanged since); agent log https://embodied-agent-interface-v2-internal.github.io/runs/behavior-1k/codex-gpt6_luna-xhigh/collecting_aluminum_cans/unlimited/, verifier replay https://huggingface.co/datasets/eai-v2-internal/agent-runs/resolve/main/behavior-1k/codex-gpt6_luna-xhigh/collecting_aluminum_cans/unlimited/replay.mp4; not re-run on this version (static validation only). Human review per family (protocol §2.7)."
instance_seed = 301 # scene = environment/instance.json (2026 challenge public_test instance 301 + its state hash); identity = dataset.toml digest
diff --git a/tasks/behavior-collecting-aluminum-cans-i00-standard/task.toml b/tasks/behavior-collecting-aluminum-cans-i00-standard/task.toml
index 6b030d1663..08ca195dde 100644
--- a/tasks/behavior-collecting-aluminum-cans-i00-standard/task.toml
+++ b/tasks/behavior-collecting-aluminum-cans-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "behavior"
robot = "r1pro"
family = "collecting_aluminum_cans"
mode = "standard" # robot as a service: RGB-D + proprioception, no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance. Solvable: model trial behavior-collecting-aluminum-can__VRvxwLv (behavior-lunaxh-t8-0927, the privileged twin behavior-collecting-aluminum-cans-i00-privileged on the same frozen instance and BDDL goal) graded success=1, deterministic=1 by the separate verifier on a pre-v1.0 version of this task (dev/qineng 20a382a8; grader verdict and frozen instance unchanged since); agent log https://embodied-agent-interface-v2-internal.github.io/runs/behavior-1k/codex-gpt6_luna-xhigh/collecting_aluminum_cans/unlimited/, verifier replay https://huggingface.co/datasets/eai-v2-internal/agent-runs/resolve/main/behavior-1k/codex-gpt6_luna-xhigh/collecting_aluminum_cans/unlimited/replay.mp4; not re-run on this version (static validation only). Human review per family (protocol §2.7)."
diff --git a/tasks/behavior-collecting-childrens-toys-i00-privileged/task.toml b/tasks/behavior-collecting-childrens-toys-i00-privileged/task.toml
index d70b8b045a..288ff5a87b 100644
--- a/tasks/behavior-collecting-childrens-toys-i00-privileged/task.toml
+++ b/tasks/behavior-collecting-childrens-toys-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "behavior" # BEHAVIOR-1K v3.9.2 = OmniGibson 3.9.2 on Isa
robot = "r1pro"
family = "collecting_childrens_toys"
mode = "privileged" # privileged state, unlimited resets, thin primitives (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (640 s on average); demonstration video (640 s): https://player.vimeo.com/video/1114058872 (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-04); human review per family (protocol §2.7)."
instance_seed = 301 # scene = environment/instance.json (2026 challenge public_test instance 301 + its state hash); identity = dataset.toml digest
diff --git a/tasks/behavior-collecting-childrens-toys-i00-standard/task.toml b/tasks/behavior-collecting-childrens-toys-i00-standard/task.toml
index e327c67988..442a83db35 100644
--- a/tasks/behavior-collecting-childrens-toys-i00-standard/task.toml
+++ b/tasks/behavior-collecting-childrens-toys-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "behavior"
robot = "r1pro"
family = "collecting_childrens_toys"
mode = "standard" # robot as a service: RGB-D + proprioception, no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (640 s on average); demonstration video (640 s): https://player.vimeo.com/video/1114058872 (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-04); human review per family (protocol §2.7)."
diff --git a/tasks/behavior-composting-waste-i00-privileged/task.toml b/tasks/behavior-composting-waste-i00-privileged/task.toml
index 20ab8eb54e..4dc68eb9ed 100644
--- a/tasks/behavior-composting-waste-i00-privileged/task.toml
+++ b/tasks/behavior-composting-waste-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "behavior" # BEHAVIOR-1K v3.9.2 = OmniGibson 3.9.2 on Isa
robot = "r1pro"
family = "composting_waste"
mode = "privileged" # privileged state, unlimited resets, thin primitives (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance. Solvable: model trial behavior-composting-waste-i00__qFjjebW (behavior-lunaxh-t8-0927, privileged) graded success=1, deterministic=1 by the separate verifier on a pre-v1.0 version of this task (dev/qineng 20a382a8; grader verdict and frozen instance unchanged since); agent log https://embodied-agent-interface-v2-internal.github.io/runs/behavior-1k/codex-gpt6_luna-xhigh/composting_waste/unlimited/, verifier replay https://huggingface.co/datasets/eai-v2-internal/agent-runs/resolve/main/behavior-1k/codex-gpt6_luna-xhigh/composting_waste/unlimited/replay.mp4; not re-run on this version (static validation only). Human review per family (protocol §2.7)."
instance_seed = 301 # scene = environment/instance.json (2026 challenge public_test instance 301 + its state hash); identity = dataset.toml digest
diff --git a/tasks/behavior-composting-waste-i00-standard/task.toml b/tasks/behavior-composting-waste-i00-standard/task.toml
index 6372ffe864..f164f9f126 100644
--- a/tasks/behavior-composting-waste-i00-standard/task.toml
+++ b/tasks/behavior-composting-waste-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "behavior"
robot = "r1pro"
family = "composting_waste"
mode = "standard" # robot as a service: RGB-D + proprioception, no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance. Solvable: model trial behavior-composting-waste-i00__qFjjebW (behavior-lunaxh-t8-0927, the privileged twin behavior-composting-waste-i00-privileged on the same frozen instance and BDDL goal) graded success=1, deterministic=1 by the separate verifier on a pre-v1.0 version of this task (dev/qineng 20a382a8; grader verdict and frozen instance unchanged since); agent log https://embodied-agent-interface-v2-internal.github.io/runs/behavior-1k/codex-gpt6_luna-xhigh/composting_waste/unlimited/, verifier replay https://huggingface.co/datasets/eai-v2-internal/agent-runs/resolve/main/behavior-1k/codex-gpt6_luna-xhigh/composting_waste/unlimited/replay.mp4; not re-run on this version (static validation only). Human review per family (protocol §2.7)."
diff --git a/tasks/behavior-cook-a-brisket-i00-privileged/task.toml b/tasks/behavior-cook-a-brisket-i00-privileged/task.toml
index 16baef4502..d161f798e5 100644
--- a/tasks/behavior-cook-a-brisket-i00-privileged/task.toml
+++ b/tasks/behavior-cook-a-brisket-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "behavior" # BEHAVIOR-1K v3.9.2 = OmniGibson 3.9.2 on Isa
robot = "r1pro"
family = "cook_a_brisket"
mode = "privileged" # privileged state, unlimited resets, thin primitives (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (244 s on average); demonstration video (251 s): https://www.youtube.com/embed/RCSy8B0_s5Y (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-07); human review in PR #71 (protocol §2.7)."
instance_seed = 301 # scene = environment/instance.json (2026 challenge public_test instance 301 + its state hash); identity = dataset.toml digest
diff --git a/tasks/behavior-cook-a-brisket-i00-standard/task.toml b/tasks/behavior-cook-a-brisket-i00-standard/task.toml
index bb783df03e..c56387aa61 100644
--- a/tasks/behavior-cook-a-brisket-i00-standard/task.toml
+++ b/tasks/behavior-cook-a-brisket-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "behavior"
robot = "r1pro"
family = "cook_a_brisket"
mode = "standard" # robot as a service: RGB-D + proprioception, no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (244 s on average); demonstration video (251 s): https://www.youtube.com/embed/RCSy8B0_s5Y (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-07); human review in PR #71 (protocol §2.7)."
diff --git a/tasks/behavior-cook-a-frozen-pie-i00-privileged/task.toml b/tasks/behavior-cook-a-frozen-pie-i00-privileged/task.toml
index 07ec37570e..7f4b16d3b5 100644
--- a/tasks/behavior-cook-a-frozen-pie-i00-privileged/task.toml
+++ b/tasks/behavior-cook-a-frozen-pie-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "behavior" # BEHAVIOR-1K v3.9.2 = OmniGibson 3.9.2 on Isa
robot = "r1pro"
family = "cook_a_frozen_pie"
mode = "privileged" # privileged state, unlimited resets, thin primitives (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (289 s on average); demonstration video (323 s): https://www.youtube.com/embed/H4z3xTJn058 (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-07); human review in PR #75 (protocol §2.7)."
instance_seed = 301 # scene = environment/instance.json (2026 challenge public_test instance 301 + its state hash); identity = dataset.toml digest
diff --git a/tasks/behavior-cook-a-frozen-pie-i00-standard/task.toml b/tasks/behavior-cook-a-frozen-pie-i00-standard/task.toml
index 4eed8893c4..1def342570 100644
--- a/tasks/behavior-cook-a-frozen-pie-i00-standard/task.toml
+++ b/tasks/behavior-cook-a-frozen-pie-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "behavior"
robot = "r1pro"
family = "cook_a_frozen_pie"
mode = "standard" # robot as a service: RGB-D + proprioception, no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (289 s on average); demonstration video (323 s): https://www.youtube.com/embed/H4z3xTJn058 (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-07); human review in PR #75 (protocol §2.7)."
diff --git a/tasks/behavior-cook-bacon-i00-privileged/task.toml b/tasks/behavior-cook-bacon-i00-privileged/task.toml
index 8ffe403f12..0cb4eb982e 100644
--- a/tasks/behavior-cook-bacon-i00-privileged/task.toml
+++ b/tasks/behavior-cook-bacon-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "behavior" # BEHAVIOR-1K v3.9.2 = OmniGibson 3.9.2 on Isa
robot = "r1pro"
family = "cook_bacon"
mode = "privileged" # privileged state, unlimited resets, thin primitives (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (256 s on average); demonstration video (256 s): https://player.vimeo.com/video/1114061497 (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-07); human review in PR #75 (protocol §2.7)."
instance_seed = 301 # scene = environment/instance.json (2026 challenge public_test instance 301 + its state hash); identity = dataset.toml digest
diff --git a/tasks/behavior-cook-bacon-i00-standard/task.toml b/tasks/behavior-cook-bacon-i00-standard/task.toml
index 2fe9817430..d3db8401df 100644
--- a/tasks/behavior-cook-bacon-i00-standard/task.toml
+++ b/tasks/behavior-cook-bacon-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "behavior"
robot = "r1pro"
family = "cook_bacon"
mode = "standard" # robot as a service: RGB-D + proprioception, no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (256 s on average); demonstration video (256 s): https://player.vimeo.com/video/1114061497 (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-07); human review in PR #75 (protocol §2.7)."
diff --git a/tasks/behavior-cook-broccolini-i00-privileged/task.toml b/tasks/behavior-cook-broccolini-i00-privileged/task.toml
index d08d7ab5fa..806d50d1ae 100644
--- a/tasks/behavior-cook-broccolini-i00-privileged/task.toml
+++ b/tasks/behavior-cook-broccolini-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "behavior" # BEHAVIOR-1K v3.9.2 = OmniGibson 3.9.2 on Isa
robot = "r1pro"
family = "cook_broccolini"
mode = "privileged" # privileged state, unlimited resets, thin primitives (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (154 s on average); demonstration video (137 s): https://www.youtube.com/embed/NXoYiAs_KCs (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-07); human review in PR #71 (protocol §2.7)."
instance_seed = 301 # scene = environment/instance.json (2026 challenge public_test instance 301 + its state hash); identity = dataset.toml digest
diff --git a/tasks/behavior-cook-broccolini-i00-standard/task.toml b/tasks/behavior-cook-broccolini-i00-standard/task.toml
index 5af29e8c80..2a490bf0a4 100644
--- a/tasks/behavior-cook-broccolini-i00-standard/task.toml
+++ b/tasks/behavior-cook-broccolini-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "behavior"
robot = "r1pro"
family = "cook_broccolini"
mode = "standard" # robot as a service: RGB-D + proprioception, no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (154 s on average); demonstration video (137 s): https://www.youtube.com/embed/NXoYiAs_KCs (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-07); human review in PR #71 (protocol §2.7)."
diff --git a/tasks/behavior-cook-brussels-sprouts-i00-privileged/task.toml b/tasks/behavior-cook-brussels-sprouts-i00-privileged/task.toml
index b0bd1d3ab9..f9565e36b9 100644
--- a/tasks/behavior-cook-brussels-sprouts-i00-privileged/task.toml
+++ b/tasks/behavior-cook-brussels-sprouts-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "behavior" # BEHAVIOR-1K v3.9.2 = OmniGibson 3.9.2 on Isa
robot = "r1pro"
family = "cook_brussels_sprouts"
mode = "privileged" # privileged state, unlimited resets, thin primitives (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (522 s on average); demonstration video (687 s): https://www.youtube.com/embed/9XPaPQmWAgI (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-04); human review per family (protocol §2.7)."
instance_seed = 301 # scene = environment/instance.json (2026 challenge public_test instance 301 + its state hash); identity = dataset.toml digest
diff --git a/tasks/behavior-cook-brussels-sprouts-i00-standard/task.toml b/tasks/behavior-cook-brussels-sprouts-i00-standard/task.toml
index d9292edfb5..58e05b7744 100644
--- a/tasks/behavior-cook-brussels-sprouts-i00-standard/task.toml
+++ b/tasks/behavior-cook-brussels-sprouts-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "behavior"
robot = "r1pro"
family = "cook_brussels_sprouts"
mode = "standard" # robot as a service: RGB-D + proprioception, no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (522 s on average); demonstration video (687 s): https://www.youtube.com/embed/9XPaPQmWAgI (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-04); human review per family (protocol §2.7)."
diff --git a/tasks/behavior-cook-cabbage-i00-privileged/task.toml b/tasks/behavior-cook-cabbage-i00-privileged/task.toml
index db348bec4b..8307ccab6c 100644
--- a/tasks/behavior-cook-cabbage-i00-privileged/task.toml
+++ b/tasks/behavior-cook-cabbage-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "behavior" # BEHAVIOR-1K v3.9.2 = OmniGibson 3.9.2 on Isa
robot = "r1pro"
family = "cook_cabbage"
mode = "privileged" # privileged state, unlimited resets, thin primitives (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (471 s on average); demonstration video (471 s): https://player.vimeo.com/video/1114060935 (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-04); human review per family (protocol §2.7)."
instance_seed = 301 # scene = environment/instance.json (2026 challenge public_test instance 301 + its state hash); identity = dataset.toml digest
diff --git a/tasks/behavior-cook-cabbage-i00-standard/task.toml b/tasks/behavior-cook-cabbage-i00-standard/task.toml
index 705b50ac92..a3fb2373d7 100644
--- a/tasks/behavior-cook-cabbage-i00-standard/task.toml
+++ b/tasks/behavior-cook-cabbage-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "behavior"
robot = "r1pro"
family = "cook_cabbage"
mode = "standard" # robot as a service: RGB-D + proprioception, no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (471 s on average); demonstration video (471 s): https://player.vimeo.com/video/1114060935 (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-04); human review per family (protocol §2.7)."
diff --git a/tasks/behavior-cook-hot-dogs-i00-privileged/task.toml b/tasks/behavior-cook-hot-dogs-i00-privileged/task.toml
index a8c0490cc2..e8433905df 100644
--- a/tasks/behavior-cook-hot-dogs-i00-privileged/task.toml
+++ b/tasks/behavior-cook-hot-dogs-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "behavior" # BEHAVIOR-1K v3.9.2 = OmniGibson 3.9.2 on Isa
robot = "r1pro"
family = "cook_hot_dogs"
mode = "privileged" # privileged state, unlimited resets, thin primitives (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (305 s on average); demonstration video (305 s): https://player.vimeo.com/video/1114061371 (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-07); human review in PR #75 (protocol §2.7)."
instance_seed = 301 # scene = environment/instance.json (2026 challenge public_test instance 301 + its state hash); identity = dataset.toml digest
diff --git a/tasks/behavior-cook-hot-dogs-i00-standard/task.toml b/tasks/behavior-cook-hot-dogs-i00-standard/task.toml
index 5ba9e6a771..9b98ad386c 100644
--- a/tasks/behavior-cook-hot-dogs-i00-standard/task.toml
+++ b/tasks/behavior-cook-hot-dogs-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "behavior"
robot = "r1pro"
family = "cook_hot_dogs"
mode = "standard" # robot as a service: RGB-D + proprioception, no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (305 s on average); demonstration video (305 s): https://player.vimeo.com/video/1114061371 (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-07); human review in PR #75 (protocol §2.7)."
diff --git a/tasks/behavior-dispose-of-batteries-i00-privileged/task.toml b/tasks/behavior-dispose-of-batteries-i00-privileged/task.toml
index 52f553de43..00b4cae669 100644
--- a/tasks/behavior-dispose-of-batteries-i00-privileged/task.toml
+++ b/tasks/behavior-dispose-of-batteries-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "behavior" # BEHAVIOR-1K v3.9.2 = OmniGibson 3.9.2 on Isa
robot = "r1pro"
family = "dispose_of_batteries"
mode = "privileged" # privileged state, unlimited resets, thin primitives (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (481 s on average); demonstration video (441 s): https://www.youtube.com/embed/TmJ7dNwDyws (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-04); human review per family (protocol §2.7)."
instance_seed = 301 # scene = environment/instance.json (2026 challenge public_test instance 301 + its state hash); identity = dataset.toml digest
diff --git a/tasks/behavior-dispose-of-batteries-i00-standard/task.toml b/tasks/behavior-dispose-of-batteries-i00-standard/task.toml
index 16c7a4644c..9b4931a19a 100644
--- a/tasks/behavior-dispose-of-batteries-i00-standard/task.toml
+++ b/tasks/behavior-dispose-of-batteries-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "behavior"
robot = "r1pro"
family = "dispose_of_batteries"
mode = "standard" # robot as a service: RGB-D + proprioception, no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (481 s on average); demonstration video (441 s): https://www.youtube.com/embed/TmJ7dNwDyws (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-04); human review per family (protocol §2.7)."
diff --git a/tasks/behavior-dispose-of-glass-i00-privileged/task.toml b/tasks/behavior-dispose-of-glass-i00-privileged/task.toml
index 8a8a0b7532..4c0d4e5843 100644
--- a/tasks/behavior-dispose-of-glass-i00-privileged/task.toml
+++ b/tasks/behavior-dispose-of-glass-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "behavior" # BEHAVIOR-1K v3.9.2 = OmniGibson 3.9.2 on Isa
robot = "r1pro"
family = "dispose_of_glass"
mode = "privileged" # privileged state, unlimited resets, thin primitives (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance. Solvable: model trial behavior-dispose-of-glass-i00__GAuQ2wP (behavior-lunaxh-t8-0927, privileged) graded success=1, deterministic=1 by the separate verifier on a pre-v1.0 version of this task (dev/qineng 912567fa; grader verdict and frozen instance unchanged since); agent log https://embodied-agent-interface-v2-internal.github.io/runs/behavior-1k/codex-gpt6_luna-xhigh/dispose_of_glass/unlimited/, verifier replay https://huggingface.co/datasets/eai-v2-internal/agent-runs/resolve/main/behavior-1k/codex-gpt6_luna-xhigh/dispose_of_glass/unlimited/replay.mp4; not re-run on this version (static validation only). Human review per family (protocol §2.7)."
instance_seed = 301 # scene = environment/instance.json (2026 challenge public_test instance 301 + its state hash); identity = dataset.toml digest
diff --git a/tasks/behavior-dispose-of-glass-i00-standard/task.toml b/tasks/behavior-dispose-of-glass-i00-standard/task.toml
index 1845010307..ebc7e5f240 100644
--- a/tasks/behavior-dispose-of-glass-i00-standard/task.toml
+++ b/tasks/behavior-dispose-of-glass-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "behavior"
robot = "r1pro"
family = "dispose_of_glass"
mode = "standard" # robot as a service: RGB-D + proprioception, no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance. Solvable: model trial behavior-dispose-of-glass-i00__GAuQ2wP (behavior-lunaxh-t8-0927, the privileged twin behavior-dispose-of-glass-i00-privileged on the same frozen instance and BDDL goal) graded success=1, deterministic=1 by the separate verifier on a pre-v1.0 version of this task (dev/qineng 912567fa; grader verdict and frozen instance unchanged since); agent log https://embodied-agent-interface-v2-internal.github.io/runs/behavior-1k/codex-gpt6_luna-xhigh/dispose_of_glass/unlimited/, verifier replay https://huggingface.co/datasets/eai-v2-internal/agent-runs/resolve/main/behavior-1k/codex-gpt6_luna-xhigh/dispose_of_glass/unlimited/replay.mp4; not re-run on this version (static validation only). Human review per family (protocol §2.7)."
diff --git a/tasks/behavior-freeze-fruit-i00-privileged/task.toml b/tasks/behavior-freeze-fruit-i00-privileged/task.toml
index 7ea842e3bb..bd0ddc0c95 100644
--- a/tasks/behavior-freeze-fruit-i00-privileged/task.toml
+++ b/tasks/behavior-freeze-fruit-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "behavior" # BEHAVIOR-1K v3.9.2 = OmniGibson 3.9.2 on Isa
robot = "r1pro"
family = "freeze_fruit"
mode = "privileged" # privileged state, unlimited resets, thin primitives (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (421 s on average); demonstration video (339 s): https://www.youtube.com/embed/6nM9_B5OoRY (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-04); human review per family (protocol §2.7)."
instance_seed = 301 # scene = environment/instance.json (2026 challenge public_test instance 301 + its state hash); identity = dataset.toml digest
diff --git a/tasks/behavior-freeze-fruit-i00-standard/task.toml b/tasks/behavior-freeze-fruit-i00-standard/task.toml
index 5846a4dbcd..d643d0bd0c 100644
--- a/tasks/behavior-freeze-fruit-i00-standard/task.toml
+++ b/tasks/behavior-freeze-fruit-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "behavior"
robot = "r1pro"
family = "freeze_fruit"
mode = "standard" # robot as a service: RGB-D + proprioception, no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (421 s on average); demonstration video (339 s): https://www.youtube.com/embed/6nM9_B5OoRY (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-04); human review per family (protocol §2.7)."
diff --git a/tasks/behavior-freeze-pies-i00-privileged/task.toml b/tasks/behavior-freeze-pies-i00-privileged/task.toml
index d9972b78ef..c903e8e708 100644
--- a/tasks/behavior-freeze-pies-i00-privileged/task.toml
+++ b/tasks/behavior-freeze-pies-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "behavior" # BEHAVIOR-1K v3.9.2 = OmniGibson 3.9.2 on Isa
robot = "r1pro"
family = "freeze_pies"
mode = "privileged" # privileged state, unlimited resets, thin primitives (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (415 s on average); demonstration video (415 s): https://player.vimeo.com/video/1114061625 (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-04); human review per family (protocol §2.7)."
instance_seed = 301 # scene = environment/instance.json (2026 challenge public_test instance 301 + its state hash); identity = dataset.toml digest
diff --git a/tasks/behavior-freeze-pies-i00-standard/task.toml b/tasks/behavior-freeze-pies-i00-standard/task.toml
index cd80316753..b1cb5b0123 100644
--- a/tasks/behavior-freeze-pies-i00-standard/task.toml
+++ b/tasks/behavior-freeze-pies-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "behavior"
robot = "r1pro"
family = "freeze_pies"
mode = "standard" # robot as a service: RGB-D + proprioception, no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (415 s on average); demonstration video (415 s): https://player.vimeo.com/video/1114061625 (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-04); human review per family (protocol §2.7)."
diff --git a/tasks/behavior-getting-organized-for-work-i00-privileged/task.toml b/tasks/behavior-getting-organized-for-work-i00-privileged/task.toml
index 57c4fe6fe6..7aa7596be0 100644
--- a/tasks/behavior-getting-organized-for-work-i00-privileged/task.toml
+++ b/tasks/behavior-getting-organized-for-work-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "behavior" # BEHAVIOR-1K v3.9.2 = OmniGibson 3.9.2 on Isa
robot = "r1pro"
family = "getting_organized_for_work"
mode = "privileged" # privileged state, unlimited resets, thin primitives (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (522 s on average); demonstration video (522 s): https://player.vimeo.com/video/1114059683 (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-04); human review per family (protocol §2.7)."
instance_seed = 301 # scene = environment/instance.json (2026 challenge public_test instance 301 + its state hash); identity = dataset.toml digest
diff --git a/tasks/behavior-getting-organized-for-work-i00-standard/task.toml b/tasks/behavior-getting-organized-for-work-i00-standard/task.toml
index 5261f6a215..78940183c6 100644
--- a/tasks/behavior-getting-organized-for-work-i00-standard/task.toml
+++ b/tasks/behavior-getting-organized-for-work-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "behavior"
robot = "r1pro"
family = "getting_organized_for_work"
mode = "standard" # robot as a service: RGB-D + proprioception, no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (522 s on average); demonstration video (522 s): https://player.vimeo.com/video/1114059683 (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-04); human review per family (protocol §2.7)."
diff --git a/tasks/behavior-halve-an-egg-i00-privileged/task.toml b/tasks/behavior-halve-an-egg-i00-privileged/task.toml
index 4607852ccf..7ebe7851ab 100644
--- a/tasks/behavior-halve-an-egg-i00-privileged/task.toml
+++ b/tasks/behavior-halve-an-egg-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "behavior" # BEHAVIOR-1K v3.9.2 = OmniGibson 3.9.2 on Isa
robot = "r1pro"
family = "halve_an_egg"
mode = "privileged" # privileged state, unlimited resets, thin primitives (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (213 s on average); demonstration video (212 s): https://www.youtube.com/embed/c1YJezdD7oM (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-07); human review in PR #75 (protocol §2.7)."
instance_seed = 301 # scene = environment/instance.json (2026 challenge public_test instance 301 + its state hash); identity = dataset.toml digest
diff --git a/tasks/behavior-halve-an-egg-i00-standard/task.toml b/tasks/behavior-halve-an-egg-i00-standard/task.toml
index 6ad66cf885..0e53be5b06 100644
--- a/tasks/behavior-halve-an-egg-i00-standard/task.toml
+++ b/tasks/behavior-halve-an-egg-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "behavior"
robot = "r1pro"
family = "halve_an_egg"
mode = "standard" # robot as a service: RGB-D + proprioception, no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (213 s on average); demonstration video (212 s): https://www.youtube.com/embed/c1YJezdD7oM (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-07); human review in PR #75 (protocol §2.7)."
diff --git a/tasks/behavior-hanging-pictures-i00-privileged/task.toml b/tasks/behavior-hanging-pictures-i00-privileged/task.toml
index 6c7babd9c6..afd6467a39 100644
--- a/tasks/behavior-hanging-pictures-i00-privileged/task.toml
+++ b/tasks/behavior-hanging-pictures-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "behavior" # BEHAVIOR-1K v3.9.2 = OmniGibson 3.9.2 on Isa
robot = "r1pro"
family = "hanging_pictures"
mode = "privileged" # privileged state, unlimited resets, thin primitives (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (80 s on average); demonstration video (80 s): https://player.vimeo.com/video/1114060330 (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-07); human review in PR #71 (protocol §2.7)."
instance_seed = 301 # scene = environment/instance.json (2026 challenge public_test instance 301 + its state hash); identity = dataset.toml digest
diff --git a/tasks/behavior-hanging-pictures-i00-standard/task.toml b/tasks/behavior-hanging-pictures-i00-standard/task.toml
index 8f61f07e80..3f28a48736 100644
--- a/tasks/behavior-hanging-pictures-i00-standard/task.toml
+++ b/tasks/behavior-hanging-pictures-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "behavior"
robot = "r1pro"
family = "hanging_pictures"
mode = "standard" # robot as a service: RGB-D + proprioception, no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (80 s on average); demonstration video (80 s): https://player.vimeo.com/video/1114060330 (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-07); human review in PR #71 (protocol §2.7)."
diff --git a/tasks/behavior-hiding-Easter-eggs-i00-privileged/task.toml b/tasks/behavior-hiding-Easter-eggs-i00-privileged/task.toml
index 15aeb00b71..fd81c45b09 100644
--- a/tasks/behavior-hiding-Easter-eggs-i00-privileged/task.toml
+++ b/tasks/behavior-hiding-Easter-eggs-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "behavior" # BEHAVIOR-1K v3.9.2 = OmniGibson 3.9.2 on Isa
robot = "r1pro"
family = "hiding_Easter_eggs"
mode = "privileged" # privileged state, unlimited resets, thin primitives (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance. Solvable: model trial behavior-hiding-Easter-eggs-i00__QL9YpHL (behavior-lunaxh-t8-0927, privileged) graded success=1, deterministic=1 by the separate verifier on a pre-v1.0 version of this task (dev/qineng 20a382a8; grader verdict and frozen instance unchanged since); agent log https://embodied-agent-interface-v2-internal.github.io/runs/behavior-1k/codex-gpt6_luna-xhigh/hiding_Easter_eggs/unlimited/, verifier replay https://huggingface.co/datasets/eai-v2-internal/agent-runs/resolve/main/behavior-1k/codex-gpt6_luna-xhigh/hiding_Easter_eggs/unlimited/replay.mp4; not re-run on this version (static validation only). Human review per family (protocol §2.7)."
instance_seed = 301 # scene = environment/instance.json (2026 challenge public_test instance 301 + its state hash); identity = dataset.toml digest
diff --git a/tasks/behavior-hiding-Easter-eggs-i00-standard/task.toml b/tasks/behavior-hiding-Easter-eggs-i00-standard/task.toml
index d554c064a1..3d65c8dd69 100644
--- a/tasks/behavior-hiding-Easter-eggs-i00-standard/task.toml
+++ b/tasks/behavior-hiding-Easter-eggs-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "behavior"
robot = "r1pro"
family = "hiding_Easter_eggs"
mode = "standard" # robot as a service: RGB-D + proprioception, no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance. Solvable: model trial behavior-hiding-Easter-eggs-i00__QL9YpHL (behavior-lunaxh-t8-0927, the privileged twin behavior-hiding-Easter-eggs-i00-privileged on the same frozen instance and BDDL goal) graded success=1, deterministic=1 by the separate verifier on a pre-v1.0 version of this task (dev/qineng 20a382a8; grader verdict and frozen instance unchanged since); agent log https://embodied-agent-interface-v2-internal.github.io/runs/behavior-1k/codex-gpt6_luna-xhigh/hiding_Easter_eggs/unlimited/, verifier replay https://huggingface.co/datasets/eai-v2-internal/agent-runs/resolve/main/behavior-1k/codex-gpt6_luna-xhigh/hiding_Easter_eggs/unlimited/replay.mp4; not re-run on this version (static validation only). Human review per family (protocol §2.7)."
diff --git a/tasks/behavior-installing-a-fax-machine-i00-privileged/task.toml b/tasks/behavior-installing-a-fax-machine-i00-privileged/task.toml
index 55564f8093..76419fc2c9 100644
--- a/tasks/behavior-installing-a-fax-machine-i00-privileged/task.toml
+++ b/tasks/behavior-installing-a-fax-machine-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "behavior" # BEHAVIOR-1K v3.9.2 = OmniGibson 3.9.2 on Isa
robot = "r1pro"
family = "installing_a_fax_machine"
mode = "privileged" # privileged state, unlimited resets, thin primitives (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (136 s on average); demonstration video (148 s): https://www.youtube.com/embed/v1CrUuIN0lY (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-07); human review in PR #75 (protocol §2.7)."
instance_seed = 301 # scene = environment/instance.json (2026 challenge public_test instance 301 + its state hash); identity = dataset.toml digest
diff --git a/tasks/behavior-installing-a-fax-machine-i00-standard/task.toml b/tasks/behavior-installing-a-fax-machine-i00-standard/task.toml
index 3d357acc1f..6539783036 100644
--- a/tasks/behavior-installing-a-fax-machine-i00-standard/task.toml
+++ b/tasks/behavior-installing-a-fax-machine-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "behavior"
robot = "r1pro"
family = "installing_a_fax_machine"
mode = "standard" # robot as a service: RGB-D + proprioception, no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (136 s on average); demonstration video (148 s): https://www.youtube.com/embed/v1CrUuIN0lY (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-07); human review in PR #75 (protocol §2.7)."
diff --git a/tasks/behavior-installing-a-modem-i00-privileged/task.toml b/tasks/behavior-installing-a-modem-i00-privileged/task.toml
index 2b1f488feb..8bacbd02d6 100644
--- a/tasks/behavior-installing-a-modem-i00-privileged/task.toml
+++ b/tasks/behavior-installing-a-modem-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "behavior" # BEHAVIOR-1K v3.9.2 = OmniGibson 3.9.2 on Isa
robot = "r1pro"
family = "installing_a_modem"
mode = "privileged" # privileged state, unlimited resets, thin primitives (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (80 s on average); demonstration video (63 s): https://www.youtube.com/embed/zHEfUkrCWfc (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-07); human review in PR #75 (protocol §2.7)."
instance_seed = 301 # scene = environment/instance.json (2026 challenge public_test instance 301 + its state hash); identity = dataset.toml digest
diff --git a/tasks/behavior-installing-a-modem-i00-standard/task.toml b/tasks/behavior-installing-a-modem-i00-standard/task.toml
index 4209b446d2..632b54944d 100644
--- a/tasks/behavior-installing-a-modem-i00-standard/task.toml
+++ b/tasks/behavior-installing-a-modem-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "behavior"
robot = "r1pro"
family = "installing_a_modem"
mode = "standard" # robot as a service: RGB-D + proprioception, no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (80 s on average); demonstration video (63 s): https://www.youtube.com/embed/zHEfUkrCWfc (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-07); human review in PR #75 (protocol §2.7)."
diff --git a/tasks/behavior-installing-a-scanner-i00-privileged/task.toml b/tasks/behavior-installing-a-scanner-i00-privileged/task.toml
index d854bb8d46..51a4b9ba24 100644
--- a/tasks/behavior-installing-a-scanner-i00-privileged/task.toml
+++ b/tasks/behavior-installing-a-scanner-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "behavior" # BEHAVIOR-1K v3.9.2 = OmniGibson 3.9.2 on Isa
robot = "r1pro"
family = "installing_a_scanner"
mode = "privileged" # privileged state, unlimited resets, thin primitives (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance. Solvable: model trial behavior-installing-a-scanner-i0__pytvepD (behavior-lunaxh-t8-0927, privileged) graded success=1, deterministic=1 by the separate verifier on a pre-v1.0 version of this task (dev/qineng 912567fa; grader verdict and frozen instance unchanged since); agent log https://embodied-agent-interface-v2-internal.github.io/runs/behavior-1k/codex-gpt6_luna-xhigh/installing_a_scanner/unlimited/, verifier replay https://huggingface.co/datasets/eai-v2-internal/agent-runs/resolve/main/behavior-1k/codex-gpt6_luna-xhigh/installing_a_scanner/unlimited/replay.mp4; not re-run on this version (static validation only). Human review per family (protocol §2.7)."
instance_seed = 301 # scene = environment/instance.json (2026 challenge public_test instance 301 + its state hash); identity = dataset.toml digest
diff --git a/tasks/behavior-installing-a-scanner-i00-standard/task.toml b/tasks/behavior-installing-a-scanner-i00-standard/task.toml
index d5f65c32c9..186538d8ec 100644
--- a/tasks/behavior-installing-a-scanner-i00-standard/task.toml
+++ b/tasks/behavior-installing-a-scanner-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "behavior"
robot = "r1pro"
family = "installing_a_scanner"
mode = "standard" # robot as a service: RGB-D + proprioception, no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance. Solvable: model trial behavior-installing-a-scanner-i0__pytvepD (behavior-lunaxh-t8-0927, the privileged twin behavior-installing-a-scanner-i00-privileged on the same frozen instance and BDDL goal) graded success=1, deterministic=1 by the separate verifier on a pre-v1.0 version of this task (dev/qineng 912567fa; grader verdict and frozen instance unchanged since); agent log https://embodied-agent-interface-v2-internal.github.io/runs/behavior-1k/codex-gpt6_luna-xhigh/installing_a_scanner/unlimited/, verifier replay https://huggingface.co/datasets/eai-v2-internal/agent-runs/resolve/main/behavior-1k/codex-gpt6_luna-xhigh/installing_a_scanner/unlimited/replay.mp4; not re-run on this version (static validation only). Human review per family (protocol §2.7)."
diff --git a/tasks/behavior-installing-smoke-detectors-i00-privileged/task.toml b/tasks/behavior-installing-smoke-detectors-i00-privileged/task.toml
index 7cd00ec890..24e7f6410e 100644
--- a/tasks/behavior-installing-smoke-detectors-i00-privileged/task.toml
+++ b/tasks/behavior-installing-smoke-detectors-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "behavior" # BEHAVIOR-1K v3.9.2 = OmniGibson 3.9.2 on Isa
robot = "r1pro"
family = "installing_smoke_detectors"
mode = "privileged" # privileged state, unlimited resets, thin primitives (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (86 s on average); demonstration video (68 s): https://www.youtube.com/embed/gRtLcTllXqQ (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-07); human review in PR #75 (protocol §2.7)."
instance_seed = 301 # scene = environment/instance.json (2026 challenge public_test instance 301 + its state hash); identity = dataset.toml digest
diff --git a/tasks/behavior-installing-smoke-detectors-i00-standard/task.toml b/tasks/behavior-installing-smoke-detectors-i00-standard/task.toml
index 8fd8ad3c2c..b097098e57 100644
--- a/tasks/behavior-installing-smoke-detectors-i00-standard/task.toml
+++ b/tasks/behavior-installing-smoke-detectors-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "behavior"
robot = "r1pro"
family = "installing_smoke_detectors"
mode = "standard" # robot as a service: RGB-D + proprioception, no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (86 s on average); demonstration video (68 s): https://www.youtube.com/embed/gRtLcTllXqQ (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-07); human review in PR #75 (protocol §2.7)."
diff --git a/tasks/behavior-laying-tile-floors-i00-privileged/task.toml b/tasks/behavior-laying-tile-floors-i00-privileged/task.toml
index 1d7d8d87f5..3c84860c87 100644
--- a/tasks/behavior-laying-tile-floors-i00-privileged/task.toml
+++ b/tasks/behavior-laying-tile-floors-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "behavior" # BEHAVIOR-1K v3.9.2 = OmniGibson 3.9.2 on Isa
robot = "r1pro"
family = "laying_tile_floors"
mode = "privileged" # privileged state, unlimited resets, thin primitives (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (478 s on average); demonstration video (520 s): https://www.youtube.com/embed/a3OjMFaaSs8 (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-04); human review per family (protocol §2.7)."
instance_seed = 301 # scene = environment/instance.json (2026 challenge public_test instance 301 + its state hash); identity = dataset.toml digest
diff --git a/tasks/behavior-laying-tile-floors-i00-standard/task.toml b/tasks/behavior-laying-tile-floors-i00-standard/task.toml
index 508c7a22b4..3cfd461630 100644
--- a/tasks/behavior-laying-tile-floors-i00-standard/task.toml
+++ b/tasks/behavior-laying-tile-floors-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "behavior"
robot = "r1pro"
family = "laying_tile_floors"
mode = "standard" # robot as a service: RGB-D + proprioception, no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (478 s on average); demonstration video (520 s): https://www.youtube.com/embed/a3OjMFaaSs8 (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-04); human review per family (protocol §2.7)."
diff --git a/tasks/behavior-loading-the-car-i00-privileged/task.toml b/tasks/behavior-loading-the-car-i00-privileged/task.toml
index 13ae5f5628..cf559a8bcf 100644
--- a/tasks/behavior-loading-the-car-i00-privileged/task.toml
+++ b/tasks/behavior-loading-the-car-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "behavior" # BEHAVIOR-1K v3.9.2 = OmniGibson 3.9.2 on Isa
robot = "r1pro"
family = "loading_the_car"
mode = "privileged" # privileged state, unlimited resets, thin primitives (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (641 s on average); demonstration video (641 s): https://player.vimeo.com/video/1114057519 (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-04); human review per family (protocol §2.7)."
instance_seed = 301 # scene = environment/instance.json (2026 challenge public_test instance 301 + its state hash); identity = dataset.toml digest
diff --git a/tasks/behavior-loading-the-car-i00-standard/task.toml b/tasks/behavior-loading-the-car-i00-standard/task.toml
index 55ccca6073..ca1089cd76 100644
--- a/tasks/behavior-loading-the-car-i00-standard/task.toml
+++ b/tasks/behavior-loading-the-car-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "behavior"
robot = "r1pro"
family = "loading_the_car"
mode = "standard" # robot as a service: RGB-D + proprioception, no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (641 s on average); demonstration video (641 s): https://player.vimeo.com/video/1114057519 (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-04); human review per family (protocol §2.7)."
diff --git a/tasks/behavior-make-cabinet-doors-i00-privileged/task.toml b/tasks/behavior-make-cabinet-doors-i00-privileged/task.toml
index 7619b46a20..3d219953b4 100644
--- a/tasks/behavior-make-cabinet-doors-i00-privileged/task.toml
+++ b/tasks/behavior-make-cabinet-doors-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "behavior" # BEHAVIOR-1K v3.9.2 = OmniGibson 3.9.2 on Isa
robot = "r1pro"
family = "make_cabinet_doors"
mode = "privileged" # privileged state, unlimited resets, thin primitives (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (115 s on average); demonstration video (120 s): https://www.youtube.com/embed/rqqgwEPTLwk (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-07); human review in PR #75 (protocol §2.7)."
instance_seed = 301 # scene = environment/instance.json (2026 challenge public_test instance 301 + its state hash); identity = dataset.toml digest
diff --git a/tasks/behavior-make-cabinet-doors-i00-standard/task.toml b/tasks/behavior-make-cabinet-doors-i00-standard/task.toml
index c2f9a8a2ac..d1d8871534 100644
--- a/tasks/behavior-make-cabinet-doors-i00-standard/task.toml
+++ b/tasks/behavior-make-cabinet-doors-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "behavior"
robot = "r1pro"
family = "make_cabinet_doors"
mode = "standard" # robot as a service: RGB-D + proprioception, no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (115 s on average); demonstration video (120 s): https://www.youtube.com/embed/rqqgwEPTLwk (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-07); human review in PR #75 (protocol §2.7)."
diff --git a/tasks/behavior-make-gift-bags-for-baby-showers-i00-privileged/task.toml b/tasks/behavior-make-gift-bags-for-baby-showers-i00-privileged/task.toml
index 8aa418a35f..08ac0bbb74 100644
--- a/tasks/behavior-make-gift-bags-for-baby-showers-i00-privileged/task.toml
+++ b/tasks/behavior-make-gift-bags-for-baby-showers-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "behavior" # BEHAVIOR-1K v3.9.2 = OmniGibson 3.9.2 on Isa
robot = "r1pro"
family = "make_gift_bags_for_baby_showers"
mode = "privileged" # privileged state, unlimited resets, thin primitives (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (322 s on average); demonstration video (242 s): https://www.youtube.com/embed/EY42jyNoEh8 (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-07); human review in PR #71 (protocol §2.7)."
instance_seed = 301 # scene = environment/instance.json (2026 challenge public_test instance 301 + its state hash); identity = dataset.toml digest
diff --git a/tasks/behavior-make-gift-bags-for-baby-showers-i00-standard/task.toml b/tasks/behavior-make-gift-bags-for-baby-showers-i00-standard/task.toml
index 342092d4b8..3f17c3ddc0 100644
--- a/tasks/behavior-make-gift-bags-for-baby-showers-i00-standard/task.toml
+++ b/tasks/behavior-make-gift-bags-for-baby-showers-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "behavior"
robot = "r1pro"
family = "make_gift_bags_for_baby_showers"
mode = "standard" # robot as a service: RGB-D + proprioception, no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (322 s on average); demonstration video (242 s): https://www.youtube.com/embed/EY42jyNoEh8 (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-07); human review in PR #71 (protocol §2.7)."
diff --git a/tasks/behavior-make-microwave-popcorn-i00-privileged/task.toml b/tasks/behavior-make-microwave-popcorn-i00-privileged/task.toml
index 3ed1c5bdd5..3f986eed2f 100644
--- a/tasks/behavior-make-microwave-popcorn-i00-privileged/task.toml
+++ b/tasks/behavior-make-microwave-popcorn-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "behavior" # BEHAVIOR-1K v3.9.2 = OmniGibson 3.9.2 on Isa
robot = "r1pro"
family = "make_microwave_popcorn"
mode = "privileged" # privileged state, unlimited resets, thin primitives (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (108 s on average); demonstration video (108 s): https://player.vimeo.com/video/1114060842 (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-07); human review in PR #75 (protocol §2.7)."
instance_seed = 301 # scene = environment/instance.json (2026 challenge public_test instance 301 + its state hash); identity = dataset.toml digest
diff --git a/tasks/behavior-make-microwave-popcorn-i00-standard/task.toml b/tasks/behavior-make-microwave-popcorn-i00-standard/task.toml
index ac1d63eab1..f42b3506a0 100644
--- a/tasks/behavior-make-microwave-popcorn-i00-standard/task.toml
+++ b/tasks/behavior-make-microwave-popcorn-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "behavior"
robot = "r1pro"
family = "make_microwave_popcorn"
mode = "standard" # robot as a service: RGB-D + proprioception, no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (108 s on average); demonstration video (108 s): https://player.vimeo.com/video/1114060842 (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-07); human review in PR #75 (protocol §2.7)."
diff --git a/tasks/behavior-make-pizza-i00-privileged/task.toml b/tasks/behavior-make-pizza-i00-privileged/task.toml
index 2d6a2c1727..4e8a071a2d 100644
--- a/tasks/behavior-make-pizza-i00-privileged/task.toml
+++ b/tasks/behavior-make-pizza-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "behavior" # BEHAVIOR-1K v3.9.2 = OmniGibson 3.9.2 on Isa
robot = "r1pro"
family = "make_pizza"
mode = "privileged" # privileged state, unlimited resets, thin primitives (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (640 s on average); demonstration video (640 s): https://player.vimeo.com/video/1114061917 (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-04); human review per family (protocol §2.7)."
instance_seed = 301 # scene = environment/instance.json (2026 challenge public_test instance 301 + its state hash); identity = dataset.toml digest
diff --git a/tasks/behavior-make-pizza-i00-standard/task.toml b/tasks/behavior-make-pizza-i00-standard/task.toml
index b7f778d993..3282cd931d 100644
--- a/tasks/behavior-make-pizza-i00-standard/task.toml
+++ b/tasks/behavior-make-pizza-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "behavior"
robot = "r1pro"
family = "make_pizza"
mode = "standard" # robot as a service: RGB-D + proprioception, no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (640 s on average); demonstration video (640 s): https://player.vimeo.com/video/1114061917 (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-04); human review per family (protocol §2.7)."
diff --git a/tasks/behavior-make-rose-centerpieces-i00-privileged/task.toml b/tasks/behavior-make-rose-centerpieces-i00-privileged/task.toml
index f463995cc0..088733b22c 100644
--- a/tasks/behavior-make-rose-centerpieces-i00-privileged/task.toml
+++ b/tasks/behavior-make-rose-centerpieces-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "behavior" # BEHAVIOR-1K v3.9.2 = OmniGibson 3.9.2 on Isa
robot = "r1pro"
family = "make_rose_centerpieces"
mode = "privileged" # privileged state, unlimited resets, thin primitives (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (151 s on average); demonstration video (111 s): https://www.youtube.com/embed/Nb8kdrf7c_M (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-07); human review in PR #71 (protocol §2.7)."
instance_seed = 301 # scene = environment/instance.json (2026 challenge public_test instance 301 + its state hash); identity = dataset.toml digest
diff --git a/tasks/behavior-make-rose-centerpieces-i00-standard/task.toml b/tasks/behavior-make-rose-centerpieces-i00-standard/task.toml
index 98eba48512..762028f832 100644
--- a/tasks/behavior-make-rose-centerpieces-i00-standard/task.toml
+++ b/tasks/behavior-make-rose-centerpieces-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "behavior"
robot = "r1pro"
family = "make_rose_centerpieces"
mode = "standard" # robot as a service: RGB-D + proprioception, no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (151 s on average); demonstration video (111 s): https://www.youtube.com/embed/Nb8kdrf7c_M (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-07); human review in PR #71 (protocol §2.7)."
diff --git a/tasks/behavior-moving-boxes-to-storage-i00-privileged/task.toml b/tasks/behavior-moving-boxes-to-storage-i00-privileged/task.toml
index 535029d1b8..3e31cb9bb9 100644
--- a/tasks/behavior-moving-boxes-to-storage-i00-privileged/task.toml
+++ b/tasks/behavior-moving-boxes-to-storage-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "behavior" # BEHAVIOR-1K v3.9.2 = OmniGibson 3.9.2 on Isa
robot = "r1pro"
family = "moving_boxes_to_storage"
mode = "privileged" # privileged state, unlimited resets, thin primitives (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (487 s on average); demonstration video (487 s): https://player.vimeo.com/video/1114058315 (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-04); human review per family (protocol §2.7)."
instance_seed = 301 # scene = environment/instance.json (2026 challenge public_test instance 301 + its state hash); identity = dataset.toml digest
diff --git a/tasks/behavior-moving-boxes-to-storage-i00-standard/task.toml b/tasks/behavior-moving-boxes-to-storage-i00-standard/task.toml
index 7e8d4d79ca..3e3695cce1 100644
--- a/tasks/behavior-moving-boxes-to-storage-i00-standard/task.toml
+++ b/tasks/behavior-moving-boxes-to-storage-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "behavior"
robot = "r1pro"
family = "moving_boxes_to_storage"
mode = "standard" # robot as a service: RGB-D + proprioception, no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (487 s on average); demonstration video (487 s): https://player.vimeo.com/video/1114058315 (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-04); human review per family (protocol §2.7)."
diff --git a/tasks/behavior-organizing-art-supplies-i00-privileged/task.toml b/tasks/behavior-organizing-art-supplies-i00-privileged/task.toml
index b0fa08dc7f..91b4ca13f6 100644
--- a/tasks/behavior-organizing-art-supplies-i00-privileged/task.toml
+++ b/tasks/behavior-organizing-art-supplies-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "behavior" # BEHAVIOR-1K v3.9.2 = OmniGibson 3.9.2 on Isa
robot = "r1pro"
family = "organizing_art_supplies"
mode = "privileged" # privileged state, unlimited resets, thin primitives (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (206 s on average); demonstration video (228 s): https://www.youtube.com/embed/M1D9oCKBJZg (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-07); human review in PR #71 (protocol §2.7)."
instance_seed = 301 # scene = environment/instance.json (2026 challenge public_test instance 301 + its state hash); identity = dataset.toml digest
diff --git a/tasks/behavior-organizing-art-supplies-i00-standard/task.toml b/tasks/behavior-organizing-art-supplies-i00-standard/task.toml
index 4b7fc55194..c8a554b66d 100644
--- a/tasks/behavior-organizing-art-supplies-i00-standard/task.toml
+++ b/tasks/behavior-organizing-art-supplies-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "behavior"
robot = "r1pro"
family = "organizing_art_supplies"
mode = "standard" # robot as a service: RGB-D + proprioception, no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (206 s on average); demonstration video (228 s): https://www.youtube.com/embed/M1D9oCKBJZg (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-07); human review in PR #71 (protocol §2.7)."
diff --git a/tasks/behavior-organizing-school-stuff-i00-privileged/task.toml b/tasks/behavior-organizing-school-stuff-i00-privileged/task.toml
index e0108e048d..4ad8a70945 100644
--- a/tasks/behavior-organizing-school-stuff-i00-privileged/task.toml
+++ b/tasks/behavior-organizing-school-stuff-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "behavior" # BEHAVIOR-1K v3.9.2 = OmniGibson 3.9.2 on Isa
robot = "r1pro"
family = "organizing_school_stuff"
mode = "privileged" # privileged state, unlimited resets, thin primitives (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (486 s on average); demonstration video (426 s): https://www.youtube.com/embed/GO2Bw-0OY7M (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-04); human review per family (protocol §2.7)."
instance_seed = 301 # scene = environment/instance.json (2026 challenge public_test instance 301 + its state hash); identity = dataset.toml digest
diff --git a/tasks/behavior-organizing-school-stuff-i00-standard/task.toml b/tasks/behavior-organizing-school-stuff-i00-standard/task.toml
index 21f4a7de01..c5d17f5ab9 100644
--- a/tasks/behavior-organizing-school-stuff-i00-standard/task.toml
+++ b/tasks/behavior-organizing-school-stuff-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "behavior"
robot = "r1pro"
family = "organizing_school_stuff"
mode = "standard" # robot as a service: RGB-D + proprioception, no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (486 s on average); demonstration video (426 s): https://www.youtube.com/embed/GO2Bw-0OY7M (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-04); human review per family (protocol §2.7)."
diff --git a/tasks/behavior-outfit-a-basic-toolbox-i00-privileged/task.toml b/tasks/behavior-outfit-a-basic-toolbox-i00-privileged/task.toml
index b44547a1f1..0ff830d288 100644
--- a/tasks/behavior-outfit-a-basic-toolbox-i00-privileged/task.toml
+++ b/tasks/behavior-outfit-a-basic-toolbox-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "behavior" # BEHAVIOR-1K v3.9.2 = OmniGibson 3.9.2 on Isa
robot = "r1pro"
family = "outfit_a_basic_toolbox"
mode = "privileged" # privileged state, unlimited resets, thin primitives (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (355 s on average); demonstration video (355 s): https://player.vimeo.com/video/1114058641 (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-04); human review per family (protocol §2.7)."
instance_seed = 301 # scene = environment/instance.json (2026 challenge public_test instance 301 + its state hash); identity = dataset.toml digest
diff --git a/tasks/behavior-outfit-a-basic-toolbox-i00-standard/task.toml b/tasks/behavior-outfit-a-basic-toolbox-i00-standard/task.toml
index 2f05024ba0..3ea5f95997 100644
--- a/tasks/behavior-outfit-a-basic-toolbox-i00-standard/task.toml
+++ b/tasks/behavior-outfit-a-basic-toolbox-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "behavior"
robot = "r1pro"
family = "outfit_a_basic_toolbox"
mode = "standard" # robot as a service: RGB-D + proprioception, no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (355 s on average); demonstration video (355 s): https://player.vimeo.com/video/1114058641 (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-04); human review per family (protocol §2.7)."
diff --git a/tasks/behavior-packing-meal-for-delivery-i00-privileged/task.toml b/tasks/behavior-packing-meal-for-delivery-i00-privileged/task.toml
index 793c89a3cb..a1858fd969 100644
--- a/tasks/behavior-packing-meal-for-delivery-i00-privileged/task.toml
+++ b/tasks/behavior-packing-meal-for-delivery-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "behavior" # BEHAVIOR-1K v3.9.2 = OmniGibson 3.9.2 on Isa
robot = "r1pro"
family = "packing_meal_for_delivery"
mode = "privileged" # privileged state, unlimited resets, thin primitives (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (295 s on average); demonstration video (300 s): https://www.youtube.com/embed/DirqFshrOAA (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-07); human review in PR #71 (protocol §2.7)."
instance_seed = 301 # scene = environment/instance.json (2026 challenge public_test instance 301 + its state hash); identity = dataset.toml digest
diff --git a/tasks/behavior-packing-meal-for-delivery-i00-standard/task.toml b/tasks/behavior-packing-meal-for-delivery-i00-standard/task.toml
index 37d6d799e2..b2b04ae8ba 100644
--- a/tasks/behavior-packing-meal-for-delivery-i00-standard/task.toml
+++ b/tasks/behavior-packing-meal-for-delivery-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "behavior"
robot = "r1pro"
family = "packing_meal_for_delivery"
mode = "standard" # robot as a service: RGB-D + proprioception, no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (295 s on average); demonstration video (300 s): https://www.youtube.com/embed/DirqFshrOAA (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-07); human review in PR #71 (protocol §2.7)."
diff --git a/tasks/behavior-picking-up-toys-i00-privileged/task.toml b/tasks/behavior-picking-up-toys-i00-privileged/task.toml
index 4e231cbdd2..cc40db1a7c 100644
--- a/tasks/behavior-picking-up-toys-i00-privileged/task.toml
+++ b/tasks/behavior-picking-up-toys-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "behavior" # BEHAVIOR-1K v3.9.2 = OmniGibson 3.9.2 on Isa
robot = "r1pro"
family = "picking_up_toys"
mode = "privileged" # privileged state, unlimited resets, thin primitives (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (630 s on average); demonstration video (630 s): https://player.vimeo.com/video/1114054920 (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-04); human review per family (protocol §2.7)."
instance_seed = 301 # scene = environment/instance.json (2026 challenge public_test instance 301 + its state hash); identity = dataset.toml digest
diff --git a/tasks/behavior-picking-up-toys-i00-standard/task.toml b/tasks/behavior-picking-up-toys-i00-standard/task.toml
index 016af631a8..eca0926c01 100644
--- a/tasks/behavior-picking-up-toys-i00-standard/task.toml
+++ b/tasks/behavior-picking-up-toys-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "behavior"
robot = "r1pro"
family = "picking_up_toys"
mode = "standard" # robot as a service: RGB-D + proprioception, no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (630 s on average); demonstration video (630 s): https://player.vimeo.com/video/1114054920 (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-04); human review per family (protocol §2.7)."
diff --git a/tasks/behavior-picking-up-trash-i00-privileged/task.toml b/tasks/behavior-picking-up-trash-i00-privileged/task.toml
index 75e92c7785..24eed48de7 100644
--- a/tasks/behavior-picking-up-trash-i00-privileged/task.toml
+++ b/tasks/behavior-picking-up-trash-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "behavior" # BEHAVIOR-1K v3.9.2 = OmniGibson 3.9.2 on Isa
robot = "r1pro"
family = "picking_up_trash"
mode = "privileged" # privileged state, unlimited resets, thin primitives (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "partial"
oracle_evidence = "Reference solution solution/oracle_picking_up_trash.py on the frozen task-level primitives solution/rcb_behavior_ref.py. Recorded Harbor oracle runs (laptop, pre-v1.0 uncommitted trees): 2026-09-24 behavior-picking-up-trash-i00__FKLU2nS (job behavior-picking-up-trash-i00-oracle-0924-230430) scored 1 (5115 control steps, deterministic 1, instance_exact 1); the run before it, behavior-picking-up-trash-i00__MjRvy7Y (an older placement routine), scored 0; the 2026-10-01 review counts 1 success in 3 runs of that version. Model trial behavior-picking-up-trash-i00__JRcB9fb (behavior-lunaxh-0925-2057, ws) scored 1. The tooling and verifier files changed since; not re-run on this version (static validation only). Measured 2026-10-10 on main 332bf8080 (board 0067 oracle sweep, AWS 8-vCPU L40S host rcb-gpu9, Harbor oracle agent, twice): the reference solution's own run put 2 of the 3 cans in (the third grasp closed on nothing; q_score 0.667, 4390 control steps), the verifier replayed it faithfully (deterministic 1) and scored 0; both runs were bit-identical. The frozen scene rebuilt there within tolerance but not bit-exactly (instance_exact 0), so the solution's online planning took another path than on the recording host: this reference solution does not succeed on that host class."
instance_seed = 301 # scene = environment/instance.json (2026 challenge public_test instance 301 + its state hash); identity = dataset.toml digest
diff --git a/tasks/behavior-picking-up-trash-i00-standard/task.toml b/tasks/behavior-picking-up-trash-i00-standard/task.toml
index ca55f08272..fb373b8cd0 100644
--- a/tasks/behavior-picking-up-trash-i00-standard/task.toml
+++ b/tasks/behavior-picking-up-trash-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "behavior"
robot = "r1pro"
family = "picking_up_trash"
mode = "standard" # robot as a service: RGB-D + proprioception, no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "partial"
oracle_evidence = "solution/ streams the privileged oracle's joint targets (oracle_qpos.npy, 5115 rows) through the client. Recorded Harbor oracle runs: 3 of 3 scored 1 on ws3 (job v22-ws3-oracle-picking-up-trash-1741, best behavior-picking-up-trash-i00-li__eg8GdNj) on the dev/qineng 9eb18edf version of this task (rcb-limited/2.2, no crash recovery); 2026-09-25 behavior-picking-up-trash-i00-li__fyYUXt4 (laptop, an older version) scored 1 with live_matches_replay 1. The sim service, collect hook, grader and oracle client changed since (crash recovery); not re-run on this version (static validation only). Measured 2026-10-10 on main 332bf8080 (board 0067 oracle sweep, AWS L40S host rcb-gpu11): scored 1 (5114 control steps; live, both replays and deterministic 1) once the oracle waited long enough for the service (the simulator took 620-682 s to start on those hosts, longer than the client's 600 s default wait at the time, which gave two 0-step runs; the default is 3600 s since)."
diff --git a/tasks/behavior-polishing-shoes-i00-privileged/task.toml b/tasks/behavior-polishing-shoes-i00-privileged/task.toml
index 2b836ad06e..1091d3d03c 100644
--- a/tasks/behavior-polishing-shoes-i00-privileged/task.toml
+++ b/tasks/behavior-polishing-shoes-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "behavior" # BEHAVIOR-1K v3.9.2 = OmniGibson 3.9.2 on Isa
robot = "r1pro"
family = "polishing_shoes"
mode = "privileged" # privileged state, unlimited resets, thin primitives (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (358 s on average); demonstration video (363 s): https://www.youtube.com/embed/bSCbEvVJykI (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-04); human review per family (protocol §2.7)."
instance_seed = 301 # scene = environment/instance.json (2026 challenge public_test instance 301 + its state hash); identity = dataset.toml digest
diff --git a/tasks/behavior-polishing-shoes-i00-standard/task.toml b/tasks/behavior-polishing-shoes-i00-standard/task.toml
index bca9cab6ae..dd0b39124a 100644
--- a/tasks/behavior-polishing-shoes-i00-standard/task.toml
+++ b/tasks/behavior-polishing-shoes-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "behavior"
robot = "r1pro"
family = "polishing_shoes"
mode = "standard" # robot as a service: RGB-D + proprioception, no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (358 s on average); demonstration video (363 s): https://www.youtube.com/embed/bSCbEvVJykI (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-04); human review per family (protocol §2.7)."
diff --git a/tasks/behavior-preparing-lunch-box-i00-privileged/task.toml b/tasks/behavior-preparing-lunch-box-i00-privileged/task.toml
index 61b1297f62..9fecd202d8 100644
--- a/tasks/behavior-preparing-lunch-box-i00-privileged/task.toml
+++ b/tasks/behavior-preparing-lunch-box-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "behavior" # BEHAVIOR-1K v3.9.2 = OmniGibson 3.9.2 on Isa
robot = "r1pro"
family = "preparing_lunch_box"
mode = "privileged" # privileged state, unlimited resets, thin primitives (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (275 s on average); demonstration video (275 s): https://player.vimeo.com/video/1114057330 (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-07); human review in PR #71 (protocol §2.7)."
instance_seed = 301 # scene = environment/instance.json (2026 challenge public_test instance 301 + its state hash); identity = dataset.toml digest
diff --git a/tasks/behavior-preparing-lunch-box-i00-standard/task.toml b/tasks/behavior-preparing-lunch-box-i00-standard/task.toml
index 5f5e5728cc..1deb26c6dd 100644
--- a/tasks/behavior-preparing-lunch-box-i00-standard/task.toml
+++ b/tasks/behavior-preparing-lunch-box-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "behavior"
robot = "r1pro"
family = "preparing_lunch_box"
mode = "standard" # robot as a service: RGB-D + proprioception, no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (275 s on average); demonstration video (275 s): https://player.vimeo.com/video/1114057330 (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-07); human review in PR #71 (protocol §2.7)."
diff --git a/tasks/behavior-put-together-a-basic-pruning-kit-i00-privileged/task.toml b/tasks/behavior-put-together-a-basic-pruning-kit-i00-privileged/task.toml
index 21f44f483f..0a00d86375 100644
--- a/tasks/behavior-put-together-a-basic-pruning-kit-i00-privileged/task.toml
+++ b/tasks/behavior-put-together-a-basic-pruning-kit-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "behavior" # BEHAVIOR-1K v3.9.2 = OmniGibson 3.9.2 on Isa
robot = "r1pro"
family = "put_together_a_basic_pruning_kit"
mode = "privileged" # privileged state, unlimited resets, thin primitives (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (375 s on average); demonstration video (448 s): https://www.youtube.com/embed/UmGRVx-OCac (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-04); human review per family (protocol §2.7)."
instance_seed = 301 # scene = environment/instance.json (2026 challenge public_test instance 301 + its state hash); identity = dataset.toml digest
diff --git a/tasks/behavior-put-together-a-basic-pruning-kit-i00-standard/task.toml b/tasks/behavior-put-together-a-basic-pruning-kit-i00-standard/task.toml
index 2e742e2582..ed44e50746 100644
--- a/tasks/behavior-put-together-a-basic-pruning-kit-i00-standard/task.toml
+++ b/tasks/behavior-put-together-a-basic-pruning-kit-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "behavior"
robot = "r1pro"
family = "put_together_a_basic_pruning_kit"
mode = "standard" # robot as a service: RGB-D + proprioception, no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (375 s on average); demonstration video (448 s): https://www.youtube.com/embed/UmGRVx-OCac (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-04); human review per family (protocol §2.7)."
diff --git a/tasks/behavior-putting-away-Halloween-decorations-i00-privileged/task.toml b/tasks/behavior-putting-away-Halloween-decorations-i00-privileged/task.toml
index 20b6a64a94..1a491ab5b4 100644
--- a/tasks/behavior-putting-away-Halloween-decorations-i00-privileged/task.toml
+++ b/tasks/behavior-putting-away-Halloween-decorations-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "behavior" # BEHAVIOR-1K v3.9.2 = OmniGibson 3.9.2 on Isa
robot = "r1pro"
family = "putting_away_Halloween_decorations"
mode = "privileged" # privileged state, unlimited resets, thin primitives (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (460 s on average); demonstration video (460 s): https://player.vimeo.com/video/1114054409 (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-04); human review per family (protocol §2.7)."
instance_seed = 301 # scene = environment/instance.json (2026 challenge public_test instance 301 + its state hash); identity = dataset.toml digest
diff --git a/tasks/behavior-putting-away-Halloween-decorations-i00-standard/task.toml b/tasks/behavior-putting-away-Halloween-decorations-i00-standard/task.toml
index d85270a7ad..cca672884c 100644
--- a/tasks/behavior-putting-away-Halloween-decorations-i00-standard/task.toml
+++ b/tasks/behavior-putting-away-Halloween-decorations-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "behavior"
robot = "r1pro"
family = "putting_away_Halloween_decorations"
mode = "standard" # robot as a service: RGB-D + proprioception, no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (460 s on average); demonstration video (460 s): https://player.vimeo.com/video/1114054409 (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-04); human review per family (protocol §2.7)."
diff --git a/tasks/behavior-putting-away-toys-i00-privileged/task.toml b/tasks/behavior-putting-away-toys-i00-privileged/task.toml
index 782d79c014..f099dc46dd 100644
--- a/tasks/behavior-putting-away-toys-i00-privileged/task.toml
+++ b/tasks/behavior-putting-away-toys-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "behavior" # BEHAVIOR-1K v3.9.2 = OmniGibson 3.9.2 on Isa
robot = "r1pro"
family = "putting_away_toys"
mode = "privileged" # privileged state, unlimited resets, thin primitives (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (377 s on average); demonstration video (438 s): https://www.youtube.com/embed/StbSZCTn7D4 (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-04); human review per family (protocol §2.7)."
instance_seed = 301 # scene = environment/instance.json (2026 challenge public_test instance 301 + its state hash); identity = dataset.toml digest
diff --git a/tasks/behavior-putting-away-toys-i00-standard/task.toml b/tasks/behavior-putting-away-toys-i00-standard/task.toml
index cc44620d09..429b29f52e 100644
--- a/tasks/behavior-putting-away-toys-i00-standard/task.toml
+++ b/tasks/behavior-putting-away-toys-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "behavior"
robot = "r1pro"
family = "putting_away_toys"
mode = "standard" # robot as a service: RGB-D + proprioception, no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (377 s on average); demonstration video (438 s): https://www.youtube.com/embed/StbSZCTn7D4 (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-04); human review per family (protocol §2.7)."
diff --git a/tasks/behavior-putting-dirty-dishes-in-sink-i00-privileged/task.toml b/tasks/behavior-putting-dirty-dishes-in-sink-i00-privileged/task.toml
index 06f53b1e6b..23a744b1c4 100644
--- a/tasks/behavior-putting-dirty-dishes-in-sink-i00-privileged/task.toml
+++ b/tasks/behavior-putting-dirty-dishes-in-sink-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "behavior" # BEHAVIOR-1K v3.9.2 = OmniGibson 3.9.2 on Isa
robot = "r1pro"
family = "putting_dirty_dishes_in_sink"
mode = "privileged" # privileged state, unlimited resets, thin primitives (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (405 s on average); demonstration video (337 s): https://www.youtube.com/embed/SCX0NPFlMhI (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-04); human review per family (protocol §2.7)."
instance_seed = 301 # scene = environment/instance.json (2026 challenge public_test instance 301 + its state hash); identity = dataset.toml digest
diff --git a/tasks/behavior-putting-dirty-dishes-in-sink-i00-standard/task.toml b/tasks/behavior-putting-dirty-dishes-in-sink-i00-standard/task.toml
index e85986dfa2..a1dd8e29d3 100644
--- a/tasks/behavior-putting-dirty-dishes-in-sink-i00-standard/task.toml
+++ b/tasks/behavior-putting-dirty-dishes-in-sink-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "behavior"
robot = "r1pro"
family = "putting_dirty_dishes_in_sink"
mode = "standard" # robot as a service: RGB-D + proprioception, no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (405 s on average); demonstration video (337 s): https://www.youtube.com/embed/SCX0NPFlMhI (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-04); human review per family (protocol §2.7)."
diff --git a/tasks/behavior-putting-dishes-away-after-cleaning-i00-privileged/task.toml b/tasks/behavior-putting-dishes-away-after-cleaning-i00-privileged/task.toml
index d2b9923478..7677c57c5a 100644
--- a/tasks/behavior-putting-dishes-away-after-cleaning-i00-privileged/task.toml
+++ b/tasks/behavior-putting-dishes-away-after-cleaning-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "behavior" # BEHAVIOR-1K v3.9.2 = OmniGibson 3.9.2 on Isa
robot = "r1pro"
family = "putting_dishes_away_after_cleaning"
mode = "privileged" # privileged state, unlimited resets, thin primitives (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (365 s on average); demonstration video (365 s): https://player.vimeo.com/video/1114057082 (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-04); human review per family (protocol §2.7)."
instance_seed = 301 # scene = environment/instance.json (2026 challenge public_test instance 301 + its state hash); identity = dataset.toml digest
diff --git a/tasks/behavior-putting-dishes-away-after-cleaning-i00-standard/task.toml b/tasks/behavior-putting-dishes-away-after-cleaning-i00-standard/task.toml
index c4c53da381..645ac68919 100644
--- a/tasks/behavior-putting-dishes-away-after-cleaning-i00-standard/task.toml
+++ b/tasks/behavior-putting-dishes-away-after-cleaning-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "behavior"
robot = "r1pro"
family = "putting_dishes_away_after_cleaning"
mode = "standard" # robot as a service: RGB-D + proprioception, no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (365 s on average); demonstration video (365 s): https://player.vimeo.com/video/1114057082 (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-04); human review per family (protocol §2.7)."
diff --git a/tasks/behavior-putting-shoes-on-rack-i00-privileged/task.toml b/tasks/behavior-putting-shoes-on-rack-i00-privileged/task.toml
index adb92e1bcd..92a5b2dd3e 100644
--- a/tasks/behavior-putting-shoes-on-rack-i00-privileged/task.toml
+++ b/tasks/behavior-putting-shoes-on-rack-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "behavior" # BEHAVIOR-1K v3.9.2 = OmniGibson 3.9.2 on Isa
robot = "r1pro"
family = "putting_shoes_on_rack"
mode = "privileged" # privileged state, unlimited resets, thin primitives (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "partial"
oracle_evidence = "Reference solution solution/oracle_putting_shoes_on_rack.py on the frozen task-level primitives solution/rcb_behavior_ref.py. Recorded Harbor oracle run (laptop, a pre-v1.0 uncommitted tree): 2026-09-24 behavior-putting-shoes-on-rack-i__LkouR9y (job behavior-putting-shoes-on-rack-i00-oracle-0924-234514) scored 1 (6283 control steps, deterministic 1, instance_exact 1), 1 of 1. Model trial behavior-putting-shoes-on-rack-i__KozJnde (behavior-lunaxh-0925-2057, ws) scored 1. The tooling and verifier files changed since; not re-run on this version (static validation only). Measured 2026-10-10 on main 332bf8080 (board 0067 oracle sweep, AWS 8-vCPU L40S host rcb-gpu13, Harbor oracle agent, once): the reference solution's own run ended without success (q_score 0.8, 8921 control steps against the recorded 6283); the frozen scene rebuilt there within tolerance but not bit-exactly (instance_exact 0), so the solution's online planning took another path: this reference solution does not succeed on that host class."
instance_seed = 301 # scene = environment/instance.json (2026 challenge public_test instance 301 + its state hash); identity = dataset.toml digest
diff --git a/tasks/behavior-putting-shoes-on-rack-i00-standard/task.toml b/tasks/behavior-putting-shoes-on-rack-i00-standard/task.toml
index 879e803da0..05fe62d25b 100644
--- a/tasks/behavior-putting-shoes-on-rack-i00-standard/task.toml
+++ b/tasks/behavior-putting-shoes-on-rack-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "behavior"
robot = "r1pro"
family = "putting_shoes_on_rack"
mode = "standard" # robot as a service: RGB-D + proprioception, no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "partial"
oracle_evidence = "solution/ streams the privileged oracle's joint targets (oracle_qpos.npy, 6283 rows) through the client. Recorded Harbor oracle run (laptop, a pre-v1.0 version): 2026-09-25 behavior-putting-shoes-on-rack-i__XB2NUoX (job behavior-putting-shoes-on-rack-i00-limited-oracle-0925-015349) scored 1 with live_matches_replay 1 (6283 control steps), 1 of 1. The sim service, collect hook, grader and oracle client changed since (crash recovery); not re-run on this version (static validation only)."
diff --git a/tasks/behavior-putting-up-Christmas-decorations-inside-i00-privileged/task.toml b/tasks/behavior-putting-up-Christmas-decorations-inside-i00-privileged/task.toml
index 74550f44c1..8b703cbff3 100644
--- a/tasks/behavior-putting-up-Christmas-decorations-inside-i00-privileged/task.toml
+++ b/tasks/behavior-putting-up-Christmas-decorations-inside-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "behavior" # BEHAVIOR-1K v3.9.2 = OmniGibson 3.9.2 on Isa
robot = "r1pro"
family = "putting_up_Christmas_decorations_inside"
mode = "privileged" # privileged state, unlimited resets, thin primitives (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (457 s on average); demonstration video (457 s): https://player.vimeo.com/video/1114056731 (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-04); human review per family (protocol §2.7)."
instance_seed = 301 # scene = environment/instance.json (2026 challenge public_test instance 301 + its state hash); identity = dataset.toml digest
diff --git a/tasks/behavior-putting-up-Christmas-decorations-inside-i00-standard/task.toml b/tasks/behavior-putting-up-Christmas-decorations-inside-i00-standard/task.toml
index 8f32c50eea..05bf04494c 100644
--- a/tasks/behavior-putting-up-Christmas-decorations-inside-i00-standard/task.toml
+++ b/tasks/behavior-putting-up-Christmas-decorations-inside-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "behavior"
robot = "r1pro"
family = "putting_up_Christmas_decorations_inside"
mode = "standard" # robot as a service: RGB-D + proprioception, no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (457 s on average); demonstration video (457 s): https://player.vimeo.com/video/1114056731 (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-04); human review per family (protocol §2.7)."
diff --git a/tasks/behavior-re-shelving-library-books-i00-privileged/task.toml b/tasks/behavior-re-shelving-library-books-i00-privileged/task.toml
index d3b8560571..6da524e2c9 100644
--- a/tasks/behavior-re-shelving-library-books-i00-privileged/task.toml
+++ b/tasks/behavior-re-shelving-library-books-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "behavior" # BEHAVIOR-1K v3.9.2 = OmniGibson 3.9.2 on Isa
robot = "r1pro"
family = "re_shelving_library_books"
mode = "privileged" # privileged state, unlimited resets, thin primitives (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (318 s on average); demonstration video (281 s): https://www.youtube.com/embed/AQdhTrye7NQ (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-07); human review in PR #75 (protocol §2.7)."
instance_seed = 301 # scene = environment/instance.json (2026 challenge public_test instance 301 + its state hash); identity = dataset.toml digest
diff --git a/tasks/behavior-re-shelving-library-books-i00-standard/task.toml b/tasks/behavior-re-shelving-library-books-i00-standard/task.toml
index 73259939b3..67a6bc3941 100644
--- a/tasks/behavior-re-shelving-library-books-i00-standard/task.toml
+++ b/tasks/behavior-re-shelving-library-books-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "behavior"
robot = "r1pro"
family = "re_shelving_library_books"
mode = "standard" # robot as a service: RGB-D + proprioception, no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (318 s on average); demonstration video (281 s): https://www.youtube.com/embed/AQdhTrye7NQ (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-07); human review in PR #75 (protocol §2.7)."
diff --git a/tasks/behavior-rearrange-your-room-i00-privileged/task.toml b/tasks/behavior-rearrange-your-room-i00-privileged/task.toml
index dec20a2044..026fd50344 100644
--- a/tasks/behavior-rearrange-your-room-i00-privileged/task.toml
+++ b/tasks/behavior-rearrange-your-room-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "behavior" # BEHAVIOR-1K v3.9.2 = OmniGibson 3.9.2 on Isa
robot = "r1pro"
family = "rearrange_your_room"
mode = "privileged" # privileged state, unlimited resets, thin primitives (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (428 s on average); demonstration video (515 s): https://www.youtube.com/embed/PuP9DeUF4K4 (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-04); human review per family (protocol §2.7)."
instance_seed = 301 # scene = environment/instance.json (2026 challenge public_test instance 301 + its state hash); identity = dataset.toml digest
diff --git a/tasks/behavior-rearrange-your-room-i00-standard/task.toml b/tasks/behavior-rearrange-your-room-i00-standard/task.toml
index 4edc0ccedc..ccb2b17ca9 100644
--- a/tasks/behavior-rearrange-your-room-i00-standard/task.toml
+++ b/tasks/behavior-rearrange-your-room-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "behavior"
robot = "r1pro"
family = "rearrange_your_room"
mode = "standard" # robot as a service: RGB-D + proprioception, no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (428 s on average); demonstration video (515 s): https://www.youtube.com/embed/PuP9DeUF4K4 (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-04); human review per family (protocol §2.7)."
diff --git a/tasks/behavior-rearranging-kitchen-furniture-i00-privileged/task.toml b/tasks/behavior-rearranging-kitchen-furniture-i00-privileged/task.toml
index ebce97d1c2..9af6a218e8 100644
--- a/tasks/behavior-rearranging-kitchen-furniture-i00-privileged/task.toml
+++ b/tasks/behavior-rearranging-kitchen-furniture-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "behavior" # BEHAVIOR-1K v3.9.2 = OmniGibson 3.9.2 on Isa
robot = "r1pro"
family = "rearranging_kitchen_furniture"
mode = "privileged" # privileged state, unlimited resets, thin primitives (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (298 s on average); demonstration video (298 s): https://player.vimeo.com/video/1114056046 (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-07); human review in PR #71 (protocol §2.7)."
instance_seed = 301 # scene = environment/instance.json (2026 challenge public_test instance 301 + its state hash); identity = dataset.toml digest
diff --git a/tasks/behavior-rearranging-kitchen-furniture-i00-standard/task.toml b/tasks/behavior-rearranging-kitchen-furniture-i00-standard/task.toml
index 9ffc3d8dba..34fdaf3f7f 100644
--- a/tasks/behavior-rearranging-kitchen-furniture-i00-standard/task.toml
+++ b/tasks/behavior-rearranging-kitchen-furniture-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "behavior"
robot = "r1pro"
family = "rearranging_kitchen_furniture"
mode = "standard" # robot as a service: RGB-D + proprioception, no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (298 s on average); demonstration video (298 s): https://player.vimeo.com/video/1114056046 (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-07); human review in PR #71 (protocol §2.7)."
diff --git a/tasks/behavior-scrubbing-bathroom-floor-i00-privileged/task.toml b/tasks/behavior-scrubbing-bathroom-floor-i00-privileged/task.toml
index 124b61e325..0059dbd917 100644
--- a/tasks/behavior-scrubbing-bathroom-floor-i00-privileged/task.toml
+++ b/tasks/behavior-scrubbing-bathroom-floor-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "behavior" # BEHAVIOR-1K v3.9.2 = OmniGibson 3.9.2 on Isa
robot = "r1pro"
family = "scrubbing_bathroom_floor"
mode = "privileged" # privileged state, unlimited resets, thin primitives (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (105 s on average); demonstration video (123 s): https://www.youtube.com/embed/u9wtUgoB8o8 (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-07); human review in PR #75 (protocol §2.7)."
instance_seed = 301 # scene = environment/instance.json (2026 challenge public_test instance 301 + its state hash); identity = dataset.toml digest
diff --git a/tasks/behavior-scrubbing-bathroom-floor-i00-standard/task.toml b/tasks/behavior-scrubbing-bathroom-floor-i00-standard/task.toml
index da00242ed3..f689d4e6b7 100644
--- a/tasks/behavior-scrubbing-bathroom-floor-i00-standard/task.toml
+++ b/tasks/behavior-scrubbing-bathroom-floor-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "behavior"
robot = "r1pro"
family = "scrubbing_bathroom_floor"
mode = "standard" # robot as a service: RGB-D + proprioception, no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (105 s on average); demonstration video (123 s): https://www.youtube.com/embed/u9wtUgoB8o8 (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-07); human review in PR #75 (protocol §2.7)."
diff --git a/tasks/behavior-set-up-a-coffee-station-in-your-kitchen-i00-privileged/task.toml b/tasks/behavior-set-up-a-coffee-station-in-your-kitchen-i00-privileged/task.toml
index 1a6a990571..91686e93d4 100644
--- a/tasks/behavior-set-up-a-coffee-station-in-your-kitchen-i00-privileged/task.toml
+++ b/tasks/behavior-set-up-a-coffee-station-in-your-kitchen-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "behavior" # BEHAVIOR-1K v3.9.2 = OmniGibson 3.9.2 on Isa
robot = "r1pro"
family = "set_up_a_coffee_station_in_your_kitchen"
mode = "privileged" # privileged state, unlimited resets, thin primitives (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (209 s on average); demonstration video (209 s): https://player.vimeo.com/video/1114056931 (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-07); human review in PR #71 (protocol §2.7)."
instance_seed = 301 # scene = environment/instance.json (2026 challenge public_test instance 301 + its state hash); identity = dataset.toml digest
diff --git a/tasks/behavior-set-up-a-coffee-station-in-your-kitchen-i00-standard/task.toml b/tasks/behavior-set-up-a-coffee-station-in-your-kitchen-i00-standard/task.toml
index 1c6c5eaf1e..eb8b74ea13 100644
--- a/tasks/behavior-set-up-a-coffee-station-in-your-kitchen-i00-standard/task.toml
+++ b/tasks/behavior-set-up-a-coffee-station-in-your-kitchen-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "behavior"
robot = "r1pro"
family = "set_up_a_coffee_station_in_your_kitchen"
mode = "standard" # robot as a service: RGB-D + proprioception, no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (209 s on average); demonstration video (209 s): https://player.vimeo.com/video/1114056931 (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-07); human review in PR #71 (protocol §2.7)."
diff --git a/tasks/behavior-setting-mousetraps-i00-privileged/task.toml b/tasks/behavior-setting-mousetraps-i00-privileged/task.toml
index 216963db0c..448c9410b5 100644
--- a/tasks/behavior-setting-mousetraps-i00-privileged/task.toml
+++ b/tasks/behavior-setting-mousetraps-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "behavior" # BEHAVIOR-1K v3.9.2 = OmniGibson 3.9.2 on Isa
robot = "r1pro"
family = "setting_mousetraps"
mode = "privileged" # privileged state, unlimited resets, thin primitives (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (340 s on average); demonstration video (339 s): https://player.vimeo.com/video/1114054686 (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-04); human review per family (protocol §2.7)."
instance_seed = 301 # scene = environment/instance.json (2026 challenge public_test instance 301 + its state hash); identity = dataset.toml digest
diff --git a/tasks/behavior-setting-mousetraps-i00-standard/task.toml b/tasks/behavior-setting-mousetraps-i00-standard/task.toml
index 1aad4bbbe0..9abb39815a 100644
--- a/tasks/behavior-setting-mousetraps-i00-standard/task.toml
+++ b/tasks/behavior-setting-mousetraps-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "behavior"
robot = "r1pro"
family = "setting_mousetraps"
mode = "standard" # robot as a service: RGB-D + proprioception, no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (340 s on average); demonstration video (339 s): https://player.vimeo.com/video/1114054686 (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-04); human review per family (protocol §2.7)."
diff --git a/tasks/behavior-setting-the-fire-i00-privileged/task.toml b/tasks/behavior-setting-the-fire-i00-privileged/task.toml
index c0bfbd8ee3..26147e45c0 100644
--- a/tasks/behavior-setting-the-fire-i00-privileged/task.toml
+++ b/tasks/behavior-setting-the-fire-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "behavior" # BEHAVIOR-1K v3.9.2 = OmniGibson 3.9.2 on Isa
robot = "r1pro"
family = "setting_the_fire"
mode = "privileged" # privileged state, unlimited resets, thin primitives (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (304 s on average); demonstration video (304 s): https://player.vimeo.com/video/1114059945 (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-04); human review per family (protocol §2.7)."
instance_seed = 301 # scene = environment/instance.json (2026 challenge public_test instance 301 + its state hash); identity = dataset.toml digest
diff --git a/tasks/behavior-setting-the-fire-i00-standard/task.toml b/tasks/behavior-setting-the-fire-i00-standard/task.toml
index 67361681ec..0d70bcba32 100644
--- a/tasks/behavior-setting-the-fire-i00-standard/task.toml
+++ b/tasks/behavior-setting-the-fire-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "behavior"
robot = "r1pro"
family = "setting_the_fire"
mode = "standard" # robot as a service: RGB-D + proprioception, no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (304 s on average); demonstration video (304 s): https://player.vimeo.com/video/1114059945 (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-04); human review per family (protocol §2.7)."
diff --git a/tasks/behavior-setting-the-table-i00-privileged/task.toml b/tasks/behavior-setting-the-table-i00-privileged/task.toml
index 91c3bd6ca9..5ae9de1f8a 100644
--- a/tasks/behavior-setting-the-table-i00-privileged/task.toml
+++ b/tasks/behavior-setting-the-table-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "behavior" # BEHAVIOR-1K v3.9.2 = OmniGibson 3.9.2 on Isa
robot = "r1pro"
family = "setting_the_table"
mode = "privileged" # privileged state, unlimited resets, thin primitives (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (594 s on average); demonstration video (727 s): https://www.youtube.com/embed/Tkf34ap-z8U (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-04); human review per family (protocol §2.7)."
instance_seed = 301 # scene = environment/instance.json (2026 challenge public_test instance 301 + its state hash); identity = dataset.toml digest
diff --git a/tasks/behavior-setting-the-table-i00-standard/task.toml b/tasks/behavior-setting-the-table-i00-standard/task.toml
index 606494724e..8f306b5a06 100644
--- a/tasks/behavior-setting-the-table-i00-standard/task.toml
+++ b/tasks/behavior-setting-the-table-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "behavior"
robot = "r1pro"
family = "setting_the_table"
mode = "standard" # robot as a service: RGB-D + proprioception, no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (594 s on average); demonstration video (727 s): https://www.youtube.com/embed/Tkf34ap-z8U (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-04); human review per family (protocol §2.7)."
diff --git a/tasks/behavior-setup-a-bar-for-a-cocktail-party-i00-privileged/task.toml b/tasks/behavior-setup-a-bar-for-a-cocktail-party-i00-privileged/task.toml
index 50487a27f4..cb7820eb73 100644
--- a/tasks/behavior-setup-a-bar-for-a-cocktail-party-i00-privileged/task.toml
+++ b/tasks/behavior-setup-a-bar-for-a-cocktail-party-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "behavior" # BEHAVIOR-1K v3.9.2 = OmniGibson 3.9.2 on Isa
robot = "r1pro"
family = "setup_a_bar_for_a_cocktail_party"
mode = "privileged" # privileged state, unlimited resets, thin primitives (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (447 s on average); demonstration video (350 s): https://www.youtube.com/embed/3V0ySKyGKRw (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-04); human review per family (protocol §2.7)."
instance_seed = 301 # scene = environment/instance.json (2026 challenge public_test instance 301 + its state hash); identity = dataset.toml digest
diff --git a/tasks/behavior-setup-a-bar-for-a-cocktail-party-i00-standard/task.toml b/tasks/behavior-setup-a-bar-for-a-cocktail-party-i00-standard/task.toml
index 3fd4363792..818a5daa76 100644
--- a/tasks/behavior-setup-a-bar-for-a-cocktail-party-i00-standard/task.toml
+++ b/tasks/behavior-setup-a-bar-for-a-cocktail-party-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "behavior"
robot = "r1pro"
family = "setup_a_bar_for_a_cocktail_party"
mode = "standard" # robot as a service: RGB-D + proprioception, no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (447 s on average); demonstration video (350 s): https://www.youtube.com/embed/3V0ySKyGKRw (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-04); human review per family (protocol §2.7)."
diff --git a/tasks/behavior-slicing-vegetables-i00-privileged/task.toml b/tasks/behavior-slicing-vegetables-i00-privileged/task.toml
index dad02f5bca..1cec232a52 100644
--- a/tasks/behavior-slicing-vegetables-i00-privileged/task.toml
+++ b/tasks/behavior-slicing-vegetables-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "behavior" # BEHAVIOR-1K v3.9.2 = OmniGibson 3.9.2 on Isa
robot = "r1pro"
family = "slicing_vegetables"
mode = "privileged" # privileged state, unlimited resets, thin primitives (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (495 s on average); demonstration video (495 s): https://player.vimeo.com/video/1114061183 (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-04); human review per family (protocol §2.7)."
instance_seed = 301 # scene = environment/instance.json (2026 challenge public_test instance 301 + its state hash); identity = dataset.toml digest
diff --git a/tasks/behavior-slicing-vegetables-i00-standard/task.toml b/tasks/behavior-slicing-vegetables-i00-standard/task.toml
index 1b25544406..086bc844f6 100644
--- a/tasks/behavior-slicing-vegetables-i00-standard/task.toml
+++ b/tasks/behavior-slicing-vegetables-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "behavior"
robot = "r1pro"
family = "slicing_vegetables"
mode = "standard" # robot as a service: RGB-D + proprioception, no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (495 s on average); demonstration video (495 s): https://player.vimeo.com/video/1114061183 (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-04); human review per family (protocol §2.7)."
diff --git a/tasks/behavior-sorting-books-on-shelf-i00-privileged/task.toml b/tasks/behavior-sorting-books-on-shelf-i00-privileged/task.toml
index bebed8f4a5..be4c7afb9b 100644
--- a/tasks/behavior-sorting-books-on-shelf-i00-privileged/task.toml
+++ b/tasks/behavior-sorting-books-on-shelf-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "behavior" # BEHAVIOR-1K v3.9.2 = OmniGibson 3.9.2 on Isa
robot = "r1pro"
family = "sorting_books_on_shelf"
mode = "privileged" # privileged state, unlimited resets, thin primitives (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (258 s on average); demonstration video (200 s): https://www.youtube.com/embed/wXUyDWyvpRA (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-07); human review in PR #71 (protocol §2.7)."
instance_seed = 301 # scene = environment/instance.json (2026 challenge public_test instance 301 + its state hash); identity = dataset.toml digest
diff --git a/tasks/behavior-sorting-books-on-shelf-i00-standard/task.toml b/tasks/behavior-sorting-books-on-shelf-i00-standard/task.toml
index eec36ae249..6ba6960a8f 100644
--- a/tasks/behavior-sorting-books-on-shelf-i00-standard/task.toml
+++ b/tasks/behavior-sorting-books-on-shelf-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "behavior"
robot = "r1pro"
family = "sorting_books_on_shelf"
mode = "standard" # robot as a service: RGB-D + proprioception, no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (258 s on average); demonstration video (200 s): https://www.youtube.com/embed/wXUyDWyvpRA (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-07); human review in PR #71 (protocol §2.7)."
diff --git a/tasks/behavior-sorting-bottles-cans-and-paper-i00-privileged/task.toml b/tasks/behavior-sorting-bottles-cans-and-paper-i00-privileged/task.toml
index 1c632ee839..b773348e8e 100644
--- a/tasks/behavior-sorting-bottles-cans-and-paper-i00-privileged/task.toml
+++ b/tasks/behavior-sorting-bottles-cans-and-paper-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "behavior" # BEHAVIOR-1K v3.9.2 = OmniGibson 3.9.2 on Isa
robot = "r1pro"
family = "sorting_bottles_cans_and_paper"
mode = "privileged" # privileged state, unlimited resets, thin primitives (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (342 s on average); demonstration video (312 s): https://www.youtube.com/embed/h8-XZ85UxSU (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-07); human review in PR #75 (protocol §2.7)."
instance_seed = 301 # scene = environment/instance.json (2026 challenge public_test instance 301 + its state hash); identity = dataset.toml digest
diff --git a/tasks/behavior-sorting-bottles-cans-and-paper-i00-standard/task.toml b/tasks/behavior-sorting-bottles-cans-and-paper-i00-standard/task.toml
index f1abcc1214..021367816f 100644
--- a/tasks/behavior-sorting-bottles-cans-and-paper-i00-standard/task.toml
+++ b/tasks/behavior-sorting-bottles-cans-and-paper-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "behavior"
robot = "r1pro"
family = "sorting_bottles_cans_and_paper"
mode = "standard" # robot as a service: RGB-D + proprioception, no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (342 s on average); demonstration video (312 s): https://www.youtube.com/embed/h8-XZ85UxSU (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-07); human review in PR #75 (protocol §2.7)."
diff --git a/tasks/behavior-sorting-household-items-i00-privileged/task.toml b/tasks/behavior-sorting-household-items-i00-privileged/task.toml
index ba2f522c15..f24f766940 100644
--- a/tasks/behavior-sorting-household-items-i00-privileged/task.toml
+++ b/tasks/behavior-sorting-household-items-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "behavior" # BEHAVIOR-1K v3.9.2 = OmniGibson 3.9.2 on Isa
robot = "r1pro"
family = "sorting_household_items"
mode = "privileged" # privileged state, unlimited resets, thin primitives (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (527 s on average); demonstration video (527 s): https://player.vimeo.com/video/1114059574 (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-04); human review per family (protocol §2.7)."
instance_seed = 301 # scene = environment/instance.json (2026 challenge public_test instance 301 + its state hash); identity = dataset.toml digest
diff --git a/tasks/behavior-sorting-household-items-i00-standard/task.toml b/tasks/behavior-sorting-household-items-i00-standard/task.toml
index 6d2b09105c..ee1ec9de5d 100644
--- a/tasks/behavior-sorting-household-items-i00-standard/task.toml
+++ b/tasks/behavior-sorting-household-items-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "behavior"
robot = "r1pro"
family = "sorting_household_items"
mode = "standard" # robot as a service: RGB-D + proprioception, no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (527 s on average); demonstration video (527 s): https://player.vimeo.com/video/1114059574 (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-04); human review per family (protocol §2.7)."
diff --git a/tasks/behavior-sorting-vegetables-i00-privileged/task.toml b/tasks/behavior-sorting-vegetables-i00-privileged/task.toml
index 19586cc6bb..b223ded9fd 100644
--- a/tasks/behavior-sorting-vegetables-i00-privileged/task.toml
+++ b/tasks/behavior-sorting-vegetables-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "behavior" # BEHAVIOR-1K v3.9.2 = OmniGibson 3.9.2 on Isa
robot = "r1pro"
family = "sorting_vegetables"
mode = "privileged" # privileged state, unlimited resets, thin primitives (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (397 s on average); demonstration video (397 s): https://player.vimeo.com/video/1114058741 (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-04); human review per family (protocol §2.7)."
instance_seed = 301 # scene = environment/instance.json (2026 challenge public_test instance 301 + its state hash); identity = dataset.toml digest
diff --git a/tasks/behavior-sorting-vegetables-i00-standard/task.toml b/tasks/behavior-sorting-vegetables-i00-standard/task.toml
index 074d343782..2bd7186789 100644
--- a/tasks/behavior-sorting-vegetables-i00-standard/task.toml
+++ b/tasks/behavior-sorting-vegetables-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "behavior"
robot = "r1pro"
family = "sorting_vegetables"
mode = "standard" # robot as a service: RGB-D + proprioception, no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (397 s on average); demonstration video (397 s): https://player.vimeo.com/video/1114058741 (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-04); human review per family (protocol §2.7)."
diff --git a/tasks/behavior-spraying-for-bugs-i00-privileged/task.toml b/tasks/behavior-spraying-for-bugs-i00-privileged/task.toml
index ef9f4ca82f..c15428ba55 100644
--- a/tasks/behavior-spraying-for-bugs-i00-privileged/task.toml
+++ b/tasks/behavior-spraying-for-bugs-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "behavior" # BEHAVIOR-1K v3.9.2 = OmniGibson 3.9.2 on Isa
robot = "r1pro"
family = "spraying_for_bugs"
mode = "privileged" # privileged state, unlimited resets, thin primitives (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (216 s on average); demonstration video (216 s): https://player.vimeo.com/video/1114060681 (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-07); human review in PR #75 (protocol §2.7)."
instance_seed = 301 # scene = environment/instance.json (2026 challenge public_test instance 301 + its state hash); identity = dataset.toml digest
diff --git a/tasks/behavior-spraying-for-bugs-i00-standard/task.toml b/tasks/behavior-spraying-for-bugs-i00-standard/task.toml
index 1a776c5a5e..4f35f07806 100644
--- a/tasks/behavior-spraying-for-bugs-i00-standard/task.toml
+++ b/tasks/behavior-spraying-for-bugs-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "behavior"
robot = "r1pro"
family = "spraying_for_bugs"
mode = "standard" # robot as a service: RGB-D + proprioception, no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (216 s on average); demonstration video (216 s): https://player.vimeo.com/video/1114060681 (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-07); human review in PR #75 (protocol §2.7)."
diff --git a/tasks/behavior-spraying-fruit-trees-i00-privileged/task.toml b/tasks/behavior-spraying-fruit-trees-i00-privileged/task.toml
index 14e1366d77..3d7a6013a7 100644
--- a/tasks/behavior-spraying-fruit-trees-i00-privileged/task.toml
+++ b/tasks/behavior-spraying-fruit-trees-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "behavior" # BEHAVIOR-1K v3.9.2 = OmniGibson 3.9.2 on Isa
robot = "r1pro"
family = "spraying_fruit_trees"
mode = "privileged" # privileged state, unlimited resets, thin primitives (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (278 s on average); demonstration video (278 s): https://player.vimeo.com/video/1114060748 (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-07); human review in PR #75 (protocol §2.7)."
instance_seed = 301 # scene = environment/instance.json (2026 challenge public_test instance 301 + its state hash); identity = dataset.toml digest
diff --git a/tasks/behavior-spraying-fruit-trees-i00-standard/task.toml b/tasks/behavior-spraying-fruit-trees-i00-standard/task.toml
index 66b3306e75..229c9a92f6 100644
--- a/tasks/behavior-spraying-fruit-trees-i00-standard/task.toml
+++ b/tasks/behavior-spraying-fruit-trees-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "behavior"
robot = "r1pro"
family = "spraying_fruit_trees"
mode = "standard" # robot as a service: RGB-D + proprioception, no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (278 s on average); demonstration video (278 s): https://player.vimeo.com/video/1114060748 (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-07); human review in PR #75 (protocol §2.7)."
diff --git a/tasks/behavior-stacking-wood-i00-privileged/task.toml b/tasks/behavior-stacking-wood-i00-privileged/task.toml
index f010d393c4..9c54597b03 100644
--- a/tasks/behavior-stacking-wood-i00-privileged/task.toml
+++ b/tasks/behavior-stacking-wood-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "behavior" # BEHAVIOR-1K v3.9.2 = OmniGibson 3.9.2 on Isa
robot = "r1pro"
family = "stacking_wood"
mode = "privileged" # privileged state, unlimited resets, thin primitives (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance. Solvable: model trial behavior-stacking-wood-i00__48hThC9 (behavior-lunaxh-t8-0927, privileged) graded success=1, deterministic=1 by the separate verifier on a pre-v1.0 version of this task (dev/qineng 912567fa; grader verdict and frozen instance unchanged since); agent log https://embodied-agent-interface-v2-internal.github.io/runs/behavior-1k/codex-gpt6_luna-xhigh/stacking_wood/unlimited/, verifier replay https://huggingface.co/datasets/eai-v2-internal/agent-runs/resolve/main/behavior-1k/codex-gpt6_luna-xhigh/stacking_wood/unlimited/replay.mp4; not re-run on this version (static validation only). Human review per family (protocol §2.7)."
instance_seed = 301 # scene = environment/instance.json (2026 challenge public_test instance 301 + its state hash); identity = dataset.toml digest
diff --git a/tasks/behavior-stacking-wood-i00-standard/task.toml b/tasks/behavior-stacking-wood-i00-standard/task.toml
index 9d53c829b0..b34098aaa0 100644
--- a/tasks/behavior-stacking-wood-i00-standard/task.toml
+++ b/tasks/behavior-stacking-wood-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "behavior"
robot = "r1pro"
family = "stacking_wood"
mode = "standard" # robot as a service: RGB-D + proprioception, no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance. Solvable: model trial behavior-stacking-wood-i00__48hThC9 (behavior-lunaxh-t8-0927, the privileged twin behavior-stacking-wood-i00-privileged on the same frozen instance and BDDL goal) graded success=1, deterministic=1 by the separate verifier on a pre-v1.0 version of this task (dev/qineng 912567fa; grader verdict and frozen instance unchanged since); agent log https://embodied-agent-interface-v2-internal.github.io/runs/behavior-1k/codex-gpt6_luna-xhigh/stacking_wood/unlimited/, verifier replay https://huggingface.co/datasets/eai-v2-internal/agent-runs/resolve/main/behavior-1k/codex-gpt6_luna-xhigh/stacking_wood/unlimited/replay.mp4; not re-run on this version (static validation only). Human review per family (protocol §2.7)."
diff --git a/tasks/behavior-store-batteries-i00-privileged/task.toml b/tasks/behavior-store-batteries-i00-privileged/task.toml
index 8fd1c0327d..b045c57398 100644
--- a/tasks/behavior-store-batteries-i00-privileged/task.toml
+++ b/tasks/behavior-store-batteries-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "behavior" # BEHAVIOR-1K v3.9.2 = OmniGibson 3.9.2 on Isa
robot = "r1pro"
family = "store_batteries"
mode = "privileged" # privileged state, unlimited resets, thin primitives (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (257 s on average); demonstration video (328 s): https://www.youtube.com/embed/jMSkFQy95FY (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-07); human review in PR #75 (protocol §2.7)."
instance_seed = 301 # scene = environment/instance.json (2026 challenge public_test instance 301 + its state hash); identity = dataset.toml digest
diff --git a/tasks/behavior-store-batteries-i00-standard/task.toml b/tasks/behavior-store-batteries-i00-standard/task.toml
index 7861f66062..e011e42c6d 100644
--- a/tasks/behavior-store-batteries-i00-standard/task.toml
+++ b/tasks/behavior-store-batteries-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "behavior"
robot = "r1pro"
family = "store_batteries"
mode = "standard" # robot as a service: RGB-D + proprioception, no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (257 s on average); demonstration video (328 s): https://www.youtube.com/embed/jMSkFQy95FY (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-07); human review in PR #75 (protocol §2.7)."
diff --git a/tasks/behavior-store-honey-i00-privileged/task.toml b/tasks/behavior-store-honey-i00-privileged/task.toml
index bca3341a0b..1597342caf 100644
--- a/tasks/behavior-store-honey-i00-privileged/task.toml
+++ b/tasks/behavior-store-honey-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "behavior" # BEHAVIOR-1K v3.9.2 = OmniGibson 3.9.2 on Isa
robot = "r1pro"
family = "store_honey"
mode = "privileged" # privileged state, unlimited resets, thin primitives (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (226 s on average); demonstration video (193 s): https://www.youtube.com/embed/oFu38YPgVEU (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-07); human review in PR #71 (protocol §2.7)."
instance_seed = 301 # scene = environment/instance.json (2026 challenge public_test instance 301 + its state hash); identity = dataset.toml digest
diff --git a/tasks/behavior-store-honey-i00-standard/task.toml b/tasks/behavior-store-honey-i00-standard/task.toml
index 23e01437cc..499c6fc9df 100644
--- a/tasks/behavior-store-honey-i00-standard/task.toml
+++ b/tasks/behavior-store-honey-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "behavior"
robot = "r1pro"
family = "store_honey"
mode = "standard" # robot as a service: RGB-D + proprioception, no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (226 s on average); demonstration video (193 s): https://www.youtube.com/embed/oFu38YPgVEU (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-07); human review in PR #71 (protocol §2.7)."
diff --git a/tasks/behavior-store-produce-i00-privileged/task.toml b/tasks/behavior-store-produce-i00-privileged/task.toml
index 18d6b979dd..a809b5d0f7 100644
--- a/tasks/behavior-store-produce-i00-privileged/task.toml
+++ b/tasks/behavior-store-produce-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "behavior" # BEHAVIOR-1K v3.9.2 = OmniGibson 3.9.2 on Isa
robot = "r1pro"
family = "store_produce"
mode = "privileged" # privileged state, unlimited resets, thin primitives (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (317 s on average); demonstration video (316 s): https://www.youtube.com/embed/jppOziwq7xI (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-07); human review in PR #71 (protocol §2.7)."
instance_seed = 301 # scene = environment/instance.json (2026 challenge public_test instance 301 + its state hash); identity = dataset.toml digest
diff --git a/tasks/behavior-store-produce-i00-standard/task.toml b/tasks/behavior-store-produce-i00-standard/task.toml
index 78806dc769..8ad904c87c 100644
--- a/tasks/behavior-store-produce-i00-standard/task.toml
+++ b/tasks/behavior-store-produce-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "behavior"
robot = "r1pro"
family = "store_produce"
mode = "standard" # robot as a service: RGB-D + proprioception, no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (317 s on average); demonstration video (316 s): https://www.youtube.com/embed/jppOziwq7xI (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-07); human review in PR #71 (protocol §2.7)."
diff --git a/tasks/behavior-storing-food-i00-privileged/task.toml b/tasks/behavior-storing-food-i00-privileged/task.toml
index ffdfc7811b..2a4a167ac5 100644
--- a/tasks/behavior-storing-food-i00-privileged/task.toml
+++ b/tasks/behavior-storing-food-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "behavior" # BEHAVIOR-1K v3.9.2 = OmniGibson 3.9.2 on Isa
robot = "r1pro"
family = "storing_food"
mode = "privileged" # privileged state, unlimited resets, thin primitives (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (662 s on average); demonstration video (662 s): https://player.vimeo.com/video/1114059272 (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-04); human review per family (protocol §2.7)."
instance_seed = 301 # scene = environment/instance.json (2026 challenge public_test instance 301 + its state hash); identity = dataset.toml digest
diff --git a/tasks/behavior-storing-food-i00-standard/task.toml b/tasks/behavior-storing-food-i00-standard/task.toml
index 8f8e468091..f7ceab2c78 100644
--- a/tasks/behavior-storing-food-i00-standard/task.toml
+++ b/tasks/behavior-storing-food-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "behavior"
robot = "r1pro"
family = "storing_food"
mode = "standard" # robot as a service: RGB-D + proprioception, no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (662 s on average); demonstration video (662 s): https://player.vimeo.com/video/1114059272 (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-04); human review per family (protocol §2.7)."
diff --git a/tasks/behavior-sweeping-garage-i00-privileged/task.toml b/tasks/behavior-sweeping-garage-i00-privileged/task.toml
index ef6e90f832..b11f4b11bd 100644
--- a/tasks/behavior-sweeping-garage-i00-privileged/task.toml
+++ b/tasks/behavior-sweeping-garage-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "behavior" # BEHAVIOR-1K v3.9.2 = OmniGibson 3.9.2 on Isa
robot = "r1pro"
family = "sweeping_garage"
mode = "privileged" # privileged state, unlimited resets, thin primitives (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (149 s on average); demonstration video (126 s): https://www.youtube.com/embed/0bdsrbEeFEc (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-07); human review in PR #75 (protocol §2.7)."
instance_seed = 301 # scene = environment/instance.json (2026 challenge public_test instance 301 + its state hash); identity = dataset.toml digest
diff --git a/tasks/behavior-sweeping-garage-i00-standard/task.toml b/tasks/behavior-sweeping-garage-i00-standard/task.toml
index 5edee5e8ef..02a101719c 100644
--- a/tasks/behavior-sweeping-garage-i00-standard/task.toml
+++ b/tasks/behavior-sweeping-garage-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "behavior"
robot = "r1pro"
family = "sweeping_garage"
mode = "standard" # robot as a service: RGB-D + proprioception, no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (149 s on average); demonstration video (126 s): https://www.youtube.com/embed/0bdsrbEeFEc (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-07); human review in PR #75 (protocol §2.7)."
diff --git a/tasks/behavior-thawing-frozen-food-i00-privileged/task.toml b/tasks/behavior-thawing-frozen-food-i00-privileged/task.toml
index 4947913a68..5f75ca9da1 100644
--- a/tasks/behavior-thawing-frozen-food-i00-privileged/task.toml
+++ b/tasks/behavior-thawing-frozen-food-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "behavior" # BEHAVIOR-1K v3.9.2 = OmniGibson 3.9.2 on Isa
robot = "r1pro"
family = "thawing_frozen_food"
mode = "privileged" # privileged state, unlimited resets, thin primitives (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (274 s on average); demonstration video (348 s): https://www.youtube.com/embed/FF3HiYSHGdc (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-04); human review per family (protocol §2.7)."
instance_seed = 301 # scene = environment/instance.json (2026 challenge public_test instance 301 + its state hash); identity = dataset.toml digest
diff --git a/tasks/behavior-thawing-frozen-food-i00-standard/task.toml b/tasks/behavior-thawing-frozen-food-i00-standard/task.toml
index 6f27947811..92e99c0a61 100644
--- a/tasks/behavior-thawing-frozen-food-i00-standard/task.toml
+++ b/tasks/behavior-thawing-frozen-food-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "behavior"
robot = "r1pro"
family = "thawing_frozen_food"
mode = "standard" # robot as a service: RGB-D + proprioception, no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (274 s on average); demonstration video (348 s): https://www.youtube.com/embed/FF3HiYSHGdc (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-04); human review per family (protocol §2.7)."
diff --git a/tasks/behavior-tidying-bathroom-i00-privileged/task.toml b/tasks/behavior-tidying-bathroom-i00-privileged/task.toml
index b8331b148a..0dd3c47e99 100644
--- a/tasks/behavior-tidying-bathroom-i00-privileged/task.toml
+++ b/tasks/behavior-tidying-bathroom-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "behavior" # BEHAVIOR-1K v3.9.2 = OmniGibson 3.9.2 on Isa
robot = "r1pro"
family = "tidying_bathroom"
mode = "privileged" # privileged state, unlimited resets, thin primitives (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (433 s on average); demonstration video (541 s): https://www.youtube.com/embed/EdKc0cKLITw (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-04); human review per family (protocol §2.7)."
instance_seed = 301 # scene = environment/instance.json (2026 challenge public_test instance 301 + its state hash); identity = dataset.toml digest
diff --git a/tasks/behavior-tidying-bathroom-i00-standard/task.toml b/tasks/behavior-tidying-bathroom-i00-standard/task.toml
index bad774be26..9c87dab169 100644
--- a/tasks/behavior-tidying-bathroom-i00-standard/task.toml
+++ b/tasks/behavior-tidying-bathroom-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "behavior"
robot = "r1pro"
family = "tidying_bathroom"
mode = "standard" # robot as a service: RGB-D + proprioception, no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (433 s on average); demonstration video (541 s): https://www.youtube.com/embed/EdKc0cKLITw (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-04); human review per family (protocol §2.7)."
diff --git a/tasks/behavior-tidying-bedroom-i00-privileged/task.toml b/tasks/behavior-tidying-bedroom-i00-privileged/task.toml
index f278d4e846..9cad5ab6b2 100644
--- a/tasks/behavior-tidying-bedroom-i00-privileged/task.toml
+++ b/tasks/behavior-tidying-bedroom-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "behavior" # BEHAVIOR-1K v3.9.2 = OmniGibson 3.9.2 on Isa
robot = "r1pro"
family = "tidying_bedroom"
mode = "privileged" # privileged state, unlimited resets, thin primitives (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance. Solvable: model trial behavior-tidying-bedroom-i00__VcXRVyJ (behavior-lunaxh-t8-0927, privileged) graded success=1, deterministic=1 by the separate verifier on a pre-v1.0 version of this task (dev/qineng 2d486a91; grader verdict and frozen instance unchanged since); agent log https://embodied-agent-interface-v2-internal.github.io/runs/behavior-1k/codex-gpt6_luna-xhigh/tidying_bedroom/unlimited/, verifier replay https://huggingface.co/datasets/eai-v2-internal/agent-runs/resolve/main/behavior-1k/codex-gpt6_luna-xhigh/tidying_bedroom/unlimited/replay.mp4; not re-run on this version (static validation only). Human review per family (protocol §2.7)."
instance_seed = 301 # scene = environment/instance.json (2026 challenge public_test instance 301 + its state hash); identity = dataset.toml digest
diff --git a/tasks/behavior-tidying-bedroom-i00-standard/task.toml b/tasks/behavior-tidying-bedroom-i00-standard/task.toml
index 7f49edc417..fb9277d7fe 100644
--- a/tasks/behavior-tidying-bedroom-i00-standard/task.toml
+++ b/tasks/behavior-tidying-bedroom-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "behavior"
robot = "r1pro"
family = "tidying_bedroom"
mode = "standard" # robot as a service: RGB-D + proprioception, no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance. Solvable: model trial behavior-tidying-bedroom-i00__VcXRVyJ (behavior-lunaxh-t8-0927, the privileged twin behavior-tidying-bedroom-i00-privileged on the same frozen instance and BDDL goal) graded success=1, deterministic=1 by the separate verifier on a pre-v1.0 version of this task (dev/qineng 2d486a91; grader verdict and frozen instance unchanged since); agent log https://embodied-agent-interface-v2-internal.github.io/runs/behavior-1k/codex-gpt6_luna-xhigh/tidying_bedroom/unlimited/, verifier replay https://huggingface.co/datasets/eai-v2-internal/agent-runs/resolve/main/behavior-1k/codex-gpt6_luna-xhigh/tidying_bedroom/unlimited/replay.mp4; not re-run on this version (static validation only). Human review per family (protocol §2.7)."
diff --git a/tasks/behavior-tidying-living-room-i00-privileged/task.toml b/tasks/behavior-tidying-living-room-i00-privileged/task.toml
index 5ad8901bcb..496a4c9791 100644
--- a/tasks/behavior-tidying-living-room-i00-privileged/task.toml
+++ b/tasks/behavior-tidying-living-room-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "behavior" # BEHAVIOR-1K v3.9.2 = OmniGibson 3.9.2 on Isa
robot = "r1pro"
family = "tidying_living_room"
mode = "privileged" # privileged state, unlimited resets, thin primitives (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (419 s on average); demonstration video (316 s): https://www.youtube.com/embed/aTCNKnjR03c (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-04); human review per family (protocol §2.7)."
instance_seed = 301 # scene = environment/instance.json (2026 challenge public_test instance 301 + its state hash); identity = dataset.toml digest
diff --git a/tasks/behavior-tidying-living-room-i00-standard/task.toml b/tasks/behavior-tidying-living-room-i00-standard/task.toml
index 07c07d5aa8..0ae3129b01 100644
--- a/tasks/behavior-tidying-living-room-i00-standard/task.toml
+++ b/tasks/behavior-tidying-living-room-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "behavior"
robot = "r1pro"
family = "tidying_living_room"
mode = "standard" # robot as a service: RGB-D + proprioception, no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (419 s on average); demonstration video (316 s): https://www.youtube.com/embed/aTCNKnjR03c (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-04); human review per family (protocol §2.7)."
diff --git a/tasks/behavior-turning-on-radio-i00-privileged/task.toml b/tasks/behavior-turning-on-radio-i00-privileged/task.toml
index 8abedf870f..104395ec12 100644
--- a/tasks/behavior-turning-on-radio-i00-privileged/task.toml
+++ b/tasks/behavior-turning-on-radio-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "behavior" # BEHAVIOR-1K v3.9.2 = OmniGibson 3.9.2 on Isa
robot = "r1pro"
family = "turning_on_radio"
mode = "privileged" # privileged state, unlimited resets, thin primitives (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance. Solvable: model trial behavior-turning-on-radio-i00__up5rnyp (behavior-lunaxh-t8-0927, privileged) graded success=1, deterministic=1 by the separate verifier on a pre-v1.0 version of this task (dev/qineng 912567fa; grader verdict and frozen instance unchanged since); agent log https://embodied-agent-interface-v2-internal.github.io/runs/behavior-1k/codex-gpt6_luna-xhigh/turning_on_radio/unlimited/, verifier replay https://huggingface.co/datasets/eai-v2-internal/agent-runs/resolve/main/behavior-1k/codex-gpt6_luna-xhigh/turning_on_radio/unlimited/replay.mp4; not re-run on this version (static validation only). Human review per family (protocol §2.7)."
instance_seed = 301 # scene = environment/instance.json (2026 challenge public_test instance 301 + its state hash); identity = dataset.toml digest
diff --git a/tasks/behavior-turning-on-radio-i00-standard/task.toml b/tasks/behavior-turning-on-radio-i00-standard/task.toml
index 885ddde9da..ee0bfaf7cb 100644
--- a/tasks/behavior-turning-on-radio-i00-standard/task.toml
+++ b/tasks/behavior-turning-on-radio-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "behavior"
robot = "r1pro"
family = "turning_on_radio"
mode = "standard" # robot as a service: RGB-D + proprioception, no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance. Solvable: model trial behavior-turning-on-radio-i00__up5rnyp (behavior-lunaxh-t8-0927, the privileged twin behavior-turning-on-radio-i00-privileged on the same frozen instance and BDDL goal) graded success=1, deterministic=1 by the separate verifier on a pre-v1.0 version of this task (dev/qineng 912567fa; grader verdict and frozen instance unchanged since); agent log https://embodied-agent-interface-v2-internal.github.io/runs/behavior-1k/codex-gpt6_luna-xhigh/turning_on_radio/unlimited/, verifier replay https://huggingface.co/datasets/eai-v2-internal/agent-runs/resolve/main/behavior-1k/codex-gpt6_luna-xhigh/turning_on_radio/unlimited/replay.mp4; not re-run on this version (static validation only). Human review per family (protocol §2.7)."
diff --git a/tasks/behavior-turning-out-all-lights-before-sleep-i00-privileged/task.toml b/tasks/behavior-turning-out-all-lights-before-sleep-i00-privileged/task.toml
index 5850f9edf1..7df032ad76 100644
--- a/tasks/behavior-turning-out-all-lights-before-sleep-i00-privileged/task.toml
+++ b/tasks/behavior-turning-out-all-lights-before-sleep-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "behavior" # BEHAVIOR-1K v3.9.2 = OmniGibson 3.9.2 on Isa
robot = "r1pro"
family = "turning_out_all_lights_before_sleep"
mode = "privileged" # privileged state, unlimited resets, thin primitives (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (334 s on average); demonstration video (337 s): https://www.youtube.com/embed/AN8l_cduvqU (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-07); human review in PR #75 (protocol §2.7)."
instance_seed = 301 # scene = environment/instance.json (2026 challenge public_test instance 301 + its state hash); identity = dataset.toml digest
diff --git a/tasks/behavior-turning-out-all-lights-before-sleep-i00-standard/task.toml b/tasks/behavior-turning-out-all-lights-before-sleep-i00-standard/task.toml
index 87078379f4..c89984fa7a 100644
--- a/tasks/behavior-turning-out-all-lights-before-sleep-i00-standard/task.toml
+++ b/tasks/behavior-turning-out-all-lights-before-sleep-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "behavior"
robot = "r1pro"
family = "turning_out_all_lights_before_sleep"
mode = "standard" # robot as a service: RGB-D + proprioception, no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (334 s on average); demonstration video (337 s): https://www.youtube.com/embed/AN8l_cduvqU (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-07); human review in PR #75 (protocol §2.7)."
diff --git a/tasks/behavior-unloading-the-car-i00-privileged/task.toml b/tasks/behavior-unloading-the-car-i00-privileged/task.toml
index 0eb0f47aaa..3d484629b0 100644
--- a/tasks/behavior-unloading-the-car-i00-privileged/task.toml
+++ b/tasks/behavior-unloading-the-car-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "behavior" # BEHAVIOR-1K v3.9.2 = OmniGibson 3.9.2 on Isa
robot = "r1pro"
family = "unloading_the_car"
mode = "privileged" # privileged state, unlimited resets, thin primitives (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (377 s on average); demonstration video (378 s): https://www.youtube.com/embed/IBV8nn39hTM (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-04); human review per family (protocol §2.7)."
instance_seed = 301 # scene = environment/instance.json (2026 challenge public_test instance 301 + its state hash); identity = dataset.toml digest
diff --git a/tasks/behavior-unloading-the-car-i00-standard/task.toml b/tasks/behavior-unloading-the-car-i00-standard/task.toml
index e3ff5bb0a0..ec0d9a9a1d 100644
--- a/tasks/behavior-unloading-the-car-i00-standard/task.toml
+++ b/tasks/behavior-unloading-the-car-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "behavior"
robot = "r1pro"
family = "unloading_the_car"
mode = "standard" # robot as a service: RGB-D + proprioception, no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (377 s on average); demonstration video (378 s): https://www.youtube.com/embed/IBV8nn39hTM (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-04); human review per family (protocol §2.7)."
diff --git a/tasks/behavior-vacuuming-floors-i00-privileged/task.toml b/tasks/behavior-vacuuming-floors-i00-privileged/task.toml
index 7366545943..991adacd20 100644
--- a/tasks/behavior-vacuuming-floors-i00-privileged/task.toml
+++ b/tasks/behavior-vacuuming-floors-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "behavior" # BEHAVIOR-1K v3.9.2 = OmniGibson 3.9.2 on Isa
robot = "r1pro"
family = "vacuuming_floors"
mode = "privileged" # privileged state, unlimited resets, thin primitives (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (80 s on average); demonstration video (65 s): https://www.youtube.com/embed/RqYuT-aj21A (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-07); human review in PR #71 (protocol §2.7)."
instance_seed = 301 # scene = environment/instance.json (2026 challenge public_test instance 301 + its state hash); identity = dataset.toml digest
diff --git a/tasks/behavior-vacuuming-floors-i00-standard/task.toml b/tasks/behavior-vacuuming-floors-i00-standard/task.toml
index 6ff23b5b6d..e861c5242b 100644
--- a/tasks/behavior-vacuuming-floors-i00-standard/task.toml
+++ b/tasks/behavior-vacuuming-floors-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "behavior"
robot = "r1pro"
family = "vacuuming_floors"
mode = "standard" # robot as a service: RGB-D + proprioception, no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (80 s on average); demonstration video (65 s): https://www.youtube.com/embed/RqYuT-aj21A (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-07); human review in PR #71 (protocol §2.7)."
diff --git a/tasks/behavior-wash-a-baseball-cap-i00-privileged/task.toml b/tasks/behavior-wash-a-baseball-cap-i00-privileged/task.toml
index e92154269b..3fe1514621 100644
--- a/tasks/behavior-wash-a-baseball-cap-i00-privileged/task.toml
+++ b/tasks/behavior-wash-a-baseball-cap-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "behavior" # BEHAVIOR-1K v3.9.2 = OmniGibson 3.9.2 on Isa
robot = "r1pro"
family = "wash_a_baseball_cap"
mode = "privileged" # privileged state, unlimited resets, thin primitives (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (278 s on average); demonstration video (278 s): https://player.vimeo.com/video/1114060147 (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-07); human review in PR #75 (protocol §2.7)."
instance_seed = 301 # scene = environment/instance.json (2026 challenge public_test instance 301 + its state hash); identity = dataset.toml digest
diff --git a/tasks/behavior-wash-a-baseball-cap-i00-standard/task.toml b/tasks/behavior-wash-a-baseball-cap-i00-standard/task.toml
index 6bba136b07..be9ce5e822 100644
--- a/tasks/behavior-wash-a-baseball-cap-i00-standard/task.toml
+++ b/tasks/behavior-wash-a-baseball-cap-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "behavior"
robot = "r1pro"
family = "wash_a_baseball_cap"
mode = "standard" # robot as a service: RGB-D + proprioception, no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (278 s on average); demonstration video (278 s): https://player.vimeo.com/video/1114060147 (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-07); human review in PR #75 (protocol §2.7)."
diff --git a/tasks/behavior-wash-dog-toys-i00-privileged/task.toml b/tasks/behavior-wash-dog-toys-i00-privileged/task.toml
index 8a44ec2a47..69d29a23c1 100644
--- a/tasks/behavior-wash-dog-toys-i00-privileged/task.toml
+++ b/tasks/behavior-wash-dog-toys-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "behavior" # BEHAVIOR-1K v3.9.2 = OmniGibson 3.9.2 on Isa
robot = "r1pro"
family = "wash_dog_toys"
mode = "privileged" # privileged state, unlimited resets, thin primitives (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (374 s on average); demonstration video (374 s): https://player.vimeo.com/video/1114060219 (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-04); human review per family (protocol §2.7)."
instance_seed = 301 # scene = environment/instance.json (2026 challenge public_test instance 301 + its state hash); identity = dataset.toml digest
diff --git a/tasks/behavior-wash-dog-toys-i00-standard/task.toml b/tasks/behavior-wash-dog-toys-i00-standard/task.toml
index c5f4452b79..85df679b31 100644
--- a/tasks/behavior-wash-dog-toys-i00-standard/task.toml
+++ b/tasks/behavior-wash-dog-toys-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "behavior"
robot = "r1pro"
family = "wash_dog_toys"
mode = "standard" # robot as a service: RGB-D + proprioception, no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution and no verified positive example yet: BEHAVIOR-1K ships no expert or planner that replays on this frozen instance, and no model trial has solved this task. Solvable by a human with this robot: the 2026 BEHAVIOR Challenge publishes 200 teleoperated demonstrations of this activity on the R1Pro in this scene model (374 s on average); demonstration video (374 s): https://player.vimeo.com/video/1114060219 (challenge task list https://behavior.stanford.edu/challenge/tasks/index.html). Supplementary task (added 2026-10-04); human review per family (protocol §2.7)."
diff --git a/tasks/bigym2-dishwasher-close-i00-privileged/task.toml b/tasks/bigym2-dishwasher-close-i00-privileged/task.toml
index 091d243c18..49c5eb5567 100644
--- a/tasks/bigym2-dishwasher-close-i00-privileged/task.toml
+++ b/tasks/bigym2-dishwasher-close-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "bigym2" # BiGym 2.0 (bigym 1.0.0 at 93ad4b3): MuJoCo
robot = "unitree-g1-dex1"
family = "dishwasher-close"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
# scene = environment/instance.json (upstream's task reset with this seed + its t = 0 state hash); identity = the dataset.toml digest
instance_seed = 0
expert_time_estimate_hours = 2.0
diff --git a/tasks/bigym2-dishwasher-close-i00-standard-realtime/README.md b/tasks/bigym2-dishwasher-close-i00-standard-realtime/README.md
index ff80167700..57a4a7019a 100644
--- a/tasks/bigym2-dishwasher-close-i00-standard-realtime/README.md
+++ b/tasks/bigym2-dishwasher-close-i00-standard-realtime/README.md
@@ -9,6 +9,7 @@
| Family / instance | `dishwasher-close` / seed 0 |
| Oracle | `none` — no reference solution (BiGym 2.0 ships no expert or planner; its demonstrations and its agents' programs stay off the agent's network); positive example graded 1 by the separate verifier: bigym2-dishwasher-close-i00-priv__BtaZZZZ (run codex-gpt6_luna-medium, 2026-10-08, on an earlier text of the privileged task; regraded by its current verifier); human review in PR #123 |
| Agent time budget | 3600 s; tokens: protocol default (PROTOCOL §5) |
+| Real time | slowdown S = 1 (the world runs at 1/S of real speed: one control step every S x control_dt s of wall time; `RCB_SLOWDOWN`'s default); reset budget 0 (`RCB_RESET_BUDGET`'s default) (PROTOCOL §4b) |
| Base image | `ghcr.io/mll-lab-nu/eai-bigym2:0.1.1` |
## Task
diff --git a/tasks/bigym2-dishwasher-close-i00-standard-realtime/task.toml b/tasks/bigym2-dishwasher-close-i00-standard-realtime/task.toml
index 4243668c9f..c902d9092d 100644
--- a/tasks/bigym2-dishwasher-close-i00-standard-realtime/task.toml
+++ b/tasks/bigym2-dishwasher-close-i00-standard-realtime/task.toml
@@ -26,6 +26,8 @@ simulator = "bigym2"
robot = "unitree-g1-dex1"
family = "dishwasher-close"
mode = "standard-realtime"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
+slowdown = 1 # standard-realtime: the world runs at 1/S of real speed (docs/PROTOCOL.md §4b); RCB_SLOWDOWN's default
protocol = "eai-standard/2.3"
instance_seed = 0
expert_time_estimate_hours = 3.0
diff --git a/tasks/bigym2-dishwasher-close-i00-standard/task.toml b/tasks/bigym2-dishwasher-close-i00-standard/task.toml
index ce14849702..566502b21a 100644
--- a/tasks/bigym2-dishwasher-close-i00-standard/task.toml
+++ b/tasks/bigym2-dishwasher-close-i00-standard/task.toml
@@ -24,6 +24,7 @@ simulator = "bigym2"
robot = "unitree-g1-dex1"
family = "dishwasher-close"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
instance_seed = 0
expert_time_estimate_hours = 3.0
diff --git a/tasks/bigym2-dishwasher-load-plates-i00-privileged/task.toml b/tasks/bigym2-dishwasher-load-plates-i00-privileged/task.toml
index ae6583657a..e9c556b3be 100644
--- a/tasks/bigym2-dishwasher-load-plates-i00-privileged/task.toml
+++ b/tasks/bigym2-dishwasher-load-plates-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "bigym2" # BiGym 2.0 (bigym 1.0.0 at 93ad4b3): MuJoCo
robot = "unitree-g1-dex1"
family = "dishwasher-load-plates"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
# scene = environment/instance.json (upstream's task reset with this seed + its t = 0 state hash); identity = the dataset.toml digest
instance_seed = 0
expert_time_estimate_hours = 2.0
diff --git a/tasks/bigym2-dishwasher-load-plates-i00-standard-realtime/README.md b/tasks/bigym2-dishwasher-load-plates-i00-standard-realtime/README.md
index 5ed76e4213..818b71f9d4 100644
--- a/tasks/bigym2-dishwasher-load-plates-i00-standard-realtime/README.md
+++ b/tasks/bigym2-dishwasher-load-plates-i00-standard-realtime/README.md
@@ -9,6 +9,7 @@
| Family / instance | `dishwasher-load-plates` / seed 0 |
| Oracle | `none` — no reference solution (BiGym 2.0 ships no expert or planner; its demonstrations and its agents' programs stay off the agent's network); no positive example yet; human review in PR #123 |
| Agent time budget | 3600 s; tokens: protocol default (PROTOCOL §5) |
+| Real time | slowdown S = 1 (the world runs at 1/S of real speed: one control step every S x control_dt s of wall time; `RCB_SLOWDOWN`'s default); reset budget 0 (`RCB_RESET_BUDGET`'s default) (PROTOCOL §4b) |
| Base image | `ghcr.io/mll-lab-nu/eai-bigym2:0.1.1` |
## Task
diff --git a/tasks/bigym2-dishwasher-load-plates-i00-standard-realtime/task.toml b/tasks/bigym2-dishwasher-load-plates-i00-standard-realtime/task.toml
index 73aafb5a00..184997ecd0 100644
--- a/tasks/bigym2-dishwasher-load-plates-i00-standard-realtime/task.toml
+++ b/tasks/bigym2-dishwasher-load-plates-i00-standard-realtime/task.toml
@@ -26,6 +26,8 @@ simulator = "bigym2"
robot = "unitree-g1-dex1"
family = "dishwasher-load-plates"
mode = "standard-realtime"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
+slowdown = 1 # standard-realtime: the world runs at 1/S of real speed (docs/PROTOCOL.md §4b); RCB_SLOWDOWN's default
protocol = "eai-standard/2.3"
instance_seed = 0
expert_time_estimate_hours = 3.0
diff --git a/tasks/bigym2-dishwasher-load-plates-i00-standard/task.toml b/tasks/bigym2-dishwasher-load-plates-i00-standard/task.toml
index b9bf5087f6..9c505cbe18 100644
--- a/tasks/bigym2-dishwasher-load-plates-i00-standard/task.toml
+++ b/tasks/bigym2-dishwasher-load-plates-i00-standard/task.toml
@@ -24,6 +24,7 @@ simulator = "bigym2"
robot = "unitree-g1-dex1"
family = "dishwasher-load-plates"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
instance_seed = 0
expert_time_estimate_hours = 3.0
diff --git a/tasks/bigym2-drawer-top-open-i00-privileged/task.toml b/tasks/bigym2-drawer-top-open-i00-privileged/task.toml
index e195a48b9a..aa01dd55af 100644
--- a/tasks/bigym2-drawer-top-open-i00-privileged/task.toml
+++ b/tasks/bigym2-drawer-top-open-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "bigym2" # BiGym 2.0 (bigym 1.0.0 at 93ad4b3): MuJoCo
robot = "unitree-g1-dex1"
family = "drawer-top-open"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
# scene = environment/instance.json (upstream's task reset with this seed + its t = 0 state hash); identity = the dataset.toml digest
instance_seed = 0
expert_time_estimate_hours = 2.0
diff --git a/tasks/bigym2-drawer-top-open-i00-standard-realtime/README.md b/tasks/bigym2-drawer-top-open-i00-standard-realtime/README.md
index bb4540a098..95abd4d407 100644
--- a/tasks/bigym2-drawer-top-open-i00-standard-realtime/README.md
+++ b/tasks/bigym2-drawer-top-open-i00-standard-realtime/README.md
@@ -9,6 +9,7 @@
| Family / instance | `drawer-top-open` / seed 0 |
| Oracle | `none` — no reference solution (BiGym 2.0 ships no expert or planner; its demonstrations and its agents' programs stay off the agent's network); positive example graded 1 by the separate verifier: bigym2-drawer-top-open-i00-privi__BR52Quk (run codex-gpt6_luna-medium, 2026-10-08, on an earlier text of the privileged task; regraded by its current verifier); human review in PR #123 |
| Agent time budget | 3600 s; tokens: protocol default (PROTOCOL §5) |
+| Real time | slowdown S = 1 (the world runs at 1/S of real speed: one control step every S x control_dt s of wall time; `RCB_SLOWDOWN`'s default); reset budget 0 (`RCB_RESET_BUDGET`'s default) (PROTOCOL §4b) |
| Base image | `ghcr.io/mll-lab-nu/eai-bigym2:0.1.1` |
## Task
diff --git a/tasks/bigym2-drawer-top-open-i00-standard-realtime/task.toml b/tasks/bigym2-drawer-top-open-i00-standard-realtime/task.toml
index 5bb262999d..c3fffc1a31 100644
--- a/tasks/bigym2-drawer-top-open-i00-standard-realtime/task.toml
+++ b/tasks/bigym2-drawer-top-open-i00-standard-realtime/task.toml
@@ -26,6 +26,8 @@ simulator = "bigym2"
robot = "unitree-g1-dex1"
family = "drawer-top-open"
mode = "standard-realtime"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
+slowdown = 1 # standard-realtime: the world runs at 1/S of real speed (docs/PROTOCOL.md §4b); RCB_SLOWDOWN's default
protocol = "eai-standard/2.3"
instance_seed = 0
expert_time_estimate_hours = 3.0
diff --git a/tasks/bigym2-drawer-top-open-i00-standard/task.toml b/tasks/bigym2-drawer-top-open-i00-standard/task.toml
index 70bb15f91e..324b9ff7e8 100644
--- a/tasks/bigym2-drawer-top-open-i00-standard/task.toml
+++ b/tasks/bigym2-drawer-top-open-i00-standard/task.toml
@@ -24,6 +24,7 @@ simulator = "bigym2"
robot = "unitree-g1-dex1"
family = "drawer-top-open"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
instance_seed = 0
expert_time_estimate_hours = 3.0
diff --git a/tasks/bigym2-flip-cup-i00-privileged/task.toml b/tasks/bigym2-flip-cup-i00-privileged/task.toml
index e7ede393b4..f58cbf9d18 100644
--- a/tasks/bigym2-flip-cup-i00-privileged/task.toml
+++ b/tasks/bigym2-flip-cup-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "bigym2" # BiGym 2.0 (bigym 1.0.0 at 93ad4b3): MuJoCo
robot = "unitree-g1-dex1"
family = "flip-cup"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
# scene = environment/instance.json (upstream's task reset with this seed + its t = 0 state hash); identity = the dataset.toml digest
instance_seed = 0
expert_time_estimate_hours = 2.0
diff --git a/tasks/bigym2-flip-cup-i00-standard-realtime/README.md b/tasks/bigym2-flip-cup-i00-standard-realtime/README.md
index 8d08707f9f..e40aa643a4 100644
--- a/tasks/bigym2-flip-cup-i00-standard-realtime/README.md
+++ b/tasks/bigym2-flip-cup-i00-standard-realtime/README.md
@@ -9,6 +9,7 @@
| Family / instance | `flip-cup` / seed 0 |
| Oracle | `none` — no reference solution (BiGym 2.0 ships no expert or planner; its demonstrations and its agents' programs stay off the agent's network); positive example graded 1 by the separate verifier: bigym2-flip-cup-i00-privileged__ebButi4 (run codex-gpt6_luna-medium, 2026-10-08, on an earlier text of the privileged task; regraded by its current verifier); human review in PR #123 |
| Agent time budget | 3600 s; tokens: protocol default (PROTOCOL §5) |
+| Real time | slowdown S = 1 (the world runs at 1/S of real speed: one control step every S x control_dt s of wall time; `RCB_SLOWDOWN`'s default); reset budget 0 (`RCB_RESET_BUDGET`'s default) (PROTOCOL §4b) |
| Base image | `ghcr.io/mll-lab-nu/eai-bigym2:0.1.1` |
## Task
diff --git a/tasks/bigym2-flip-cup-i00-standard-realtime/task.toml b/tasks/bigym2-flip-cup-i00-standard-realtime/task.toml
index b475ceb5fc..dc4bddc1c9 100644
--- a/tasks/bigym2-flip-cup-i00-standard-realtime/task.toml
+++ b/tasks/bigym2-flip-cup-i00-standard-realtime/task.toml
@@ -26,6 +26,8 @@ simulator = "bigym2"
robot = "unitree-g1-dex1"
family = "flip-cup"
mode = "standard-realtime"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
+slowdown = 1 # standard-realtime: the world runs at 1/S of real speed (docs/PROTOCOL.md §4b); RCB_SLOWDOWN's default
protocol = "eai-standard/2.3"
instance_seed = 0
expert_time_estimate_hours = 3.0
diff --git a/tasks/bigym2-flip-cup-i00-standard/task.toml b/tasks/bigym2-flip-cup-i00-standard/task.toml
index ca82cb94f6..3e9ed99544 100644
--- a/tasks/bigym2-flip-cup-i00-standard/task.toml
+++ b/tasks/bigym2-flip-cup-i00-standard/task.toml
@@ -24,6 +24,7 @@ simulator = "bigym2"
robot = "unitree-g1-dex1"
family = "flip-cup"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
instance_seed = 0
expert_time_estimate_hours = 3.0
diff --git a/tasks/bigym2-flip-cutlery-i00-privileged/task.toml b/tasks/bigym2-flip-cutlery-i00-privileged/task.toml
index 4cd18ca047..b8c304d7d3 100644
--- a/tasks/bigym2-flip-cutlery-i00-privileged/task.toml
+++ b/tasks/bigym2-flip-cutlery-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "bigym2" # BiGym 2.0 (bigym 1.0.0 at 93ad4b3): MuJoCo
robot = "unitree-g1-dex1"
family = "flip-cutlery"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
# scene = environment/instance.json (upstream's task reset with this seed + its t = 0 state hash); identity = the dataset.toml digest
instance_seed = 0
expert_time_estimate_hours = 2.0
diff --git a/tasks/bigym2-flip-cutlery-i00-standard-realtime/README.md b/tasks/bigym2-flip-cutlery-i00-standard-realtime/README.md
index d6cbeb6820..86a366f0cf 100644
--- a/tasks/bigym2-flip-cutlery-i00-standard-realtime/README.md
+++ b/tasks/bigym2-flip-cutlery-i00-standard-realtime/README.md
@@ -9,6 +9,7 @@
| Family / instance | `flip-cutlery` / seed 0 |
| Oracle | `none` — no reference solution (BiGym 2.0 ships no expert or planner; its demonstrations and its agents' programs stay off the agent's network); positive example graded 1 by the separate verifier: bigym2-flip-cutlery-i00-privileg__GQkcmXU (run codex-gpt6_luna-medium, 2026-10-08, on an earlier text of the privileged task; regraded by its current verifier); human review in PR #123 |
| Agent time budget | 3600 s; tokens: protocol default (PROTOCOL §5) |
+| Real time | slowdown S = 1 (the world runs at 1/S of real speed: one control step every S x control_dt s of wall time; `RCB_SLOWDOWN`'s default); reset budget 0 (`RCB_RESET_BUDGET`'s default) (PROTOCOL §4b) |
| Base image | `ghcr.io/mll-lab-nu/eai-bigym2:0.1.1` |
## Task
diff --git a/tasks/bigym2-flip-cutlery-i00-standard-realtime/task.toml b/tasks/bigym2-flip-cutlery-i00-standard-realtime/task.toml
index 3f21863168..392673223e 100644
--- a/tasks/bigym2-flip-cutlery-i00-standard-realtime/task.toml
+++ b/tasks/bigym2-flip-cutlery-i00-standard-realtime/task.toml
@@ -26,6 +26,8 @@ simulator = "bigym2"
robot = "unitree-g1-dex1"
family = "flip-cutlery"
mode = "standard-realtime"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
+slowdown = 1 # standard-realtime: the world runs at 1/S of real speed (docs/PROTOCOL.md §4b); RCB_SLOWDOWN's default
protocol = "eai-standard/2.3"
instance_seed = 0
expert_time_estimate_hours = 3.0
diff --git a/tasks/bigym2-flip-cutlery-i00-standard/task.toml b/tasks/bigym2-flip-cutlery-i00-standard/task.toml
index 965bf3ea7a..cf0c6576a2 100644
--- a/tasks/bigym2-flip-cutlery-i00-standard/task.toml
+++ b/tasks/bigym2-flip-cutlery-i00-standard/task.toml
@@ -24,6 +24,7 @@ simulator = "bigym2"
robot = "unitree-g1-dex1"
family = "flip-cutlery"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
instance_seed = 0
expert_time_estimate_hours = 3.0
diff --git a/tasks/bigym2-move-two-plates-i00-privileged/task.toml b/tasks/bigym2-move-two-plates-i00-privileged/task.toml
index 0a82e0d9dc..0aaccd32dd 100644
--- a/tasks/bigym2-move-two-plates-i00-privileged/task.toml
+++ b/tasks/bigym2-move-two-plates-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "bigym2" # BiGym 2.0 (bigym 1.0.0 at 93ad4b3): MuJoCo
robot = "unitree-g1-dex1"
family = "move-two-plates"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
# scene = environment/instance.json (upstream's task reset with this seed + its t = 0 state hash); identity = the dataset.toml digest
instance_seed = 0
expert_time_estimate_hours = 2.0
diff --git a/tasks/bigym2-move-two-plates-i00-standard-realtime/README.md b/tasks/bigym2-move-two-plates-i00-standard-realtime/README.md
index 1d9d6f05f0..984d65efc2 100644
--- a/tasks/bigym2-move-two-plates-i00-standard-realtime/README.md
+++ b/tasks/bigym2-move-two-plates-i00-standard-realtime/README.md
@@ -9,6 +9,7 @@
| Family / instance | `move-two-plates` / seed 0 |
| Oracle | `none` — no reference solution (BiGym 2.0 ships no expert or planner; its demonstrations and its agents' programs stay off the agent's network); no positive example yet; human review in PR #123 |
| Agent time budget | 3600 s; tokens: protocol default (PROTOCOL §5) |
+| Real time | slowdown S = 1 (the world runs at 1/S of real speed: one control step every S x control_dt s of wall time; `RCB_SLOWDOWN`'s default); reset budget 0 (`RCB_RESET_BUDGET`'s default) (PROTOCOL §4b) |
| Base image | `ghcr.io/mll-lab-nu/eai-bigym2:0.1.1` |
## Task
diff --git a/tasks/bigym2-move-two-plates-i00-standard-realtime/task.toml b/tasks/bigym2-move-two-plates-i00-standard-realtime/task.toml
index 0f3d622a62..2f98ae5ddf 100644
--- a/tasks/bigym2-move-two-plates-i00-standard-realtime/task.toml
+++ b/tasks/bigym2-move-two-plates-i00-standard-realtime/task.toml
@@ -26,6 +26,8 @@ simulator = "bigym2"
robot = "unitree-g1-dex1"
family = "move-two-plates"
mode = "standard-realtime"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
+slowdown = 1 # standard-realtime: the world runs at 1/S of real speed (docs/PROTOCOL.md §4b); RCB_SLOWDOWN's default
protocol = "eai-standard/2.3"
instance_seed = 0
expert_time_estimate_hours = 3.0
diff --git a/tasks/bigym2-move-two-plates-i00-standard/task.toml b/tasks/bigym2-move-two-plates-i00-standard/task.toml
index 517ffc2d71..6a9c699228 100644
--- a/tasks/bigym2-move-two-plates-i00-standard/task.toml
+++ b/tasks/bigym2-move-two-plates-i00-standard/task.toml
@@ -24,6 +24,7 @@ simulator = "bigym2"
robot = "unitree-g1-dex1"
family = "move-two-plates"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
instance_seed = 0
expert_time_estimate_hours = 3.0
diff --git a/tasks/bigym2-pick-box-i00-privileged/task.toml b/tasks/bigym2-pick-box-i00-privileged/task.toml
index dff0fb729b..4b1da6de94 100644
--- a/tasks/bigym2-pick-box-i00-privileged/task.toml
+++ b/tasks/bigym2-pick-box-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "bigym2" # BiGym 2.0 (bigym 1.0.0 at 93ad4b3): MuJoCo
robot = "unitree-g1-dex1"
family = "pick-box"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
# scene = environment/instance.json (upstream's task reset with this seed + its t = 0 state hash); identity = the dataset.toml digest
instance_seed = 0
expert_time_estimate_hours = 2.0
diff --git a/tasks/bigym2-pick-box-i00-standard-realtime/README.md b/tasks/bigym2-pick-box-i00-standard-realtime/README.md
index 1cddc4a87f..8a806e3cae 100644
--- a/tasks/bigym2-pick-box-i00-standard-realtime/README.md
+++ b/tasks/bigym2-pick-box-i00-standard-realtime/README.md
@@ -9,6 +9,7 @@
| Family / instance | `pick-box` / seed 0 |
| Oracle | `none` — no reference solution (BiGym 2.0 ships no expert or planner; its demonstrations and its agents' programs stay off the agent's network); no positive example yet; human review in PR #123 |
| Agent time budget | 3600 s; tokens: protocol default (PROTOCOL §5) |
+| Real time | slowdown S = 1 (the world runs at 1/S of real speed: one control step every S x control_dt s of wall time; `RCB_SLOWDOWN`'s default); reset budget 0 (`RCB_RESET_BUDGET`'s default) (PROTOCOL §4b) |
| Base image | `ghcr.io/mll-lab-nu/eai-bigym2:0.1.1` |
## Task
diff --git a/tasks/bigym2-pick-box-i00-standard-realtime/task.toml b/tasks/bigym2-pick-box-i00-standard-realtime/task.toml
index b3f91db38c..ed9e7a5a43 100644
--- a/tasks/bigym2-pick-box-i00-standard-realtime/task.toml
+++ b/tasks/bigym2-pick-box-i00-standard-realtime/task.toml
@@ -26,6 +26,8 @@ simulator = "bigym2"
robot = "unitree-g1-dex1"
family = "pick-box"
mode = "standard-realtime"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
+slowdown = 1 # standard-realtime: the world runs at 1/S of real speed (docs/PROTOCOL.md §4b); RCB_SLOWDOWN's default
protocol = "eai-standard/2.3"
instance_seed = 0
expert_time_estimate_hours = 3.0
diff --git a/tasks/bigym2-pick-box-i00-standard/task.toml b/tasks/bigym2-pick-box-i00-standard/task.toml
index 7bdb6db47e..e79473c992 100644
--- a/tasks/bigym2-pick-box-i00-standard/task.toml
+++ b/tasks/bigym2-pick-box-i00-standard/task.toml
@@ -24,6 +24,7 @@ simulator = "bigym2"
robot = "unitree-g1-dex1"
family = "pick-box"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
instance_seed = 0
expert_time_estimate_hours = 3.0
diff --git a/tasks/bigym2-put-cups-i00-privileged/task.toml b/tasks/bigym2-put-cups-i00-privileged/task.toml
index 989179268d..cccff2ecf2 100644
--- a/tasks/bigym2-put-cups-i00-privileged/task.toml
+++ b/tasks/bigym2-put-cups-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "bigym2" # BiGym 2.0 (bigym 1.0.0 at 93ad4b3): MuJoCo
robot = "unitree-g1-dex1"
family = "put-cups"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
# scene = environment/instance.json (upstream's task reset with this seed + its t = 0 state hash); identity = the dataset.toml digest
instance_seed = 0
expert_time_estimate_hours = 2.0
diff --git a/tasks/bigym2-put-cups-i00-standard-realtime/README.md b/tasks/bigym2-put-cups-i00-standard-realtime/README.md
index 5deb47c60a..c6b7dcaf8d 100644
--- a/tasks/bigym2-put-cups-i00-standard-realtime/README.md
+++ b/tasks/bigym2-put-cups-i00-standard-realtime/README.md
@@ -9,6 +9,7 @@
| Family / instance | `put-cups` / seed 0 |
| Oracle | `none` — no reference solution (BiGym 2.0 ships no expert or planner; its demonstrations and its agents' programs stay off the agent's network); no positive example yet; human review in PR #123 |
| Agent time budget | 3600 s; tokens: protocol default (PROTOCOL §5) |
+| Real time | slowdown S = 1 (the world runs at 1/S of real speed: one control step every S x control_dt s of wall time; `RCB_SLOWDOWN`'s default); reset budget 0 (`RCB_RESET_BUDGET`'s default) (PROTOCOL §4b) |
| Base image | `ghcr.io/mll-lab-nu/eai-bigym2:0.1.1` |
## Task
diff --git a/tasks/bigym2-put-cups-i00-standard-realtime/task.toml b/tasks/bigym2-put-cups-i00-standard-realtime/task.toml
index ee1decbbcc..91b710109c 100644
--- a/tasks/bigym2-put-cups-i00-standard-realtime/task.toml
+++ b/tasks/bigym2-put-cups-i00-standard-realtime/task.toml
@@ -26,6 +26,8 @@ simulator = "bigym2"
robot = "unitree-g1-dex1"
family = "put-cups"
mode = "standard-realtime"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
+slowdown = 1 # standard-realtime: the world runs at 1/S of real speed (docs/PROTOCOL.md §4b); RCB_SLOWDOWN's default
protocol = "eai-standard/2.3"
instance_seed = 0
expert_time_estimate_hours = 3.0
diff --git a/tasks/bigym2-put-cups-i00-standard/task.toml b/tasks/bigym2-put-cups-i00-standard/task.toml
index abf3f275df..b6a3208a14 100644
--- a/tasks/bigym2-put-cups-i00-standard/task.toml
+++ b/tasks/bigym2-put-cups-i00-standard/task.toml
@@ -24,6 +24,7 @@ simulator = "bigym2"
robot = "unitree-g1-dex1"
family = "put-cups"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
instance_seed = 0
expert_time_estimate_hours = 3.0
diff --git a/tasks/bigym2-reach-target-dual-i00-privileged/task.toml b/tasks/bigym2-reach-target-dual-i00-privileged/task.toml
index e60de3723a..2c7be73f20 100644
--- a/tasks/bigym2-reach-target-dual-i00-privileged/task.toml
+++ b/tasks/bigym2-reach-target-dual-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "bigym2" # BiGym 2.0 (bigym 1.0.0 at 93ad4b3): MuJoCo
robot = "unitree-g1-dex1"
family = "reach-target-dual"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
# scene = environment/instance.json (upstream's task reset with this seed + its t = 0 state hash); identity = the dataset.toml digest
instance_seed = 0
expert_time_estimate_hours = 2.0
diff --git a/tasks/bigym2-reach-target-dual-i00-standard-realtime/README.md b/tasks/bigym2-reach-target-dual-i00-standard-realtime/README.md
index d6d2ee1820..bc3c111353 100644
--- a/tasks/bigym2-reach-target-dual-i00-standard-realtime/README.md
+++ b/tasks/bigym2-reach-target-dual-i00-standard-realtime/README.md
@@ -9,6 +9,7 @@
| Family / instance | `reach-target-dual` / seed 0 |
| Oracle | `none` — no reference solution (BiGym 2.0 ships no expert or planner; its demonstrations and its agents' programs stay off the agent's network); positive example graded 1 by the separate verifier: bigym2-reach-target-dual-i00-pri__fxPRBTg (run codex-gpt6_luna-medium, 2026-10-08, on an earlier text of the privileged task; regraded by its current verifier); human review in PR #123 |
| Agent time budget | 3600 s; tokens: protocol default (PROTOCOL §5) |
+| Real time | slowdown S = 1 (the world runs at 1/S of real speed: one control step every S x control_dt s of wall time; `RCB_SLOWDOWN`'s default); reset budget 0 (`RCB_RESET_BUDGET`'s default) (PROTOCOL §4b) |
| Base image | `ghcr.io/mll-lab-nu/eai-bigym2:0.1.1` |
## Task
diff --git a/tasks/bigym2-reach-target-dual-i00-standard-realtime/task.toml b/tasks/bigym2-reach-target-dual-i00-standard-realtime/task.toml
index 11cd6d54b4..33c0f0c48b 100644
--- a/tasks/bigym2-reach-target-dual-i00-standard-realtime/task.toml
+++ b/tasks/bigym2-reach-target-dual-i00-standard-realtime/task.toml
@@ -26,6 +26,8 @@ simulator = "bigym2"
robot = "unitree-g1-dex1"
family = "reach-target-dual"
mode = "standard-realtime"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
+slowdown = 1 # standard-realtime: the world runs at 1/S of real speed (docs/PROTOCOL.md §4b); RCB_SLOWDOWN's default
protocol = "eai-standard/2.3"
instance_seed = 0
expert_time_estimate_hours = 3.0
diff --git a/tasks/bigym2-reach-target-dual-i00-standard/task.toml b/tasks/bigym2-reach-target-dual-i00-standard/task.toml
index 000df8d79d..70f3e38b25 100644
--- a/tasks/bigym2-reach-target-dual-i00-standard/task.toml
+++ b/tasks/bigym2-reach-target-dual-i00-standard/task.toml
@@ -24,6 +24,7 @@ simulator = "bigym2"
robot = "unitree-g1-dex1"
family = "reach-target-dual"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
instance_seed = 0
expert_time_estimate_hours = 3.0
diff --git a/tasks/bigym2-sandwich-remove-i00-privileged/task.toml b/tasks/bigym2-sandwich-remove-i00-privileged/task.toml
index 96e261fdeb..4b784c1491 100644
--- a/tasks/bigym2-sandwich-remove-i00-privileged/task.toml
+++ b/tasks/bigym2-sandwich-remove-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "bigym2" # BiGym 2.0 (bigym 1.0.0 at 93ad4b3): MuJoCo
robot = "unitree-g1-dex1"
family = "sandwich-remove"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
# scene = environment/instance.json (upstream's task reset with this seed + its t = 0 state hash); identity = the dataset.toml digest
instance_seed = 0
expert_time_estimate_hours = 2.0
diff --git a/tasks/bigym2-sandwich-remove-i00-standard-realtime/README.md b/tasks/bigym2-sandwich-remove-i00-standard-realtime/README.md
index ab8b745cb2..e36eabeb7a 100644
--- a/tasks/bigym2-sandwich-remove-i00-standard-realtime/README.md
+++ b/tasks/bigym2-sandwich-remove-i00-standard-realtime/README.md
@@ -9,6 +9,7 @@
| Family / instance | `sandwich-remove` / seed 0 |
| Oracle | `none` — no reference solution (BiGym 2.0 ships no expert or planner; its demonstrations and its agents' programs stay off the agent's network); no positive example yet; human review in PR #123 |
| Agent time budget | 3600 s; tokens: protocol default (PROTOCOL §5) |
+| Real time | slowdown S = 1 (the world runs at 1/S of real speed: one control step every S x control_dt s of wall time; `RCB_SLOWDOWN`'s default); reset budget 0 (`RCB_RESET_BUDGET`'s default) (PROTOCOL §4b) |
| Base image | `ghcr.io/mll-lab-nu/eai-bigym2:0.1.1` |
## Task
diff --git a/tasks/bigym2-sandwich-remove-i00-standard-realtime/task.toml b/tasks/bigym2-sandwich-remove-i00-standard-realtime/task.toml
index 2ae6e5d398..5ae0e60c83 100644
--- a/tasks/bigym2-sandwich-remove-i00-standard-realtime/task.toml
+++ b/tasks/bigym2-sandwich-remove-i00-standard-realtime/task.toml
@@ -26,6 +26,8 @@ simulator = "bigym2"
robot = "unitree-g1-dex1"
family = "sandwich-remove"
mode = "standard-realtime"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
+slowdown = 1 # standard-realtime: the world runs at 1/S of real speed (docs/PROTOCOL.md §4b); RCB_SLOWDOWN's default
protocol = "eai-standard/2.3"
instance_seed = 0
expert_time_estimate_hours = 3.0
diff --git a/tasks/bigym2-sandwich-remove-i00-standard/task.toml b/tasks/bigym2-sandwich-remove-i00-standard/task.toml
index 2376d82a05..9494894608 100644
--- a/tasks/bigym2-sandwich-remove-i00-standard/task.toml
+++ b/tasks/bigym2-sandwich-remove-i00-standard/task.toml
@@ -24,6 +24,7 @@ simulator = "bigym2"
robot = "unitree-g1-dex1"
family = "sandwich-remove"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
instance_seed = 0
expert_time_estimate_hours = 3.0
diff --git a/tasks/bigym2-saucepan-to-hob-i00-privileged/task.toml b/tasks/bigym2-saucepan-to-hob-i00-privileged/task.toml
index e53a2c49e6..e3cbce2ddb 100644
--- a/tasks/bigym2-saucepan-to-hob-i00-privileged/task.toml
+++ b/tasks/bigym2-saucepan-to-hob-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "bigym2" # BiGym 2.0 (bigym 1.0.0 at 93ad4b3): MuJoCo
robot = "unitree-g1-dex1"
family = "saucepan-to-hob"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
# scene = environment/instance.json (upstream's task reset with this seed + its t = 0 state hash); identity = the dataset.toml digest
instance_seed = 0
expert_time_estimate_hours = 2.0
diff --git a/tasks/bigym2-saucepan-to-hob-i00-standard-realtime/README.md b/tasks/bigym2-saucepan-to-hob-i00-standard-realtime/README.md
index 1a7ab4e298..44513fbac5 100644
--- a/tasks/bigym2-saucepan-to-hob-i00-standard-realtime/README.md
+++ b/tasks/bigym2-saucepan-to-hob-i00-standard-realtime/README.md
@@ -9,6 +9,7 @@
| Family / instance | `saucepan-to-hob` / seed 0 |
| Oracle | `none` — no reference solution (BiGym 2.0 ships no expert or planner; its demonstrations and its agents' programs stay off the agent's network); no positive example yet; human review in PR #123 |
| Agent time budget | 3600 s; tokens: protocol default (PROTOCOL §5) |
+| Real time | slowdown S = 1 (the world runs at 1/S of real speed: one control step every S x control_dt s of wall time; `RCB_SLOWDOWN`'s default); reset budget 0 (`RCB_RESET_BUDGET`'s default) (PROTOCOL §4b) |
| Base image | `ghcr.io/mll-lab-nu/eai-bigym2:0.1.1` |
## Task
diff --git a/tasks/bigym2-saucepan-to-hob-i00-standard-realtime/task.toml b/tasks/bigym2-saucepan-to-hob-i00-standard-realtime/task.toml
index c668efebb6..8242ca9818 100644
--- a/tasks/bigym2-saucepan-to-hob-i00-standard-realtime/task.toml
+++ b/tasks/bigym2-saucepan-to-hob-i00-standard-realtime/task.toml
@@ -26,6 +26,8 @@ simulator = "bigym2"
robot = "unitree-g1-dex1"
family = "saucepan-to-hob"
mode = "standard-realtime"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
+slowdown = 1 # standard-realtime: the world runs at 1/S of real speed (docs/PROTOCOL.md §4b); RCB_SLOWDOWN's default
protocol = "eai-standard/2.3"
instance_seed = 0
expert_time_estimate_hours = 3.0
diff --git a/tasks/bigym2-saucepan-to-hob-i00-standard/task.toml b/tasks/bigym2-saucepan-to-hob-i00-standard/task.toml
index c616168aec..d8093b04d2 100644
--- a/tasks/bigym2-saucepan-to-hob-i00-standard/task.toml
+++ b/tasks/bigym2-saucepan-to-hob-i00-standard/task.toml
@@ -24,6 +24,7 @@ simulator = "bigym2"
robot = "unitree-g1-dex1"
family = "saucepan-to-hob"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
instance_seed = 0
expert_time_estimate_hours = 3.0
diff --git a/tasks/bigym2-stack-blocks-i00-privileged/task.toml b/tasks/bigym2-stack-blocks-i00-privileged/task.toml
index fe5439be0a..6f7f4cf013 100644
--- a/tasks/bigym2-stack-blocks-i00-privileged/task.toml
+++ b/tasks/bigym2-stack-blocks-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "bigym2" # BiGym 2.0 (bigym 1.0.0 at 93ad4b3): MuJoCo
robot = "unitree-g1-dex1"
family = "stack-blocks"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
# scene = environment/instance.json (upstream's task reset with this seed + its t = 0 state hash); identity = the dataset.toml digest
instance_seed = 0
expert_time_estimate_hours = 2.0
diff --git a/tasks/bigym2-stack-blocks-i00-standard-realtime/README.md b/tasks/bigym2-stack-blocks-i00-standard-realtime/README.md
index 73d3b79f2c..dd00733305 100644
--- a/tasks/bigym2-stack-blocks-i00-standard-realtime/README.md
+++ b/tasks/bigym2-stack-blocks-i00-standard-realtime/README.md
@@ -9,6 +9,7 @@
| Family / instance | `stack-blocks` / seed 0 |
| Oracle | `none` — no reference solution (BiGym 2.0 ships no expert or planner; its demonstrations and its agents' programs stay off the agent's network); positive example graded 1 by the separate verifier: bigym2-stack-blocks-i00-privileg__Rz4Pgdi (run codex-gpt6_luna-medium, 2026-10-08, on an earlier text of the privileged task; regraded by its current verifier); human review in PR #123 |
| Agent time budget | 3600 s; tokens: protocol default (PROTOCOL §5) |
+| Real time | slowdown S = 1 (the world runs at 1/S of real speed: one control step every S x control_dt s of wall time; `RCB_SLOWDOWN`'s default); reset budget 0 (`RCB_RESET_BUDGET`'s default) (PROTOCOL §4b) |
| Base image | `ghcr.io/mll-lab-nu/eai-bigym2:0.1.1` |
## Task
diff --git a/tasks/bigym2-stack-blocks-i00-standard-realtime/task.toml b/tasks/bigym2-stack-blocks-i00-standard-realtime/task.toml
index 47a06ad3af..13098ca91f 100644
--- a/tasks/bigym2-stack-blocks-i00-standard-realtime/task.toml
+++ b/tasks/bigym2-stack-blocks-i00-standard-realtime/task.toml
@@ -26,6 +26,8 @@ simulator = "bigym2"
robot = "unitree-g1-dex1"
family = "stack-blocks"
mode = "standard-realtime"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
+slowdown = 1 # standard-realtime: the world runs at 1/S of real speed (docs/PROTOCOL.md §4b); RCB_SLOWDOWN's default
protocol = "eai-standard/2.3"
instance_seed = 0
expert_time_estimate_hours = 3.0
diff --git a/tasks/bigym2-stack-blocks-i00-standard/task.toml b/tasks/bigym2-stack-blocks-i00-standard/task.toml
index d9d54053ff..d56b408967 100644
--- a/tasks/bigym2-stack-blocks-i00-standard/task.toml
+++ b/tasks/bigym2-stack-blocks-i00-standard/task.toml
@@ -24,6 +24,7 @@ simulator = "bigym2"
robot = "unitree-g1-dex1"
family = "stack-blocks"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
instance_seed = 0
expert_time_estimate_hours = 3.0
diff --git a/tasks/bigym2-wall-cupboard-open-i00-privileged/task.toml b/tasks/bigym2-wall-cupboard-open-i00-privileged/task.toml
index 45ab7d1ca9..0ffe5b07e9 100644
--- a/tasks/bigym2-wall-cupboard-open-i00-privileged/task.toml
+++ b/tasks/bigym2-wall-cupboard-open-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "bigym2" # BiGym 2.0 (bigym 1.0.0 at 93ad4b3): MuJoCo
robot = "unitree-g1-dex1"
family = "wall-cupboard-open"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
# scene = environment/instance.json (upstream's task reset with this seed + its t = 0 state hash); identity = the dataset.toml digest
instance_seed = 0
expert_time_estimate_hours = 2.0
diff --git a/tasks/bigym2-wall-cupboard-open-i00-standard-realtime/README.md b/tasks/bigym2-wall-cupboard-open-i00-standard-realtime/README.md
index f301cac165..ab6db111d1 100644
--- a/tasks/bigym2-wall-cupboard-open-i00-standard-realtime/README.md
+++ b/tasks/bigym2-wall-cupboard-open-i00-standard-realtime/README.md
@@ -9,6 +9,7 @@
| Family / instance | `wall-cupboard-open` / seed 0 |
| Oracle | `none` — no reference solution (BiGym 2.0 ships no expert or planner; its demonstrations and its agents' programs stay off the agent's network); positive example graded 1 by the separate verifier: bigym2-wall-cupboard-open-i00-pr__mcuR6nu (run codex-gpt6_luna-medium, 2026-10-08, on an earlier text of the privileged task; regraded by its current verifier); human review in PR #123 |
| Agent time budget | 3600 s; tokens: protocol default (PROTOCOL §5) |
+| Real time | slowdown S = 1 (the world runs at 1/S of real speed: one control step every S x control_dt s of wall time; `RCB_SLOWDOWN`'s default); reset budget 0 (`RCB_RESET_BUDGET`'s default) (PROTOCOL §4b) |
| Base image | `ghcr.io/mll-lab-nu/eai-bigym2:0.1.1` |
## Task
diff --git a/tasks/bigym2-wall-cupboard-open-i00-standard-realtime/task.toml b/tasks/bigym2-wall-cupboard-open-i00-standard-realtime/task.toml
index f8a2e7da16..76f72001d2 100644
--- a/tasks/bigym2-wall-cupboard-open-i00-standard-realtime/task.toml
+++ b/tasks/bigym2-wall-cupboard-open-i00-standard-realtime/task.toml
@@ -26,6 +26,8 @@ simulator = "bigym2"
robot = "unitree-g1-dex1"
family = "wall-cupboard-open"
mode = "standard-realtime"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
+slowdown = 1 # standard-realtime: the world runs at 1/S of real speed (docs/PROTOCOL.md §4b); RCB_SLOWDOWN's default
protocol = "eai-standard/2.3"
instance_seed = 0
expert_time_estimate_hours = 3.0
diff --git a/tasks/bigym2-wall-cupboard-open-i00-standard/task.toml b/tasks/bigym2-wall-cupboard-open-i00-standard/task.toml
index e16f314554..7e87864e83 100644
--- a/tasks/bigym2-wall-cupboard-open-i00-standard/task.toml
+++ b/tasks/bigym2-wall-cupboard-open-i00-standard/task.toml
@@ -24,6 +24,7 @@ simulator = "bigym2"
robot = "unitree-g1-dex1"
family = "wall-cupboard-open"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
instance_seed = 0
expert_time_estimate_hours = 3.0
diff --git a/tasks/dexjoco-bimanual-assembly-i00-privileged/README.md b/tasks/dexjoco-bimanual-assembly-i00-privileged/README.md
index 026ee8e2c1..ee4ef22dd0 100644
--- a/tasks/dexjoco-bimanual-assembly-i00-privileged/README.md
+++ b/tasks/dexjoco-bimanual-assembly-i00-privileged/README.md
@@ -29,8 +29,10 @@ limited to the model APIs.
## Verification
Open-loop replay of `/app/output/trajectory.npz` from `tests/instance.npz`, twice in a fresh process in the verifier's own
-sandbox (no network); pass iff both episodes end in success, their final-state hashes agree and the rebuilt instance
-matches its recorded hash (`tests/grade.py`); plus `/logs/verifier/replay.mp4`, rendered after the verdict is written.
+sandbox (no network); pass iff both episodes end in success and their final-state hashes agree (`tests/grade.py`); plus
+`/logs/verifier/replay.mp4`, rendered after the verdict is written. A rebuilt instance that does not match its recorded
+hash (the image no longer rebuilds the scene) replays nothing and gives no verdict: `instance_mismatch`, `grader_error`
+and `infra_error` = 1, not the agent's 0.
A file that is not a readable trajectory (not an npz, no `task` / `actions`, a wrong shape, NaN) or has more than
50,000 rows scores 0 as `bad_trajectory`: the cap is what two replays of the slowest family fit into the verifier's
3600 s. A verifier that runs fewer than 4 OpenBLAS threads (fewer than 4 CPUs visible to it) replays
@@ -65,3 +67,15 @@ real-time sleep, episode cap and hanoi's wall-clock auto-reset are disabled in t
### Provenance
Generated by `scripts/dexjoco/generate.py` from `scripts/dexjoco/instances/bimanual_assembly/` (`make_instances.py`).
+
+### Change Log
+- 2026-10-14 (review of board 0105, B-M3 and B6-3): `environment/instance.npz` and `tests/instance.npz` no longer name
+ the demonstration the scene was restored from: their `meta` entry had `source` (the dataset) and `demo` (the
+ demonstration's id), and the reference solution is that demonstration's own actions, so the agent could read at
+ `/app/instance.npz` where a solution was (PROTOCOL §2.7). Only `meta` changed: every array of the scene is the same
+ (instance sha `e39abc9b3d77cd72`), so the scene, the pass rule and the reference are unchanged and the task keeps its
+ name. The tooling (`rcb_dexjoco.py`, both copies) refuses to save an instance whose `meta` says where its scene came
+ from. `tests/grade.py`: a rebuilt instance that does not match its recorded hash is now no verdict
+ (`instance_mismatch`, `grader_error`, `infra_error` = 1, nothing replayed); before, it was the agent's 0 with
+ `infra_error` = 0 and `instance_ok` = 0. Past results: the demonstration itself was out of reach under the egress
+ allowlist (Hugging Face is not on it); a past 0 with `instance_ok` = 0 was an infrastructure failure.
diff --git a/tasks/dexjoco-bimanual-assembly-i00-privileged/environment/instance.npz b/tasks/dexjoco-bimanual-assembly-i00-privileged/environment/instance.npz
index d8c094224d..97687642a1 100644
Binary files a/tasks/dexjoco-bimanual-assembly-i00-privileged/environment/instance.npz and b/tasks/dexjoco-bimanual-assembly-i00-privileged/environment/instance.npz differ
diff --git a/tasks/dexjoco-bimanual-assembly-i00-privileged/environment/rcb_dexjoco.py b/tasks/dexjoco-bimanual-assembly-i00-privileged/environment/rcb_dexjoco.py
index f2e2db848f..8eb183fe2a 100644
--- a/tasks/dexjoco-bimanual-assembly-i00-privileged/environment/rcb_dexjoco.py
+++ b/tasks/dexjoco-bimanual-assembly-i00-privileged/environment/rcb_dexjoco.py
@@ -54,6 +54,7 @@ import argparse
import copy
import hashlib
import json
+import re
import sys
import time
from pathlib import Path
@@ -222,6 +223,21 @@ def water_aim_deg(model, data) -> float:
return float(np.degrees(np.arccos(np.clip(c, -1.0, 1.0))))
+# instance.npz is in the agent's container (/app/instance.npz), and a task's reference solution is the actions of the
+# recording its instance was restored from: its `meta` must not say where the scene came from (PROTOCOL §2.7, no pointer
+# to a solution where the agent can reach). save_instance refuses such keys, and values that name a recording.
+PROVENANCE_KEYS = ("source", "demo", "dataset", "episode", "recording", "reference", "solution", "zarr", "url")
+PROVENANCE_VALUE = re.compile(r"demo|episode|dataset|recording|replay\.zarr|huggingface|https?://", re.I)
+
+
+def check_instance_meta(meta: dict) -> None:
+ """ValueError if an instance's `meta` says where its scene came from (keys or values; see PROVENANCE_KEYS)."""
+ bad = [k for k, v in meta.items() if any(p in str(k).lower() for p in PROVENANCE_KEYS)
+ or PROVENANCE_VALUE.search(json.dumps(v))]
+ if bad:
+ raise ValueError(f"instance meta {bad} names where the scene came from: the agent reads instance.npz")
+
+
def _log(*a: Any) -> None:
print("[rcb_dexjoco]", *a, file=sys.stderr, flush=True)
@@ -349,6 +365,7 @@ class Session:
return h.hexdigest()[:16]
def save_instance(self, path: str | Path, meta: dict | None = None) -> None:
+ check_instance_meta(meta or {})
snap = self.snapshot()
np.savez_compressed(path, task=self.task, physics=snap["physics"], mocap_pos=snap["mocap_pos"],
**{"model__" + k: v for k, v in snap["model"].items()},
@@ -803,15 +820,14 @@ def cmd_make_instance(a) -> None:
restore_initial_state(s.env, a.task, s.cfg, state0)
s._snapshot = s.snapshot()
s._episode_reset()
- meta = {"source": "DexJoCo-Datasets-Raw demo, initial state restored with dexjoco.tasks.state_restorers",
- "demo": a.demo_name or os.path.basename(os.path.dirname(os.path.abspath(a.demo))), "dexjoco": "8d23b0f"}
- s.save_instance(a.out, meta)
- _log("instance", a.out, "sha", s.state_sha())
+ demo = a.demo_name or os.path.basename(os.path.dirname(os.path.abspath(a.demo)))
+ s.save_instance(a.out, {"dexjoco": "8d23b0f"}) # no provenance in the instance (check_instance_meta)
+ _log("instance", a.out, "sha", s.state_sha(), "from", demo)
if a.oracle_out:
A = np.asarray(root["data"]["action"], dtype=np.float64)
s2 = Session(a.task, instance=a.out)
st = s2.replay(A)
- s2.save_trajectory(A[:st["replayed_rows"]], a.oracle_out, {"source": "demo actions", "demo": meta["demo"]})
+ s2.save_trajectory(A[:st["replayed_rows"]], a.oracle_out, {"source": "demo actions", "demo": demo})
print(json.dumps({"oracle": a.oracle_out, **{k: st[k] for k in ("success", "reason", "replayed_rows", "rows", "state_sha")}}))
diff --git a/tasks/dexjoco-bimanual-assembly-i00-privileged/task.toml b/tasks/dexjoco-bimanual-assembly-i00-privileged/task.toml
index e502b9f9c8..4084932778 100644
--- a/tasks/dexjoco-bimanual-assembly-i00-privileged/task.toml
+++ b/tasks/dexjoco-bimanual-assembly-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "dexjoco"
robot = "franka-allegro-bimanual"
family = "bimanual_assembly"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
expert_time_estimate_hours = 2.5
diff --git a/tasks/dexjoco-bimanual-assembly-i00-privileged/tests/grade.py b/tasks/dexjoco-bimanual-assembly-i00-privileged/tests/grade.py
index c754f999ad..3cfbab9c42 100644
--- a/tasks/dexjoco-bimanual-assembly-i00-privileged/tests/grade.py
+++ b/tasks/dexjoco-bimanual-assembly-i00-privileged/tests/grade.py
@@ -6,8 +6,10 @@ write Harbor rewards.
Writes /logs/verifier/reward.txt (1/0), reward.json (numeric fields only; Harbor rejects strings), a replay video
/logs/verifier/replay.mp4 (front camera + wrist camera) and replay.last.png. Replays `--repeat` times in this fresh
process; success requires every replay to end its episode in success (DexJoCo's own rule: the episode ends at the first
-success or task failure) AND all final-state hashes to agree (MuJoCo on the CPU is bit-deterministic), and the frozen
-instance rebuilt here from tests/instance.npz must match its recorded hash.
+success or task failure) AND all final-state hashes to agree (MuJoCo on the CPU is bit-deterministic). The frozen
+instance rebuilt here from tests/instance.npz must match its recorded hash; one that does not (the image no longer
+rebuilds the scene) is no verdict on the agent: `instance_mismatch`, `grader_error` and `infra_error` = 1, nothing
+replayed.
A file the agent got wrong is the agent's 0, `bad_trajectory` (not `grader_error`): not an npz, pickled or object
arrays, no `task` or `actions` entry, a wrong shape, NaN/inf, or more than rcb_dexjoco.MAX_TRAJECTORY_ROWS rows (the
@@ -119,6 +121,10 @@ def main() -> int:
print("too few CPUs: " + rcb_dexjoco.threads_message(threads), file=sys.stderr)
write(False, {"grader_error": 1, "infra_error": 1, "too_few_cpus": 1, "n_actions": n_rows, **threads, **cost}); return 0
s = Session(a.family, instance=a.instance)
+ if not s.instance_exact: # the image no longer rebuilds the frozen scene: not the agent's 0 (infra, no verdict)
+ print(f"instance mismatch: rebuilt state sha {s.state_sha()} != recorded {s.instance_sha}", file=sys.stderr)
+ write(False, {"grader_error": 1, "infra_error": 1, "instance_mismatch": 1, "instance_ok": 0,
+ "n_actions": n_rows, **threads, **cost}); return 0
if acts.shape[1] != s.action_dim:
write(False, {"bad_trajectory": 1, "n_actions": n_rows, **cost}); return 0
t0 = time.time()
@@ -130,10 +136,9 @@ def main() -> int:
write(False, {"bad_trajectory": 1, "n_actions": n_rows, **cost}); return 0
all_success = all(o["success"] for o in outs)
deterministic = len({o["state_sha"] for o in outs}) == 1
- instance_ok = bool(s.instance_exact)
- ok = all_success and deterministic and instance_ok
+ ok = all_success and deterministic
out = {"success": int(ok), "all_success": int(all_success), "deterministic": int(deterministic),
- "instance_ok": int(instance_ok), "n_actions": int(len(acts)), "sim_steps": int(outs[0]["steps"]),
+ "instance_ok": 1, "n_actions": int(len(acts)), "sim_steps": int(outs[0]["steps"]),
"task_failure": int(bool(outs[0]["reason"]) and not outs[0]["success"]),
"replay_wall_s": round(time.time() - t0, 1), "grader_error": 0, "video_rendered": 0, **threads, **cost}
write(ok, out) # the verdict is on disk before the video; the video only adds video_rendered
diff --git a/tasks/dexjoco-bimanual-assembly-i00-privileged/tests/instance.npz b/tasks/dexjoco-bimanual-assembly-i00-privileged/tests/instance.npz
index d8c094224d..97687642a1 100644
Binary files a/tasks/dexjoco-bimanual-assembly-i00-privileged/tests/instance.npz and b/tasks/dexjoco-bimanual-assembly-i00-privileged/tests/instance.npz differ
diff --git a/tasks/dexjoco-bimanual-assembly-i00-privileged/tests/rcb_dexjoco.py b/tasks/dexjoco-bimanual-assembly-i00-privileged/tests/rcb_dexjoco.py
index f2e2db848f..8eb183fe2a 100644
--- a/tasks/dexjoco-bimanual-assembly-i00-privileged/tests/rcb_dexjoco.py
+++ b/tasks/dexjoco-bimanual-assembly-i00-privileged/tests/rcb_dexjoco.py
@@ -54,6 +54,7 @@ import argparse
import copy
import hashlib
import json
+import re
import sys
import time
from pathlib import Path
@@ -222,6 +223,21 @@ def water_aim_deg(model, data) -> float:
return float(np.degrees(np.arccos(np.clip(c, -1.0, 1.0))))
+# instance.npz is in the agent's container (/app/instance.npz), and a task's reference solution is the actions of the
+# recording its instance was restored from: its `meta` must not say where the scene came from (PROTOCOL §2.7, no pointer
+# to a solution where the agent can reach). save_instance refuses such keys, and values that name a recording.
+PROVENANCE_KEYS = ("source", "demo", "dataset", "episode", "recording", "reference", "solution", "zarr", "url")
+PROVENANCE_VALUE = re.compile(r"demo|episode|dataset|recording|replay\.zarr|huggingface|https?://", re.I)
+
+
+def check_instance_meta(meta: dict) -> None:
+ """ValueError if an instance's `meta` says where its scene came from (keys or values; see PROVENANCE_KEYS)."""
+ bad = [k for k, v in meta.items() if any(p in str(k).lower() for p in PROVENANCE_KEYS)
+ or PROVENANCE_VALUE.search(json.dumps(v))]
+ if bad:
+ raise ValueError(f"instance meta {bad} names where the scene came from: the agent reads instance.npz")
+
+
def _log(*a: Any) -> None:
print("[rcb_dexjoco]", *a, file=sys.stderr, flush=True)
@@ -349,6 +365,7 @@ class Session:
return h.hexdigest()[:16]
def save_instance(self, path: str | Path, meta: dict | None = None) -> None:
+ check_instance_meta(meta or {})
snap = self.snapshot()
np.savez_compressed(path, task=self.task, physics=snap["physics"], mocap_pos=snap["mocap_pos"],
**{"model__" + k: v for k, v in snap["model"].items()},
@@ -803,15 +820,14 @@ def cmd_make_instance(a) -> None:
restore_initial_state(s.env, a.task, s.cfg, state0)
s._snapshot = s.snapshot()
s._episode_reset()
- meta = {"source": "DexJoCo-Datasets-Raw demo, initial state restored with dexjoco.tasks.state_restorers",
- "demo": a.demo_name or os.path.basename(os.path.dirname(os.path.abspath(a.demo))), "dexjoco": "8d23b0f"}
- s.save_instance(a.out, meta)
- _log("instance", a.out, "sha", s.state_sha())
+ demo = a.demo_name or os.path.basename(os.path.dirname(os.path.abspath(a.demo)))
+ s.save_instance(a.out, {"dexjoco": "8d23b0f"}) # no provenance in the instance (check_instance_meta)
+ _log("instance", a.out, "sha", s.state_sha(), "from", demo)
if a.oracle_out:
A = np.asarray(root["data"]["action"], dtype=np.float64)
s2 = Session(a.task, instance=a.out)
st = s2.replay(A)
- s2.save_trajectory(A[:st["replayed_rows"]], a.oracle_out, {"source": "demo actions", "demo": meta["demo"]})
+ s2.save_trajectory(A[:st["replayed_rows"]], a.oracle_out, {"source": "demo actions", "demo": demo})
print(json.dumps({"oracle": a.oracle_out, **{k: st[k] for k in ("success", "reason", "replayed_rows", "rows", "state_sha")}}))
diff --git a/tasks/dexjoco-bimanual-assembly-i00-standard/README.md b/tasks/dexjoco-bimanual-assembly-i00-standard/README.md
index b022c1391f..454dd9b032 100644
--- a/tasks/dexjoco-bimanual-assembly-i00-standard/README.md
+++ b/tasks/dexjoco-bimanual-assembly-i00-standard/README.md
@@ -72,6 +72,13 @@ Generated by `scripts/dexjoco/generate.py`, as its privileged twin.
returned once. A service that has given up on the episode (too few CPUs, or it could not resume after its
crashes) answers every request with 409 and the reason, and `Robot()` / `wait_ready()` raise `EpisodeOver` at
once; before, nothing answered and they waited out their timeout (up to an hour).
+- 2026-10-14 (review of board 0105, B-M3): `environment/sim/instance.npz` and `environment/sim/rcb_dexjoco.py`. The
+ frozen instance's `meta` entry no longer names the demonstration the scene was restored from (`source`, `demo`;
+ the reference solution is that demonstration's actions), and the tooling refuses to save an instance that does.
+ The sidecar's files are out of the agent's reach, so nothing reached the agent from this copy; it changed with
+ the privileged twin's, from the one committed instance. Every array of the scene is the same (instance sha
+ `e39abc9b3d77cd72`): the scene, the pass rule and the grader are unchanged, so the task keeps its name and earlier
+ results stand.
- 2026-10-13 (owner, boards 0085 and 0087): the sim service's clock (`environment/sim/rcb_limited.py`,
`environment/sim/server.py`, the docstring of `environment/rcb_client.py`). `meta["time_s"]` and
`meta["time_left_s"]` count from the agent's first request to the service (any endpoint), and a restarted service
diff --git a/tasks/dexjoco-bimanual-assembly-i00-standard/environment/sim/instance.npz b/tasks/dexjoco-bimanual-assembly-i00-standard/environment/sim/instance.npz
index d8c094224d..97687642a1 100644
Binary files a/tasks/dexjoco-bimanual-assembly-i00-standard/environment/sim/instance.npz and b/tasks/dexjoco-bimanual-assembly-i00-standard/environment/sim/instance.npz differ
diff --git a/tasks/dexjoco-bimanual-assembly-i00-standard/environment/sim/rcb_dexjoco.py b/tasks/dexjoco-bimanual-assembly-i00-standard/environment/sim/rcb_dexjoco.py
index f2e2db848f..8eb183fe2a 100644
--- a/tasks/dexjoco-bimanual-assembly-i00-standard/environment/sim/rcb_dexjoco.py
+++ b/tasks/dexjoco-bimanual-assembly-i00-standard/environment/sim/rcb_dexjoco.py
@@ -54,6 +54,7 @@ import argparse
import copy
import hashlib
import json
+import re
import sys
import time
from pathlib import Path
@@ -222,6 +223,21 @@ def water_aim_deg(model, data) -> float:
return float(np.degrees(np.arccos(np.clip(c, -1.0, 1.0))))
+# instance.npz is in the agent's container (/app/instance.npz), and a task's reference solution is the actions of the
+# recording its instance was restored from: its `meta` must not say where the scene came from (PROTOCOL §2.7, no pointer
+# to a solution where the agent can reach). save_instance refuses such keys, and values that name a recording.
+PROVENANCE_KEYS = ("source", "demo", "dataset", "episode", "recording", "reference", "solution", "zarr", "url")
+PROVENANCE_VALUE = re.compile(r"demo|episode|dataset|recording|replay\.zarr|huggingface|https?://", re.I)
+
+
+def check_instance_meta(meta: dict) -> None:
+ """ValueError if an instance's `meta` says where its scene came from (keys or values; see PROVENANCE_KEYS)."""
+ bad = [k for k, v in meta.items() if any(p in str(k).lower() for p in PROVENANCE_KEYS)
+ or PROVENANCE_VALUE.search(json.dumps(v))]
+ if bad:
+ raise ValueError(f"instance meta {bad} names where the scene came from: the agent reads instance.npz")
+
+
def _log(*a: Any) -> None:
print("[rcb_dexjoco]", *a, file=sys.stderr, flush=True)
@@ -349,6 +365,7 @@ class Session:
return h.hexdigest()[:16]
def save_instance(self, path: str | Path, meta: dict | None = None) -> None:
+ check_instance_meta(meta or {})
snap = self.snapshot()
np.savez_compressed(path, task=self.task, physics=snap["physics"], mocap_pos=snap["mocap_pos"],
**{"model__" + k: v for k, v in snap["model"].items()},
@@ -803,15 +820,14 @@ def cmd_make_instance(a) -> None:
restore_initial_state(s.env, a.task, s.cfg, state0)
s._snapshot = s.snapshot()
s._episode_reset()
- meta = {"source": "DexJoCo-Datasets-Raw demo, initial state restored with dexjoco.tasks.state_restorers",
- "demo": a.demo_name or os.path.basename(os.path.dirname(os.path.abspath(a.demo))), "dexjoco": "8d23b0f"}
- s.save_instance(a.out, meta)
- _log("instance", a.out, "sha", s.state_sha())
+ demo = a.demo_name or os.path.basename(os.path.dirname(os.path.abspath(a.demo)))
+ s.save_instance(a.out, {"dexjoco": "8d23b0f"}) # no provenance in the instance (check_instance_meta)
+ _log("instance", a.out, "sha", s.state_sha(), "from", demo)
if a.oracle_out:
A = np.asarray(root["data"]["action"], dtype=np.float64)
s2 = Session(a.task, instance=a.out)
st = s2.replay(A)
- s2.save_trajectory(A[:st["replayed_rows"]], a.oracle_out, {"source": "demo actions", "demo": meta["demo"]})
+ s2.save_trajectory(A[:st["replayed_rows"]], a.oracle_out, {"source": "demo actions", "demo": demo})
print(json.dumps({"oracle": a.oracle_out, **{k: st[k] for k in ("success", "reason", "replayed_rows", "rows", "state_sha")}}))
diff --git a/tasks/dexjoco-bimanual-assembly-i00-standard/task.toml b/tasks/dexjoco-bimanual-assembly-i00-standard/task.toml
index 643942762b..245537effb 100644
--- a/tasks/dexjoco-bimanual-assembly-i00-standard/task.toml
+++ b/tasks/dexjoco-bimanual-assembly-i00-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "dexjoco"
robot = "franka-allegro-bimanual"
family = "bimanual_assembly"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "full"
expert_time_estimate_hours = 2.5
diff --git a/tasks/dexjoco-bimanual-hanoi-i00-privileged/README.md b/tasks/dexjoco-bimanual-hanoi-i00-privileged/README.md
index 699ff8fbb9..2385901185 100644
--- a/tasks/dexjoco-bimanual-hanoi-i00-privileged/README.md
+++ b/tasks/dexjoco-bimanual-hanoi-i00-privileged/README.md
@@ -29,8 +29,10 @@ limited to the model APIs.
## Verification
Open-loop replay of `/app/output/trajectory.npz` from `tests/instance.npz`, twice in a fresh process in the verifier's own
-sandbox (no network); pass iff both episodes end in success, their final-state hashes agree and the rebuilt instance
-matches its recorded hash (`tests/grade.py`); plus `/logs/verifier/replay.mp4`, rendered after the verdict is written.
+sandbox (no network); pass iff both episodes end in success and their final-state hashes agree (`tests/grade.py`); plus
+`/logs/verifier/replay.mp4`, rendered after the verdict is written. A rebuilt instance that does not match its recorded
+hash (the image no longer rebuilds the scene) replays nothing and gives no verdict: `instance_mismatch`, `grader_error`
+and `infra_error` = 1, not the agent's 0.
A file that is not a readable trajectory (not an npz, no `task` / `actions`, a wrong shape, NaN) or has more than
50,000 rows scores 0 as `bad_trajectory`: the cap is what two replays of the slowest family fit into the verifier's
3600 s. A verifier that runs fewer than 4 OpenBLAS threads (fewer than 4 CPUs visible to it) replays
@@ -65,3 +67,15 @@ real-time sleep, episode cap and hanoi's wall-clock auto-reset are disabled in t
### Provenance
Generated by `scripts/dexjoco/generate.py` from `scripts/dexjoco/instances/bimanual_hanoi/` (`make_instances.py`).
+
+### Change Log
+- 2026-10-14 (review of board 0105, B-M3 and B6-3): `environment/instance.npz` and `tests/instance.npz` no longer name
+ the demonstration the scene was restored from: their `meta` entry had `source` (the dataset) and `demo` (the
+ demonstration's id), and the reference solution is that demonstration's own actions, so the agent could read at
+ `/app/instance.npz` where a solution was (PROTOCOL §2.7). Only `meta` changed: every array of the scene is the same
+ (instance sha `93165d7fd0c8813c`), so the scene, the pass rule and the reference are unchanged and the task keeps its
+ name. The tooling (`rcb_dexjoco.py`, both copies) refuses to save an instance whose `meta` says where its scene came
+ from. `tests/grade.py`: a rebuilt instance that does not match its recorded hash is now no verdict
+ (`instance_mismatch`, `grader_error`, `infra_error` = 1, nothing replayed); before, it was the agent's 0 with
+ `infra_error` = 0 and `instance_ok` = 0. Past results: the demonstration itself was out of reach under the egress
+ allowlist (Hugging Face is not on it); a past 0 with `instance_ok` = 0 was an infrastructure failure.
diff --git a/tasks/dexjoco-bimanual-hanoi-i00-privileged/environment/instance.npz b/tasks/dexjoco-bimanual-hanoi-i00-privileged/environment/instance.npz
index 0f1065cea8..552afa8a8c 100644
Binary files a/tasks/dexjoco-bimanual-hanoi-i00-privileged/environment/instance.npz and b/tasks/dexjoco-bimanual-hanoi-i00-privileged/environment/instance.npz differ
diff --git a/tasks/dexjoco-bimanual-hanoi-i00-privileged/environment/rcb_dexjoco.py b/tasks/dexjoco-bimanual-hanoi-i00-privileged/environment/rcb_dexjoco.py
index f2e2db848f..8eb183fe2a 100644
--- a/tasks/dexjoco-bimanual-hanoi-i00-privileged/environment/rcb_dexjoco.py
+++ b/tasks/dexjoco-bimanual-hanoi-i00-privileged/environment/rcb_dexjoco.py
@@ -54,6 +54,7 @@ import argparse
import copy
import hashlib
import json
+import re
import sys
import time
from pathlib import Path
@@ -222,6 +223,21 @@ def water_aim_deg(model, data) -> float:
return float(np.degrees(np.arccos(np.clip(c, -1.0, 1.0))))
+# instance.npz is in the agent's container (/app/instance.npz), and a task's reference solution is the actions of the
+# recording its instance was restored from: its `meta` must not say where the scene came from (PROTOCOL §2.7, no pointer
+# to a solution where the agent can reach). save_instance refuses such keys, and values that name a recording.
+PROVENANCE_KEYS = ("source", "demo", "dataset", "episode", "recording", "reference", "solution", "zarr", "url")
+PROVENANCE_VALUE = re.compile(r"demo|episode|dataset|recording|replay\.zarr|huggingface|https?://", re.I)
+
+
+def check_instance_meta(meta: dict) -> None:
+ """ValueError if an instance's `meta` says where its scene came from (keys or values; see PROVENANCE_KEYS)."""
+ bad = [k for k, v in meta.items() if any(p in str(k).lower() for p in PROVENANCE_KEYS)
+ or PROVENANCE_VALUE.search(json.dumps(v))]
+ if bad:
+ raise ValueError(f"instance meta {bad} names where the scene came from: the agent reads instance.npz")
+
+
def _log(*a: Any) -> None:
print("[rcb_dexjoco]", *a, file=sys.stderr, flush=True)
@@ -349,6 +365,7 @@ class Session:
return h.hexdigest()[:16]
def save_instance(self, path: str | Path, meta: dict | None = None) -> None:
+ check_instance_meta(meta or {})
snap = self.snapshot()
np.savez_compressed(path, task=self.task, physics=snap["physics"], mocap_pos=snap["mocap_pos"],
**{"model__" + k: v for k, v in snap["model"].items()},
@@ -803,15 +820,14 @@ def cmd_make_instance(a) -> None:
restore_initial_state(s.env, a.task, s.cfg, state0)
s._snapshot = s.snapshot()
s._episode_reset()
- meta = {"source": "DexJoCo-Datasets-Raw demo, initial state restored with dexjoco.tasks.state_restorers",
- "demo": a.demo_name or os.path.basename(os.path.dirname(os.path.abspath(a.demo))), "dexjoco": "8d23b0f"}
- s.save_instance(a.out, meta)
- _log("instance", a.out, "sha", s.state_sha())
+ demo = a.demo_name or os.path.basename(os.path.dirname(os.path.abspath(a.demo)))
+ s.save_instance(a.out, {"dexjoco": "8d23b0f"}) # no provenance in the instance (check_instance_meta)
+ _log("instance", a.out, "sha", s.state_sha(), "from", demo)
if a.oracle_out:
A = np.asarray(root["data"]["action"], dtype=np.float64)
s2 = Session(a.task, instance=a.out)
st = s2.replay(A)
- s2.save_trajectory(A[:st["replayed_rows"]], a.oracle_out, {"source": "demo actions", "demo": meta["demo"]})
+ s2.save_trajectory(A[:st["replayed_rows"]], a.oracle_out, {"source": "demo actions", "demo": demo})
print(json.dumps({"oracle": a.oracle_out, **{k: st[k] for k in ("success", "reason", "replayed_rows", "rows", "state_sha")}}))
diff --git a/tasks/dexjoco-bimanual-hanoi-i00-privileged/task.toml b/tasks/dexjoco-bimanual-hanoi-i00-privileged/task.toml
index 6ffe9b1271..9cb8bc6210 100644
--- a/tasks/dexjoco-bimanual-hanoi-i00-privileged/task.toml
+++ b/tasks/dexjoco-bimanual-hanoi-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "dexjoco"
robot = "franka-allegro-bimanual"
family = "bimanual_hanoi"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
expert_time_estimate_hours = 2.0
diff --git a/tasks/dexjoco-bimanual-hanoi-i00-privileged/tests/grade.py b/tasks/dexjoco-bimanual-hanoi-i00-privileged/tests/grade.py
index c754f999ad..3cfbab9c42 100644
--- a/tasks/dexjoco-bimanual-hanoi-i00-privileged/tests/grade.py
+++ b/tasks/dexjoco-bimanual-hanoi-i00-privileged/tests/grade.py
@@ -6,8 +6,10 @@ write Harbor rewards.
Writes /logs/verifier/reward.txt (1/0), reward.json (numeric fields only; Harbor rejects strings), a replay video
/logs/verifier/replay.mp4 (front camera + wrist camera) and replay.last.png. Replays `--repeat` times in this fresh
process; success requires every replay to end its episode in success (DexJoCo's own rule: the episode ends at the first
-success or task failure) AND all final-state hashes to agree (MuJoCo on the CPU is bit-deterministic), and the frozen
-instance rebuilt here from tests/instance.npz must match its recorded hash.
+success or task failure) AND all final-state hashes to agree (MuJoCo on the CPU is bit-deterministic). The frozen
+instance rebuilt here from tests/instance.npz must match its recorded hash; one that does not (the image no longer
+rebuilds the scene) is no verdict on the agent: `instance_mismatch`, `grader_error` and `infra_error` = 1, nothing
+replayed.
A file the agent got wrong is the agent's 0, `bad_trajectory` (not `grader_error`): not an npz, pickled or object
arrays, no `task` or `actions` entry, a wrong shape, NaN/inf, or more than rcb_dexjoco.MAX_TRAJECTORY_ROWS rows (the
@@ -119,6 +121,10 @@ def main() -> int:
print("too few CPUs: " + rcb_dexjoco.threads_message(threads), file=sys.stderr)
write(False, {"grader_error": 1, "infra_error": 1, "too_few_cpus": 1, "n_actions": n_rows, **threads, **cost}); return 0
s = Session(a.family, instance=a.instance)
+ if not s.instance_exact: # the image no longer rebuilds the frozen scene: not the agent's 0 (infra, no verdict)
+ print(f"instance mismatch: rebuilt state sha {s.state_sha()} != recorded {s.instance_sha}", file=sys.stderr)
+ write(False, {"grader_error": 1, "infra_error": 1, "instance_mismatch": 1, "instance_ok": 0,
+ "n_actions": n_rows, **threads, **cost}); return 0
if acts.shape[1] != s.action_dim:
write(False, {"bad_trajectory": 1, "n_actions": n_rows, **cost}); return 0
t0 = time.time()
@@ -130,10 +136,9 @@ def main() -> int:
write(False, {"bad_trajectory": 1, "n_actions": n_rows, **cost}); return 0
all_success = all(o["success"] for o in outs)
deterministic = len({o["state_sha"] for o in outs}) == 1
- instance_ok = bool(s.instance_exact)
- ok = all_success and deterministic and instance_ok
+ ok = all_success and deterministic
out = {"success": int(ok), "all_success": int(all_success), "deterministic": int(deterministic),
- "instance_ok": int(instance_ok), "n_actions": int(len(acts)), "sim_steps": int(outs[0]["steps"]),
+ "instance_ok": 1, "n_actions": int(len(acts)), "sim_steps": int(outs[0]["steps"]),
"task_failure": int(bool(outs[0]["reason"]) and not outs[0]["success"]),
"replay_wall_s": round(time.time() - t0, 1), "grader_error": 0, "video_rendered": 0, **threads, **cost}
write(ok, out) # the verdict is on disk before the video; the video only adds video_rendered
diff --git a/tasks/dexjoco-bimanual-hanoi-i00-privileged/tests/instance.npz b/tasks/dexjoco-bimanual-hanoi-i00-privileged/tests/instance.npz
index 0f1065cea8..552afa8a8c 100644
Binary files a/tasks/dexjoco-bimanual-hanoi-i00-privileged/tests/instance.npz and b/tasks/dexjoco-bimanual-hanoi-i00-privileged/tests/instance.npz differ
diff --git a/tasks/dexjoco-bimanual-hanoi-i00-privileged/tests/rcb_dexjoco.py b/tasks/dexjoco-bimanual-hanoi-i00-privileged/tests/rcb_dexjoco.py
index f2e2db848f..8eb183fe2a 100644
--- a/tasks/dexjoco-bimanual-hanoi-i00-privileged/tests/rcb_dexjoco.py
+++ b/tasks/dexjoco-bimanual-hanoi-i00-privileged/tests/rcb_dexjoco.py
@@ -54,6 +54,7 @@ import argparse
import copy
import hashlib
import json
+import re
import sys
import time
from pathlib import Path
@@ -222,6 +223,21 @@ def water_aim_deg(model, data) -> float:
return float(np.degrees(np.arccos(np.clip(c, -1.0, 1.0))))
+# instance.npz is in the agent's container (/app/instance.npz), and a task's reference solution is the actions of the
+# recording its instance was restored from: its `meta` must not say where the scene came from (PROTOCOL §2.7, no pointer
+# to a solution where the agent can reach). save_instance refuses such keys, and values that name a recording.
+PROVENANCE_KEYS = ("source", "demo", "dataset", "episode", "recording", "reference", "solution", "zarr", "url")
+PROVENANCE_VALUE = re.compile(r"demo|episode|dataset|recording|replay\.zarr|huggingface|https?://", re.I)
+
+
+def check_instance_meta(meta: dict) -> None:
+ """ValueError if an instance's `meta` says where its scene came from (keys or values; see PROVENANCE_KEYS)."""
+ bad = [k for k, v in meta.items() if any(p in str(k).lower() for p in PROVENANCE_KEYS)
+ or PROVENANCE_VALUE.search(json.dumps(v))]
+ if bad:
+ raise ValueError(f"instance meta {bad} names where the scene came from: the agent reads instance.npz")
+
+
def _log(*a: Any) -> None:
print("[rcb_dexjoco]", *a, file=sys.stderr, flush=True)
@@ -349,6 +365,7 @@ class Session:
return h.hexdigest()[:16]
def save_instance(self, path: str | Path, meta: dict | None = None) -> None:
+ check_instance_meta(meta or {})
snap = self.snapshot()
np.savez_compressed(path, task=self.task, physics=snap["physics"], mocap_pos=snap["mocap_pos"],
**{"model__" + k: v for k, v in snap["model"].items()},
@@ -803,15 +820,14 @@ def cmd_make_instance(a) -> None:
restore_initial_state(s.env, a.task, s.cfg, state0)
s._snapshot = s.snapshot()
s._episode_reset()
- meta = {"source": "DexJoCo-Datasets-Raw demo, initial state restored with dexjoco.tasks.state_restorers",
- "demo": a.demo_name or os.path.basename(os.path.dirname(os.path.abspath(a.demo))), "dexjoco": "8d23b0f"}
- s.save_instance(a.out, meta)
- _log("instance", a.out, "sha", s.state_sha())
+ demo = a.demo_name or os.path.basename(os.path.dirname(os.path.abspath(a.demo)))
+ s.save_instance(a.out, {"dexjoco": "8d23b0f"}) # no provenance in the instance (check_instance_meta)
+ _log("instance", a.out, "sha", s.state_sha(), "from", demo)
if a.oracle_out:
A = np.asarray(root["data"]["action"], dtype=np.float64)
s2 = Session(a.task, instance=a.out)
st = s2.replay(A)
- s2.save_trajectory(A[:st["replayed_rows"]], a.oracle_out, {"source": "demo actions", "demo": meta["demo"]})
+ s2.save_trajectory(A[:st["replayed_rows"]], a.oracle_out, {"source": "demo actions", "demo": demo})
print(json.dumps({"oracle": a.oracle_out, **{k: st[k] for k in ("success", "reason", "replayed_rows", "rows", "state_sha")}}))
diff --git a/tasks/dexjoco-bimanual-hanoi-i00-standard/README.md b/tasks/dexjoco-bimanual-hanoi-i00-standard/README.md
index d4f5f3dc86..9fc4563c3a 100644
--- a/tasks/dexjoco-bimanual-hanoi-i00-standard/README.md
+++ b/tasks/dexjoco-bimanual-hanoi-i00-standard/README.md
@@ -72,6 +72,13 @@ Generated by `scripts/dexjoco/generate.py`, as its privileged twin.
returned once. A service that has given up on the episode (too few CPUs, or it could not resume after its
crashes) answers every request with 409 and the reason, and `Robot()` / `wait_ready()` raise `EpisodeOver` at
once; before, nothing answered and they waited out their timeout (up to an hour).
+- 2026-10-14 (review of board 0105, B-M3): `environment/sim/instance.npz` and `environment/sim/rcb_dexjoco.py`. The
+ frozen instance's `meta` entry no longer names the demonstration the scene was restored from (`source`, `demo`;
+ the reference solution is that demonstration's actions), and the tooling refuses to save an instance that does.
+ The sidecar's files are out of the agent's reach, so nothing reached the agent from this copy; it changed with
+ the privileged twin's, from the one committed instance. Every array of the scene is the same (instance sha
+ `93165d7fd0c8813c`): the scene, the pass rule and the grader are unchanged, so the task keeps its name and earlier
+ results stand.
- 2026-10-13 (owner, boards 0085 and 0087): the sim service's clock (`environment/sim/rcb_limited.py`,
`environment/sim/server.py`, the docstring of `environment/rcb_client.py`). `meta["time_s"]` and
`meta["time_left_s"]` count from the agent's first request to the service (any endpoint), and a restarted service
diff --git a/tasks/dexjoco-bimanual-hanoi-i00-standard/environment/sim/instance.npz b/tasks/dexjoco-bimanual-hanoi-i00-standard/environment/sim/instance.npz
index 0f1065cea8..552afa8a8c 100644
Binary files a/tasks/dexjoco-bimanual-hanoi-i00-standard/environment/sim/instance.npz and b/tasks/dexjoco-bimanual-hanoi-i00-standard/environment/sim/instance.npz differ
diff --git a/tasks/dexjoco-bimanual-hanoi-i00-standard/environment/sim/rcb_dexjoco.py b/tasks/dexjoco-bimanual-hanoi-i00-standard/environment/sim/rcb_dexjoco.py
index f2e2db848f..8eb183fe2a 100644
--- a/tasks/dexjoco-bimanual-hanoi-i00-standard/environment/sim/rcb_dexjoco.py
+++ b/tasks/dexjoco-bimanual-hanoi-i00-standard/environment/sim/rcb_dexjoco.py
@@ -54,6 +54,7 @@ import argparse
import copy
import hashlib
import json
+import re
import sys
import time
from pathlib import Path
@@ -222,6 +223,21 @@ def water_aim_deg(model, data) -> float:
return float(np.degrees(np.arccos(np.clip(c, -1.0, 1.0))))
+# instance.npz is in the agent's container (/app/instance.npz), and a task's reference solution is the actions of the
+# recording its instance was restored from: its `meta` must not say where the scene came from (PROTOCOL §2.7, no pointer
+# to a solution where the agent can reach). save_instance refuses such keys, and values that name a recording.
+PROVENANCE_KEYS = ("source", "demo", "dataset", "episode", "recording", "reference", "solution", "zarr", "url")
+PROVENANCE_VALUE = re.compile(r"demo|episode|dataset|recording|replay\.zarr|huggingface|https?://", re.I)
+
+
+def check_instance_meta(meta: dict) -> None:
+ """ValueError if an instance's `meta` says where its scene came from (keys or values; see PROVENANCE_KEYS)."""
+ bad = [k for k, v in meta.items() if any(p in str(k).lower() for p in PROVENANCE_KEYS)
+ or PROVENANCE_VALUE.search(json.dumps(v))]
+ if bad:
+ raise ValueError(f"instance meta {bad} names where the scene came from: the agent reads instance.npz")
+
+
def _log(*a: Any) -> None:
print("[rcb_dexjoco]", *a, file=sys.stderr, flush=True)
@@ -349,6 +365,7 @@ class Session:
return h.hexdigest()[:16]
def save_instance(self, path: str | Path, meta: dict | None = None) -> None:
+ check_instance_meta(meta or {})
snap = self.snapshot()
np.savez_compressed(path, task=self.task, physics=snap["physics"], mocap_pos=snap["mocap_pos"],
**{"model__" + k: v for k, v in snap["model"].items()},
@@ -803,15 +820,14 @@ def cmd_make_instance(a) -> None:
restore_initial_state(s.env, a.task, s.cfg, state0)
s._snapshot = s.snapshot()
s._episode_reset()
- meta = {"source": "DexJoCo-Datasets-Raw demo, initial state restored with dexjoco.tasks.state_restorers",
- "demo": a.demo_name or os.path.basename(os.path.dirname(os.path.abspath(a.demo))), "dexjoco": "8d23b0f"}
- s.save_instance(a.out, meta)
- _log("instance", a.out, "sha", s.state_sha())
+ demo = a.demo_name or os.path.basename(os.path.dirname(os.path.abspath(a.demo)))
+ s.save_instance(a.out, {"dexjoco": "8d23b0f"}) # no provenance in the instance (check_instance_meta)
+ _log("instance", a.out, "sha", s.state_sha(), "from", demo)
if a.oracle_out:
A = np.asarray(root["data"]["action"], dtype=np.float64)
s2 = Session(a.task, instance=a.out)
st = s2.replay(A)
- s2.save_trajectory(A[:st["replayed_rows"]], a.oracle_out, {"source": "demo actions", "demo": meta["demo"]})
+ s2.save_trajectory(A[:st["replayed_rows"]], a.oracle_out, {"source": "demo actions", "demo": demo})
print(json.dumps({"oracle": a.oracle_out, **{k: st[k] for k in ("success", "reason", "replayed_rows", "rows", "state_sha")}}))
diff --git a/tasks/dexjoco-bimanual-hanoi-i00-standard/task.toml b/tasks/dexjoco-bimanual-hanoi-i00-standard/task.toml
index dc1b72f9c0..c7f89cef4a 100644
--- a/tasks/dexjoco-bimanual-hanoi-i00-standard/task.toml
+++ b/tasks/dexjoco-bimanual-hanoi-i00-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "dexjoco"
robot = "franka-allegro-bimanual"
family = "bimanual_hanoi"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "full"
expert_time_estimate_hours = 2.0
diff --git a/tasks/dexjoco-bimanual-microwave-cook-i00-privileged/README.md b/tasks/dexjoco-bimanual-microwave-cook-i00-privileged/README.md
index 45324563b6..abf710c878 100644
--- a/tasks/dexjoco-bimanual-microwave-cook-i00-privileged/README.md
+++ b/tasks/dexjoco-bimanual-microwave-cook-i00-privileged/README.md
@@ -29,8 +29,10 @@ limited to the model APIs.
## Verification
Open-loop replay of `/app/output/trajectory.npz` from `tests/instance.npz`, twice in a fresh process in the verifier's own
-sandbox (no network); pass iff both episodes end in success, their final-state hashes agree and the rebuilt instance
-matches its recorded hash (`tests/grade.py`); plus `/logs/verifier/replay.mp4`, rendered after the verdict is written.
+sandbox (no network); pass iff both episodes end in success and their final-state hashes agree (`tests/grade.py`); plus
+`/logs/verifier/replay.mp4`, rendered after the verdict is written. A rebuilt instance that does not match its recorded
+hash (the image no longer rebuilds the scene) replays nothing and gives no verdict: `instance_mismatch`, `grader_error`
+and `infra_error` = 1, not the agent's 0.
A file that is not a readable trajectory (not an npz, no `task` / `actions`, a wrong shape, NaN) or has more than
50,000 rows scores 0 as `bad_trajectory`: the cap is what two replays of the slowest family fit into the verifier's
3600 s. A verifier that runs fewer than 4 OpenBLAS threads (fewer than 4 CPUs visible to it) replays
@@ -65,3 +67,15 @@ real-time sleep, episode cap and hanoi's wall-clock auto-reset are disabled in t
### Provenance
Generated by `scripts/dexjoco/generate.py` from `scripts/dexjoco/instances/bimanual_microwave_cook/` (`make_instances.py`).
+
+### Change Log
+- 2026-10-14 (review of board 0105, B-M3 and B6-3): `environment/instance.npz` and `tests/instance.npz` no longer name
+ the demonstration the scene was restored from: their `meta` entry had `source` (the dataset) and `demo` (the
+ demonstration's id), and the reference solution is that demonstration's own actions, so the agent could read at
+ `/app/instance.npz` where a solution was (PROTOCOL §2.7). Only `meta` changed: every array of the scene is the same
+ (instance sha `6c16f3cfa1c737c7`), so the scene, the pass rule and the reference are unchanged and the task keeps its
+ name. The tooling (`rcb_dexjoco.py`, both copies) refuses to save an instance whose `meta` says where its scene came
+ from. `tests/grade.py`: a rebuilt instance that does not match its recorded hash is now no verdict
+ (`instance_mismatch`, `grader_error`, `infra_error` = 1, nothing replayed); before, it was the agent's 0 with
+ `infra_error` = 0 and `instance_ok` = 0. Past results: the demonstration itself was out of reach under the egress
+ allowlist (Hugging Face is not on it); a past 0 with `instance_ok` = 0 was an infrastructure failure.
diff --git a/tasks/dexjoco-bimanual-microwave-cook-i00-privileged/environment/instance.npz b/tasks/dexjoco-bimanual-microwave-cook-i00-privileged/environment/instance.npz
index d07075df98..db769fb23f 100644
Binary files a/tasks/dexjoco-bimanual-microwave-cook-i00-privileged/environment/instance.npz and b/tasks/dexjoco-bimanual-microwave-cook-i00-privileged/environment/instance.npz differ
diff --git a/tasks/dexjoco-bimanual-microwave-cook-i00-privileged/environment/rcb_dexjoco.py b/tasks/dexjoco-bimanual-microwave-cook-i00-privileged/environment/rcb_dexjoco.py
index f2e2db848f..8eb183fe2a 100644
--- a/tasks/dexjoco-bimanual-microwave-cook-i00-privileged/environment/rcb_dexjoco.py
+++ b/tasks/dexjoco-bimanual-microwave-cook-i00-privileged/environment/rcb_dexjoco.py
@@ -54,6 +54,7 @@ import argparse
import copy
import hashlib
import json
+import re
import sys
import time
from pathlib import Path
@@ -222,6 +223,21 @@ def water_aim_deg(model, data) -> float:
return float(np.degrees(np.arccos(np.clip(c, -1.0, 1.0))))
+# instance.npz is in the agent's container (/app/instance.npz), and a task's reference solution is the actions of the
+# recording its instance was restored from: its `meta` must not say where the scene came from (PROTOCOL §2.7, no pointer
+# to a solution where the agent can reach). save_instance refuses such keys, and values that name a recording.
+PROVENANCE_KEYS = ("source", "demo", "dataset", "episode", "recording", "reference", "solution", "zarr", "url")
+PROVENANCE_VALUE = re.compile(r"demo|episode|dataset|recording|replay\.zarr|huggingface|https?://", re.I)
+
+
+def check_instance_meta(meta: dict) -> None:
+ """ValueError if an instance's `meta` says where its scene came from (keys or values; see PROVENANCE_KEYS)."""
+ bad = [k for k, v in meta.items() if any(p in str(k).lower() for p in PROVENANCE_KEYS)
+ or PROVENANCE_VALUE.search(json.dumps(v))]
+ if bad:
+ raise ValueError(f"instance meta {bad} names where the scene came from: the agent reads instance.npz")
+
+
def _log(*a: Any) -> None:
print("[rcb_dexjoco]", *a, file=sys.stderr, flush=True)
@@ -349,6 +365,7 @@ class Session:
return h.hexdigest()[:16]
def save_instance(self, path: str | Path, meta: dict | None = None) -> None:
+ check_instance_meta(meta or {})
snap = self.snapshot()
np.savez_compressed(path, task=self.task, physics=snap["physics"], mocap_pos=snap["mocap_pos"],
**{"model__" + k: v for k, v in snap["model"].items()},
@@ -803,15 +820,14 @@ def cmd_make_instance(a) -> None:
restore_initial_state(s.env, a.task, s.cfg, state0)
s._snapshot = s.snapshot()
s._episode_reset()
- meta = {"source": "DexJoCo-Datasets-Raw demo, initial state restored with dexjoco.tasks.state_restorers",
- "demo": a.demo_name or os.path.basename(os.path.dirname(os.path.abspath(a.demo))), "dexjoco": "8d23b0f"}
- s.save_instance(a.out, meta)
- _log("instance", a.out, "sha", s.state_sha())
+ demo = a.demo_name or os.path.basename(os.path.dirname(os.path.abspath(a.demo)))
+ s.save_instance(a.out, {"dexjoco": "8d23b0f"}) # no provenance in the instance (check_instance_meta)
+ _log("instance", a.out, "sha", s.state_sha(), "from", demo)
if a.oracle_out:
A = np.asarray(root["data"]["action"], dtype=np.float64)
s2 = Session(a.task, instance=a.out)
st = s2.replay(A)
- s2.save_trajectory(A[:st["replayed_rows"]], a.oracle_out, {"source": "demo actions", "demo": meta["demo"]})
+ s2.save_trajectory(A[:st["replayed_rows"]], a.oracle_out, {"source": "demo actions", "demo": demo})
print(json.dumps({"oracle": a.oracle_out, **{k: st[k] for k in ("success", "reason", "replayed_rows", "rows", "state_sha")}}))
diff --git a/tasks/dexjoco-bimanual-microwave-cook-i00-privileged/task.toml b/tasks/dexjoco-bimanual-microwave-cook-i00-privileged/task.toml
index 0774080a59..825e025224 100644
--- a/tasks/dexjoco-bimanual-microwave-cook-i00-privileged/task.toml
+++ b/tasks/dexjoco-bimanual-microwave-cook-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "dexjoco"
robot = "franka-allegro-bimanual"
family = "bimanual_microwave_cook"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
expert_time_estimate_hours = 3.0
diff --git a/tasks/dexjoco-bimanual-microwave-cook-i00-privileged/tests/grade.py b/tasks/dexjoco-bimanual-microwave-cook-i00-privileged/tests/grade.py
index c754f999ad..3cfbab9c42 100644
--- a/tasks/dexjoco-bimanual-microwave-cook-i00-privileged/tests/grade.py
+++ b/tasks/dexjoco-bimanual-microwave-cook-i00-privileged/tests/grade.py
@@ -6,8 +6,10 @@ write Harbor rewards.
Writes /logs/verifier/reward.txt (1/0), reward.json (numeric fields only; Harbor rejects strings), a replay video
/logs/verifier/replay.mp4 (front camera + wrist camera) and replay.last.png. Replays `--repeat` times in this fresh
process; success requires every replay to end its episode in success (DexJoCo's own rule: the episode ends at the first
-success or task failure) AND all final-state hashes to agree (MuJoCo on the CPU is bit-deterministic), and the frozen
-instance rebuilt here from tests/instance.npz must match its recorded hash.
+success or task failure) AND all final-state hashes to agree (MuJoCo on the CPU is bit-deterministic). The frozen
+instance rebuilt here from tests/instance.npz must match its recorded hash; one that does not (the image no longer
+rebuilds the scene) is no verdict on the agent: `instance_mismatch`, `grader_error` and `infra_error` = 1, nothing
+replayed.
A file the agent got wrong is the agent's 0, `bad_trajectory` (not `grader_error`): not an npz, pickled or object
arrays, no `task` or `actions` entry, a wrong shape, NaN/inf, or more than rcb_dexjoco.MAX_TRAJECTORY_ROWS rows (the
@@ -119,6 +121,10 @@ def main() -> int:
print("too few CPUs: " + rcb_dexjoco.threads_message(threads), file=sys.stderr)
write(False, {"grader_error": 1, "infra_error": 1, "too_few_cpus": 1, "n_actions": n_rows, **threads, **cost}); return 0
s = Session(a.family, instance=a.instance)
+ if not s.instance_exact: # the image no longer rebuilds the frozen scene: not the agent's 0 (infra, no verdict)
+ print(f"instance mismatch: rebuilt state sha {s.state_sha()} != recorded {s.instance_sha}", file=sys.stderr)
+ write(False, {"grader_error": 1, "infra_error": 1, "instance_mismatch": 1, "instance_ok": 0,
+ "n_actions": n_rows, **threads, **cost}); return 0
if acts.shape[1] != s.action_dim:
write(False, {"bad_trajectory": 1, "n_actions": n_rows, **cost}); return 0
t0 = time.time()
@@ -130,10 +136,9 @@ def main() -> int:
write(False, {"bad_trajectory": 1, "n_actions": n_rows, **cost}); return 0
all_success = all(o["success"] for o in outs)
deterministic = len({o["state_sha"] for o in outs}) == 1
- instance_ok = bool(s.instance_exact)
- ok = all_success and deterministic and instance_ok
+ ok = all_success and deterministic
out = {"success": int(ok), "all_success": int(all_success), "deterministic": int(deterministic),
- "instance_ok": int(instance_ok), "n_actions": int(len(acts)), "sim_steps": int(outs[0]["steps"]),
+ "instance_ok": 1, "n_actions": int(len(acts)), "sim_steps": int(outs[0]["steps"]),
"task_failure": int(bool(outs[0]["reason"]) and not outs[0]["success"]),
"replay_wall_s": round(time.time() - t0, 1), "grader_error": 0, "video_rendered": 0, **threads, **cost}
write(ok, out) # the verdict is on disk before the video; the video only adds video_rendered
diff --git a/tasks/dexjoco-bimanual-microwave-cook-i00-privileged/tests/instance.npz b/tasks/dexjoco-bimanual-microwave-cook-i00-privileged/tests/instance.npz
index d07075df98..db769fb23f 100644
Binary files a/tasks/dexjoco-bimanual-microwave-cook-i00-privileged/tests/instance.npz and b/tasks/dexjoco-bimanual-microwave-cook-i00-privileged/tests/instance.npz differ
diff --git a/tasks/dexjoco-bimanual-microwave-cook-i00-privileged/tests/rcb_dexjoco.py b/tasks/dexjoco-bimanual-microwave-cook-i00-privileged/tests/rcb_dexjoco.py
index f2e2db848f..8eb183fe2a 100644
--- a/tasks/dexjoco-bimanual-microwave-cook-i00-privileged/tests/rcb_dexjoco.py
+++ b/tasks/dexjoco-bimanual-microwave-cook-i00-privileged/tests/rcb_dexjoco.py
@@ -54,6 +54,7 @@ import argparse
import copy
import hashlib
import json
+import re
import sys
import time
from pathlib import Path
@@ -222,6 +223,21 @@ def water_aim_deg(model, data) -> float:
return float(np.degrees(np.arccos(np.clip(c, -1.0, 1.0))))
+# instance.npz is in the agent's container (/app/instance.npz), and a task's reference solution is the actions of the
+# recording its instance was restored from: its `meta` must not say where the scene came from (PROTOCOL §2.7, no pointer
+# to a solution where the agent can reach). save_instance refuses such keys, and values that name a recording.
+PROVENANCE_KEYS = ("source", "demo", "dataset", "episode", "recording", "reference", "solution", "zarr", "url")
+PROVENANCE_VALUE = re.compile(r"demo|episode|dataset|recording|replay\.zarr|huggingface|https?://", re.I)
+
+
+def check_instance_meta(meta: dict) -> None:
+ """ValueError if an instance's `meta` says where its scene came from (keys or values; see PROVENANCE_KEYS)."""
+ bad = [k for k, v in meta.items() if any(p in str(k).lower() for p in PROVENANCE_KEYS)
+ or PROVENANCE_VALUE.search(json.dumps(v))]
+ if bad:
+ raise ValueError(f"instance meta {bad} names where the scene came from: the agent reads instance.npz")
+
+
def _log(*a: Any) -> None:
print("[rcb_dexjoco]", *a, file=sys.stderr, flush=True)
@@ -349,6 +365,7 @@ class Session:
return h.hexdigest()[:16]
def save_instance(self, path: str | Path, meta: dict | None = None) -> None:
+ check_instance_meta(meta or {})
snap = self.snapshot()
np.savez_compressed(path, task=self.task, physics=snap["physics"], mocap_pos=snap["mocap_pos"],
**{"model__" + k: v for k, v in snap["model"].items()},
@@ -803,15 +820,14 @@ def cmd_make_instance(a) -> None:
restore_initial_state(s.env, a.task, s.cfg, state0)
s._snapshot = s.snapshot()
s._episode_reset()
- meta = {"source": "DexJoCo-Datasets-Raw demo, initial state restored with dexjoco.tasks.state_restorers",
- "demo": a.demo_name or os.path.basename(os.path.dirname(os.path.abspath(a.demo))), "dexjoco": "8d23b0f"}
- s.save_instance(a.out, meta)
- _log("instance", a.out, "sha", s.state_sha())
+ demo = a.demo_name or os.path.basename(os.path.dirname(os.path.abspath(a.demo)))
+ s.save_instance(a.out, {"dexjoco": "8d23b0f"}) # no provenance in the instance (check_instance_meta)
+ _log("instance", a.out, "sha", s.state_sha(), "from", demo)
if a.oracle_out:
A = np.asarray(root["data"]["action"], dtype=np.float64)
s2 = Session(a.task, instance=a.out)
st = s2.replay(A)
- s2.save_trajectory(A[:st["replayed_rows"]], a.oracle_out, {"source": "demo actions", "demo": meta["demo"]})
+ s2.save_trajectory(A[:st["replayed_rows"]], a.oracle_out, {"source": "demo actions", "demo": demo})
print(json.dumps({"oracle": a.oracle_out, **{k: st[k] for k in ("success", "reason", "replayed_rows", "rows", "state_sha")}}))
diff --git a/tasks/dexjoco-bimanual-microwave-cook-i00-standard/README.md b/tasks/dexjoco-bimanual-microwave-cook-i00-standard/README.md
index 631a347572..b94d1cadef 100644
--- a/tasks/dexjoco-bimanual-microwave-cook-i00-standard/README.md
+++ b/tasks/dexjoco-bimanual-microwave-cook-i00-standard/README.md
@@ -72,6 +72,13 @@ Generated by `scripts/dexjoco/generate.py`, as its privileged twin.
returned once. A service that has given up on the episode (too few CPUs, or it could not resume after its
crashes) answers every request with 409 and the reason, and `Robot()` / `wait_ready()` raise `EpisodeOver` at
once; before, nothing answered and they waited out their timeout (up to an hour).
+- 2026-10-14 (review of board 0105, B-M3): `environment/sim/instance.npz` and `environment/sim/rcb_dexjoco.py`. The
+ frozen instance's `meta` entry no longer names the demonstration the scene was restored from (`source`, `demo`;
+ the reference solution is that demonstration's actions), and the tooling refuses to save an instance that does.
+ The sidecar's files are out of the agent's reach, so nothing reached the agent from this copy; it changed with
+ the privileged twin's, from the one committed instance. Every array of the scene is the same (instance sha
+ `6c16f3cfa1c737c7`): the scene, the pass rule and the grader are unchanged, so the task keeps its name and earlier
+ results stand.
- 2026-10-13 (owner, boards 0085 and 0087): the sim service's clock (`environment/sim/rcb_limited.py`,
`environment/sim/server.py`, the docstring of `environment/rcb_client.py`). `meta["time_s"]` and
`meta["time_left_s"]` count from the agent's first request to the service (any endpoint), and a restarted service
diff --git a/tasks/dexjoco-bimanual-microwave-cook-i00-standard/environment/sim/instance.npz b/tasks/dexjoco-bimanual-microwave-cook-i00-standard/environment/sim/instance.npz
index d07075df98..db769fb23f 100644
Binary files a/tasks/dexjoco-bimanual-microwave-cook-i00-standard/environment/sim/instance.npz and b/tasks/dexjoco-bimanual-microwave-cook-i00-standard/environment/sim/instance.npz differ
diff --git a/tasks/dexjoco-bimanual-microwave-cook-i00-standard/environment/sim/rcb_dexjoco.py b/tasks/dexjoco-bimanual-microwave-cook-i00-standard/environment/sim/rcb_dexjoco.py
index f2e2db848f..8eb183fe2a 100644
--- a/tasks/dexjoco-bimanual-microwave-cook-i00-standard/environment/sim/rcb_dexjoco.py
+++ b/tasks/dexjoco-bimanual-microwave-cook-i00-standard/environment/sim/rcb_dexjoco.py
@@ -54,6 +54,7 @@ import argparse
import copy
import hashlib
import json
+import re
import sys
import time
from pathlib import Path
@@ -222,6 +223,21 @@ def water_aim_deg(model, data) -> float:
return float(np.degrees(np.arccos(np.clip(c, -1.0, 1.0))))
+# instance.npz is in the agent's container (/app/instance.npz), and a task's reference solution is the actions of the
+# recording its instance was restored from: its `meta` must not say where the scene came from (PROTOCOL §2.7, no pointer
+# to a solution where the agent can reach). save_instance refuses such keys, and values that name a recording.
+PROVENANCE_KEYS = ("source", "demo", "dataset", "episode", "recording", "reference", "solution", "zarr", "url")
+PROVENANCE_VALUE = re.compile(r"demo|episode|dataset|recording|replay\.zarr|huggingface|https?://", re.I)
+
+
+def check_instance_meta(meta: dict) -> None:
+ """ValueError if an instance's `meta` says where its scene came from (keys or values; see PROVENANCE_KEYS)."""
+ bad = [k for k, v in meta.items() if any(p in str(k).lower() for p in PROVENANCE_KEYS)
+ or PROVENANCE_VALUE.search(json.dumps(v))]
+ if bad:
+ raise ValueError(f"instance meta {bad} names where the scene came from: the agent reads instance.npz")
+
+
def _log(*a: Any) -> None:
print("[rcb_dexjoco]", *a, file=sys.stderr, flush=True)
@@ -349,6 +365,7 @@ class Session:
return h.hexdigest()[:16]
def save_instance(self, path: str | Path, meta: dict | None = None) -> None:
+ check_instance_meta(meta or {})
snap = self.snapshot()
np.savez_compressed(path, task=self.task, physics=snap["physics"], mocap_pos=snap["mocap_pos"],
**{"model__" + k: v for k, v in snap["model"].items()},
@@ -803,15 +820,14 @@ def cmd_make_instance(a) -> None:
restore_initial_state(s.env, a.task, s.cfg, state0)
s._snapshot = s.snapshot()
s._episode_reset()
- meta = {"source": "DexJoCo-Datasets-Raw demo, initial state restored with dexjoco.tasks.state_restorers",
- "demo": a.demo_name or os.path.basename(os.path.dirname(os.path.abspath(a.demo))), "dexjoco": "8d23b0f"}
- s.save_instance(a.out, meta)
- _log("instance", a.out, "sha", s.state_sha())
+ demo = a.demo_name or os.path.basename(os.path.dirname(os.path.abspath(a.demo)))
+ s.save_instance(a.out, {"dexjoco": "8d23b0f"}) # no provenance in the instance (check_instance_meta)
+ _log("instance", a.out, "sha", s.state_sha(), "from", demo)
if a.oracle_out:
A = np.asarray(root["data"]["action"], dtype=np.float64)
s2 = Session(a.task, instance=a.out)
st = s2.replay(A)
- s2.save_trajectory(A[:st["replayed_rows"]], a.oracle_out, {"source": "demo actions", "demo": meta["demo"]})
+ s2.save_trajectory(A[:st["replayed_rows"]], a.oracle_out, {"source": "demo actions", "demo": demo})
print(json.dumps({"oracle": a.oracle_out, **{k: st[k] for k in ("success", "reason", "replayed_rows", "rows", "state_sha")}}))
diff --git a/tasks/dexjoco-bimanual-microwave-cook-i00-standard/task.toml b/tasks/dexjoco-bimanual-microwave-cook-i00-standard/task.toml
index 6b1a0cccd0..1933cbbedd 100644
--- a/tasks/dexjoco-bimanual-microwave-cook-i00-standard/task.toml
+++ b/tasks/dexjoco-bimanual-microwave-cook-i00-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "dexjoco"
robot = "franka-allegro-bimanual"
family = "bimanual_microwave_cook"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "full"
expert_time_estimate_hours = 3.0
diff --git a/tasks/dexjoco-bimanual-photograph-i00-privileged/README.md b/tasks/dexjoco-bimanual-photograph-i00-privileged/README.md
index a8bd1e45fd..64ac01c2bf 100644
--- a/tasks/dexjoco-bimanual-photograph-i00-privileged/README.md
+++ b/tasks/dexjoco-bimanual-photograph-i00-privileged/README.md
@@ -29,8 +29,10 @@ limited to the model APIs.
## Verification
Open-loop replay of `/app/output/trajectory.npz` from `tests/instance.npz`, twice in a fresh process in the verifier's own
-sandbox (no network); pass iff both episodes end in success, their final-state hashes agree and the rebuilt instance
-matches its recorded hash (`tests/grade.py`); plus `/logs/verifier/replay.mp4`, rendered after the verdict is written.
+sandbox (no network); pass iff both episodes end in success and their final-state hashes agree (`tests/grade.py`); plus
+`/logs/verifier/replay.mp4`, rendered after the verdict is written. A rebuilt instance that does not match its recorded
+hash (the image no longer rebuilds the scene) replays nothing and gives no verdict: `instance_mismatch`, `grader_error`
+and `infra_error` = 1, not the agent's 0.
A file that is not a readable trajectory (not an npz, no `task` / `actions`, a wrong shape, NaN) or has more than
50,000 rows scores 0 as `bad_trajectory`: the cap is what two replays of the slowest family fit into the verifier's
3600 s. A verifier that runs fewer than 4 OpenBLAS threads (fewer than 4 CPUs visible to it) replays
@@ -65,3 +67,15 @@ real-time sleep, episode cap and hanoi's wall-clock auto-reset are disabled in t
### Provenance
Generated by `scripts/dexjoco/generate.py` from `scripts/dexjoco/instances/bimanual_photograph/` (`make_instances.py`).
+
+### Change Log
+- 2026-10-14 (review of board 0105, B-M3 and B6-3): `environment/instance.npz` and `tests/instance.npz` no longer name
+ the demonstration the scene was restored from: their `meta` entry had `source` (the dataset) and `demo` (the
+ demonstration's id), and the reference solution is that demonstration's own actions, so the agent could read at
+ `/app/instance.npz` where a solution was (PROTOCOL §2.7). Only `meta` changed: every array of the scene is the same
+ (instance sha `b8b1dbe8dba6c733`), so the scene, the pass rule and the reference are unchanged and the task keeps its
+ name. The tooling (`rcb_dexjoco.py`, both copies) refuses to save an instance whose `meta` says where its scene came
+ from. `tests/grade.py`: a rebuilt instance that does not match its recorded hash is now no verdict
+ (`instance_mismatch`, `grader_error`, `infra_error` = 1, nothing replayed); before, it was the agent's 0 with
+ `infra_error` = 0 and `instance_ok` = 0. Past results: the demonstration itself was out of reach under the egress
+ allowlist (Hugging Face is not on it); a past 0 with `instance_ok` = 0 was an infrastructure failure.
diff --git a/tasks/dexjoco-bimanual-photograph-i00-privileged/environment/instance.npz b/tasks/dexjoco-bimanual-photograph-i00-privileged/environment/instance.npz
index 6e5ca5b70e..0aacad571e 100644
Binary files a/tasks/dexjoco-bimanual-photograph-i00-privileged/environment/instance.npz and b/tasks/dexjoco-bimanual-photograph-i00-privileged/environment/instance.npz differ
diff --git a/tasks/dexjoco-bimanual-photograph-i00-privileged/environment/rcb_dexjoco.py b/tasks/dexjoco-bimanual-photograph-i00-privileged/environment/rcb_dexjoco.py
index f2e2db848f..8eb183fe2a 100644
--- a/tasks/dexjoco-bimanual-photograph-i00-privileged/environment/rcb_dexjoco.py
+++ b/tasks/dexjoco-bimanual-photograph-i00-privileged/environment/rcb_dexjoco.py
@@ -54,6 +54,7 @@ import argparse
import copy
import hashlib
import json
+import re
import sys
import time
from pathlib import Path
@@ -222,6 +223,21 @@ def water_aim_deg(model, data) -> float:
return float(np.degrees(np.arccos(np.clip(c, -1.0, 1.0))))
+# instance.npz is in the agent's container (/app/instance.npz), and a task's reference solution is the actions of the
+# recording its instance was restored from: its `meta` must not say where the scene came from (PROTOCOL §2.7, no pointer
+# to a solution where the agent can reach). save_instance refuses such keys, and values that name a recording.
+PROVENANCE_KEYS = ("source", "demo", "dataset", "episode", "recording", "reference", "solution", "zarr", "url")
+PROVENANCE_VALUE = re.compile(r"demo|episode|dataset|recording|replay\.zarr|huggingface|https?://", re.I)
+
+
+def check_instance_meta(meta: dict) -> None:
+ """ValueError if an instance's `meta` says where its scene came from (keys or values; see PROVENANCE_KEYS)."""
+ bad = [k for k, v in meta.items() if any(p in str(k).lower() for p in PROVENANCE_KEYS)
+ or PROVENANCE_VALUE.search(json.dumps(v))]
+ if bad:
+ raise ValueError(f"instance meta {bad} names where the scene came from: the agent reads instance.npz")
+
+
def _log(*a: Any) -> None:
print("[rcb_dexjoco]", *a, file=sys.stderr, flush=True)
@@ -349,6 +365,7 @@ class Session:
return h.hexdigest()[:16]
def save_instance(self, path: str | Path, meta: dict | None = None) -> None:
+ check_instance_meta(meta or {})
snap = self.snapshot()
np.savez_compressed(path, task=self.task, physics=snap["physics"], mocap_pos=snap["mocap_pos"],
**{"model__" + k: v for k, v in snap["model"].items()},
@@ -803,15 +820,14 @@ def cmd_make_instance(a) -> None:
restore_initial_state(s.env, a.task, s.cfg, state0)
s._snapshot = s.snapshot()
s._episode_reset()
- meta = {"source": "DexJoCo-Datasets-Raw demo, initial state restored with dexjoco.tasks.state_restorers",
- "demo": a.demo_name or os.path.basename(os.path.dirname(os.path.abspath(a.demo))), "dexjoco": "8d23b0f"}
- s.save_instance(a.out, meta)
- _log("instance", a.out, "sha", s.state_sha())
+ demo = a.demo_name or os.path.basename(os.path.dirname(os.path.abspath(a.demo)))
+ s.save_instance(a.out, {"dexjoco": "8d23b0f"}) # no provenance in the instance (check_instance_meta)
+ _log("instance", a.out, "sha", s.state_sha(), "from", demo)
if a.oracle_out:
A = np.asarray(root["data"]["action"], dtype=np.float64)
s2 = Session(a.task, instance=a.out)
st = s2.replay(A)
- s2.save_trajectory(A[:st["replayed_rows"]], a.oracle_out, {"source": "demo actions", "demo": meta["demo"]})
+ s2.save_trajectory(A[:st["replayed_rows"]], a.oracle_out, {"source": "demo actions", "demo": demo})
print(json.dumps({"oracle": a.oracle_out, **{k: st[k] for k in ("success", "reason", "replayed_rows", "rows", "state_sha")}}))
diff --git a/tasks/dexjoco-bimanual-photograph-i00-privileged/task.toml b/tasks/dexjoco-bimanual-photograph-i00-privileged/task.toml
index 03aff7eac4..8846357510 100644
--- a/tasks/dexjoco-bimanual-photograph-i00-privileged/task.toml
+++ b/tasks/dexjoco-bimanual-photograph-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "dexjoco"
robot = "franka-allegro-bimanual"
family = "bimanual_photograph"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
expert_time_estimate_hours = 2.5
diff --git a/tasks/dexjoco-bimanual-photograph-i00-privileged/tests/grade.py b/tasks/dexjoco-bimanual-photograph-i00-privileged/tests/grade.py
index c754f999ad..3cfbab9c42 100644
--- a/tasks/dexjoco-bimanual-photograph-i00-privileged/tests/grade.py
+++ b/tasks/dexjoco-bimanual-photograph-i00-privileged/tests/grade.py
@@ -6,8 +6,10 @@ write Harbor rewards.
Writes /logs/verifier/reward.txt (1/0), reward.json (numeric fields only; Harbor rejects strings), a replay video
/logs/verifier/replay.mp4 (front camera + wrist camera) and replay.last.png. Replays `--repeat` times in this fresh
process; success requires every replay to end its episode in success (DexJoCo's own rule: the episode ends at the first
-success or task failure) AND all final-state hashes to agree (MuJoCo on the CPU is bit-deterministic), and the frozen
-instance rebuilt here from tests/instance.npz must match its recorded hash.
+success or task failure) AND all final-state hashes to agree (MuJoCo on the CPU is bit-deterministic). The frozen
+instance rebuilt here from tests/instance.npz must match its recorded hash; one that does not (the image no longer
+rebuilds the scene) is no verdict on the agent: `instance_mismatch`, `grader_error` and `infra_error` = 1, nothing
+replayed.
A file the agent got wrong is the agent's 0, `bad_trajectory` (not `grader_error`): not an npz, pickled or object
arrays, no `task` or `actions` entry, a wrong shape, NaN/inf, or more than rcb_dexjoco.MAX_TRAJECTORY_ROWS rows (the
@@ -119,6 +121,10 @@ def main() -> int:
print("too few CPUs: " + rcb_dexjoco.threads_message(threads), file=sys.stderr)
write(False, {"grader_error": 1, "infra_error": 1, "too_few_cpus": 1, "n_actions": n_rows, **threads, **cost}); return 0
s = Session(a.family, instance=a.instance)
+ if not s.instance_exact: # the image no longer rebuilds the frozen scene: not the agent's 0 (infra, no verdict)
+ print(f"instance mismatch: rebuilt state sha {s.state_sha()} != recorded {s.instance_sha}", file=sys.stderr)
+ write(False, {"grader_error": 1, "infra_error": 1, "instance_mismatch": 1, "instance_ok": 0,
+ "n_actions": n_rows, **threads, **cost}); return 0
if acts.shape[1] != s.action_dim:
write(False, {"bad_trajectory": 1, "n_actions": n_rows, **cost}); return 0
t0 = time.time()
@@ -130,10 +136,9 @@ def main() -> int:
write(False, {"bad_trajectory": 1, "n_actions": n_rows, **cost}); return 0
all_success = all(o["success"] for o in outs)
deterministic = len({o["state_sha"] for o in outs}) == 1
- instance_ok = bool(s.instance_exact)
- ok = all_success and deterministic and instance_ok
+ ok = all_success and deterministic
out = {"success": int(ok), "all_success": int(all_success), "deterministic": int(deterministic),
- "instance_ok": int(instance_ok), "n_actions": int(len(acts)), "sim_steps": int(outs[0]["steps"]),
+ "instance_ok": 1, "n_actions": int(len(acts)), "sim_steps": int(outs[0]["steps"]),
"task_failure": int(bool(outs[0]["reason"]) and not outs[0]["success"]),
"replay_wall_s": round(time.time() - t0, 1), "grader_error": 0, "video_rendered": 0, **threads, **cost}
write(ok, out) # the verdict is on disk before the video; the video only adds video_rendered
diff --git a/tasks/dexjoco-bimanual-photograph-i00-privileged/tests/instance.npz b/tasks/dexjoco-bimanual-photograph-i00-privileged/tests/instance.npz
index 6e5ca5b70e..0aacad571e 100644
Binary files a/tasks/dexjoco-bimanual-photograph-i00-privileged/tests/instance.npz and b/tasks/dexjoco-bimanual-photograph-i00-privileged/tests/instance.npz differ
diff --git a/tasks/dexjoco-bimanual-photograph-i00-privileged/tests/rcb_dexjoco.py b/tasks/dexjoco-bimanual-photograph-i00-privileged/tests/rcb_dexjoco.py
index f2e2db848f..8eb183fe2a 100644
--- a/tasks/dexjoco-bimanual-photograph-i00-privileged/tests/rcb_dexjoco.py
+++ b/tasks/dexjoco-bimanual-photograph-i00-privileged/tests/rcb_dexjoco.py
@@ -54,6 +54,7 @@ import argparse
import copy
import hashlib
import json
+import re
import sys
import time
from pathlib import Path
@@ -222,6 +223,21 @@ def water_aim_deg(model, data) -> float:
return float(np.degrees(np.arccos(np.clip(c, -1.0, 1.0))))
+# instance.npz is in the agent's container (/app/instance.npz), and a task's reference solution is the actions of the
+# recording its instance was restored from: its `meta` must not say where the scene came from (PROTOCOL §2.7, no pointer
+# to a solution where the agent can reach). save_instance refuses such keys, and values that name a recording.
+PROVENANCE_KEYS = ("source", "demo", "dataset", "episode", "recording", "reference", "solution", "zarr", "url")
+PROVENANCE_VALUE = re.compile(r"demo|episode|dataset|recording|replay\.zarr|huggingface|https?://", re.I)
+
+
+def check_instance_meta(meta: dict) -> None:
+ """ValueError if an instance's `meta` says where its scene came from (keys or values; see PROVENANCE_KEYS)."""
+ bad = [k for k, v in meta.items() if any(p in str(k).lower() for p in PROVENANCE_KEYS)
+ or PROVENANCE_VALUE.search(json.dumps(v))]
+ if bad:
+ raise ValueError(f"instance meta {bad} names where the scene came from: the agent reads instance.npz")
+
+
def _log(*a: Any) -> None:
print("[rcb_dexjoco]", *a, file=sys.stderr, flush=True)
@@ -349,6 +365,7 @@ class Session:
return h.hexdigest()[:16]
def save_instance(self, path: str | Path, meta: dict | None = None) -> None:
+ check_instance_meta(meta or {})
snap = self.snapshot()
np.savez_compressed(path, task=self.task, physics=snap["physics"], mocap_pos=snap["mocap_pos"],
**{"model__" + k: v for k, v in snap["model"].items()},
@@ -803,15 +820,14 @@ def cmd_make_instance(a) -> None:
restore_initial_state(s.env, a.task, s.cfg, state0)
s._snapshot = s.snapshot()
s._episode_reset()
- meta = {"source": "DexJoCo-Datasets-Raw demo, initial state restored with dexjoco.tasks.state_restorers",
- "demo": a.demo_name or os.path.basename(os.path.dirname(os.path.abspath(a.demo))), "dexjoco": "8d23b0f"}
- s.save_instance(a.out, meta)
- _log("instance", a.out, "sha", s.state_sha())
+ demo = a.demo_name or os.path.basename(os.path.dirname(os.path.abspath(a.demo)))
+ s.save_instance(a.out, {"dexjoco": "8d23b0f"}) # no provenance in the instance (check_instance_meta)
+ _log("instance", a.out, "sha", s.state_sha(), "from", demo)
if a.oracle_out:
A = np.asarray(root["data"]["action"], dtype=np.float64)
s2 = Session(a.task, instance=a.out)
st = s2.replay(A)
- s2.save_trajectory(A[:st["replayed_rows"]], a.oracle_out, {"source": "demo actions", "demo": meta["demo"]})
+ s2.save_trajectory(A[:st["replayed_rows"]], a.oracle_out, {"source": "demo actions", "demo": demo})
print(json.dumps({"oracle": a.oracle_out, **{k: st[k] for k in ("success", "reason", "replayed_rows", "rows", "state_sha")}}))
diff --git a/tasks/dexjoco-bimanual-photograph-i00-standard/README.md b/tasks/dexjoco-bimanual-photograph-i00-standard/README.md
index acfd1f21cb..335b295cdd 100644
--- a/tasks/dexjoco-bimanual-photograph-i00-standard/README.md
+++ b/tasks/dexjoco-bimanual-photograph-i00-standard/README.md
@@ -79,6 +79,13 @@ Generated by `scripts/dexjoco/generate.py`, as its privileged twin.
pose, which standard mode must not give (PROTOCOL §4.5). It now gives images and depth only: no mount,
`extrinsic_cv` / `cam2world_gl` null in `meta["cameras"]`, NaN per frame. Standard-mode results of this task
from before this date could read the object's pose and are reported separately.
+- 2026-10-14 (review of board 0105, B-M3): `environment/sim/instance.npz` and `environment/sim/rcb_dexjoco.py`. The
+ frozen instance's `meta` entry no longer names the demonstration the scene was restored from (`source`, `demo`;
+ the reference solution is that demonstration's actions), and the tooling refuses to save an instance that does.
+ The sidecar's files are out of the agent's reach, so nothing reached the agent from this copy; it changed with
+ the privileged twin's, from the one committed instance. Every array of the scene is the same (instance sha
+ `b8b1dbe8dba6c733`): the scene, the pass rule and the grader are unchanged, so the task keeps its name and earlier
+ results stand.
- 2026-10-13 (owner, boards 0085 and 0087): the sim service's clock (`environment/sim/rcb_limited.py`,
`environment/sim/server.py`, the docstring of `environment/rcb_client.py`). `meta["time_s"]` and
`meta["time_left_s"]` count from the agent's first request to the service (any endpoint), and a restarted service
diff --git a/tasks/dexjoco-bimanual-photograph-i00-standard/environment/sim/instance.npz b/tasks/dexjoco-bimanual-photograph-i00-standard/environment/sim/instance.npz
index 6e5ca5b70e..0aacad571e 100644
Binary files a/tasks/dexjoco-bimanual-photograph-i00-standard/environment/sim/instance.npz and b/tasks/dexjoco-bimanual-photograph-i00-standard/environment/sim/instance.npz differ
diff --git a/tasks/dexjoco-bimanual-photograph-i00-standard/environment/sim/rcb_dexjoco.py b/tasks/dexjoco-bimanual-photograph-i00-standard/environment/sim/rcb_dexjoco.py
index f2e2db848f..8eb183fe2a 100644
--- a/tasks/dexjoco-bimanual-photograph-i00-standard/environment/sim/rcb_dexjoco.py
+++ b/tasks/dexjoco-bimanual-photograph-i00-standard/environment/sim/rcb_dexjoco.py
@@ -54,6 +54,7 @@ import argparse
import copy
import hashlib
import json
+import re
import sys
import time
from pathlib import Path
@@ -222,6 +223,21 @@ def water_aim_deg(model, data) -> float:
return float(np.degrees(np.arccos(np.clip(c, -1.0, 1.0))))
+# instance.npz is in the agent's container (/app/instance.npz), and a task's reference solution is the actions of the
+# recording its instance was restored from: its `meta` must not say where the scene came from (PROTOCOL §2.7, no pointer
+# to a solution where the agent can reach). save_instance refuses such keys, and values that name a recording.
+PROVENANCE_KEYS = ("source", "demo", "dataset", "episode", "recording", "reference", "solution", "zarr", "url")
+PROVENANCE_VALUE = re.compile(r"demo|episode|dataset|recording|replay\.zarr|huggingface|https?://", re.I)
+
+
+def check_instance_meta(meta: dict) -> None:
+ """ValueError if an instance's `meta` says where its scene came from (keys or values; see PROVENANCE_KEYS)."""
+ bad = [k for k, v in meta.items() if any(p in str(k).lower() for p in PROVENANCE_KEYS)
+ or PROVENANCE_VALUE.search(json.dumps(v))]
+ if bad:
+ raise ValueError(f"instance meta {bad} names where the scene came from: the agent reads instance.npz")
+
+
def _log(*a: Any) -> None:
print("[rcb_dexjoco]", *a, file=sys.stderr, flush=True)
@@ -349,6 +365,7 @@ class Session:
return h.hexdigest()[:16]
def save_instance(self, path: str | Path, meta: dict | None = None) -> None:
+ check_instance_meta(meta or {})
snap = self.snapshot()
np.savez_compressed(path, task=self.task, physics=snap["physics"], mocap_pos=snap["mocap_pos"],
**{"model__" + k: v for k, v in snap["model"].items()},
@@ -803,15 +820,14 @@ def cmd_make_instance(a) -> None:
restore_initial_state(s.env, a.task, s.cfg, state0)
s._snapshot = s.snapshot()
s._episode_reset()
- meta = {"source": "DexJoCo-Datasets-Raw demo, initial state restored with dexjoco.tasks.state_restorers",
- "demo": a.demo_name or os.path.basename(os.path.dirname(os.path.abspath(a.demo))), "dexjoco": "8d23b0f"}
- s.save_instance(a.out, meta)
- _log("instance", a.out, "sha", s.state_sha())
+ demo = a.demo_name or os.path.basename(os.path.dirname(os.path.abspath(a.demo)))
+ s.save_instance(a.out, {"dexjoco": "8d23b0f"}) # no provenance in the instance (check_instance_meta)
+ _log("instance", a.out, "sha", s.state_sha(), "from", demo)
if a.oracle_out:
A = np.asarray(root["data"]["action"], dtype=np.float64)
s2 = Session(a.task, instance=a.out)
st = s2.replay(A)
- s2.save_trajectory(A[:st["replayed_rows"]], a.oracle_out, {"source": "demo actions", "demo": meta["demo"]})
+ s2.save_trajectory(A[:st["replayed_rows"]], a.oracle_out, {"source": "demo actions", "demo": demo})
print(json.dumps({"oracle": a.oracle_out, **{k: st[k] for k in ("success", "reason", "replayed_rows", "rows", "state_sha")}}))
diff --git a/tasks/dexjoco-bimanual-photograph-i00-standard/task.toml b/tasks/dexjoco-bimanual-photograph-i00-standard/task.toml
index ede261d8d0..d58147694c 100644
--- a/tasks/dexjoco-bimanual-photograph-i00-standard/task.toml
+++ b/tasks/dexjoco-bimanual-photograph-i00-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "dexjoco"
robot = "franka-allegro-bimanual"
family = "bimanual_photograph"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "full"
expert_time_estimate_hours = 2.5
diff --git a/tasks/dexjoco-bimanual-unlock-ipad-i00-privileged/README.md b/tasks/dexjoco-bimanual-unlock-ipad-i00-privileged/README.md
index d764fc6024..9b3c7ff129 100644
--- a/tasks/dexjoco-bimanual-unlock-ipad-i00-privileged/README.md
+++ b/tasks/dexjoco-bimanual-unlock-ipad-i00-privileged/README.md
@@ -29,8 +29,10 @@ limited to the model APIs.
## Verification
Open-loop replay of `/app/output/trajectory.npz` from `tests/instance.npz`, twice in a fresh process in the verifier's own
-sandbox (no network); pass iff both episodes end in success, their final-state hashes agree and the rebuilt instance
-matches its recorded hash (`tests/grade.py`); plus `/logs/verifier/replay.mp4`, rendered after the verdict is written.
+sandbox (no network); pass iff both episodes end in success and their final-state hashes agree (`tests/grade.py`); plus
+`/logs/verifier/replay.mp4`, rendered after the verdict is written. A rebuilt instance that does not match its recorded
+hash (the image no longer rebuilds the scene) replays nothing and gives no verdict: `instance_mismatch`, `grader_error`
+and `infra_error` = 1, not the agent's 0.
A file that is not a readable trajectory (not an npz, no `task` / `actions`, a wrong shape, NaN) or has more than
50,000 rows scores 0 as `bad_trajectory`: the cap is what two replays of the slowest family fit into the verifier's
3600 s. A verifier that runs fewer than 4 OpenBLAS threads (fewer than 4 CPUs visible to it) replays
@@ -65,3 +67,15 @@ real-time sleep, episode cap and hanoi's wall-clock auto-reset are disabled in t
### Provenance
Generated by `scripts/dexjoco/generate.py` from `scripts/dexjoco/instances/bimanual_unlock_ipad/` (`make_instances.py`).
+
+### Change Log
+- 2026-10-14 (review of board 0105, B-M3 and B6-3): `environment/instance.npz` and `tests/instance.npz` no longer name
+ the demonstration the scene was restored from: their `meta` entry had `source` (the dataset) and `demo` (the
+ demonstration's id), and the reference solution is that demonstration's own actions, so the agent could read at
+ `/app/instance.npz` where a solution was (PROTOCOL §2.7). Only `meta` changed: every array of the scene is the same
+ (instance sha `d38c827ac3bb030e`), so the scene, the pass rule and the reference are unchanged and the task keeps its
+ name. The tooling (`rcb_dexjoco.py`, both copies) refuses to save an instance whose `meta` says where its scene came
+ from. `tests/grade.py`: a rebuilt instance that does not match its recorded hash is now no verdict
+ (`instance_mismatch`, `grader_error`, `infra_error` = 1, nothing replayed); before, it was the agent's 0 with
+ `infra_error` = 0 and `instance_ok` = 0. Past results: the demonstration itself was out of reach under the egress
+ allowlist (Hugging Face is not on it); a past 0 with `instance_ok` = 0 was an infrastructure failure.
diff --git a/tasks/dexjoco-bimanual-unlock-ipad-i00-privileged/environment/instance.npz b/tasks/dexjoco-bimanual-unlock-ipad-i00-privileged/environment/instance.npz
index 8bc4e3d0c4..a8a691967b 100644
Binary files a/tasks/dexjoco-bimanual-unlock-ipad-i00-privileged/environment/instance.npz and b/tasks/dexjoco-bimanual-unlock-ipad-i00-privileged/environment/instance.npz differ
diff --git a/tasks/dexjoco-bimanual-unlock-ipad-i00-privileged/environment/rcb_dexjoco.py b/tasks/dexjoco-bimanual-unlock-ipad-i00-privileged/environment/rcb_dexjoco.py
index f2e2db848f..8eb183fe2a 100644
--- a/tasks/dexjoco-bimanual-unlock-ipad-i00-privileged/environment/rcb_dexjoco.py
+++ b/tasks/dexjoco-bimanual-unlock-ipad-i00-privileged/environment/rcb_dexjoco.py
@@ -54,6 +54,7 @@ import argparse
import copy
import hashlib
import json
+import re
import sys
import time
from pathlib import Path
@@ -222,6 +223,21 @@ def water_aim_deg(model, data) -> float:
return float(np.degrees(np.arccos(np.clip(c, -1.0, 1.0))))
+# instance.npz is in the agent's container (/app/instance.npz), and a task's reference solution is the actions of the
+# recording its instance was restored from: its `meta` must not say where the scene came from (PROTOCOL §2.7, no pointer
+# to a solution where the agent can reach). save_instance refuses such keys, and values that name a recording.
+PROVENANCE_KEYS = ("source", "demo", "dataset", "episode", "recording", "reference", "solution", "zarr", "url")
+PROVENANCE_VALUE = re.compile(r"demo|episode|dataset|recording|replay\.zarr|huggingface|https?://", re.I)
+
+
+def check_instance_meta(meta: dict) -> None:
+ """ValueError if an instance's `meta` says where its scene came from (keys or values; see PROVENANCE_KEYS)."""
+ bad = [k for k, v in meta.items() if any(p in str(k).lower() for p in PROVENANCE_KEYS)
+ or PROVENANCE_VALUE.search(json.dumps(v))]
+ if bad:
+ raise ValueError(f"instance meta {bad} names where the scene came from: the agent reads instance.npz")
+
+
def _log(*a: Any) -> None:
print("[rcb_dexjoco]", *a, file=sys.stderr, flush=True)
@@ -349,6 +365,7 @@ class Session:
return h.hexdigest()[:16]
def save_instance(self, path: str | Path, meta: dict | None = None) -> None:
+ check_instance_meta(meta or {})
snap = self.snapshot()
np.savez_compressed(path, task=self.task, physics=snap["physics"], mocap_pos=snap["mocap_pos"],
**{"model__" + k: v for k, v in snap["model"].items()},
@@ -803,15 +820,14 @@ def cmd_make_instance(a) -> None:
restore_initial_state(s.env, a.task, s.cfg, state0)
s._snapshot = s.snapshot()
s._episode_reset()
- meta = {"source": "DexJoCo-Datasets-Raw demo, initial state restored with dexjoco.tasks.state_restorers",
- "demo": a.demo_name or os.path.basename(os.path.dirname(os.path.abspath(a.demo))), "dexjoco": "8d23b0f"}
- s.save_instance(a.out, meta)
- _log("instance", a.out, "sha", s.state_sha())
+ demo = a.demo_name or os.path.basename(os.path.dirname(os.path.abspath(a.demo)))
+ s.save_instance(a.out, {"dexjoco": "8d23b0f"}) # no provenance in the instance (check_instance_meta)
+ _log("instance", a.out, "sha", s.state_sha(), "from", demo)
if a.oracle_out:
A = np.asarray(root["data"]["action"], dtype=np.float64)
s2 = Session(a.task, instance=a.out)
st = s2.replay(A)
- s2.save_trajectory(A[:st["replayed_rows"]], a.oracle_out, {"source": "demo actions", "demo": meta["demo"]})
+ s2.save_trajectory(A[:st["replayed_rows"]], a.oracle_out, {"source": "demo actions", "demo": demo})
print(json.dumps({"oracle": a.oracle_out, **{k: st[k] for k in ("success", "reason", "replayed_rows", "rows", "state_sha")}}))
diff --git a/tasks/dexjoco-bimanual-unlock-ipad-i00-privileged/task.toml b/tasks/dexjoco-bimanual-unlock-ipad-i00-privileged/task.toml
index 01edca00d5..8a42f8cde2 100644
--- a/tasks/dexjoco-bimanual-unlock-ipad-i00-privileged/task.toml
+++ b/tasks/dexjoco-bimanual-unlock-ipad-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "dexjoco"
robot = "franka-allegro-bimanual"
family = "bimanual_unlock_ipad"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
expert_time_estimate_hours = 2.5
diff --git a/tasks/dexjoco-bimanual-unlock-ipad-i00-privileged/tests/grade.py b/tasks/dexjoco-bimanual-unlock-ipad-i00-privileged/tests/grade.py
index c754f999ad..3cfbab9c42 100644
--- a/tasks/dexjoco-bimanual-unlock-ipad-i00-privileged/tests/grade.py
+++ b/tasks/dexjoco-bimanual-unlock-ipad-i00-privileged/tests/grade.py
@@ -6,8 +6,10 @@ write Harbor rewards.
Writes /logs/verifier/reward.txt (1/0), reward.json (numeric fields only; Harbor rejects strings), a replay video
/logs/verifier/replay.mp4 (front camera + wrist camera) and replay.last.png. Replays `--repeat` times in this fresh
process; success requires every replay to end its episode in success (DexJoCo's own rule: the episode ends at the first
-success or task failure) AND all final-state hashes to agree (MuJoCo on the CPU is bit-deterministic), and the frozen
-instance rebuilt here from tests/instance.npz must match its recorded hash.
+success or task failure) AND all final-state hashes to agree (MuJoCo on the CPU is bit-deterministic). The frozen
+instance rebuilt here from tests/instance.npz must match its recorded hash; one that does not (the image no longer
+rebuilds the scene) is no verdict on the agent: `instance_mismatch`, `grader_error` and `infra_error` = 1, nothing
+replayed.
A file the agent got wrong is the agent's 0, `bad_trajectory` (not `grader_error`): not an npz, pickled or object
arrays, no `task` or `actions` entry, a wrong shape, NaN/inf, or more than rcb_dexjoco.MAX_TRAJECTORY_ROWS rows (the
@@ -119,6 +121,10 @@ def main() -> int:
print("too few CPUs: " + rcb_dexjoco.threads_message(threads), file=sys.stderr)
write(False, {"grader_error": 1, "infra_error": 1, "too_few_cpus": 1, "n_actions": n_rows, **threads, **cost}); return 0
s = Session(a.family, instance=a.instance)
+ if not s.instance_exact: # the image no longer rebuilds the frozen scene: not the agent's 0 (infra, no verdict)
+ print(f"instance mismatch: rebuilt state sha {s.state_sha()} != recorded {s.instance_sha}", file=sys.stderr)
+ write(False, {"grader_error": 1, "infra_error": 1, "instance_mismatch": 1, "instance_ok": 0,
+ "n_actions": n_rows, **threads, **cost}); return 0
if acts.shape[1] != s.action_dim:
write(False, {"bad_trajectory": 1, "n_actions": n_rows, **cost}); return 0
t0 = time.time()
@@ -130,10 +136,9 @@ def main() -> int:
write(False, {"bad_trajectory": 1, "n_actions": n_rows, **cost}); return 0
all_success = all(o["success"] for o in outs)
deterministic = len({o["state_sha"] for o in outs}) == 1
- instance_ok = bool(s.instance_exact)
- ok = all_success and deterministic and instance_ok
+ ok = all_success and deterministic
out = {"success": int(ok), "all_success": int(all_success), "deterministic": int(deterministic),
- "instance_ok": int(instance_ok), "n_actions": int(len(acts)), "sim_steps": int(outs[0]["steps"]),
+ "instance_ok": 1, "n_actions": int(len(acts)), "sim_steps": int(outs[0]["steps"]),
"task_failure": int(bool(outs[0]["reason"]) and not outs[0]["success"]),
"replay_wall_s": round(time.time() - t0, 1), "grader_error": 0, "video_rendered": 0, **threads, **cost}
write(ok, out) # the verdict is on disk before the video; the video only adds video_rendered
diff --git a/tasks/dexjoco-bimanual-unlock-ipad-i00-privileged/tests/instance.npz b/tasks/dexjoco-bimanual-unlock-ipad-i00-privileged/tests/instance.npz
index 8bc4e3d0c4..a8a691967b 100644
Binary files a/tasks/dexjoco-bimanual-unlock-ipad-i00-privileged/tests/instance.npz and b/tasks/dexjoco-bimanual-unlock-ipad-i00-privileged/tests/instance.npz differ
diff --git a/tasks/dexjoco-bimanual-unlock-ipad-i00-privileged/tests/rcb_dexjoco.py b/tasks/dexjoco-bimanual-unlock-ipad-i00-privileged/tests/rcb_dexjoco.py
index f2e2db848f..8eb183fe2a 100644
--- a/tasks/dexjoco-bimanual-unlock-ipad-i00-privileged/tests/rcb_dexjoco.py
+++ b/tasks/dexjoco-bimanual-unlock-ipad-i00-privileged/tests/rcb_dexjoco.py
@@ -54,6 +54,7 @@ import argparse
import copy
import hashlib
import json
+import re
import sys
import time
from pathlib import Path
@@ -222,6 +223,21 @@ def water_aim_deg(model, data) -> float:
return float(np.degrees(np.arccos(np.clip(c, -1.0, 1.0))))
+# instance.npz is in the agent's container (/app/instance.npz), and a task's reference solution is the actions of the
+# recording its instance was restored from: its `meta` must not say where the scene came from (PROTOCOL §2.7, no pointer
+# to a solution where the agent can reach). save_instance refuses such keys, and values that name a recording.
+PROVENANCE_KEYS = ("source", "demo", "dataset", "episode", "recording", "reference", "solution", "zarr", "url")
+PROVENANCE_VALUE = re.compile(r"demo|episode|dataset|recording|replay\.zarr|huggingface|https?://", re.I)
+
+
+def check_instance_meta(meta: dict) -> None:
+ """ValueError if an instance's `meta` says where its scene came from (keys or values; see PROVENANCE_KEYS)."""
+ bad = [k for k, v in meta.items() if any(p in str(k).lower() for p in PROVENANCE_KEYS)
+ or PROVENANCE_VALUE.search(json.dumps(v))]
+ if bad:
+ raise ValueError(f"instance meta {bad} names where the scene came from: the agent reads instance.npz")
+
+
def _log(*a: Any) -> None:
print("[rcb_dexjoco]", *a, file=sys.stderr, flush=True)
@@ -349,6 +365,7 @@ class Session:
return h.hexdigest()[:16]
def save_instance(self, path: str | Path, meta: dict | None = None) -> None:
+ check_instance_meta(meta or {})
snap = self.snapshot()
np.savez_compressed(path, task=self.task, physics=snap["physics"], mocap_pos=snap["mocap_pos"],
**{"model__" + k: v for k, v in snap["model"].items()},
@@ -803,15 +820,14 @@ def cmd_make_instance(a) -> None:
restore_initial_state(s.env, a.task, s.cfg, state0)
s._snapshot = s.snapshot()
s._episode_reset()
- meta = {"source": "DexJoCo-Datasets-Raw demo, initial state restored with dexjoco.tasks.state_restorers",
- "demo": a.demo_name or os.path.basename(os.path.dirname(os.path.abspath(a.demo))), "dexjoco": "8d23b0f"}
- s.save_instance(a.out, meta)
- _log("instance", a.out, "sha", s.state_sha())
+ demo = a.demo_name or os.path.basename(os.path.dirname(os.path.abspath(a.demo)))
+ s.save_instance(a.out, {"dexjoco": "8d23b0f"}) # no provenance in the instance (check_instance_meta)
+ _log("instance", a.out, "sha", s.state_sha(), "from", demo)
if a.oracle_out:
A = np.asarray(root["data"]["action"], dtype=np.float64)
s2 = Session(a.task, instance=a.out)
st = s2.replay(A)
- s2.save_trajectory(A[:st["replayed_rows"]], a.oracle_out, {"source": "demo actions", "demo": meta["demo"]})
+ s2.save_trajectory(A[:st["replayed_rows"]], a.oracle_out, {"source": "demo actions", "demo": demo})
print(json.dumps({"oracle": a.oracle_out, **{k: st[k] for k in ("success", "reason", "replayed_rows", "rows", "state_sha")}}))
diff --git a/tasks/dexjoco-bimanual-unlock-ipad-i00-standard/README.md b/tasks/dexjoco-bimanual-unlock-ipad-i00-standard/README.md
index 2e33f5c801..50908b8772 100644
--- a/tasks/dexjoco-bimanual-unlock-ipad-i00-standard/README.md
+++ b/tasks/dexjoco-bimanual-unlock-ipad-i00-standard/README.md
@@ -72,6 +72,13 @@ Generated by `scripts/dexjoco/generate.py`, as its privileged twin.
returned once. A service that has given up on the episode (too few CPUs, or it could not resume after its
crashes) answers every request with 409 and the reason, and `Robot()` / `wait_ready()` raise `EpisodeOver` at
once; before, nothing answered and they waited out their timeout (up to an hour).
+- 2026-10-14 (review of board 0105, B-M3): `environment/sim/instance.npz` and `environment/sim/rcb_dexjoco.py`. The
+ frozen instance's `meta` entry no longer names the demonstration the scene was restored from (`source`, `demo`;
+ the reference solution is that demonstration's actions), and the tooling refuses to save an instance that does.
+ The sidecar's files are out of the agent's reach, so nothing reached the agent from this copy; it changed with
+ the privileged twin's, from the one committed instance. Every array of the scene is the same (instance sha
+ `d38c827ac3bb030e`): the scene, the pass rule and the grader are unchanged, so the task keeps its name and earlier
+ results stand.
- 2026-10-13 (owner, boards 0085 and 0087): the sim service's clock (`environment/sim/rcb_limited.py`,
`environment/sim/server.py`, the docstring of `environment/rcb_client.py`). `meta["time_s"]` and
`meta["time_left_s"]` count from the agent's first request to the service (any endpoint), and a restarted service
diff --git a/tasks/dexjoco-bimanual-unlock-ipad-i00-standard/environment/sim/instance.npz b/tasks/dexjoco-bimanual-unlock-ipad-i00-standard/environment/sim/instance.npz
index 8bc4e3d0c4..a8a691967b 100644
Binary files a/tasks/dexjoco-bimanual-unlock-ipad-i00-standard/environment/sim/instance.npz and b/tasks/dexjoco-bimanual-unlock-ipad-i00-standard/environment/sim/instance.npz differ
diff --git a/tasks/dexjoco-bimanual-unlock-ipad-i00-standard/environment/sim/rcb_dexjoco.py b/tasks/dexjoco-bimanual-unlock-ipad-i00-standard/environment/sim/rcb_dexjoco.py
index f2e2db848f..8eb183fe2a 100644
--- a/tasks/dexjoco-bimanual-unlock-ipad-i00-standard/environment/sim/rcb_dexjoco.py
+++ b/tasks/dexjoco-bimanual-unlock-ipad-i00-standard/environment/sim/rcb_dexjoco.py
@@ -54,6 +54,7 @@ import argparse
import copy
import hashlib
import json
+import re
import sys
import time
from pathlib import Path
@@ -222,6 +223,21 @@ def water_aim_deg(model, data) -> float:
return float(np.degrees(np.arccos(np.clip(c, -1.0, 1.0))))
+# instance.npz is in the agent's container (/app/instance.npz), and a task's reference solution is the actions of the
+# recording its instance was restored from: its `meta` must not say where the scene came from (PROTOCOL §2.7, no pointer
+# to a solution where the agent can reach). save_instance refuses such keys, and values that name a recording.
+PROVENANCE_KEYS = ("source", "demo", "dataset", "episode", "recording", "reference", "solution", "zarr", "url")
+PROVENANCE_VALUE = re.compile(r"demo|episode|dataset|recording|replay\.zarr|huggingface|https?://", re.I)
+
+
+def check_instance_meta(meta: dict) -> None:
+ """ValueError if an instance's `meta` says where its scene came from (keys or values; see PROVENANCE_KEYS)."""
+ bad = [k for k, v in meta.items() if any(p in str(k).lower() for p in PROVENANCE_KEYS)
+ or PROVENANCE_VALUE.search(json.dumps(v))]
+ if bad:
+ raise ValueError(f"instance meta {bad} names where the scene came from: the agent reads instance.npz")
+
+
def _log(*a: Any) -> None:
print("[rcb_dexjoco]", *a, file=sys.stderr, flush=True)
@@ -349,6 +365,7 @@ class Session:
return h.hexdigest()[:16]
def save_instance(self, path: str | Path, meta: dict | None = None) -> None:
+ check_instance_meta(meta or {})
snap = self.snapshot()
np.savez_compressed(path, task=self.task, physics=snap["physics"], mocap_pos=snap["mocap_pos"],
**{"model__" + k: v for k, v in snap["model"].items()},
@@ -803,15 +820,14 @@ def cmd_make_instance(a) -> None:
restore_initial_state(s.env, a.task, s.cfg, state0)
s._snapshot = s.snapshot()
s._episode_reset()
- meta = {"source": "DexJoCo-Datasets-Raw demo, initial state restored with dexjoco.tasks.state_restorers",
- "demo": a.demo_name or os.path.basename(os.path.dirname(os.path.abspath(a.demo))), "dexjoco": "8d23b0f"}
- s.save_instance(a.out, meta)
- _log("instance", a.out, "sha", s.state_sha())
+ demo = a.demo_name or os.path.basename(os.path.dirname(os.path.abspath(a.demo)))
+ s.save_instance(a.out, {"dexjoco": "8d23b0f"}) # no provenance in the instance (check_instance_meta)
+ _log("instance", a.out, "sha", s.state_sha(), "from", demo)
if a.oracle_out:
A = np.asarray(root["data"]["action"], dtype=np.float64)
s2 = Session(a.task, instance=a.out)
st = s2.replay(A)
- s2.save_trajectory(A[:st["replayed_rows"]], a.oracle_out, {"source": "demo actions", "demo": meta["demo"]})
+ s2.save_trajectory(A[:st["replayed_rows"]], a.oracle_out, {"source": "demo actions", "demo": demo})
print(json.dumps({"oracle": a.oracle_out, **{k: st[k] for k in ("success", "reason", "replayed_rows", "rows", "state_sha")}}))
diff --git a/tasks/dexjoco-bimanual-unlock-ipad-i00-standard/task.toml b/tasks/dexjoco-bimanual-unlock-ipad-i00-standard/task.toml
index 02a2c23ee4..625ddcac82 100644
--- a/tasks/dexjoco-bimanual-unlock-ipad-i00-standard/task.toml
+++ b/tasks/dexjoco-bimanual-unlock-ipad-i00-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "dexjoco"
robot = "franka-allegro-bimanual"
family = "bimanual_unlock_ipad"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "full"
expert_time_estimate_hours = 2.5
diff --git a/tasks/dexjoco-click-mouse-i00-privileged/README.md b/tasks/dexjoco-click-mouse-i00-privileged/README.md
index be043661dd..87b8309624 100644
--- a/tasks/dexjoco-click-mouse-i00-privileged/README.md
+++ b/tasks/dexjoco-click-mouse-i00-privileged/README.md
@@ -29,8 +29,10 @@ limited to the model APIs.
## Verification
Open-loop replay of `/app/output/trajectory.npz` from `tests/instance.npz`, twice in a fresh process in the verifier's own
-sandbox (no network); pass iff both episodes end in success, their final-state hashes agree and the rebuilt instance
-matches its recorded hash (`tests/grade.py`); plus `/logs/verifier/replay.mp4`, rendered after the verdict is written.
+sandbox (no network); pass iff both episodes end in success and their final-state hashes agree (`tests/grade.py`); plus
+`/logs/verifier/replay.mp4`, rendered after the verdict is written. A rebuilt instance that does not match its recorded
+hash (the image no longer rebuilds the scene) replays nothing and gives no verdict: `instance_mismatch`, `grader_error`
+and `infra_error` = 1, not the agent's 0.
A file that is not a readable trajectory (not an npz, no `task` / `actions`, a wrong shape, NaN) or has more than
50,000 rows scores 0 as `bad_trajectory`: the cap is what two replays of the slowest family fit into the verifier's
3600 s. A verifier that runs fewer than 4 OpenBLAS threads (fewer than 4 CPUs visible to it) replays
@@ -65,3 +67,15 @@ real-time sleep, episode cap and hanoi's wall-clock auto-reset are disabled in t
### Provenance
Generated by `scripts/dexjoco/generate.py` from `scripts/dexjoco/instances/click_mouse/` (`make_instances.py`).
+
+### Change Log
+- 2026-10-14 (review of board 0105, B-M3 and B6-3): `environment/instance.npz` and `tests/instance.npz` no longer name
+ the demonstration the scene was restored from: their `meta` entry had `source` (the dataset) and `demo` (the
+ demonstration's id), and the reference solution is that demonstration's own actions, so the agent could read at
+ `/app/instance.npz` where a solution was (PROTOCOL §2.7). Only `meta` changed: every array of the scene is the same
+ (instance sha `ad73bedaf42257f0`), so the scene, the pass rule and the reference are unchanged and the task keeps its
+ name. The tooling (`rcb_dexjoco.py`, both copies) refuses to save an instance whose `meta` says where its scene came
+ from. `tests/grade.py`: a rebuilt instance that does not match its recorded hash is now no verdict
+ (`instance_mismatch`, `grader_error`, `infra_error` = 1, nothing replayed); before, it was the agent's 0 with
+ `infra_error` = 0 and `instance_ok` = 0. Past results: the demonstration itself was out of reach under the egress
+ allowlist (Hugging Face is not on it); a past 0 with `instance_ok` = 0 was an infrastructure failure.
diff --git a/tasks/dexjoco-click-mouse-i00-privileged/environment/instance.npz b/tasks/dexjoco-click-mouse-i00-privileged/environment/instance.npz
index 274a4b7352..ad7d5c22cf 100644
Binary files a/tasks/dexjoco-click-mouse-i00-privileged/environment/instance.npz and b/tasks/dexjoco-click-mouse-i00-privileged/environment/instance.npz differ
diff --git a/tasks/dexjoco-click-mouse-i00-privileged/environment/rcb_dexjoco.py b/tasks/dexjoco-click-mouse-i00-privileged/environment/rcb_dexjoco.py
index f2e2db848f..8eb183fe2a 100644
--- a/tasks/dexjoco-click-mouse-i00-privileged/environment/rcb_dexjoco.py
+++ b/tasks/dexjoco-click-mouse-i00-privileged/environment/rcb_dexjoco.py
@@ -54,6 +54,7 @@ import argparse
import copy
import hashlib
import json
+import re
import sys
import time
from pathlib import Path
@@ -222,6 +223,21 @@ def water_aim_deg(model, data) -> float:
return float(np.degrees(np.arccos(np.clip(c, -1.0, 1.0))))
+# instance.npz is in the agent's container (/app/instance.npz), and a task's reference solution is the actions of the
+# recording its instance was restored from: its `meta` must not say where the scene came from (PROTOCOL §2.7, no pointer
+# to a solution where the agent can reach). save_instance refuses such keys, and values that name a recording.
+PROVENANCE_KEYS = ("source", "demo", "dataset", "episode", "recording", "reference", "solution", "zarr", "url")
+PROVENANCE_VALUE = re.compile(r"demo|episode|dataset|recording|replay\.zarr|huggingface|https?://", re.I)
+
+
+def check_instance_meta(meta: dict) -> None:
+ """ValueError if an instance's `meta` says where its scene came from (keys or values; see PROVENANCE_KEYS)."""
+ bad = [k for k, v in meta.items() if any(p in str(k).lower() for p in PROVENANCE_KEYS)
+ or PROVENANCE_VALUE.search(json.dumps(v))]
+ if bad:
+ raise ValueError(f"instance meta {bad} names where the scene came from: the agent reads instance.npz")
+
+
def _log(*a: Any) -> None:
print("[rcb_dexjoco]", *a, file=sys.stderr, flush=True)
@@ -349,6 +365,7 @@ class Session:
return h.hexdigest()[:16]
def save_instance(self, path: str | Path, meta: dict | None = None) -> None:
+ check_instance_meta(meta or {})
snap = self.snapshot()
np.savez_compressed(path, task=self.task, physics=snap["physics"], mocap_pos=snap["mocap_pos"],
**{"model__" + k: v for k, v in snap["model"].items()},
@@ -803,15 +820,14 @@ def cmd_make_instance(a) -> None:
restore_initial_state(s.env, a.task, s.cfg, state0)
s._snapshot = s.snapshot()
s._episode_reset()
- meta = {"source": "DexJoCo-Datasets-Raw demo, initial state restored with dexjoco.tasks.state_restorers",
- "demo": a.demo_name or os.path.basename(os.path.dirname(os.path.abspath(a.demo))), "dexjoco": "8d23b0f"}
- s.save_instance(a.out, meta)
- _log("instance", a.out, "sha", s.state_sha())
+ demo = a.demo_name or os.path.basename(os.path.dirname(os.path.abspath(a.demo)))
+ s.save_instance(a.out, {"dexjoco": "8d23b0f"}) # no provenance in the instance (check_instance_meta)
+ _log("instance", a.out, "sha", s.state_sha(), "from", demo)
if a.oracle_out:
A = np.asarray(root["data"]["action"], dtype=np.float64)
s2 = Session(a.task, instance=a.out)
st = s2.replay(A)
- s2.save_trajectory(A[:st["replayed_rows"]], a.oracle_out, {"source": "demo actions", "demo": meta["demo"]})
+ s2.save_trajectory(A[:st["replayed_rows"]], a.oracle_out, {"source": "demo actions", "demo": demo})
print(json.dumps({"oracle": a.oracle_out, **{k: st[k] for k in ("success", "reason", "replayed_rows", "rows", "state_sha")}}))
diff --git a/tasks/dexjoco-click-mouse-i00-privileged/task.toml b/tasks/dexjoco-click-mouse-i00-privileged/task.toml
index 0b1ac67d91..cf4396da5b 100644
--- a/tasks/dexjoco-click-mouse-i00-privileged/task.toml
+++ b/tasks/dexjoco-click-mouse-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "dexjoco"
robot = "franka-allegro"
family = "click_mouse"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
expert_time_estimate_hours = 1.5
diff --git a/tasks/dexjoco-click-mouse-i00-privileged/tests/grade.py b/tasks/dexjoco-click-mouse-i00-privileged/tests/grade.py
index c754f999ad..3cfbab9c42 100644
--- a/tasks/dexjoco-click-mouse-i00-privileged/tests/grade.py
+++ b/tasks/dexjoco-click-mouse-i00-privileged/tests/grade.py
@@ -6,8 +6,10 @@ write Harbor rewards.
Writes /logs/verifier/reward.txt (1/0), reward.json (numeric fields only; Harbor rejects strings), a replay video
/logs/verifier/replay.mp4 (front camera + wrist camera) and replay.last.png. Replays `--repeat` times in this fresh
process; success requires every replay to end its episode in success (DexJoCo's own rule: the episode ends at the first
-success or task failure) AND all final-state hashes to agree (MuJoCo on the CPU is bit-deterministic), and the frozen
-instance rebuilt here from tests/instance.npz must match its recorded hash.
+success or task failure) AND all final-state hashes to agree (MuJoCo on the CPU is bit-deterministic). The frozen
+instance rebuilt here from tests/instance.npz must match its recorded hash; one that does not (the image no longer
+rebuilds the scene) is no verdict on the agent: `instance_mismatch`, `grader_error` and `infra_error` = 1, nothing
+replayed.
A file the agent got wrong is the agent's 0, `bad_trajectory` (not `grader_error`): not an npz, pickled or object
arrays, no `task` or `actions` entry, a wrong shape, NaN/inf, or more than rcb_dexjoco.MAX_TRAJECTORY_ROWS rows (the
@@ -119,6 +121,10 @@ def main() -> int:
print("too few CPUs: " + rcb_dexjoco.threads_message(threads), file=sys.stderr)
write(False, {"grader_error": 1, "infra_error": 1, "too_few_cpus": 1, "n_actions": n_rows, **threads, **cost}); return 0
s = Session(a.family, instance=a.instance)
+ if not s.instance_exact: # the image no longer rebuilds the frozen scene: not the agent's 0 (infra, no verdict)
+ print(f"instance mismatch: rebuilt state sha {s.state_sha()} != recorded {s.instance_sha}", file=sys.stderr)
+ write(False, {"grader_error": 1, "infra_error": 1, "instance_mismatch": 1, "instance_ok": 0,
+ "n_actions": n_rows, **threads, **cost}); return 0
if acts.shape[1] != s.action_dim:
write(False, {"bad_trajectory": 1, "n_actions": n_rows, **cost}); return 0
t0 = time.time()
@@ -130,10 +136,9 @@ def main() -> int:
write(False, {"bad_trajectory": 1, "n_actions": n_rows, **cost}); return 0
all_success = all(o["success"] for o in outs)
deterministic = len({o["state_sha"] for o in outs}) == 1
- instance_ok = bool(s.instance_exact)
- ok = all_success and deterministic and instance_ok
+ ok = all_success and deterministic
out = {"success": int(ok), "all_success": int(all_success), "deterministic": int(deterministic),
- "instance_ok": int(instance_ok), "n_actions": int(len(acts)), "sim_steps": int(outs[0]["steps"]),
+ "instance_ok": 1, "n_actions": int(len(acts)), "sim_steps": int(outs[0]["steps"]),
"task_failure": int(bool(outs[0]["reason"]) and not outs[0]["success"]),
"replay_wall_s": round(time.time() - t0, 1), "grader_error": 0, "video_rendered": 0, **threads, **cost}
write(ok, out) # the verdict is on disk before the video; the video only adds video_rendered
diff --git a/tasks/dexjoco-click-mouse-i00-privileged/tests/instance.npz b/tasks/dexjoco-click-mouse-i00-privileged/tests/instance.npz
index 274a4b7352..ad7d5c22cf 100644
Binary files a/tasks/dexjoco-click-mouse-i00-privileged/tests/instance.npz and b/tasks/dexjoco-click-mouse-i00-privileged/tests/instance.npz differ
diff --git a/tasks/dexjoco-click-mouse-i00-privileged/tests/rcb_dexjoco.py b/tasks/dexjoco-click-mouse-i00-privileged/tests/rcb_dexjoco.py
index f2e2db848f..8eb183fe2a 100644
--- a/tasks/dexjoco-click-mouse-i00-privileged/tests/rcb_dexjoco.py
+++ b/tasks/dexjoco-click-mouse-i00-privileged/tests/rcb_dexjoco.py
@@ -54,6 +54,7 @@ import argparse
import copy
import hashlib
import json
+import re
import sys
import time
from pathlib import Path
@@ -222,6 +223,21 @@ def water_aim_deg(model, data) -> float:
return float(np.degrees(np.arccos(np.clip(c, -1.0, 1.0))))
+# instance.npz is in the agent's container (/app/instance.npz), and a task's reference solution is the actions of the
+# recording its instance was restored from: its `meta` must not say where the scene came from (PROTOCOL §2.7, no pointer
+# to a solution where the agent can reach). save_instance refuses such keys, and values that name a recording.
+PROVENANCE_KEYS = ("source", "demo", "dataset", "episode", "recording", "reference", "solution", "zarr", "url")
+PROVENANCE_VALUE = re.compile(r"demo|episode|dataset|recording|replay\.zarr|huggingface|https?://", re.I)
+
+
+def check_instance_meta(meta: dict) -> None:
+ """ValueError if an instance's `meta` says where its scene came from (keys or values; see PROVENANCE_KEYS)."""
+ bad = [k for k, v in meta.items() if any(p in str(k).lower() for p in PROVENANCE_KEYS)
+ or PROVENANCE_VALUE.search(json.dumps(v))]
+ if bad:
+ raise ValueError(f"instance meta {bad} names where the scene came from: the agent reads instance.npz")
+
+
def _log(*a: Any) -> None:
print("[rcb_dexjoco]", *a, file=sys.stderr, flush=True)
@@ -349,6 +365,7 @@ class Session:
return h.hexdigest()[:16]
def save_instance(self, path: str | Path, meta: dict | None = None) -> None:
+ check_instance_meta(meta or {})
snap = self.snapshot()
np.savez_compressed(path, task=self.task, physics=snap["physics"], mocap_pos=snap["mocap_pos"],
**{"model__" + k: v for k, v in snap["model"].items()},
@@ -803,15 +820,14 @@ def cmd_make_instance(a) -> None:
restore_initial_state(s.env, a.task, s.cfg, state0)
s._snapshot = s.snapshot()
s._episode_reset()
- meta = {"source": "DexJoCo-Datasets-Raw demo, initial state restored with dexjoco.tasks.state_restorers",
- "demo": a.demo_name or os.path.basename(os.path.dirname(os.path.abspath(a.demo))), "dexjoco": "8d23b0f"}
- s.save_instance(a.out, meta)
- _log("instance", a.out, "sha", s.state_sha())
+ demo = a.demo_name or os.path.basename(os.path.dirname(os.path.abspath(a.demo)))
+ s.save_instance(a.out, {"dexjoco": "8d23b0f"}) # no provenance in the instance (check_instance_meta)
+ _log("instance", a.out, "sha", s.state_sha(), "from", demo)
if a.oracle_out:
A = np.asarray(root["data"]["action"], dtype=np.float64)
s2 = Session(a.task, instance=a.out)
st = s2.replay(A)
- s2.save_trajectory(A[:st["replayed_rows"]], a.oracle_out, {"source": "demo actions", "demo": meta["demo"]})
+ s2.save_trajectory(A[:st["replayed_rows"]], a.oracle_out, {"source": "demo actions", "demo": demo})
print(json.dumps({"oracle": a.oracle_out, **{k: st[k] for k in ("success", "reason", "replayed_rows", "rows", "state_sha")}}))
diff --git a/tasks/dexjoco-click-mouse-i00-standard/README.md b/tasks/dexjoco-click-mouse-i00-standard/README.md
index e5df1bf5fd..0121579ec7 100644
--- a/tasks/dexjoco-click-mouse-i00-standard/README.md
+++ b/tasks/dexjoco-click-mouse-i00-standard/README.md
@@ -72,6 +72,13 @@ Generated by `scripts/dexjoco/generate.py`, as its privileged twin.
returned once. A service that has given up on the episode (too few CPUs, or it could not resume after its
crashes) answers every request with 409 and the reason, and `Robot()` / `wait_ready()` raise `EpisodeOver` at
once; before, nothing answered and they waited out their timeout (up to an hour).
+- 2026-10-14 (review of board 0105, B-M3): `environment/sim/instance.npz` and `environment/sim/rcb_dexjoco.py`. The
+ frozen instance's `meta` entry no longer names the demonstration the scene was restored from (`source`, `demo`;
+ the reference solution is that demonstration's actions), and the tooling refuses to save an instance that does.
+ The sidecar's files are out of the agent's reach, so nothing reached the agent from this copy; it changed with
+ the privileged twin's, from the one committed instance. Every array of the scene is the same (instance sha
+ `ad73bedaf42257f0`): the scene, the pass rule and the grader are unchanged, so the task keeps its name and earlier
+ results stand.
- 2026-10-13 (owner, boards 0085 and 0087): the sim service's clock (`environment/sim/rcb_limited.py`,
`environment/sim/server.py`, the docstring of `environment/rcb_client.py`). `meta["time_s"]` and
`meta["time_left_s"]` count from the agent's first request to the service (any endpoint), and a restarted service
diff --git a/tasks/dexjoco-click-mouse-i00-standard/environment/sim/instance.npz b/tasks/dexjoco-click-mouse-i00-standard/environment/sim/instance.npz
index 274a4b7352..ad7d5c22cf 100644
Binary files a/tasks/dexjoco-click-mouse-i00-standard/environment/sim/instance.npz and b/tasks/dexjoco-click-mouse-i00-standard/environment/sim/instance.npz differ
diff --git a/tasks/dexjoco-click-mouse-i00-standard/environment/sim/rcb_dexjoco.py b/tasks/dexjoco-click-mouse-i00-standard/environment/sim/rcb_dexjoco.py
index f2e2db848f..8eb183fe2a 100644
--- a/tasks/dexjoco-click-mouse-i00-standard/environment/sim/rcb_dexjoco.py
+++ b/tasks/dexjoco-click-mouse-i00-standard/environment/sim/rcb_dexjoco.py
@@ -54,6 +54,7 @@ import argparse
import copy
import hashlib
import json
+import re
import sys
import time
from pathlib import Path
@@ -222,6 +223,21 @@ def water_aim_deg(model, data) -> float:
return float(np.degrees(np.arccos(np.clip(c, -1.0, 1.0))))
+# instance.npz is in the agent's container (/app/instance.npz), and a task's reference solution is the actions of the
+# recording its instance was restored from: its `meta` must not say where the scene came from (PROTOCOL §2.7, no pointer
+# to a solution where the agent can reach). save_instance refuses such keys, and values that name a recording.
+PROVENANCE_KEYS = ("source", "demo", "dataset", "episode", "recording", "reference", "solution", "zarr", "url")
+PROVENANCE_VALUE = re.compile(r"demo|episode|dataset|recording|replay\.zarr|huggingface|https?://", re.I)
+
+
+def check_instance_meta(meta: dict) -> None:
+ """ValueError if an instance's `meta` says where its scene came from (keys or values; see PROVENANCE_KEYS)."""
+ bad = [k for k, v in meta.items() if any(p in str(k).lower() for p in PROVENANCE_KEYS)
+ or PROVENANCE_VALUE.search(json.dumps(v))]
+ if bad:
+ raise ValueError(f"instance meta {bad} names where the scene came from: the agent reads instance.npz")
+
+
def _log(*a: Any) -> None:
print("[rcb_dexjoco]", *a, file=sys.stderr, flush=True)
@@ -349,6 +365,7 @@ class Session:
return h.hexdigest()[:16]
def save_instance(self, path: str | Path, meta: dict | None = None) -> None:
+ check_instance_meta(meta or {})
snap = self.snapshot()
np.savez_compressed(path, task=self.task, physics=snap["physics"], mocap_pos=snap["mocap_pos"],
**{"model__" + k: v for k, v in snap["model"].items()},
@@ -803,15 +820,14 @@ def cmd_make_instance(a) -> None:
restore_initial_state(s.env, a.task, s.cfg, state0)
s._snapshot = s.snapshot()
s._episode_reset()
- meta = {"source": "DexJoCo-Datasets-Raw demo, initial state restored with dexjoco.tasks.state_restorers",
- "demo": a.demo_name or os.path.basename(os.path.dirname(os.path.abspath(a.demo))), "dexjoco": "8d23b0f"}
- s.save_instance(a.out, meta)
- _log("instance", a.out, "sha", s.state_sha())
+ demo = a.demo_name or os.path.basename(os.path.dirname(os.path.abspath(a.demo)))
+ s.save_instance(a.out, {"dexjoco": "8d23b0f"}) # no provenance in the instance (check_instance_meta)
+ _log("instance", a.out, "sha", s.state_sha(), "from", demo)
if a.oracle_out:
A = np.asarray(root["data"]["action"], dtype=np.float64)
s2 = Session(a.task, instance=a.out)
st = s2.replay(A)
- s2.save_trajectory(A[:st["replayed_rows"]], a.oracle_out, {"source": "demo actions", "demo": meta["demo"]})
+ s2.save_trajectory(A[:st["replayed_rows"]], a.oracle_out, {"source": "demo actions", "demo": demo})
print(json.dumps({"oracle": a.oracle_out, **{k: st[k] for k in ("success", "reason", "replayed_rows", "rows", "state_sha")}}))
diff --git a/tasks/dexjoco-click-mouse-i00-standard/task.toml b/tasks/dexjoco-click-mouse-i00-standard/task.toml
index 520b78eec4..d895e4da19 100644
--- a/tasks/dexjoco-click-mouse-i00-standard/task.toml
+++ b/tasks/dexjoco-click-mouse-i00-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "dexjoco"
robot = "franka-allegro"
family = "click_mouse"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "full"
expert_time_estimate_hours = 1.5
diff --git a/tasks/dexjoco-fold-glasses-i00-privileged/README.md b/tasks/dexjoco-fold-glasses-i00-privileged/README.md
index f6b8c5c547..8ac336f00c 100644
--- a/tasks/dexjoco-fold-glasses-i00-privileged/README.md
+++ b/tasks/dexjoco-fold-glasses-i00-privileged/README.md
@@ -29,8 +29,10 @@ limited to the model APIs.
## Verification
Open-loop replay of `/app/output/trajectory.npz` from `tests/instance.npz`, twice in a fresh process in the verifier's own
-sandbox (no network); pass iff both episodes end in success, their final-state hashes agree and the rebuilt instance
-matches its recorded hash (`tests/grade.py`); plus `/logs/verifier/replay.mp4`, rendered after the verdict is written.
+sandbox (no network); pass iff both episodes end in success and their final-state hashes agree (`tests/grade.py`); plus
+`/logs/verifier/replay.mp4`, rendered after the verdict is written. A rebuilt instance that does not match its recorded
+hash (the image no longer rebuilds the scene) replays nothing and gives no verdict: `instance_mismatch`, `grader_error`
+and `infra_error` = 1, not the agent's 0.
A file that is not a readable trajectory (not an npz, no `task` / `actions`, a wrong shape, NaN) or has more than
50,000 rows scores 0 as `bad_trajectory`: the cap is what two replays of the slowest family fit into the verifier's
3600 s. A verifier that runs fewer than 4 OpenBLAS threads (fewer than 4 CPUs visible to it) replays
@@ -65,3 +67,15 @@ real-time sleep, episode cap and hanoi's wall-clock auto-reset are disabled in t
### Provenance
Generated by `scripts/dexjoco/generate.py` from `scripts/dexjoco/instances/fold_glasses/` (`make_instances.py`).
+
+### Change Log
+- 2026-10-14 (review of board 0105, B-M3 and B6-3): `environment/instance.npz` and `tests/instance.npz` no longer name
+ the demonstration the scene was restored from: their `meta` entry had `source` (the dataset) and `demo` (the
+ demonstration's id), and the reference solution is that demonstration's own actions, so the agent could read at
+ `/app/instance.npz` where a solution was (PROTOCOL §2.7). Only `meta` changed: every array of the scene is the same
+ (instance sha `35849caecf4a0f6d`), so the scene, the pass rule and the reference are unchanged and the task keeps its
+ name. The tooling (`rcb_dexjoco.py`, both copies) refuses to save an instance whose `meta` says where its scene came
+ from. `tests/grade.py`: a rebuilt instance that does not match its recorded hash is now no verdict
+ (`instance_mismatch`, `grader_error`, `infra_error` = 1, nothing replayed); before, it was the agent's 0 with
+ `infra_error` = 0 and `instance_ok` = 0. Past results: the demonstration itself was out of reach under the egress
+ allowlist (Hugging Face is not on it); a past 0 with `instance_ok` = 0 was an infrastructure failure.
diff --git a/tasks/dexjoco-fold-glasses-i00-privileged/environment/instance.npz b/tasks/dexjoco-fold-glasses-i00-privileged/environment/instance.npz
index ac1baebfb1..1467404f75 100644
Binary files a/tasks/dexjoco-fold-glasses-i00-privileged/environment/instance.npz and b/tasks/dexjoco-fold-glasses-i00-privileged/environment/instance.npz differ
diff --git a/tasks/dexjoco-fold-glasses-i00-privileged/environment/rcb_dexjoco.py b/tasks/dexjoco-fold-glasses-i00-privileged/environment/rcb_dexjoco.py
index f2e2db848f..8eb183fe2a 100644
--- a/tasks/dexjoco-fold-glasses-i00-privileged/environment/rcb_dexjoco.py
+++ b/tasks/dexjoco-fold-glasses-i00-privileged/environment/rcb_dexjoco.py
@@ -54,6 +54,7 @@ import argparse
import copy
import hashlib
import json
+import re
import sys
import time
from pathlib import Path
@@ -222,6 +223,21 @@ def water_aim_deg(model, data) -> float:
return float(np.degrees(np.arccos(np.clip(c, -1.0, 1.0))))
+# instance.npz is in the agent's container (/app/instance.npz), and a task's reference solution is the actions of the
+# recording its instance was restored from: its `meta` must not say where the scene came from (PROTOCOL §2.7, no pointer
+# to a solution where the agent can reach). save_instance refuses such keys, and values that name a recording.
+PROVENANCE_KEYS = ("source", "demo", "dataset", "episode", "recording", "reference", "solution", "zarr", "url")
+PROVENANCE_VALUE = re.compile(r"demo|episode|dataset|recording|replay\.zarr|huggingface|https?://", re.I)
+
+
+def check_instance_meta(meta: dict) -> None:
+ """ValueError if an instance's `meta` says where its scene came from (keys or values; see PROVENANCE_KEYS)."""
+ bad = [k for k, v in meta.items() if any(p in str(k).lower() for p in PROVENANCE_KEYS)
+ or PROVENANCE_VALUE.search(json.dumps(v))]
+ if bad:
+ raise ValueError(f"instance meta {bad} names where the scene came from: the agent reads instance.npz")
+
+
def _log(*a: Any) -> None:
print("[rcb_dexjoco]", *a, file=sys.stderr, flush=True)
@@ -349,6 +365,7 @@ class Session:
return h.hexdigest()[:16]
def save_instance(self, path: str | Path, meta: dict | None = None) -> None:
+ check_instance_meta(meta or {})
snap = self.snapshot()
np.savez_compressed(path, task=self.task, physics=snap["physics"], mocap_pos=snap["mocap_pos"],
**{"model__" + k: v for k, v in snap["model"].items()},
@@ -803,15 +820,14 @@ def cmd_make_instance(a) -> None:
restore_initial_state(s.env, a.task, s.cfg, state0)
s._snapshot = s.snapshot()
s._episode_reset()
- meta = {"source": "DexJoCo-Datasets-Raw demo, initial state restored with dexjoco.tasks.state_restorers",
- "demo": a.demo_name or os.path.basename(os.path.dirname(os.path.abspath(a.demo))), "dexjoco": "8d23b0f"}
- s.save_instance(a.out, meta)
- _log("instance", a.out, "sha", s.state_sha())
+ demo = a.demo_name or os.path.basename(os.path.dirname(os.path.abspath(a.demo)))
+ s.save_instance(a.out, {"dexjoco": "8d23b0f"}) # no provenance in the instance (check_instance_meta)
+ _log("instance", a.out, "sha", s.state_sha(), "from", demo)
if a.oracle_out:
A = np.asarray(root["data"]["action"], dtype=np.float64)
s2 = Session(a.task, instance=a.out)
st = s2.replay(A)
- s2.save_trajectory(A[:st["replayed_rows"]], a.oracle_out, {"source": "demo actions", "demo": meta["demo"]})
+ s2.save_trajectory(A[:st["replayed_rows"]], a.oracle_out, {"source": "demo actions", "demo": demo})
print(json.dumps({"oracle": a.oracle_out, **{k: st[k] for k in ("success", "reason", "replayed_rows", "rows", "state_sha")}}))
diff --git a/tasks/dexjoco-fold-glasses-i00-privileged/task.toml b/tasks/dexjoco-fold-glasses-i00-privileged/task.toml
index 4db63ab49c..b2f7150910 100644
--- a/tasks/dexjoco-fold-glasses-i00-privileged/task.toml
+++ b/tasks/dexjoco-fold-glasses-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "dexjoco"
robot = "franka-allegro"
family = "fold_glasses"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
expert_time_estimate_hours = 2.5
diff --git a/tasks/dexjoco-fold-glasses-i00-privileged/tests/grade.py b/tasks/dexjoco-fold-glasses-i00-privileged/tests/grade.py
index c754f999ad..3cfbab9c42 100644
--- a/tasks/dexjoco-fold-glasses-i00-privileged/tests/grade.py
+++ b/tasks/dexjoco-fold-glasses-i00-privileged/tests/grade.py
@@ -6,8 +6,10 @@ write Harbor rewards.
Writes /logs/verifier/reward.txt (1/0), reward.json (numeric fields only; Harbor rejects strings), a replay video
/logs/verifier/replay.mp4 (front camera + wrist camera) and replay.last.png. Replays `--repeat` times in this fresh
process; success requires every replay to end its episode in success (DexJoCo's own rule: the episode ends at the first
-success or task failure) AND all final-state hashes to agree (MuJoCo on the CPU is bit-deterministic), and the frozen
-instance rebuilt here from tests/instance.npz must match its recorded hash.
+success or task failure) AND all final-state hashes to agree (MuJoCo on the CPU is bit-deterministic). The frozen
+instance rebuilt here from tests/instance.npz must match its recorded hash; one that does not (the image no longer
+rebuilds the scene) is no verdict on the agent: `instance_mismatch`, `grader_error` and `infra_error` = 1, nothing
+replayed.
A file the agent got wrong is the agent's 0, `bad_trajectory` (not `grader_error`): not an npz, pickled or object
arrays, no `task` or `actions` entry, a wrong shape, NaN/inf, or more than rcb_dexjoco.MAX_TRAJECTORY_ROWS rows (the
@@ -119,6 +121,10 @@ def main() -> int:
print("too few CPUs: " + rcb_dexjoco.threads_message(threads), file=sys.stderr)
write(False, {"grader_error": 1, "infra_error": 1, "too_few_cpus": 1, "n_actions": n_rows, **threads, **cost}); return 0
s = Session(a.family, instance=a.instance)
+ if not s.instance_exact: # the image no longer rebuilds the frozen scene: not the agent's 0 (infra, no verdict)
+ print(f"instance mismatch: rebuilt state sha {s.state_sha()} != recorded {s.instance_sha}", file=sys.stderr)
+ write(False, {"grader_error": 1, "infra_error": 1, "instance_mismatch": 1, "instance_ok": 0,
+ "n_actions": n_rows, **threads, **cost}); return 0
if acts.shape[1] != s.action_dim:
write(False, {"bad_trajectory": 1, "n_actions": n_rows, **cost}); return 0
t0 = time.time()
@@ -130,10 +136,9 @@ def main() -> int:
write(False, {"bad_trajectory": 1, "n_actions": n_rows, **cost}); return 0
all_success = all(o["success"] for o in outs)
deterministic = len({o["state_sha"] for o in outs}) == 1
- instance_ok = bool(s.instance_exact)
- ok = all_success and deterministic and instance_ok
+ ok = all_success and deterministic
out = {"success": int(ok), "all_success": int(all_success), "deterministic": int(deterministic),
- "instance_ok": int(instance_ok), "n_actions": int(len(acts)), "sim_steps": int(outs[0]["steps"]),
+ "instance_ok": 1, "n_actions": int(len(acts)), "sim_steps": int(outs[0]["steps"]),
"task_failure": int(bool(outs[0]["reason"]) and not outs[0]["success"]),
"replay_wall_s": round(time.time() - t0, 1), "grader_error": 0, "video_rendered": 0, **threads, **cost}
write(ok, out) # the verdict is on disk before the video; the video only adds video_rendered
diff --git a/tasks/dexjoco-fold-glasses-i00-privileged/tests/instance.npz b/tasks/dexjoco-fold-glasses-i00-privileged/tests/instance.npz
index ac1baebfb1..1467404f75 100644
Binary files a/tasks/dexjoco-fold-glasses-i00-privileged/tests/instance.npz and b/tasks/dexjoco-fold-glasses-i00-privileged/tests/instance.npz differ
diff --git a/tasks/dexjoco-fold-glasses-i00-privileged/tests/rcb_dexjoco.py b/tasks/dexjoco-fold-glasses-i00-privileged/tests/rcb_dexjoco.py
index f2e2db848f..8eb183fe2a 100644
--- a/tasks/dexjoco-fold-glasses-i00-privileged/tests/rcb_dexjoco.py
+++ b/tasks/dexjoco-fold-glasses-i00-privileged/tests/rcb_dexjoco.py
@@ -54,6 +54,7 @@ import argparse
import copy
import hashlib
import json
+import re
import sys
import time
from pathlib import Path
@@ -222,6 +223,21 @@ def water_aim_deg(model, data) -> float:
return float(np.degrees(np.arccos(np.clip(c, -1.0, 1.0))))
+# instance.npz is in the agent's container (/app/instance.npz), and a task's reference solution is the actions of the
+# recording its instance was restored from: its `meta` must not say where the scene came from (PROTOCOL §2.7, no pointer
+# to a solution where the agent can reach). save_instance refuses such keys, and values that name a recording.
+PROVENANCE_KEYS = ("source", "demo", "dataset", "episode", "recording", "reference", "solution", "zarr", "url")
+PROVENANCE_VALUE = re.compile(r"demo|episode|dataset|recording|replay\.zarr|huggingface|https?://", re.I)
+
+
+def check_instance_meta(meta: dict) -> None:
+ """ValueError if an instance's `meta` says where its scene came from (keys or values; see PROVENANCE_KEYS)."""
+ bad = [k for k, v in meta.items() if any(p in str(k).lower() for p in PROVENANCE_KEYS)
+ or PROVENANCE_VALUE.search(json.dumps(v))]
+ if bad:
+ raise ValueError(f"instance meta {bad} names where the scene came from: the agent reads instance.npz")
+
+
def _log(*a: Any) -> None:
print("[rcb_dexjoco]", *a, file=sys.stderr, flush=True)
@@ -349,6 +365,7 @@ class Session:
return h.hexdigest()[:16]
def save_instance(self, path: str | Path, meta: dict | None = None) -> None:
+ check_instance_meta(meta or {})
snap = self.snapshot()
np.savez_compressed(path, task=self.task, physics=snap["physics"], mocap_pos=snap["mocap_pos"],
**{"model__" + k: v for k, v in snap["model"].items()},
@@ -803,15 +820,14 @@ def cmd_make_instance(a) -> None:
restore_initial_state(s.env, a.task, s.cfg, state0)
s._snapshot = s.snapshot()
s._episode_reset()
- meta = {"source": "DexJoCo-Datasets-Raw demo, initial state restored with dexjoco.tasks.state_restorers",
- "demo": a.demo_name or os.path.basename(os.path.dirname(os.path.abspath(a.demo))), "dexjoco": "8d23b0f"}
- s.save_instance(a.out, meta)
- _log("instance", a.out, "sha", s.state_sha())
+ demo = a.demo_name or os.path.basename(os.path.dirname(os.path.abspath(a.demo)))
+ s.save_instance(a.out, {"dexjoco": "8d23b0f"}) # no provenance in the instance (check_instance_meta)
+ _log("instance", a.out, "sha", s.state_sha(), "from", demo)
if a.oracle_out:
A = np.asarray(root["data"]["action"], dtype=np.float64)
s2 = Session(a.task, instance=a.out)
st = s2.replay(A)
- s2.save_trajectory(A[:st["replayed_rows"]], a.oracle_out, {"source": "demo actions", "demo": meta["demo"]})
+ s2.save_trajectory(A[:st["replayed_rows"]], a.oracle_out, {"source": "demo actions", "demo": demo})
print(json.dumps({"oracle": a.oracle_out, **{k: st[k] for k in ("success", "reason", "replayed_rows", "rows", "state_sha")}}))
diff --git a/tasks/dexjoco-fold-glasses-i00-standard/README.md b/tasks/dexjoco-fold-glasses-i00-standard/README.md
index 257913982f..b1014e8001 100644
--- a/tasks/dexjoco-fold-glasses-i00-standard/README.md
+++ b/tasks/dexjoco-fold-glasses-i00-standard/README.md
@@ -72,6 +72,13 @@ Generated by `scripts/dexjoco/generate.py`, as its privileged twin.
returned once. A service that has given up on the episode (too few CPUs, or it could not resume after its
crashes) answers every request with 409 and the reason, and `Robot()` / `wait_ready()` raise `EpisodeOver` at
once; before, nothing answered and they waited out their timeout (up to an hour).
+- 2026-10-14 (review of board 0105, B-M3): `environment/sim/instance.npz` and `environment/sim/rcb_dexjoco.py`. The
+ frozen instance's `meta` entry no longer names the demonstration the scene was restored from (`source`, `demo`;
+ the reference solution is that demonstration's actions), and the tooling refuses to save an instance that does.
+ The sidecar's files are out of the agent's reach, so nothing reached the agent from this copy; it changed with
+ the privileged twin's, from the one committed instance. Every array of the scene is the same (instance sha
+ `35849caecf4a0f6d`): the scene, the pass rule and the grader are unchanged, so the task keeps its name and earlier
+ results stand.
- 2026-10-13 (owner, boards 0085 and 0087): the sim service's clock (`environment/sim/rcb_limited.py`,
`environment/sim/server.py`, the docstring of `environment/rcb_client.py`). `meta["time_s"]` and
`meta["time_left_s"]` count from the agent's first request to the service (any endpoint), and a restarted service
diff --git a/tasks/dexjoco-fold-glasses-i00-standard/environment/sim/instance.npz b/tasks/dexjoco-fold-glasses-i00-standard/environment/sim/instance.npz
index ac1baebfb1..1467404f75 100644
Binary files a/tasks/dexjoco-fold-glasses-i00-standard/environment/sim/instance.npz and b/tasks/dexjoco-fold-glasses-i00-standard/environment/sim/instance.npz differ
diff --git a/tasks/dexjoco-fold-glasses-i00-standard/environment/sim/rcb_dexjoco.py b/tasks/dexjoco-fold-glasses-i00-standard/environment/sim/rcb_dexjoco.py
index f2e2db848f..8eb183fe2a 100644
--- a/tasks/dexjoco-fold-glasses-i00-standard/environment/sim/rcb_dexjoco.py
+++ b/tasks/dexjoco-fold-glasses-i00-standard/environment/sim/rcb_dexjoco.py
@@ -54,6 +54,7 @@ import argparse
import copy
import hashlib
import json
+import re
import sys
import time
from pathlib import Path
@@ -222,6 +223,21 @@ def water_aim_deg(model, data) -> float:
return float(np.degrees(np.arccos(np.clip(c, -1.0, 1.0))))
+# instance.npz is in the agent's container (/app/instance.npz), and a task's reference solution is the actions of the
+# recording its instance was restored from: its `meta` must not say where the scene came from (PROTOCOL §2.7, no pointer
+# to a solution where the agent can reach). save_instance refuses such keys, and values that name a recording.
+PROVENANCE_KEYS = ("source", "demo", "dataset", "episode", "recording", "reference", "solution", "zarr", "url")
+PROVENANCE_VALUE = re.compile(r"demo|episode|dataset|recording|replay\.zarr|huggingface|https?://", re.I)
+
+
+def check_instance_meta(meta: dict) -> None:
+ """ValueError if an instance's `meta` says where its scene came from (keys or values; see PROVENANCE_KEYS)."""
+ bad = [k for k, v in meta.items() if any(p in str(k).lower() for p in PROVENANCE_KEYS)
+ or PROVENANCE_VALUE.search(json.dumps(v))]
+ if bad:
+ raise ValueError(f"instance meta {bad} names where the scene came from: the agent reads instance.npz")
+
+
def _log(*a: Any) -> None:
print("[rcb_dexjoco]", *a, file=sys.stderr, flush=True)
@@ -349,6 +365,7 @@ class Session:
return h.hexdigest()[:16]
def save_instance(self, path: str | Path, meta: dict | None = None) -> None:
+ check_instance_meta(meta or {})
snap = self.snapshot()
np.savez_compressed(path, task=self.task, physics=snap["physics"], mocap_pos=snap["mocap_pos"],
**{"model__" + k: v for k, v in snap["model"].items()},
@@ -803,15 +820,14 @@ def cmd_make_instance(a) -> None:
restore_initial_state(s.env, a.task, s.cfg, state0)
s._snapshot = s.snapshot()
s._episode_reset()
- meta = {"source": "DexJoCo-Datasets-Raw demo, initial state restored with dexjoco.tasks.state_restorers",
- "demo": a.demo_name or os.path.basename(os.path.dirname(os.path.abspath(a.demo))), "dexjoco": "8d23b0f"}
- s.save_instance(a.out, meta)
- _log("instance", a.out, "sha", s.state_sha())
+ demo = a.demo_name or os.path.basename(os.path.dirname(os.path.abspath(a.demo)))
+ s.save_instance(a.out, {"dexjoco": "8d23b0f"}) # no provenance in the instance (check_instance_meta)
+ _log("instance", a.out, "sha", s.state_sha(), "from", demo)
if a.oracle_out:
A = np.asarray(root["data"]["action"], dtype=np.float64)
s2 = Session(a.task, instance=a.out)
st = s2.replay(A)
- s2.save_trajectory(A[:st["replayed_rows"]], a.oracle_out, {"source": "demo actions", "demo": meta["demo"]})
+ s2.save_trajectory(A[:st["replayed_rows"]], a.oracle_out, {"source": "demo actions", "demo": demo})
print(json.dumps({"oracle": a.oracle_out, **{k: st[k] for k in ("success", "reason", "replayed_rows", "rows", "state_sha")}}))
diff --git a/tasks/dexjoco-fold-glasses-i00-standard/task.toml b/tasks/dexjoco-fold-glasses-i00-standard/task.toml
index c1600ff227..82b06ca2b9 100644
--- a/tasks/dexjoco-fold-glasses-i00-standard/task.toml
+++ b/tasks/dexjoco-fold-glasses-i00-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "dexjoco"
robot = "franka-allegro"
family = "fold_glasses"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "full"
expert_time_estimate_hours = 2.5
diff --git a/tasks/dexjoco-hammer-nail-i00-privileged/README.md b/tasks/dexjoco-hammer-nail-i00-privileged/README.md
index f786d96134..63d64b134d 100644
--- a/tasks/dexjoco-hammer-nail-i00-privileged/README.md
+++ b/tasks/dexjoco-hammer-nail-i00-privileged/README.md
@@ -29,8 +29,10 @@ limited to the model APIs.
## Verification
Open-loop replay of `/app/output/trajectory.npz` from `tests/instance.npz`, twice in a fresh process in the verifier's own
-sandbox (no network); pass iff both episodes end in success, their final-state hashes agree and the rebuilt instance
-matches its recorded hash (`tests/grade.py`); plus `/logs/verifier/replay.mp4`, rendered after the verdict is written.
+sandbox (no network); pass iff both episodes end in success and their final-state hashes agree (`tests/grade.py`); plus
+`/logs/verifier/replay.mp4`, rendered after the verdict is written. A rebuilt instance that does not match its recorded
+hash (the image no longer rebuilds the scene) replays nothing and gives no verdict: `instance_mismatch`, `grader_error`
+and `infra_error` = 1, not the agent's 0.
A file that is not a readable trajectory (not an npz, no `task` / `actions`, a wrong shape, NaN) or has more than
50,000 rows scores 0 as `bad_trajectory`: the cap is what two replays of the slowest family fit into the verifier's
3600 s. A verifier that runs fewer than 4 OpenBLAS threads (fewer than 4 CPUs visible to it) replays
@@ -65,3 +67,15 @@ real-time sleep, episode cap and hanoi's wall-clock auto-reset are disabled in t
### Provenance
Generated by `scripts/dexjoco/generate.py` from `scripts/dexjoco/instances/hammer_nail/` (`make_instances.py`).
+
+### Change Log
+- 2026-10-14 (review of board 0105, B-M3 and B6-3): `environment/instance.npz` and `tests/instance.npz` no longer name
+ the demonstration the scene was restored from: their `meta` entry had `source` (the dataset) and `demo` (the
+ demonstration's id), and the reference solution is that demonstration's own actions, so the agent could read at
+ `/app/instance.npz` where a solution was (PROTOCOL §2.7). Only `meta` changed: every array of the scene is the same
+ (instance sha `095c53d8e9a1cae0`), so the scene, the pass rule and the reference are unchanged and the task keeps its
+ name. The tooling (`rcb_dexjoco.py`, both copies) refuses to save an instance whose `meta` says where its scene came
+ from. `tests/grade.py`: a rebuilt instance that does not match its recorded hash is now no verdict
+ (`instance_mismatch`, `grader_error`, `infra_error` = 1, nothing replayed); before, it was the agent's 0 with
+ `infra_error` = 0 and `instance_ok` = 0. Past results: the demonstration itself was out of reach under the egress
+ allowlist (Hugging Face is not on it); a past 0 with `instance_ok` = 0 was an infrastructure failure.
diff --git a/tasks/dexjoco-hammer-nail-i00-privileged/environment/instance.npz b/tasks/dexjoco-hammer-nail-i00-privileged/environment/instance.npz
index 443d63038d..0ef696d692 100644
Binary files a/tasks/dexjoco-hammer-nail-i00-privileged/environment/instance.npz and b/tasks/dexjoco-hammer-nail-i00-privileged/environment/instance.npz differ
diff --git a/tasks/dexjoco-hammer-nail-i00-privileged/environment/rcb_dexjoco.py b/tasks/dexjoco-hammer-nail-i00-privileged/environment/rcb_dexjoco.py
index f2e2db848f..8eb183fe2a 100644
--- a/tasks/dexjoco-hammer-nail-i00-privileged/environment/rcb_dexjoco.py
+++ b/tasks/dexjoco-hammer-nail-i00-privileged/environment/rcb_dexjoco.py
@@ -54,6 +54,7 @@ import argparse
import copy
import hashlib
import json
+import re
import sys
import time
from pathlib import Path
@@ -222,6 +223,21 @@ def water_aim_deg(model, data) -> float:
return float(np.degrees(np.arccos(np.clip(c, -1.0, 1.0))))
+# instance.npz is in the agent's container (/app/instance.npz), and a task's reference solution is the actions of the
+# recording its instance was restored from: its `meta` must not say where the scene came from (PROTOCOL §2.7, no pointer
+# to a solution where the agent can reach). save_instance refuses such keys, and values that name a recording.
+PROVENANCE_KEYS = ("source", "demo", "dataset", "episode", "recording", "reference", "solution", "zarr", "url")
+PROVENANCE_VALUE = re.compile(r"demo|episode|dataset|recording|replay\.zarr|huggingface|https?://", re.I)
+
+
+def check_instance_meta(meta: dict) -> None:
+ """ValueError if an instance's `meta` says where its scene came from (keys or values; see PROVENANCE_KEYS)."""
+ bad = [k for k, v in meta.items() if any(p in str(k).lower() for p in PROVENANCE_KEYS)
+ or PROVENANCE_VALUE.search(json.dumps(v))]
+ if bad:
+ raise ValueError(f"instance meta {bad} names where the scene came from: the agent reads instance.npz")
+
+
def _log(*a: Any) -> None:
print("[rcb_dexjoco]", *a, file=sys.stderr, flush=True)
@@ -349,6 +365,7 @@ class Session:
return h.hexdigest()[:16]
def save_instance(self, path: str | Path, meta: dict | None = None) -> None:
+ check_instance_meta(meta or {})
snap = self.snapshot()
np.savez_compressed(path, task=self.task, physics=snap["physics"], mocap_pos=snap["mocap_pos"],
**{"model__" + k: v for k, v in snap["model"].items()},
@@ -803,15 +820,14 @@ def cmd_make_instance(a) -> None:
restore_initial_state(s.env, a.task, s.cfg, state0)
s._snapshot = s.snapshot()
s._episode_reset()
- meta = {"source": "DexJoCo-Datasets-Raw demo, initial state restored with dexjoco.tasks.state_restorers",
- "demo": a.demo_name or os.path.basename(os.path.dirname(os.path.abspath(a.demo))), "dexjoco": "8d23b0f"}
- s.save_instance(a.out, meta)
- _log("instance", a.out, "sha", s.state_sha())
+ demo = a.demo_name or os.path.basename(os.path.dirname(os.path.abspath(a.demo)))
+ s.save_instance(a.out, {"dexjoco": "8d23b0f"}) # no provenance in the instance (check_instance_meta)
+ _log("instance", a.out, "sha", s.state_sha(), "from", demo)
if a.oracle_out:
A = np.asarray(root["data"]["action"], dtype=np.float64)
s2 = Session(a.task, instance=a.out)
st = s2.replay(A)
- s2.save_trajectory(A[:st["replayed_rows"]], a.oracle_out, {"source": "demo actions", "demo": meta["demo"]})
+ s2.save_trajectory(A[:st["replayed_rows"]], a.oracle_out, {"source": "demo actions", "demo": demo})
print(json.dumps({"oracle": a.oracle_out, **{k: st[k] for k in ("success", "reason", "replayed_rows", "rows", "state_sha")}}))
diff --git a/tasks/dexjoco-hammer-nail-i00-privileged/task.toml b/tasks/dexjoco-hammer-nail-i00-privileged/task.toml
index af39e692c9..2a0c2e0c67 100644
--- a/tasks/dexjoco-hammer-nail-i00-privileged/task.toml
+++ b/tasks/dexjoco-hammer-nail-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "dexjoco"
robot = "franka-allegro"
family = "hammer_nail"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
expert_time_estimate_hours = 2.0
diff --git a/tasks/dexjoco-hammer-nail-i00-privileged/tests/grade.py b/tasks/dexjoco-hammer-nail-i00-privileged/tests/grade.py
index c754f999ad..3cfbab9c42 100644
--- a/tasks/dexjoco-hammer-nail-i00-privileged/tests/grade.py
+++ b/tasks/dexjoco-hammer-nail-i00-privileged/tests/grade.py
@@ -6,8 +6,10 @@ write Harbor rewards.
Writes /logs/verifier/reward.txt (1/0), reward.json (numeric fields only; Harbor rejects strings), a replay video
/logs/verifier/replay.mp4 (front camera + wrist camera) and replay.last.png. Replays `--repeat` times in this fresh
process; success requires every replay to end its episode in success (DexJoCo's own rule: the episode ends at the first
-success or task failure) AND all final-state hashes to agree (MuJoCo on the CPU is bit-deterministic), and the frozen
-instance rebuilt here from tests/instance.npz must match its recorded hash.
+success or task failure) AND all final-state hashes to agree (MuJoCo on the CPU is bit-deterministic). The frozen
+instance rebuilt here from tests/instance.npz must match its recorded hash; one that does not (the image no longer
+rebuilds the scene) is no verdict on the agent: `instance_mismatch`, `grader_error` and `infra_error` = 1, nothing
+replayed.
A file the agent got wrong is the agent's 0, `bad_trajectory` (not `grader_error`): not an npz, pickled or object
arrays, no `task` or `actions` entry, a wrong shape, NaN/inf, or more than rcb_dexjoco.MAX_TRAJECTORY_ROWS rows (the
@@ -119,6 +121,10 @@ def main() -> int:
print("too few CPUs: " + rcb_dexjoco.threads_message(threads), file=sys.stderr)
write(False, {"grader_error": 1, "infra_error": 1, "too_few_cpus": 1, "n_actions": n_rows, **threads, **cost}); return 0
s = Session(a.family, instance=a.instance)
+ if not s.instance_exact: # the image no longer rebuilds the frozen scene: not the agent's 0 (infra, no verdict)
+ print(f"instance mismatch: rebuilt state sha {s.state_sha()} != recorded {s.instance_sha}", file=sys.stderr)
+ write(False, {"grader_error": 1, "infra_error": 1, "instance_mismatch": 1, "instance_ok": 0,
+ "n_actions": n_rows, **threads, **cost}); return 0
if acts.shape[1] != s.action_dim:
write(False, {"bad_trajectory": 1, "n_actions": n_rows, **cost}); return 0
t0 = time.time()
@@ -130,10 +136,9 @@ def main() -> int:
write(False, {"bad_trajectory": 1, "n_actions": n_rows, **cost}); return 0
all_success = all(o["success"] for o in outs)
deterministic = len({o["state_sha"] for o in outs}) == 1
- instance_ok = bool(s.instance_exact)
- ok = all_success and deterministic and instance_ok
+ ok = all_success and deterministic
out = {"success": int(ok), "all_success": int(all_success), "deterministic": int(deterministic),
- "instance_ok": int(instance_ok), "n_actions": int(len(acts)), "sim_steps": int(outs[0]["steps"]),
+ "instance_ok": 1, "n_actions": int(len(acts)), "sim_steps": int(outs[0]["steps"]),
"task_failure": int(bool(outs[0]["reason"]) and not outs[0]["success"]),
"replay_wall_s": round(time.time() - t0, 1), "grader_error": 0, "video_rendered": 0, **threads, **cost}
write(ok, out) # the verdict is on disk before the video; the video only adds video_rendered
diff --git a/tasks/dexjoco-hammer-nail-i00-privileged/tests/instance.npz b/tasks/dexjoco-hammer-nail-i00-privileged/tests/instance.npz
index 443d63038d..0ef696d692 100644
Binary files a/tasks/dexjoco-hammer-nail-i00-privileged/tests/instance.npz and b/tasks/dexjoco-hammer-nail-i00-privileged/tests/instance.npz differ
diff --git a/tasks/dexjoco-hammer-nail-i00-privileged/tests/rcb_dexjoco.py b/tasks/dexjoco-hammer-nail-i00-privileged/tests/rcb_dexjoco.py
index f2e2db848f..8eb183fe2a 100644
--- a/tasks/dexjoco-hammer-nail-i00-privileged/tests/rcb_dexjoco.py
+++ b/tasks/dexjoco-hammer-nail-i00-privileged/tests/rcb_dexjoco.py
@@ -54,6 +54,7 @@ import argparse
import copy
import hashlib
import json
+import re
import sys
import time
from pathlib import Path
@@ -222,6 +223,21 @@ def water_aim_deg(model, data) -> float:
return float(np.degrees(np.arccos(np.clip(c, -1.0, 1.0))))
+# instance.npz is in the agent's container (/app/instance.npz), and a task's reference solution is the actions of the
+# recording its instance was restored from: its `meta` must not say where the scene came from (PROTOCOL §2.7, no pointer
+# to a solution where the agent can reach). save_instance refuses such keys, and values that name a recording.
+PROVENANCE_KEYS = ("source", "demo", "dataset", "episode", "recording", "reference", "solution", "zarr", "url")
+PROVENANCE_VALUE = re.compile(r"demo|episode|dataset|recording|replay\.zarr|huggingface|https?://", re.I)
+
+
+def check_instance_meta(meta: dict) -> None:
+ """ValueError if an instance's `meta` says where its scene came from (keys or values; see PROVENANCE_KEYS)."""
+ bad = [k for k, v in meta.items() if any(p in str(k).lower() for p in PROVENANCE_KEYS)
+ or PROVENANCE_VALUE.search(json.dumps(v))]
+ if bad:
+ raise ValueError(f"instance meta {bad} names where the scene came from: the agent reads instance.npz")
+
+
def _log(*a: Any) -> None:
print("[rcb_dexjoco]", *a, file=sys.stderr, flush=True)
@@ -349,6 +365,7 @@ class Session:
return h.hexdigest()[:16]
def save_instance(self, path: str | Path, meta: dict | None = None) -> None:
+ check_instance_meta(meta or {})
snap = self.snapshot()
np.savez_compressed(path, task=self.task, physics=snap["physics"], mocap_pos=snap["mocap_pos"],
**{"model__" + k: v for k, v in snap["model"].items()},
@@ -803,15 +820,14 @@ def cmd_make_instance(a) -> None:
restore_initial_state(s.env, a.task, s.cfg, state0)
s._snapshot = s.snapshot()
s._episode_reset()
- meta = {"source": "DexJoCo-Datasets-Raw demo, initial state restored with dexjoco.tasks.state_restorers",
- "demo": a.demo_name or os.path.basename(os.path.dirname(os.path.abspath(a.demo))), "dexjoco": "8d23b0f"}
- s.save_instance(a.out, meta)
- _log("instance", a.out, "sha", s.state_sha())
+ demo = a.demo_name or os.path.basename(os.path.dirname(os.path.abspath(a.demo)))
+ s.save_instance(a.out, {"dexjoco": "8d23b0f"}) # no provenance in the instance (check_instance_meta)
+ _log("instance", a.out, "sha", s.state_sha(), "from", demo)
if a.oracle_out:
A = np.asarray(root["data"]["action"], dtype=np.float64)
s2 = Session(a.task, instance=a.out)
st = s2.replay(A)
- s2.save_trajectory(A[:st["replayed_rows"]], a.oracle_out, {"source": "demo actions", "demo": meta["demo"]})
+ s2.save_trajectory(A[:st["replayed_rows"]], a.oracle_out, {"source": "demo actions", "demo": demo})
print(json.dumps({"oracle": a.oracle_out, **{k: st[k] for k in ("success", "reason", "replayed_rows", "rows", "state_sha")}}))
diff --git a/tasks/dexjoco-hammer-nail-i00-standard/README.md b/tasks/dexjoco-hammer-nail-i00-standard/README.md
index 43c890be76..abe46e7c52 100644
--- a/tasks/dexjoco-hammer-nail-i00-standard/README.md
+++ b/tasks/dexjoco-hammer-nail-i00-standard/README.md
@@ -72,6 +72,13 @@ Generated by `scripts/dexjoco/generate.py`, as its privileged twin.
returned once. A service that has given up on the episode (too few CPUs, or it could not resume after its
crashes) answers every request with 409 and the reason, and `Robot()` / `wait_ready()` raise `EpisodeOver` at
once; before, nothing answered and they waited out their timeout (up to an hour).
+- 2026-10-14 (review of board 0105, B-M3): `environment/sim/instance.npz` and `environment/sim/rcb_dexjoco.py`. The
+ frozen instance's `meta` entry no longer names the demonstration the scene was restored from (`source`, `demo`;
+ the reference solution is that demonstration's actions), and the tooling refuses to save an instance that does.
+ The sidecar's files are out of the agent's reach, so nothing reached the agent from this copy; it changed with
+ the privileged twin's, from the one committed instance. Every array of the scene is the same (instance sha
+ `095c53d8e9a1cae0`): the scene, the pass rule and the grader are unchanged, so the task keeps its name and earlier
+ results stand.
- 2026-10-13 (owner, boards 0085 and 0087): the sim service's clock (`environment/sim/rcb_limited.py`,
`environment/sim/server.py`, the docstring of `environment/rcb_client.py`). `meta["time_s"]` and
`meta["time_left_s"]` count from the agent's first request to the service (any endpoint), and a restarted service
diff --git a/tasks/dexjoco-hammer-nail-i00-standard/environment/sim/instance.npz b/tasks/dexjoco-hammer-nail-i00-standard/environment/sim/instance.npz
index 443d63038d..0ef696d692 100644
Binary files a/tasks/dexjoco-hammer-nail-i00-standard/environment/sim/instance.npz and b/tasks/dexjoco-hammer-nail-i00-standard/environment/sim/instance.npz differ
diff --git a/tasks/dexjoco-hammer-nail-i00-standard/environment/sim/rcb_dexjoco.py b/tasks/dexjoco-hammer-nail-i00-standard/environment/sim/rcb_dexjoco.py
index f2e2db848f..8eb183fe2a 100644
--- a/tasks/dexjoco-hammer-nail-i00-standard/environment/sim/rcb_dexjoco.py
+++ b/tasks/dexjoco-hammer-nail-i00-standard/environment/sim/rcb_dexjoco.py
@@ -54,6 +54,7 @@ import argparse
import copy
import hashlib
import json
+import re
import sys
import time
from pathlib import Path
@@ -222,6 +223,21 @@ def water_aim_deg(model, data) -> float:
return float(np.degrees(np.arccos(np.clip(c, -1.0, 1.0))))
+# instance.npz is in the agent's container (/app/instance.npz), and a task's reference solution is the actions of the
+# recording its instance was restored from: its `meta` must not say where the scene came from (PROTOCOL §2.7, no pointer
+# to a solution where the agent can reach). save_instance refuses such keys, and values that name a recording.
+PROVENANCE_KEYS = ("source", "demo", "dataset", "episode", "recording", "reference", "solution", "zarr", "url")
+PROVENANCE_VALUE = re.compile(r"demo|episode|dataset|recording|replay\.zarr|huggingface|https?://", re.I)
+
+
+def check_instance_meta(meta: dict) -> None:
+ """ValueError if an instance's `meta` says where its scene came from (keys or values; see PROVENANCE_KEYS)."""
+ bad = [k for k, v in meta.items() if any(p in str(k).lower() for p in PROVENANCE_KEYS)
+ or PROVENANCE_VALUE.search(json.dumps(v))]
+ if bad:
+ raise ValueError(f"instance meta {bad} names where the scene came from: the agent reads instance.npz")
+
+
def _log(*a: Any) -> None:
print("[rcb_dexjoco]", *a, file=sys.stderr, flush=True)
@@ -349,6 +365,7 @@ class Session:
return h.hexdigest()[:16]
def save_instance(self, path: str | Path, meta: dict | None = None) -> None:
+ check_instance_meta(meta or {})
snap = self.snapshot()
np.savez_compressed(path, task=self.task, physics=snap["physics"], mocap_pos=snap["mocap_pos"],
**{"model__" + k: v for k, v in snap["model"].items()},
@@ -803,15 +820,14 @@ def cmd_make_instance(a) -> None:
restore_initial_state(s.env, a.task, s.cfg, state0)
s._snapshot = s.snapshot()
s._episode_reset()
- meta = {"source": "DexJoCo-Datasets-Raw demo, initial state restored with dexjoco.tasks.state_restorers",
- "demo": a.demo_name or os.path.basename(os.path.dirname(os.path.abspath(a.demo))), "dexjoco": "8d23b0f"}
- s.save_instance(a.out, meta)
- _log("instance", a.out, "sha", s.state_sha())
+ demo = a.demo_name or os.path.basename(os.path.dirname(os.path.abspath(a.demo)))
+ s.save_instance(a.out, {"dexjoco": "8d23b0f"}) # no provenance in the instance (check_instance_meta)
+ _log("instance", a.out, "sha", s.state_sha(), "from", demo)
if a.oracle_out:
A = np.asarray(root["data"]["action"], dtype=np.float64)
s2 = Session(a.task, instance=a.out)
st = s2.replay(A)
- s2.save_trajectory(A[:st["replayed_rows"]], a.oracle_out, {"source": "demo actions", "demo": meta["demo"]})
+ s2.save_trajectory(A[:st["replayed_rows"]], a.oracle_out, {"source": "demo actions", "demo": demo})
print(json.dumps({"oracle": a.oracle_out, **{k: st[k] for k in ("success", "reason", "replayed_rows", "rows", "state_sha")}}))
diff --git a/tasks/dexjoco-hammer-nail-i00-standard/task.toml b/tasks/dexjoco-hammer-nail-i00-standard/task.toml
index 8883546a1b..345d75e91f 100644
--- a/tasks/dexjoco-hammer-nail-i00-standard/task.toml
+++ b/tasks/dexjoco-hammer-nail-i00-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "dexjoco"
robot = "franka-allegro"
family = "hammer_nail"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "full"
expert_time_estimate_hours = 2.0
diff --git a/tasks/dexjoco-pick-bucket-i00-privileged/README.md b/tasks/dexjoco-pick-bucket-i00-privileged/README.md
index f454b1465c..cb3a26aa5d 100644
--- a/tasks/dexjoco-pick-bucket-i00-privileged/README.md
+++ b/tasks/dexjoco-pick-bucket-i00-privileged/README.md
@@ -29,8 +29,10 @@ limited to the model APIs.
## Verification
Open-loop replay of `/app/output/trajectory.npz` from `tests/instance.npz`, twice in a fresh process in the verifier's own
-sandbox (no network); pass iff both episodes end in success, their final-state hashes agree and the rebuilt instance
-matches its recorded hash (`tests/grade.py`); plus `/logs/verifier/replay.mp4`, rendered after the verdict is written.
+sandbox (no network); pass iff both episodes end in success and their final-state hashes agree (`tests/grade.py`); plus
+`/logs/verifier/replay.mp4`, rendered after the verdict is written. A rebuilt instance that does not match its recorded
+hash (the image no longer rebuilds the scene) replays nothing and gives no verdict: `instance_mismatch`, `grader_error`
+and `infra_error` = 1, not the agent's 0.
A file that is not a readable trajectory (not an npz, no `task` / `actions`, a wrong shape, NaN) or has more than
50,000 rows scores 0 as `bad_trajectory`: the cap is what two replays of the slowest family fit into the verifier's
3600 s. A verifier that runs fewer than 4 OpenBLAS threads (fewer than 4 CPUs visible to it) replays
@@ -65,3 +67,15 @@ real-time sleep, episode cap and hanoi's wall-clock auto-reset are disabled in t
### Provenance
Generated by `scripts/dexjoco/generate.py` from `scripts/dexjoco/instances/pick_bucket/` (`make_instances.py`).
+
+### Change Log
+- 2026-10-14 (review of board 0105, B-M3 and B6-3): `environment/instance.npz` and `tests/instance.npz` no longer name
+ the demonstration the scene was restored from: their `meta` entry had `source` (the dataset) and `demo` (the
+ demonstration's id), and the reference solution is that demonstration's own actions, so the agent could read at
+ `/app/instance.npz` where a solution was (PROTOCOL §2.7). Only `meta` changed: every array of the scene is the same
+ (instance sha `1cb042bf6d46ad45`), so the scene, the pass rule and the reference are unchanged and the task keeps its
+ name. The tooling (`rcb_dexjoco.py`, both copies) refuses to save an instance whose `meta` says where its scene came
+ from. `tests/grade.py`: a rebuilt instance that does not match its recorded hash is now no verdict
+ (`instance_mismatch`, `grader_error`, `infra_error` = 1, nothing replayed); before, it was the agent's 0 with
+ `infra_error` = 0 and `instance_ok` = 0. Past results: the demonstration itself was out of reach under the egress
+ allowlist (Hugging Face is not on it); a past 0 with `instance_ok` = 0 was an infrastructure failure.
diff --git a/tasks/dexjoco-pick-bucket-i00-privileged/environment/instance.npz b/tasks/dexjoco-pick-bucket-i00-privileged/environment/instance.npz
index 473da0693d..eabd7c2ed1 100644
Binary files a/tasks/dexjoco-pick-bucket-i00-privileged/environment/instance.npz and b/tasks/dexjoco-pick-bucket-i00-privileged/environment/instance.npz differ
diff --git a/tasks/dexjoco-pick-bucket-i00-privileged/environment/rcb_dexjoco.py b/tasks/dexjoco-pick-bucket-i00-privileged/environment/rcb_dexjoco.py
index f2e2db848f..8eb183fe2a 100644
--- a/tasks/dexjoco-pick-bucket-i00-privileged/environment/rcb_dexjoco.py
+++ b/tasks/dexjoco-pick-bucket-i00-privileged/environment/rcb_dexjoco.py
@@ -54,6 +54,7 @@ import argparse
import copy
import hashlib
import json
+import re
import sys
import time
from pathlib import Path
@@ -222,6 +223,21 @@ def water_aim_deg(model, data) -> float:
return float(np.degrees(np.arccos(np.clip(c, -1.0, 1.0))))
+# instance.npz is in the agent's container (/app/instance.npz), and a task's reference solution is the actions of the
+# recording its instance was restored from: its `meta` must not say where the scene came from (PROTOCOL §2.7, no pointer
+# to a solution where the agent can reach). save_instance refuses such keys, and values that name a recording.
+PROVENANCE_KEYS = ("source", "demo", "dataset", "episode", "recording", "reference", "solution", "zarr", "url")
+PROVENANCE_VALUE = re.compile(r"demo|episode|dataset|recording|replay\.zarr|huggingface|https?://", re.I)
+
+
+def check_instance_meta(meta: dict) -> None:
+ """ValueError if an instance's `meta` says where its scene came from (keys or values; see PROVENANCE_KEYS)."""
+ bad = [k for k, v in meta.items() if any(p in str(k).lower() for p in PROVENANCE_KEYS)
+ or PROVENANCE_VALUE.search(json.dumps(v))]
+ if bad:
+ raise ValueError(f"instance meta {bad} names where the scene came from: the agent reads instance.npz")
+
+
def _log(*a: Any) -> None:
print("[rcb_dexjoco]", *a, file=sys.stderr, flush=True)
@@ -349,6 +365,7 @@ class Session:
return h.hexdigest()[:16]
def save_instance(self, path: str | Path, meta: dict | None = None) -> None:
+ check_instance_meta(meta or {})
snap = self.snapshot()
np.savez_compressed(path, task=self.task, physics=snap["physics"], mocap_pos=snap["mocap_pos"],
**{"model__" + k: v for k, v in snap["model"].items()},
@@ -803,15 +820,14 @@ def cmd_make_instance(a) -> None:
restore_initial_state(s.env, a.task, s.cfg, state0)
s._snapshot = s.snapshot()
s._episode_reset()
- meta = {"source": "DexJoCo-Datasets-Raw demo, initial state restored with dexjoco.tasks.state_restorers",
- "demo": a.demo_name or os.path.basename(os.path.dirname(os.path.abspath(a.demo))), "dexjoco": "8d23b0f"}
- s.save_instance(a.out, meta)
- _log("instance", a.out, "sha", s.state_sha())
+ demo = a.demo_name or os.path.basename(os.path.dirname(os.path.abspath(a.demo)))
+ s.save_instance(a.out, {"dexjoco": "8d23b0f"}) # no provenance in the instance (check_instance_meta)
+ _log("instance", a.out, "sha", s.state_sha(), "from", demo)
if a.oracle_out:
A = np.asarray(root["data"]["action"], dtype=np.float64)
s2 = Session(a.task, instance=a.out)
st = s2.replay(A)
- s2.save_trajectory(A[:st["replayed_rows"]], a.oracle_out, {"source": "demo actions", "demo": meta["demo"]})
+ s2.save_trajectory(A[:st["replayed_rows"]], a.oracle_out, {"source": "demo actions", "demo": demo})
print(json.dumps({"oracle": a.oracle_out, **{k: st[k] for k in ("success", "reason", "replayed_rows", "rows", "state_sha")}}))
diff --git a/tasks/dexjoco-pick-bucket-i00-privileged/task.toml b/tasks/dexjoco-pick-bucket-i00-privileged/task.toml
index 06158919b6..3a35023590 100644
--- a/tasks/dexjoco-pick-bucket-i00-privileged/task.toml
+++ b/tasks/dexjoco-pick-bucket-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "dexjoco"
robot = "franka-allegro"
family = "pick_bucket"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
expert_time_estimate_hours = 2.0
diff --git a/tasks/dexjoco-pick-bucket-i00-privileged/tests/grade.py b/tasks/dexjoco-pick-bucket-i00-privileged/tests/grade.py
index c754f999ad..3cfbab9c42 100644
--- a/tasks/dexjoco-pick-bucket-i00-privileged/tests/grade.py
+++ b/tasks/dexjoco-pick-bucket-i00-privileged/tests/grade.py
@@ -6,8 +6,10 @@ write Harbor rewards.
Writes /logs/verifier/reward.txt (1/0), reward.json (numeric fields only; Harbor rejects strings), a replay video
/logs/verifier/replay.mp4 (front camera + wrist camera) and replay.last.png. Replays `--repeat` times in this fresh
process; success requires every replay to end its episode in success (DexJoCo's own rule: the episode ends at the first
-success or task failure) AND all final-state hashes to agree (MuJoCo on the CPU is bit-deterministic), and the frozen
-instance rebuilt here from tests/instance.npz must match its recorded hash.
+success or task failure) AND all final-state hashes to agree (MuJoCo on the CPU is bit-deterministic). The frozen
+instance rebuilt here from tests/instance.npz must match its recorded hash; one that does not (the image no longer
+rebuilds the scene) is no verdict on the agent: `instance_mismatch`, `grader_error` and `infra_error` = 1, nothing
+replayed.
A file the agent got wrong is the agent's 0, `bad_trajectory` (not `grader_error`): not an npz, pickled or object
arrays, no `task` or `actions` entry, a wrong shape, NaN/inf, or more than rcb_dexjoco.MAX_TRAJECTORY_ROWS rows (the
@@ -119,6 +121,10 @@ def main() -> int:
print("too few CPUs: " + rcb_dexjoco.threads_message(threads), file=sys.stderr)
write(False, {"grader_error": 1, "infra_error": 1, "too_few_cpus": 1, "n_actions": n_rows, **threads, **cost}); return 0
s = Session(a.family, instance=a.instance)
+ if not s.instance_exact: # the image no longer rebuilds the frozen scene: not the agent's 0 (infra, no verdict)
+ print(f"instance mismatch: rebuilt state sha {s.state_sha()} != recorded {s.instance_sha}", file=sys.stderr)
+ write(False, {"grader_error": 1, "infra_error": 1, "instance_mismatch": 1, "instance_ok": 0,
+ "n_actions": n_rows, **threads, **cost}); return 0
if acts.shape[1] != s.action_dim:
write(False, {"bad_trajectory": 1, "n_actions": n_rows, **cost}); return 0
t0 = time.time()
@@ -130,10 +136,9 @@ def main() -> int:
write(False, {"bad_trajectory": 1, "n_actions": n_rows, **cost}); return 0
all_success = all(o["success"] for o in outs)
deterministic = len({o["state_sha"] for o in outs}) == 1
- instance_ok = bool(s.instance_exact)
- ok = all_success and deterministic and instance_ok
+ ok = all_success and deterministic
out = {"success": int(ok), "all_success": int(all_success), "deterministic": int(deterministic),
- "instance_ok": int(instance_ok), "n_actions": int(len(acts)), "sim_steps": int(outs[0]["steps"]),
+ "instance_ok": 1, "n_actions": int(len(acts)), "sim_steps": int(outs[0]["steps"]),
"task_failure": int(bool(outs[0]["reason"]) and not outs[0]["success"]),
"replay_wall_s": round(time.time() - t0, 1), "grader_error": 0, "video_rendered": 0, **threads, **cost}
write(ok, out) # the verdict is on disk before the video; the video only adds video_rendered
diff --git a/tasks/dexjoco-pick-bucket-i00-privileged/tests/instance.npz b/tasks/dexjoco-pick-bucket-i00-privileged/tests/instance.npz
index 473da0693d..eabd7c2ed1 100644
Binary files a/tasks/dexjoco-pick-bucket-i00-privileged/tests/instance.npz and b/tasks/dexjoco-pick-bucket-i00-privileged/tests/instance.npz differ
diff --git a/tasks/dexjoco-pick-bucket-i00-privileged/tests/rcb_dexjoco.py b/tasks/dexjoco-pick-bucket-i00-privileged/tests/rcb_dexjoco.py
index f2e2db848f..8eb183fe2a 100644
--- a/tasks/dexjoco-pick-bucket-i00-privileged/tests/rcb_dexjoco.py
+++ b/tasks/dexjoco-pick-bucket-i00-privileged/tests/rcb_dexjoco.py
@@ -54,6 +54,7 @@ import argparse
import copy
import hashlib
import json
+import re
import sys
import time
from pathlib import Path
@@ -222,6 +223,21 @@ def water_aim_deg(model, data) -> float:
return float(np.degrees(np.arccos(np.clip(c, -1.0, 1.0))))
+# instance.npz is in the agent's container (/app/instance.npz), and a task's reference solution is the actions of the
+# recording its instance was restored from: its `meta` must not say where the scene came from (PROTOCOL §2.7, no pointer
+# to a solution where the agent can reach). save_instance refuses such keys, and values that name a recording.
+PROVENANCE_KEYS = ("source", "demo", "dataset", "episode", "recording", "reference", "solution", "zarr", "url")
+PROVENANCE_VALUE = re.compile(r"demo|episode|dataset|recording|replay\.zarr|huggingface|https?://", re.I)
+
+
+def check_instance_meta(meta: dict) -> None:
+ """ValueError if an instance's `meta` says where its scene came from (keys or values; see PROVENANCE_KEYS)."""
+ bad = [k for k, v in meta.items() if any(p in str(k).lower() for p in PROVENANCE_KEYS)
+ or PROVENANCE_VALUE.search(json.dumps(v))]
+ if bad:
+ raise ValueError(f"instance meta {bad} names where the scene came from: the agent reads instance.npz")
+
+
def _log(*a: Any) -> None:
print("[rcb_dexjoco]", *a, file=sys.stderr, flush=True)
@@ -349,6 +365,7 @@ class Session:
return h.hexdigest()[:16]
def save_instance(self, path: str | Path, meta: dict | None = None) -> None:
+ check_instance_meta(meta or {})
snap = self.snapshot()
np.savez_compressed(path, task=self.task, physics=snap["physics"], mocap_pos=snap["mocap_pos"],
**{"model__" + k: v for k, v in snap["model"].items()},
@@ -803,15 +820,14 @@ def cmd_make_instance(a) -> None:
restore_initial_state(s.env, a.task, s.cfg, state0)
s._snapshot = s.snapshot()
s._episode_reset()
- meta = {"source": "DexJoCo-Datasets-Raw demo, initial state restored with dexjoco.tasks.state_restorers",
- "demo": a.demo_name or os.path.basename(os.path.dirname(os.path.abspath(a.demo))), "dexjoco": "8d23b0f"}
- s.save_instance(a.out, meta)
- _log("instance", a.out, "sha", s.state_sha())
+ demo = a.demo_name or os.path.basename(os.path.dirname(os.path.abspath(a.demo)))
+ s.save_instance(a.out, {"dexjoco": "8d23b0f"}) # no provenance in the instance (check_instance_meta)
+ _log("instance", a.out, "sha", s.state_sha(), "from", demo)
if a.oracle_out:
A = np.asarray(root["data"]["action"], dtype=np.float64)
s2 = Session(a.task, instance=a.out)
st = s2.replay(A)
- s2.save_trajectory(A[:st["replayed_rows"]], a.oracle_out, {"source": "demo actions", "demo": meta["demo"]})
+ s2.save_trajectory(A[:st["replayed_rows"]], a.oracle_out, {"source": "demo actions", "demo": demo})
print(json.dumps({"oracle": a.oracle_out, **{k: st[k] for k in ("success", "reason", "replayed_rows", "rows", "state_sha")}}))
diff --git a/tasks/dexjoco-pick-bucket-i00-standard/README.md b/tasks/dexjoco-pick-bucket-i00-standard/README.md
index 9043c43edc..0b107f8aea 100644
--- a/tasks/dexjoco-pick-bucket-i00-standard/README.md
+++ b/tasks/dexjoco-pick-bucket-i00-standard/README.md
@@ -72,6 +72,13 @@ Generated by `scripts/dexjoco/generate.py`, as its privileged twin.
returned once. A service that has given up on the episode (too few CPUs, or it could not resume after its
crashes) answers every request with 409 and the reason, and `Robot()` / `wait_ready()` raise `EpisodeOver` at
once; before, nothing answered and they waited out their timeout (up to an hour).
+- 2026-10-14 (review of board 0105, B-M3): `environment/sim/instance.npz` and `environment/sim/rcb_dexjoco.py`. The
+ frozen instance's `meta` entry no longer names the demonstration the scene was restored from (`source`, `demo`;
+ the reference solution is that demonstration's actions), and the tooling refuses to save an instance that does.
+ The sidecar's files are out of the agent's reach, so nothing reached the agent from this copy; it changed with
+ the privileged twin's, from the one committed instance. Every array of the scene is the same (instance sha
+ `1cb042bf6d46ad45`): the scene, the pass rule and the grader are unchanged, so the task keeps its name and earlier
+ results stand.
- 2026-10-13 (owner, boards 0085 and 0087): the sim service's clock (`environment/sim/rcb_limited.py`,
`environment/sim/server.py`, the docstring of `environment/rcb_client.py`). `meta["time_s"]` and
`meta["time_left_s"]` count from the agent's first request to the service (any endpoint), and a restarted service
diff --git a/tasks/dexjoco-pick-bucket-i00-standard/environment/sim/instance.npz b/tasks/dexjoco-pick-bucket-i00-standard/environment/sim/instance.npz
index 473da0693d..eabd7c2ed1 100644
Binary files a/tasks/dexjoco-pick-bucket-i00-standard/environment/sim/instance.npz and b/tasks/dexjoco-pick-bucket-i00-standard/environment/sim/instance.npz differ
diff --git a/tasks/dexjoco-pick-bucket-i00-standard/environment/sim/rcb_dexjoco.py b/tasks/dexjoco-pick-bucket-i00-standard/environment/sim/rcb_dexjoco.py
index f2e2db848f..8eb183fe2a 100644
--- a/tasks/dexjoco-pick-bucket-i00-standard/environment/sim/rcb_dexjoco.py
+++ b/tasks/dexjoco-pick-bucket-i00-standard/environment/sim/rcb_dexjoco.py
@@ -54,6 +54,7 @@ import argparse
import copy
import hashlib
import json
+import re
import sys
import time
from pathlib import Path
@@ -222,6 +223,21 @@ def water_aim_deg(model, data) -> float:
return float(np.degrees(np.arccos(np.clip(c, -1.0, 1.0))))
+# instance.npz is in the agent's container (/app/instance.npz), and a task's reference solution is the actions of the
+# recording its instance was restored from: its `meta` must not say where the scene came from (PROTOCOL §2.7, no pointer
+# to a solution where the agent can reach). save_instance refuses such keys, and values that name a recording.
+PROVENANCE_KEYS = ("source", "demo", "dataset", "episode", "recording", "reference", "solution", "zarr", "url")
+PROVENANCE_VALUE = re.compile(r"demo|episode|dataset|recording|replay\.zarr|huggingface|https?://", re.I)
+
+
+def check_instance_meta(meta: dict) -> None:
+ """ValueError if an instance's `meta` says where its scene came from (keys or values; see PROVENANCE_KEYS)."""
+ bad = [k for k, v in meta.items() if any(p in str(k).lower() for p in PROVENANCE_KEYS)
+ or PROVENANCE_VALUE.search(json.dumps(v))]
+ if bad:
+ raise ValueError(f"instance meta {bad} names where the scene came from: the agent reads instance.npz")
+
+
def _log(*a: Any) -> None:
print("[rcb_dexjoco]", *a, file=sys.stderr, flush=True)
@@ -349,6 +365,7 @@ class Session:
return h.hexdigest()[:16]
def save_instance(self, path: str | Path, meta: dict | None = None) -> None:
+ check_instance_meta(meta or {})
snap = self.snapshot()
np.savez_compressed(path, task=self.task, physics=snap["physics"], mocap_pos=snap["mocap_pos"],
**{"model__" + k: v for k, v in snap["model"].items()},
@@ -803,15 +820,14 @@ def cmd_make_instance(a) -> None:
restore_initial_state(s.env, a.task, s.cfg, state0)
s._snapshot = s.snapshot()
s._episode_reset()
- meta = {"source": "DexJoCo-Datasets-Raw demo, initial state restored with dexjoco.tasks.state_restorers",
- "demo": a.demo_name or os.path.basename(os.path.dirname(os.path.abspath(a.demo))), "dexjoco": "8d23b0f"}
- s.save_instance(a.out, meta)
- _log("instance", a.out, "sha", s.state_sha())
+ demo = a.demo_name or os.path.basename(os.path.dirname(os.path.abspath(a.demo)))
+ s.save_instance(a.out, {"dexjoco": "8d23b0f"}) # no provenance in the instance (check_instance_meta)
+ _log("instance", a.out, "sha", s.state_sha(), "from", demo)
if a.oracle_out:
A = np.asarray(root["data"]["action"], dtype=np.float64)
s2 = Session(a.task, instance=a.out)
st = s2.replay(A)
- s2.save_trajectory(A[:st["replayed_rows"]], a.oracle_out, {"source": "demo actions", "demo": meta["demo"]})
+ s2.save_trajectory(A[:st["replayed_rows"]], a.oracle_out, {"source": "demo actions", "demo": demo})
print(json.dumps({"oracle": a.oracle_out, **{k: st[k] for k in ("success", "reason", "replayed_rows", "rows", "state_sha")}}))
diff --git a/tasks/dexjoco-pick-bucket-i00-standard/task.toml b/tasks/dexjoco-pick-bucket-i00-standard/task.toml
index 2e9208381e..7dfd12e9b2 100644
--- a/tasks/dexjoco-pick-bucket-i00-standard/task.toml
+++ b/tasks/dexjoco-pick-bucket-i00-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "dexjoco"
robot = "franka-allegro"
family = "pick_bucket"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "full"
expert_time_estimate_hours = 2.0
diff --git a/tasks/dexjoco-pinch-tongs-i00-privileged/README.md b/tasks/dexjoco-pinch-tongs-i00-privileged/README.md
index 13a8acbd71..453776fbd1 100644
--- a/tasks/dexjoco-pinch-tongs-i00-privileged/README.md
+++ b/tasks/dexjoco-pinch-tongs-i00-privileged/README.md
@@ -29,8 +29,10 @@ limited to the model APIs.
## Verification
Open-loop replay of `/app/output/trajectory.npz` from `tests/instance.npz`, twice in a fresh process in the verifier's own
-sandbox (no network); pass iff both episodes end in success, their final-state hashes agree and the rebuilt instance
-matches its recorded hash (`tests/grade.py`); plus `/logs/verifier/replay.mp4`, rendered after the verdict is written.
+sandbox (no network); pass iff both episodes end in success and their final-state hashes agree (`tests/grade.py`); plus
+`/logs/verifier/replay.mp4`, rendered after the verdict is written. A rebuilt instance that does not match its recorded
+hash (the image no longer rebuilds the scene) replays nothing and gives no verdict: `instance_mismatch`, `grader_error`
+and `infra_error` = 1, not the agent's 0.
A file that is not a readable trajectory (not an npz, no `task` / `actions`, a wrong shape, NaN) or has more than
50,000 rows scores 0 as `bad_trajectory`: the cap is what two replays of the slowest family fit into the verifier's
3600 s. A verifier that runs fewer than 4 OpenBLAS threads (fewer than 4 CPUs visible to it) replays
@@ -65,3 +67,15 @@ real-time sleep, episode cap and hanoi's wall-clock auto-reset are disabled in t
### Provenance
Generated by `scripts/dexjoco/generate.py` from `scripts/dexjoco/instances/pinch_tongs/` (`make_instances.py`).
+
+### Change Log
+- 2026-10-14 (review of board 0105, B-M3 and B6-3): `environment/instance.npz` and `tests/instance.npz` no longer name
+ the demonstration the scene was restored from: their `meta` entry had `source` (the dataset) and `demo` (the
+ demonstration's id), and the reference solution is that demonstration's own actions, so the agent could read at
+ `/app/instance.npz` where a solution was (PROTOCOL §2.7). Only `meta` changed: every array of the scene is the same
+ (instance sha `ad2cfca40d427e20`), so the scene, the pass rule and the reference are unchanged and the task keeps its
+ name. The tooling (`rcb_dexjoco.py`, both copies) refuses to save an instance whose `meta` says where its scene came
+ from. `tests/grade.py`: a rebuilt instance that does not match its recorded hash is now no verdict
+ (`instance_mismatch`, `grader_error`, `infra_error` = 1, nothing replayed); before, it was the agent's 0 with
+ `infra_error` = 0 and `instance_ok` = 0. Past results: the demonstration itself was out of reach under the egress
+ allowlist (Hugging Face is not on it); a past 0 with `instance_ok` = 0 was an infrastructure failure.
diff --git a/tasks/dexjoco-pinch-tongs-i00-privileged/environment/instance.npz b/tasks/dexjoco-pinch-tongs-i00-privileged/environment/instance.npz
index fb46289594..22b9c055b9 100644
Binary files a/tasks/dexjoco-pinch-tongs-i00-privileged/environment/instance.npz and b/tasks/dexjoco-pinch-tongs-i00-privileged/environment/instance.npz differ
diff --git a/tasks/dexjoco-pinch-tongs-i00-privileged/environment/rcb_dexjoco.py b/tasks/dexjoco-pinch-tongs-i00-privileged/environment/rcb_dexjoco.py
index f2e2db848f..8eb183fe2a 100644
--- a/tasks/dexjoco-pinch-tongs-i00-privileged/environment/rcb_dexjoco.py
+++ b/tasks/dexjoco-pinch-tongs-i00-privileged/environment/rcb_dexjoco.py
@@ -54,6 +54,7 @@ import argparse
import copy
import hashlib
import json
+import re
import sys
import time
from pathlib import Path
@@ -222,6 +223,21 @@ def water_aim_deg(model, data) -> float:
return float(np.degrees(np.arccos(np.clip(c, -1.0, 1.0))))
+# instance.npz is in the agent's container (/app/instance.npz), and a task's reference solution is the actions of the
+# recording its instance was restored from: its `meta` must not say where the scene came from (PROTOCOL §2.7, no pointer
+# to a solution where the agent can reach). save_instance refuses such keys, and values that name a recording.
+PROVENANCE_KEYS = ("source", "demo", "dataset", "episode", "recording", "reference", "solution", "zarr", "url")
+PROVENANCE_VALUE = re.compile(r"demo|episode|dataset|recording|replay\.zarr|huggingface|https?://", re.I)
+
+
+def check_instance_meta(meta: dict) -> None:
+ """ValueError if an instance's `meta` says where its scene came from (keys or values; see PROVENANCE_KEYS)."""
+ bad = [k for k, v in meta.items() if any(p in str(k).lower() for p in PROVENANCE_KEYS)
+ or PROVENANCE_VALUE.search(json.dumps(v))]
+ if bad:
+ raise ValueError(f"instance meta {bad} names where the scene came from: the agent reads instance.npz")
+
+
def _log(*a: Any) -> None:
print("[rcb_dexjoco]", *a, file=sys.stderr, flush=True)
@@ -349,6 +365,7 @@ class Session:
return h.hexdigest()[:16]
def save_instance(self, path: str | Path, meta: dict | None = None) -> None:
+ check_instance_meta(meta or {})
snap = self.snapshot()
np.savez_compressed(path, task=self.task, physics=snap["physics"], mocap_pos=snap["mocap_pos"],
**{"model__" + k: v for k, v in snap["model"].items()},
@@ -803,15 +820,14 @@ def cmd_make_instance(a) -> None:
restore_initial_state(s.env, a.task, s.cfg, state0)
s._snapshot = s.snapshot()
s._episode_reset()
- meta = {"source": "DexJoCo-Datasets-Raw demo, initial state restored with dexjoco.tasks.state_restorers",
- "demo": a.demo_name or os.path.basename(os.path.dirname(os.path.abspath(a.demo))), "dexjoco": "8d23b0f"}
- s.save_instance(a.out, meta)
- _log("instance", a.out, "sha", s.state_sha())
+ demo = a.demo_name or os.path.basename(os.path.dirname(os.path.abspath(a.demo)))
+ s.save_instance(a.out, {"dexjoco": "8d23b0f"}) # no provenance in the instance (check_instance_meta)
+ _log("instance", a.out, "sha", s.state_sha(), "from", demo)
if a.oracle_out:
A = np.asarray(root["data"]["action"], dtype=np.float64)
s2 = Session(a.task, instance=a.out)
st = s2.replay(A)
- s2.save_trajectory(A[:st["replayed_rows"]], a.oracle_out, {"source": "demo actions", "demo": meta["demo"]})
+ s2.save_trajectory(A[:st["replayed_rows"]], a.oracle_out, {"source": "demo actions", "demo": demo})
print(json.dumps({"oracle": a.oracle_out, **{k: st[k] for k in ("success", "reason", "replayed_rows", "rows", "state_sha")}}))
diff --git a/tasks/dexjoco-pinch-tongs-i00-privileged/task.toml b/tasks/dexjoco-pinch-tongs-i00-privileged/task.toml
index 6cd4945349..37e261f3e0 100644
--- a/tasks/dexjoco-pinch-tongs-i00-privileged/task.toml
+++ b/tasks/dexjoco-pinch-tongs-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "dexjoco"
robot = "franka-allegro"
family = "pinch_tongs"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
expert_time_estimate_hours = 2.0
diff --git a/tasks/dexjoco-pinch-tongs-i00-privileged/tests/grade.py b/tasks/dexjoco-pinch-tongs-i00-privileged/tests/grade.py
index c754f999ad..3cfbab9c42 100644
--- a/tasks/dexjoco-pinch-tongs-i00-privileged/tests/grade.py
+++ b/tasks/dexjoco-pinch-tongs-i00-privileged/tests/grade.py
@@ -6,8 +6,10 @@ write Harbor rewards.
Writes /logs/verifier/reward.txt (1/0), reward.json (numeric fields only; Harbor rejects strings), a replay video
/logs/verifier/replay.mp4 (front camera + wrist camera) and replay.last.png. Replays `--repeat` times in this fresh
process; success requires every replay to end its episode in success (DexJoCo's own rule: the episode ends at the first
-success or task failure) AND all final-state hashes to agree (MuJoCo on the CPU is bit-deterministic), and the frozen
-instance rebuilt here from tests/instance.npz must match its recorded hash.
+success or task failure) AND all final-state hashes to agree (MuJoCo on the CPU is bit-deterministic). The frozen
+instance rebuilt here from tests/instance.npz must match its recorded hash; one that does not (the image no longer
+rebuilds the scene) is no verdict on the agent: `instance_mismatch`, `grader_error` and `infra_error` = 1, nothing
+replayed.
A file the agent got wrong is the agent's 0, `bad_trajectory` (not `grader_error`): not an npz, pickled or object
arrays, no `task` or `actions` entry, a wrong shape, NaN/inf, or more than rcb_dexjoco.MAX_TRAJECTORY_ROWS rows (the
@@ -119,6 +121,10 @@ def main() -> int:
print("too few CPUs: " + rcb_dexjoco.threads_message(threads), file=sys.stderr)
write(False, {"grader_error": 1, "infra_error": 1, "too_few_cpus": 1, "n_actions": n_rows, **threads, **cost}); return 0
s = Session(a.family, instance=a.instance)
+ if not s.instance_exact: # the image no longer rebuilds the frozen scene: not the agent's 0 (infra, no verdict)
+ print(f"instance mismatch: rebuilt state sha {s.state_sha()} != recorded {s.instance_sha}", file=sys.stderr)
+ write(False, {"grader_error": 1, "infra_error": 1, "instance_mismatch": 1, "instance_ok": 0,
+ "n_actions": n_rows, **threads, **cost}); return 0
if acts.shape[1] != s.action_dim:
write(False, {"bad_trajectory": 1, "n_actions": n_rows, **cost}); return 0
t0 = time.time()
@@ -130,10 +136,9 @@ def main() -> int:
write(False, {"bad_trajectory": 1, "n_actions": n_rows, **cost}); return 0
all_success = all(o["success"] for o in outs)
deterministic = len({o["state_sha"] for o in outs}) == 1
- instance_ok = bool(s.instance_exact)
- ok = all_success and deterministic and instance_ok
+ ok = all_success and deterministic
out = {"success": int(ok), "all_success": int(all_success), "deterministic": int(deterministic),
- "instance_ok": int(instance_ok), "n_actions": int(len(acts)), "sim_steps": int(outs[0]["steps"]),
+ "instance_ok": 1, "n_actions": int(len(acts)), "sim_steps": int(outs[0]["steps"]),
"task_failure": int(bool(outs[0]["reason"]) and not outs[0]["success"]),
"replay_wall_s": round(time.time() - t0, 1), "grader_error": 0, "video_rendered": 0, **threads, **cost}
write(ok, out) # the verdict is on disk before the video; the video only adds video_rendered
diff --git a/tasks/dexjoco-pinch-tongs-i00-privileged/tests/instance.npz b/tasks/dexjoco-pinch-tongs-i00-privileged/tests/instance.npz
index fb46289594..22b9c055b9 100644
Binary files a/tasks/dexjoco-pinch-tongs-i00-privileged/tests/instance.npz and b/tasks/dexjoco-pinch-tongs-i00-privileged/tests/instance.npz differ
diff --git a/tasks/dexjoco-pinch-tongs-i00-privileged/tests/rcb_dexjoco.py b/tasks/dexjoco-pinch-tongs-i00-privileged/tests/rcb_dexjoco.py
index f2e2db848f..8eb183fe2a 100644
--- a/tasks/dexjoco-pinch-tongs-i00-privileged/tests/rcb_dexjoco.py
+++ b/tasks/dexjoco-pinch-tongs-i00-privileged/tests/rcb_dexjoco.py
@@ -54,6 +54,7 @@ import argparse
import copy
import hashlib
import json
+import re
import sys
import time
from pathlib import Path
@@ -222,6 +223,21 @@ def water_aim_deg(model, data) -> float:
return float(np.degrees(np.arccos(np.clip(c, -1.0, 1.0))))
+# instance.npz is in the agent's container (/app/instance.npz), and a task's reference solution is the actions of the
+# recording its instance was restored from: its `meta` must not say where the scene came from (PROTOCOL §2.7, no pointer
+# to a solution where the agent can reach). save_instance refuses such keys, and values that name a recording.
+PROVENANCE_KEYS = ("source", "demo", "dataset", "episode", "recording", "reference", "solution", "zarr", "url")
+PROVENANCE_VALUE = re.compile(r"demo|episode|dataset|recording|replay\.zarr|huggingface|https?://", re.I)
+
+
+def check_instance_meta(meta: dict) -> None:
+ """ValueError if an instance's `meta` says where its scene came from (keys or values; see PROVENANCE_KEYS)."""
+ bad = [k for k, v in meta.items() if any(p in str(k).lower() for p in PROVENANCE_KEYS)
+ or PROVENANCE_VALUE.search(json.dumps(v))]
+ if bad:
+ raise ValueError(f"instance meta {bad} names where the scene came from: the agent reads instance.npz")
+
+
def _log(*a: Any) -> None:
print("[rcb_dexjoco]", *a, file=sys.stderr, flush=True)
@@ -349,6 +365,7 @@ class Session:
return h.hexdigest()[:16]
def save_instance(self, path: str | Path, meta: dict | None = None) -> None:
+ check_instance_meta(meta or {})
snap = self.snapshot()
np.savez_compressed(path, task=self.task, physics=snap["physics"], mocap_pos=snap["mocap_pos"],
**{"model__" + k: v for k, v in snap["model"].items()},
@@ -803,15 +820,14 @@ def cmd_make_instance(a) -> None:
restore_initial_state(s.env, a.task, s.cfg, state0)
s._snapshot = s.snapshot()
s._episode_reset()
- meta = {"source": "DexJoCo-Datasets-Raw demo, initial state restored with dexjoco.tasks.state_restorers",
- "demo": a.demo_name or os.path.basename(os.path.dirname(os.path.abspath(a.demo))), "dexjoco": "8d23b0f"}
- s.save_instance(a.out, meta)
- _log("instance", a.out, "sha", s.state_sha())
+ demo = a.demo_name or os.path.basename(os.path.dirname(os.path.abspath(a.demo)))
+ s.save_instance(a.out, {"dexjoco": "8d23b0f"}) # no provenance in the instance (check_instance_meta)
+ _log("instance", a.out, "sha", s.state_sha(), "from", demo)
if a.oracle_out:
A = np.asarray(root["data"]["action"], dtype=np.float64)
s2 = Session(a.task, instance=a.out)
st = s2.replay(A)
- s2.save_trajectory(A[:st["replayed_rows"]], a.oracle_out, {"source": "demo actions", "demo": meta["demo"]})
+ s2.save_trajectory(A[:st["replayed_rows"]], a.oracle_out, {"source": "demo actions", "demo": demo})
print(json.dumps({"oracle": a.oracle_out, **{k: st[k] for k in ("success", "reason", "replayed_rows", "rows", "state_sha")}}))
diff --git a/tasks/dexjoco-pinch-tongs-i00-standard/README.md b/tasks/dexjoco-pinch-tongs-i00-standard/README.md
index b0240b5c17..ad171027f7 100644
--- a/tasks/dexjoco-pinch-tongs-i00-standard/README.md
+++ b/tasks/dexjoco-pinch-tongs-i00-standard/README.md
@@ -72,6 +72,13 @@ Generated by `scripts/dexjoco/generate.py`, as its privileged twin.
returned once. A service that has given up on the episode (too few CPUs, or it could not resume after its
crashes) answers every request with 409 and the reason, and `Robot()` / `wait_ready()` raise `EpisodeOver` at
once; before, nothing answered and they waited out their timeout (up to an hour).
+- 2026-10-14 (review of board 0105, B-M3): `environment/sim/instance.npz` and `environment/sim/rcb_dexjoco.py`. The
+ frozen instance's `meta` entry no longer names the demonstration the scene was restored from (`source`, `demo`;
+ the reference solution is that demonstration's actions), and the tooling refuses to save an instance that does.
+ The sidecar's files are out of the agent's reach, so nothing reached the agent from this copy; it changed with
+ the privileged twin's, from the one committed instance. Every array of the scene is the same (instance sha
+ `ad2cfca40d427e20`): the scene, the pass rule and the grader are unchanged, so the task keeps its name and earlier
+ results stand.
- 2026-10-13 (owner, boards 0085 and 0087): the sim service's clock (`environment/sim/rcb_limited.py`,
`environment/sim/server.py`, the docstring of `environment/rcb_client.py`). `meta["time_s"]` and
`meta["time_left_s"]` count from the agent's first request to the service (any endpoint), and a restarted service
diff --git a/tasks/dexjoco-pinch-tongs-i00-standard/environment/sim/instance.npz b/tasks/dexjoco-pinch-tongs-i00-standard/environment/sim/instance.npz
index fb46289594..22b9c055b9 100644
Binary files a/tasks/dexjoco-pinch-tongs-i00-standard/environment/sim/instance.npz and b/tasks/dexjoco-pinch-tongs-i00-standard/environment/sim/instance.npz differ
diff --git a/tasks/dexjoco-pinch-tongs-i00-standard/environment/sim/rcb_dexjoco.py b/tasks/dexjoco-pinch-tongs-i00-standard/environment/sim/rcb_dexjoco.py
index f2e2db848f..8eb183fe2a 100644
--- a/tasks/dexjoco-pinch-tongs-i00-standard/environment/sim/rcb_dexjoco.py
+++ b/tasks/dexjoco-pinch-tongs-i00-standard/environment/sim/rcb_dexjoco.py
@@ -54,6 +54,7 @@ import argparse
import copy
import hashlib
import json
+import re
import sys
import time
from pathlib import Path
@@ -222,6 +223,21 @@ def water_aim_deg(model, data) -> float:
return float(np.degrees(np.arccos(np.clip(c, -1.0, 1.0))))
+# instance.npz is in the agent's container (/app/instance.npz), and a task's reference solution is the actions of the
+# recording its instance was restored from: its `meta` must not say where the scene came from (PROTOCOL §2.7, no pointer
+# to a solution where the agent can reach). save_instance refuses such keys, and values that name a recording.
+PROVENANCE_KEYS = ("source", "demo", "dataset", "episode", "recording", "reference", "solution", "zarr", "url")
+PROVENANCE_VALUE = re.compile(r"demo|episode|dataset|recording|replay\.zarr|huggingface|https?://", re.I)
+
+
+def check_instance_meta(meta: dict) -> None:
+ """ValueError if an instance's `meta` says where its scene came from (keys or values; see PROVENANCE_KEYS)."""
+ bad = [k for k, v in meta.items() if any(p in str(k).lower() for p in PROVENANCE_KEYS)
+ or PROVENANCE_VALUE.search(json.dumps(v))]
+ if bad:
+ raise ValueError(f"instance meta {bad} names where the scene came from: the agent reads instance.npz")
+
+
def _log(*a: Any) -> None:
print("[rcb_dexjoco]", *a, file=sys.stderr, flush=True)
@@ -349,6 +365,7 @@ class Session:
return h.hexdigest()[:16]
def save_instance(self, path: str | Path, meta: dict | None = None) -> None:
+ check_instance_meta(meta or {})
snap = self.snapshot()
np.savez_compressed(path, task=self.task, physics=snap["physics"], mocap_pos=snap["mocap_pos"],
**{"model__" + k: v for k, v in snap["model"].items()},
@@ -803,15 +820,14 @@ def cmd_make_instance(a) -> None:
restore_initial_state(s.env, a.task, s.cfg, state0)
s._snapshot = s.snapshot()
s._episode_reset()
- meta = {"source": "DexJoCo-Datasets-Raw demo, initial state restored with dexjoco.tasks.state_restorers",
- "demo": a.demo_name or os.path.basename(os.path.dirname(os.path.abspath(a.demo))), "dexjoco": "8d23b0f"}
- s.save_instance(a.out, meta)
- _log("instance", a.out, "sha", s.state_sha())
+ demo = a.demo_name or os.path.basename(os.path.dirname(os.path.abspath(a.demo)))
+ s.save_instance(a.out, {"dexjoco": "8d23b0f"}) # no provenance in the instance (check_instance_meta)
+ _log("instance", a.out, "sha", s.state_sha(), "from", demo)
if a.oracle_out:
A = np.asarray(root["data"]["action"], dtype=np.float64)
s2 = Session(a.task, instance=a.out)
st = s2.replay(A)
- s2.save_trajectory(A[:st["replayed_rows"]], a.oracle_out, {"source": "demo actions", "demo": meta["demo"]})
+ s2.save_trajectory(A[:st["replayed_rows"]], a.oracle_out, {"source": "demo actions", "demo": demo})
print(json.dumps({"oracle": a.oracle_out, **{k: st[k] for k in ("success", "reason", "replayed_rows", "rows", "state_sha")}}))
diff --git a/tasks/dexjoco-pinch-tongs-i00-standard/task.toml b/tasks/dexjoco-pinch-tongs-i00-standard/task.toml
index 77d76110d7..76f3c9d805 100644
--- a/tasks/dexjoco-pinch-tongs-i00-standard/task.toml
+++ b/tasks/dexjoco-pinch-tongs-i00-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "dexjoco"
robot = "franka-allegro"
family = "pinch_tongs"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "full"
expert_time_estimate_hours = 2.0
diff --git a/tasks/dexjoco-water-plant-i00-privileged/README.md b/tasks/dexjoco-water-plant-i00-privileged/README.md
index 41a9f409cf..c07ec37a88 100644
--- a/tasks/dexjoco-water-plant-i00-privileged/README.md
+++ b/tasks/dexjoco-water-plant-i00-privileged/README.md
@@ -29,8 +29,10 @@ limited to the model APIs.
## Verification
Open-loop replay of `/app/output/trajectory.npz` from `tests/instance.npz`, twice in a fresh process in the verifier's own
-sandbox (no network); pass iff both episodes end in success, their final-state hashes agree and the rebuilt instance
-matches its recorded hash (`tests/grade.py`); plus `/logs/verifier/replay.mp4`, rendered after the verdict is written.
+sandbox (no network); pass iff both episodes end in success and their final-state hashes agree (`tests/grade.py`); plus
+`/logs/verifier/replay.mp4`, rendered after the verdict is written. A rebuilt instance that does not match its recorded
+hash (the image no longer rebuilds the scene) replays nothing and gives no verdict: `instance_mismatch`, `grader_error`
+and `infra_error` = 1, not the agent's 0.
A file that is not a readable trajectory (not an npz, no `task` / `actions`, a wrong shape, NaN) or has more than
50,000 rows scores 0 as `bad_trajectory`: the cap is what two replays of the slowest family fit into the verifier's
3600 s. A verifier that runs fewer than 4 OpenBLAS threads (fewer than 4 CPUs visible to it) replays
@@ -65,3 +67,15 @@ real-time sleep, episode cap and hanoi's wall-clock auto-reset are disabled in t
### Provenance
Generated by `scripts/dexjoco/generate.py` from `scripts/dexjoco/instances/water_plant/` (`make_instances.py`).
+
+### Change Log
+- 2026-10-14 (review of board 0105, B-M3 and B6-3): `environment/instance.npz` and `tests/instance.npz` no longer name
+ the demonstration the scene was restored from: their `meta` entry had `source` (the dataset) and `demo` (the
+ demonstration's id), and the reference solution is that demonstration's own actions, so the agent could read at
+ `/app/instance.npz` where a solution was (PROTOCOL §2.7). Only `meta` changed: every array of the scene is the same
+ (instance sha `04f14553ced4b0ed`), so the scene, the pass rule and the reference are unchanged and the task keeps its
+ name. The tooling (`rcb_dexjoco.py`, both copies) refuses to save an instance whose `meta` says where its scene came
+ from. `tests/grade.py`: a rebuilt instance that does not match its recorded hash is now no verdict
+ (`instance_mismatch`, `grader_error`, `infra_error` = 1, nothing replayed); before, it was the agent's 0 with
+ `infra_error` = 0 and `instance_ok` = 0. Past results: the demonstration itself was out of reach under the egress
+ allowlist (Hugging Face is not on it); a past 0 with `instance_ok` = 0 was an infrastructure failure.
diff --git a/tasks/dexjoco-water-plant-i00-privileged/environment/instance.npz b/tasks/dexjoco-water-plant-i00-privileged/environment/instance.npz
index c71cd35bcb..fd7054e2c5 100644
Binary files a/tasks/dexjoco-water-plant-i00-privileged/environment/instance.npz and b/tasks/dexjoco-water-plant-i00-privileged/environment/instance.npz differ
diff --git a/tasks/dexjoco-water-plant-i00-privileged/environment/rcb_dexjoco.py b/tasks/dexjoco-water-plant-i00-privileged/environment/rcb_dexjoco.py
index f2e2db848f..8eb183fe2a 100644
--- a/tasks/dexjoco-water-plant-i00-privileged/environment/rcb_dexjoco.py
+++ b/tasks/dexjoco-water-plant-i00-privileged/environment/rcb_dexjoco.py
@@ -54,6 +54,7 @@ import argparse
import copy
import hashlib
import json
+import re
import sys
import time
from pathlib import Path
@@ -222,6 +223,21 @@ def water_aim_deg(model, data) -> float:
return float(np.degrees(np.arccos(np.clip(c, -1.0, 1.0))))
+# instance.npz is in the agent's container (/app/instance.npz), and a task's reference solution is the actions of the
+# recording its instance was restored from: its `meta` must not say where the scene came from (PROTOCOL §2.7, no pointer
+# to a solution where the agent can reach). save_instance refuses such keys, and values that name a recording.
+PROVENANCE_KEYS = ("source", "demo", "dataset", "episode", "recording", "reference", "solution", "zarr", "url")
+PROVENANCE_VALUE = re.compile(r"demo|episode|dataset|recording|replay\.zarr|huggingface|https?://", re.I)
+
+
+def check_instance_meta(meta: dict) -> None:
+ """ValueError if an instance's `meta` says where its scene came from (keys or values; see PROVENANCE_KEYS)."""
+ bad = [k for k, v in meta.items() if any(p in str(k).lower() for p in PROVENANCE_KEYS)
+ or PROVENANCE_VALUE.search(json.dumps(v))]
+ if bad:
+ raise ValueError(f"instance meta {bad} names where the scene came from: the agent reads instance.npz")
+
+
def _log(*a: Any) -> None:
print("[rcb_dexjoco]", *a, file=sys.stderr, flush=True)
@@ -349,6 +365,7 @@ class Session:
return h.hexdigest()[:16]
def save_instance(self, path: str | Path, meta: dict | None = None) -> None:
+ check_instance_meta(meta or {})
snap = self.snapshot()
np.savez_compressed(path, task=self.task, physics=snap["physics"], mocap_pos=snap["mocap_pos"],
**{"model__" + k: v for k, v in snap["model"].items()},
@@ -803,15 +820,14 @@ def cmd_make_instance(a) -> None:
restore_initial_state(s.env, a.task, s.cfg, state0)
s._snapshot = s.snapshot()
s._episode_reset()
- meta = {"source": "DexJoCo-Datasets-Raw demo, initial state restored with dexjoco.tasks.state_restorers",
- "demo": a.demo_name or os.path.basename(os.path.dirname(os.path.abspath(a.demo))), "dexjoco": "8d23b0f"}
- s.save_instance(a.out, meta)
- _log("instance", a.out, "sha", s.state_sha())
+ demo = a.demo_name or os.path.basename(os.path.dirname(os.path.abspath(a.demo)))
+ s.save_instance(a.out, {"dexjoco": "8d23b0f"}) # no provenance in the instance (check_instance_meta)
+ _log("instance", a.out, "sha", s.state_sha(), "from", demo)
if a.oracle_out:
A = np.asarray(root["data"]["action"], dtype=np.float64)
s2 = Session(a.task, instance=a.out)
st = s2.replay(A)
- s2.save_trajectory(A[:st["replayed_rows"]], a.oracle_out, {"source": "demo actions", "demo": meta["demo"]})
+ s2.save_trajectory(A[:st["replayed_rows"]], a.oracle_out, {"source": "demo actions", "demo": demo})
print(json.dumps({"oracle": a.oracle_out, **{k: st[k] for k in ("success", "reason", "replayed_rows", "rows", "state_sha")}}))
diff --git a/tasks/dexjoco-water-plant-i00-privileged/task.toml b/tasks/dexjoco-water-plant-i00-privileged/task.toml
index 335a96b48c..7f163c44d0 100644
--- a/tasks/dexjoco-water-plant-i00-privileged/task.toml
+++ b/tasks/dexjoco-water-plant-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "dexjoco"
robot = "franka-allegro"
family = "water_plant"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
expert_time_estimate_hours = 1.5
diff --git a/tasks/dexjoco-water-plant-i00-privileged/tests/grade.py b/tasks/dexjoco-water-plant-i00-privileged/tests/grade.py
index c754f999ad..3cfbab9c42 100644
--- a/tasks/dexjoco-water-plant-i00-privileged/tests/grade.py
+++ b/tasks/dexjoco-water-plant-i00-privileged/tests/grade.py
@@ -6,8 +6,10 @@ write Harbor rewards.
Writes /logs/verifier/reward.txt (1/0), reward.json (numeric fields only; Harbor rejects strings), a replay video
/logs/verifier/replay.mp4 (front camera + wrist camera) and replay.last.png. Replays `--repeat` times in this fresh
process; success requires every replay to end its episode in success (DexJoCo's own rule: the episode ends at the first
-success or task failure) AND all final-state hashes to agree (MuJoCo on the CPU is bit-deterministic), and the frozen
-instance rebuilt here from tests/instance.npz must match its recorded hash.
+success or task failure) AND all final-state hashes to agree (MuJoCo on the CPU is bit-deterministic). The frozen
+instance rebuilt here from tests/instance.npz must match its recorded hash; one that does not (the image no longer
+rebuilds the scene) is no verdict on the agent: `instance_mismatch`, `grader_error` and `infra_error` = 1, nothing
+replayed.
A file the agent got wrong is the agent's 0, `bad_trajectory` (not `grader_error`): not an npz, pickled or object
arrays, no `task` or `actions` entry, a wrong shape, NaN/inf, or more than rcb_dexjoco.MAX_TRAJECTORY_ROWS rows (the
@@ -119,6 +121,10 @@ def main() -> int:
print("too few CPUs: " + rcb_dexjoco.threads_message(threads), file=sys.stderr)
write(False, {"grader_error": 1, "infra_error": 1, "too_few_cpus": 1, "n_actions": n_rows, **threads, **cost}); return 0
s = Session(a.family, instance=a.instance)
+ if not s.instance_exact: # the image no longer rebuilds the frozen scene: not the agent's 0 (infra, no verdict)
+ print(f"instance mismatch: rebuilt state sha {s.state_sha()} != recorded {s.instance_sha}", file=sys.stderr)
+ write(False, {"grader_error": 1, "infra_error": 1, "instance_mismatch": 1, "instance_ok": 0,
+ "n_actions": n_rows, **threads, **cost}); return 0
if acts.shape[1] != s.action_dim:
write(False, {"bad_trajectory": 1, "n_actions": n_rows, **cost}); return 0
t0 = time.time()
@@ -130,10 +136,9 @@ def main() -> int:
write(False, {"bad_trajectory": 1, "n_actions": n_rows, **cost}); return 0
all_success = all(o["success"] for o in outs)
deterministic = len({o["state_sha"] for o in outs}) == 1
- instance_ok = bool(s.instance_exact)
- ok = all_success and deterministic and instance_ok
+ ok = all_success and deterministic
out = {"success": int(ok), "all_success": int(all_success), "deterministic": int(deterministic),
- "instance_ok": int(instance_ok), "n_actions": int(len(acts)), "sim_steps": int(outs[0]["steps"]),
+ "instance_ok": 1, "n_actions": int(len(acts)), "sim_steps": int(outs[0]["steps"]),
"task_failure": int(bool(outs[0]["reason"]) and not outs[0]["success"]),
"replay_wall_s": round(time.time() - t0, 1), "grader_error": 0, "video_rendered": 0, **threads, **cost}
write(ok, out) # the verdict is on disk before the video; the video only adds video_rendered
diff --git a/tasks/dexjoco-water-plant-i00-privileged/tests/instance.npz b/tasks/dexjoco-water-plant-i00-privileged/tests/instance.npz
index c71cd35bcb..fd7054e2c5 100644
Binary files a/tasks/dexjoco-water-plant-i00-privileged/tests/instance.npz and b/tasks/dexjoco-water-plant-i00-privileged/tests/instance.npz differ
diff --git a/tasks/dexjoco-water-plant-i00-privileged/tests/rcb_dexjoco.py b/tasks/dexjoco-water-plant-i00-privileged/tests/rcb_dexjoco.py
index f2e2db848f..8eb183fe2a 100644
--- a/tasks/dexjoco-water-plant-i00-privileged/tests/rcb_dexjoco.py
+++ b/tasks/dexjoco-water-plant-i00-privileged/tests/rcb_dexjoco.py
@@ -54,6 +54,7 @@ import argparse
import copy
import hashlib
import json
+import re
import sys
import time
from pathlib import Path
@@ -222,6 +223,21 @@ def water_aim_deg(model, data) -> float:
return float(np.degrees(np.arccos(np.clip(c, -1.0, 1.0))))
+# instance.npz is in the agent's container (/app/instance.npz), and a task's reference solution is the actions of the
+# recording its instance was restored from: its `meta` must not say where the scene came from (PROTOCOL §2.7, no pointer
+# to a solution where the agent can reach). save_instance refuses such keys, and values that name a recording.
+PROVENANCE_KEYS = ("source", "demo", "dataset", "episode", "recording", "reference", "solution", "zarr", "url")
+PROVENANCE_VALUE = re.compile(r"demo|episode|dataset|recording|replay\.zarr|huggingface|https?://", re.I)
+
+
+def check_instance_meta(meta: dict) -> None:
+ """ValueError if an instance's `meta` says where its scene came from (keys or values; see PROVENANCE_KEYS)."""
+ bad = [k for k, v in meta.items() if any(p in str(k).lower() for p in PROVENANCE_KEYS)
+ or PROVENANCE_VALUE.search(json.dumps(v))]
+ if bad:
+ raise ValueError(f"instance meta {bad} names where the scene came from: the agent reads instance.npz")
+
+
def _log(*a: Any) -> None:
print("[rcb_dexjoco]", *a, file=sys.stderr, flush=True)
@@ -349,6 +365,7 @@ class Session:
return h.hexdigest()[:16]
def save_instance(self, path: str | Path, meta: dict | None = None) -> None:
+ check_instance_meta(meta or {})
snap = self.snapshot()
np.savez_compressed(path, task=self.task, physics=snap["physics"], mocap_pos=snap["mocap_pos"],
**{"model__" + k: v for k, v in snap["model"].items()},
@@ -803,15 +820,14 @@ def cmd_make_instance(a) -> None:
restore_initial_state(s.env, a.task, s.cfg, state0)
s._snapshot = s.snapshot()
s._episode_reset()
- meta = {"source": "DexJoCo-Datasets-Raw demo, initial state restored with dexjoco.tasks.state_restorers",
- "demo": a.demo_name or os.path.basename(os.path.dirname(os.path.abspath(a.demo))), "dexjoco": "8d23b0f"}
- s.save_instance(a.out, meta)
- _log("instance", a.out, "sha", s.state_sha())
+ demo = a.demo_name or os.path.basename(os.path.dirname(os.path.abspath(a.demo)))
+ s.save_instance(a.out, {"dexjoco": "8d23b0f"}) # no provenance in the instance (check_instance_meta)
+ _log("instance", a.out, "sha", s.state_sha(), "from", demo)
if a.oracle_out:
A = np.asarray(root["data"]["action"], dtype=np.float64)
s2 = Session(a.task, instance=a.out)
st = s2.replay(A)
- s2.save_trajectory(A[:st["replayed_rows"]], a.oracle_out, {"source": "demo actions", "demo": meta["demo"]})
+ s2.save_trajectory(A[:st["replayed_rows"]], a.oracle_out, {"source": "demo actions", "demo": demo})
print(json.dumps({"oracle": a.oracle_out, **{k: st[k] for k in ("success", "reason", "replayed_rows", "rows", "state_sha")}}))
diff --git a/tasks/dexjoco-water-plant-i00-standard/README.md b/tasks/dexjoco-water-plant-i00-standard/README.md
index c453d59300..f5f56d90d5 100644
--- a/tasks/dexjoco-water-plant-i00-standard/README.md
+++ b/tasks/dexjoco-water-plant-i00-standard/README.md
@@ -72,6 +72,13 @@ Generated by `scripts/dexjoco/generate.py`, as its privileged twin.
returned once. A service that has given up on the episode (too few CPUs, or it could not resume after its
crashes) answers every request with 409 and the reason, and `Robot()` / `wait_ready()` raise `EpisodeOver` at
once; before, nothing answered and they waited out their timeout (up to an hour).
+- 2026-10-14 (review of board 0105, B-M3): `environment/sim/instance.npz` and `environment/sim/rcb_dexjoco.py`. The
+ frozen instance's `meta` entry no longer names the demonstration the scene was restored from (`source`, `demo`;
+ the reference solution is that demonstration's actions), and the tooling refuses to save an instance that does.
+ The sidecar's files are out of the agent's reach, so nothing reached the agent from this copy; it changed with
+ the privileged twin's, from the one committed instance. Every array of the scene is the same (instance sha
+ `04f14553ced4b0ed`): the scene, the pass rule and the grader are unchanged, so the task keeps its name and earlier
+ results stand.
- 2026-10-13 (owner, boards 0085 and 0087): the sim service's clock (`environment/sim/rcb_limited.py`,
`environment/sim/server.py`, the docstring of `environment/rcb_client.py`). `meta["time_s"]` and
`meta["time_left_s"]` count from the agent's first request to the service (any endpoint), and a restarted service
diff --git a/tasks/dexjoco-water-plant-i00-standard/environment/sim/instance.npz b/tasks/dexjoco-water-plant-i00-standard/environment/sim/instance.npz
index c71cd35bcb..fd7054e2c5 100644
Binary files a/tasks/dexjoco-water-plant-i00-standard/environment/sim/instance.npz and b/tasks/dexjoco-water-plant-i00-standard/environment/sim/instance.npz differ
diff --git a/tasks/dexjoco-water-plant-i00-standard/environment/sim/rcb_dexjoco.py b/tasks/dexjoco-water-plant-i00-standard/environment/sim/rcb_dexjoco.py
index f2e2db848f..8eb183fe2a 100644
--- a/tasks/dexjoco-water-plant-i00-standard/environment/sim/rcb_dexjoco.py
+++ b/tasks/dexjoco-water-plant-i00-standard/environment/sim/rcb_dexjoco.py
@@ -54,6 +54,7 @@ import argparse
import copy
import hashlib
import json
+import re
import sys
import time
from pathlib import Path
@@ -222,6 +223,21 @@ def water_aim_deg(model, data) -> float:
return float(np.degrees(np.arccos(np.clip(c, -1.0, 1.0))))
+# instance.npz is in the agent's container (/app/instance.npz), and a task's reference solution is the actions of the
+# recording its instance was restored from: its `meta` must not say where the scene came from (PROTOCOL §2.7, no pointer
+# to a solution where the agent can reach). save_instance refuses such keys, and values that name a recording.
+PROVENANCE_KEYS = ("source", "demo", "dataset", "episode", "recording", "reference", "solution", "zarr", "url")
+PROVENANCE_VALUE = re.compile(r"demo|episode|dataset|recording|replay\.zarr|huggingface|https?://", re.I)
+
+
+def check_instance_meta(meta: dict) -> None:
+ """ValueError if an instance's `meta` says where its scene came from (keys or values; see PROVENANCE_KEYS)."""
+ bad = [k for k, v in meta.items() if any(p in str(k).lower() for p in PROVENANCE_KEYS)
+ or PROVENANCE_VALUE.search(json.dumps(v))]
+ if bad:
+ raise ValueError(f"instance meta {bad} names where the scene came from: the agent reads instance.npz")
+
+
def _log(*a: Any) -> None:
print("[rcb_dexjoco]", *a, file=sys.stderr, flush=True)
@@ -349,6 +365,7 @@ class Session:
return h.hexdigest()[:16]
def save_instance(self, path: str | Path, meta: dict | None = None) -> None:
+ check_instance_meta(meta or {})
snap = self.snapshot()
np.savez_compressed(path, task=self.task, physics=snap["physics"], mocap_pos=snap["mocap_pos"],
**{"model__" + k: v for k, v in snap["model"].items()},
@@ -803,15 +820,14 @@ def cmd_make_instance(a) -> None:
restore_initial_state(s.env, a.task, s.cfg, state0)
s._snapshot = s.snapshot()
s._episode_reset()
- meta = {"source": "DexJoCo-Datasets-Raw demo, initial state restored with dexjoco.tasks.state_restorers",
- "demo": a.demo_name or os.path.basename(os.path.dirname(os.path.abspath(a.demo))), "dexjoco": "8d23b0f"}
- s.save_instance(a.out, meta)
- _log("instance", a.out, "sha", s.state_sha())
+ demo = a.demo_name or os.path.basename(os.path.dirname(os.path.abspath(a.demo)))
+ s.save_instance(a.out, {"dexjoco": "8d23b0f"}) # no provenance in the instance (check_instance_meta)
+ _log("instance", a.out, "sha", s.state_sha(), "from", demo)
if a.oracle_out:
A = np.asarray(root["data"]["action"], dtype=np.float64)
s2 = Session(a.task, instance=a.out)
st = s2.replay(A)
- s2.save_trajectory(A[:st["replayed_rows"]], a.oracle_out, {"source": "demo actions", "demo": meta["demo"]})
+ s2.save_trajectory(A[:st["replayed_rows"]], a.oracle_out, {"source": "demo actions", "demo": demo})
print(json.dumps({"oracle": a.oracle_out, **{k: st[k] for k in ("success", "reason", "replayed_rows", "rows", "state_sha")}}))
diff --git a/tasks/dexjoco-water-plant-i00-standard/task.toml b/tasks/dexjoco-water-plant-i00-standard/task.toml
index 4bfa706d70..b408fa53e1 100644
--- a/tasks/dexjoco-water-plant-i00-standard/task.toml
+++ b/tasks/dexjoco-water-plant-i00-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "dexjoco"
robot = "franka-allegro"
family = "water_plant"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "full"
expert_time_estimate_hours = 1.5
diff --git a/tasks/dextoolbench_brush_blue_brush_sweep_forward-privileged/task.toml b/tasks/dextoolbench_brush_blue_brush_sweep_forward-privileged/task.toml
index d98169c11a..724910e0a2 100644
--- a/tasks/dextoolbench_brush_blue_brush_sweep_forward-privileged/task.toml
+++ b/tasks/dextoolbench_brush_blue_brush_sweep_forward-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "mujoco"
robot = "kuka-iiwa14-sharpa-ha4"
family = "dextoolbench/brush/blue_brush/sweep_forward"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "partial"
oracle_evidence = "SimToolReal's pretrained policy, replayed in this scene, reaches 8/34 goals (solution/; README of dextoolbench_brush_blue_brush_sweep_forward-privileged)"
instance_seed = 0 # DexToolBench fixes the start pose and goals (from a human demo); no randomization
diff --git a/tasks/dextoolbench_brush_blue_brush_sweep_forward-standard/task.toml b/tasks/dextoolbench_brush_blue_brush_sweep_forward-standard/task.toml
index c364cc8cd1..7e1d2d2764 100644
--- a/tasks/dextoolbench_brush_blue_brush_sweep_forward-standard/task.toml
+++ b/tasks/dextoolbench_brush_blue_brush_sweep_forward-standard/task.toml
@@ -24,6 +24,7 @@ family = "dextoolbench/brush/blue_brush/sweep_forward"
instance_seed = 0 # DexToolBench fixes the start pose and goals (from a human demo); no randomization
expert_time_estimate_hours = 4.0
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "partial"
oracle_evidence = "SimToolReal's pretrained policy, replayed in this scene, reaches 8/34 goals (solution/; README of dextoolbench_brush_blue_brush_sweep_forward-privileged)"
diff --git a/tasks/dextoolbench_brush_red_brush_sweep_forward-privileged/task.toml b/tasks/dextoolbench_brush_red_brush_sweep_forward-privileged/task.toml
index 34a7aac529..ee3174c2e9 100644
--- a/tasks/dextoolbench_brush_red_brush_sweep_forward-privileged/task.toml
+++ b/tasks/dextoolbench_brush_red_brush_sweep_forward-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "mujoco"
robot = "kuka-iiwa14-sharpa-ha4"
family = "dextoolbench/brush/red_brush/sweep_forward"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "partial"
oracle_evidence = "SimToolReal's pretrained policy, replayed in this scene, reaches 20/38 goals (solution/; README of dextoolbench_brush_red_brush_sweep_forward-privileged)"
instance_seed = 0 # DexToolBench fixes the start pose and goals (from a human demo); no randomization
diff --git a/tasks/dextoolbench_brush_red_brush_sweep_forward-standard/task.toml b/tasks/dextoolbench_brush_red_brush_sweep_forward-standard/task.toml
index 1509be6e74..a056cd26c6 100644
--- a/tasks/dextoolbench_brush_red_brush_sweep_forward-standard/task.toml
+++ b/tasks/dextoolbench_brush_red_brush_sweep_forward-standard/task.toml
@@ -24,6 +24,7 @@ family = "dextoolbench/brush/red_brush/sweep_forward"
instance_seed = 0 # DexToolBench fixes the start pose and goals (from a human demo); no randomization
expert_time_estimate_hours = 4.0
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "partial"
oracle_evidence = "SimToolReal's pretrained policy, replayed in this scene, reaches 20/38 goals (solution/; README of dextoolbench_brush_red_brush_sweep_forward-privileged)"
diff --git a/tasks/dextoolbench_brush_red_brush_sweep_right-privileged/task.toml b/tasks/dextoolbench_brush_red_brush_sweep_right-privileged/task.toml
index 8605c6b3fa..5e80942d10 100644
--- a/tasks/dextoolbench_brush_red_brush_sweep_right-privileged/task.toml
+++ b/tasks/dextoolbench_brush_red_brush_sweep_right-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "mujoco"
robot = "kuka-iiwa14-sharpa-ha4"
family = "dextoolbench/brush/red_brush/sweep_right"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "partial"
oracle_evidence = "SimToolReal's pretrained policy, replayed in this scene, reaches 19/38 goals (solution/; README of dextoolbench_brush_red_brush_sweep_right-privileged)"
instance_seed = 0 # DexToolBench fixes the start pose and goals (from a human demo); no randomization
diff --git a/tasks/dextoolbench_brush_red_brush_sweep_right-standard/task.toml b/tasks/dextoolbench_brush_red_brush_sweep_right-standard/task.toml
index ed6db1c04a..181f8d8484 100644
--- a/tasks/dextoolbench_brush_red_brush_sweep_right-standard/task.toml
+++ b/tasks/dextoolbench_brush_red_brush_sweep_right-standard/task.toml
@@ -24,6 +24,7 @@ family = "dextoolbench/brush/red_brush/sweep_right"
instance_seed = 0 # DexToolBench fixes the start pose and goals (from a human demo); no randomization
expert_time_estimate_hours = 4.0
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "partial"
oracle_evidence = "SimToolReal's pretrained policy, replayed in this scene, reaches 19/38 goals (solution/; README of dextoolbench_brush_red_brush_sweep_right-privileged)"
diff --git a/tasks/dextoolbench_eraser_handle_eraser_wipe_c-privileged/task.toml b/tasks/dextoolbench_eraser_handle_eraser_wipe_c-privileged/task.toml
index 793d15c80a..0f76abe362 100644
--- a/tasks/dextoolbench_eraser_handle_eraser_wipe_c-privileged/task.toml
+++ b/tasks/dextoolbench_eraser_handle_eraser_wipe_c-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "mujoco"
robot = "kuka-iiwa14-sharpa-ha4"
family = "dextoolbench/eraser/handle_eraser/wipe_c"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "partial"
oracle_evidence = "SimToolReal's pretrained policy, replayed in this scene, reaches 12/31 goals (solution/; README of dextoolbench_eraser_handle_eraser_wipe_c-privileged)"
instance_seed = 0 # DexToolBench fixes the start pose and goals (from a human demo); no randomization
diff --git a/tasks/dextoolbench_eraser_handle_eraser_wipe_c-standard/task.toml b/tasks/dextoolbench_eraser_handle_eraser_wipe_c-standard/task.toml
index 6dcd81cc19..7b5a577521 100644
--- a/tasks/dextoolbench_eraser_handle_eraser_wipe_c-standard/task.toml
+++ b/tasks/dextoolbench_eraser_handle_eraser_wipe_c-standard/task.toml
@@ -24,6 +24,7 @@ family = "dextoolbench/eraser/handle_eraser/wipe_c"
instance_seed = 0 # DexToolBench fixes the start pose and goals (from a human demo); no randomization
expert_time_estimate_hours = 4.0
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "partial"
oracle_evidence = "SimToolReal's pretrained policy, replayed in this scene, reaches 12/31 goals (solution/; README of dextoolbench_eraser_handle_eraser_wipe_c-privileged)"
diff --git a/tasks/dextoolbench_eraser_handle_eraser_wipe_smile-privileged/task.toml b/tasks/dextoolbench_eraser_handle_eraser_wipe_smile-privileged/task.toml
index e51641b58e..fa5ccbb735 100644
--- a/tasks/dextoolbench_eraser_handle_eraser_wipe_smile-privileged/task.toml
+++ b/tasks/dextoolbench_eraser_handle_eraser_wipe_smile-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "mujoco"
robot = "kuka-iiwa14-sharpa-ha4"
family = "dextoolbench/eraser/handle_eraser/wipe_smile"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 0 # DexToolBench fixes the start pose and goals (from a human demo); no randomization
expert_time_estimate_hours = 4.0
diff --git a/tasks/dextoolbench_eraser_handle_eraser_wipe_smile-standard/task.toml b/tasks/dextoolbench_eraser_handle_eraser_wipe_smile-standard/task.toml
index 1320493094..68462f3c50 100644
--- a/tasks/dextoolbench_eraser_handle_eraser_wipe_smile-standard/task.toml
+++ b/tasks/dextoolbench_eraser_handle_eraser_wipe_smile-standard/task.toml
@@ -24,6 +24,7 @@ family = "dextoolbench/eraser/handle_eraser/wipe_smile"
instance_seed = 0 # DexToolBench fixes the start pose and goals (from a human demo); no randomization
expert_time_estimate_hours = 4.0
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
diff --git a/tasks/dextoolbench_hammer_claw_hammer_swing_down-privileged/task.toml b/tasks/dextoolbench_hammer_claw_hammer_swing_down-privileged/task.toml
index a6ce038f03..4b6e593e1d 100644
--- a/tasks/dextoolbench_hammer_claw_hammer_swing_down-privileged/task.toml
+++ b/tasks/dextoolbench_hammer_claw_hammer_swing_down-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "mujoco"
robot = "kuka-iiwa14-sharpa-ha4"
family = "dextoolbench/hammer/claw_hammer/swing_down"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 0 # DexToolBench fixes the start pose and goals (from a human demo); no randomization
expert_time_estimate_hours = 4.0
diff --git a/tasks/dextoolbench_hammer_claw_hammer_swing_down-standard/task.toml b/tasks/dextoolbench_hammer_claw_hammer_swing_down-standard/task.toml
index c73a7d0291..3dbd99ef4b 100644
--- a/tasks/dextoolbench_hammer_claw_hammer_swing_down-standard/task.toml
+++ b/tasks/dextoolbench_hammer_claw_hammer_swing_down-standard/task.toml
@@ -24,6 +24,7 @@ family = "dextoolbench/hammer/claw_hammer/swing_down"
instance_seed = 0 # DexToolBench fixes the start pose and goals (from a human demo); no randomization
expert_time_estimate_hours = 4.0
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
diff --git a/tasks/dextoolbench_hammer_claw_hammer_swing_side-privileged/task.toml b/tasks/dextoolbench_hammer_claw_hammer_swing_side-privileged/task.toml
index e91fb860a4..0bff70f4a0 100644
--- a/tasks/dextoolbench_hammer_claw_hammer_swing_side-privileged/task.toml
+++ b/tasks/dextoolbench_hammer_claw_hammer_swing_side-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "mujoco"
robot = "kuka-iiwa14-sharpa-ha4"
family = "dextoolbench/hammer/claw_hammer/swing_side"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "partial"
oracle_evidence = "SimToolReal's pretrained policy, replayed in this scene, reaches 13/40 goals (solution/; README of dextoolbench_hammer_claw_hammer_swing_side-privileged)"
instance_seed = 0 # DexToolBench fixes the start pose and goals (from a human demo); no randomization
diff --git a/tasks/dextoolbench_hammer_claw_hammer_swing_side-standard/task.toml b/tasks/dextoolbench_hammer_claw_hammer_swing_side-standard/task.toml
index 0b3ee0ab21..fcc68f0fd5 100644
--- a/tasks/dextoolbench_hammer_claw_hammer_swing_side-standard/task.toml
+++ b/tasks/dextoolbench_hammer_claw_hammer_swing_side-standard/task.toml
@@ -24,6 +24,7 @@ family = "dextoolbench/hammer/claw_hammer/swing_side"
instance_seed = 0 # DexToolBench fixes the start pose and goals (from a human demo); no randomization
expert_time_estimate_hours = 4.0
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "partial"
oracle_evidence = "SimToolReal's pretrained policy, replayed in this scene, reaches 13/40 goals (solution/; README of dextoolbench_hammer_claw_hammer_swing_side-privileged)"
diff --git a/tasks/dextoolbench_hammer_mallet_hammer_swing_down-privileged/task.toml b/tasks/dextoolbench_hammer_mallet_hammer_swing_down-privileged/task.toml
index 332408d669..427b46942c 100644
--- a/tasks/dextoolbench_hammer_mallet_hammer_swing_down-privileged/task.toml
+++ b/tasks/dextoolbench_hammer_mallet_hammer_swing_down-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "mujoco"
robot = "kuka-iiwa14-sharpa-ha4"
family = "dextoolbench/hammer/mallet_hammer/swing_down"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "partial"
oracle_evidence = "SimToolReal's pretrained policy, replayed in this scene, reaches 12/36 goals (solution/; README of dextoolbench_hammer_mallet_hammer_swing_down-privileged)"
instance_seed = 0 # DexToolBench fixes the start pose and goals (from a human demo); no randomization
diff --git a/tasks/dextoolbench_hammer_mallet_hammer_swing_down-standard/task.toml b/tasks/dextoolbench_hammer_mallet_hammer_swing_down-standard/task.toml
index ff78ca7924..1a960fe012 100644
--- a/tasks/dextoolbench_hammer_mallet_hammer_swing_down-standard/task.toml
+++ b/tasks/dextoolbench_hammer_mallet_hammer_swing_down-standard/task.toml
@@ -24,6 +24,7 @@ family = "dextoolbench/hammer/mallet_hammer/swing_down"
instance_seed = 0 # DexToolBench fixes the start pose and goals (from a human demo); no randomization
expert_time_estimate_hours = 4.0
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "partial"
oracle_evidence = "SimToolReal's pretrained policy, replayed in this scene, reaches 12/36 goals (solution/; README of dextoolbench_hammer_mallet_hammer_swing_down-privileged)"
diff --git a/tasks/dextoolbench_hammer_mallet_hammer_swing_side-privileged/task.toml b/tasks/dextoolbench_hammer_mallet_hammer_swing_side-privileged/task.toml
index 91c31bd1b2..c0ffe80ab3 100644
--- a/tasks/dextoolbench_hammer_mallet_hammer_swing_side-privileged/task.toml
+++ b/tasks/dextoolbench_hammer_mallet_hammer_swing_side-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "mujoco"
robot = "kuka-iiwa14-sharpa-ha4"
family = "dextoolbench/hammer/mallet_hammer/swing_side"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "partial"
oracle_evidence = "SimToolReal's pretrained policy, replayed in this scene, reaches 5/32 goals (solution/; README of dextoolbench_hammer_mallet_hammer_swing_side-privileged)"
instance_seed = 0 # DexToolBench fixes the start pose and goals (from a human demo); no randomization
diff --git a/tasks/dextoolbench_hammer_mallet_hammer_swing_side-standard/task.toml b/tasks/dextoolbench_hammer_mallet_hammer_swing_side-standard/task.toml
index eac7207e20..2c39025227 100644
--- a/tasks/dextoolbench_hammer_mallet_hammer_swing_side-standard/task.toml
+++ b/tasks/dextoolbench_hammer_mallet_hammer_swing_side-standard/task.toml
@@ -24,6 +24,7 @@ family = "dextoolbench/hammer/mallet_hammer/swing_side"
instance_seed = 0 # DexToolBench fixes the start pose and goals (from a human demo); no randomization
expert_time_estimate_hours = 4.0
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "partial"
oracle_evidence = "SimToolReal's pretrained policy, replayed in this scene, reaches 5/32 goals (solution/; README of dextoolbench_hammer_mallet_hammer_swing_side-privileged)"
diff --git a/tasks/dextoolbench_marker_sharpie_marker_draw_smile-privileged/task.toml b/tasks/dextoolbench_marker_sharpie_marker_draw_smile-privileged/task.toml
index bf7f62aafe..38d0f681fc 100644
--- a/tasks/dextoolbench_marker_sharpie_marker_draw_smile-privileged/task.toml
+++ b/tasks/dextoolbench_marker_sharpie_marker_draw_smile-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "mujoco"
robot = "kuka-iiwa14-sharpa-ha4"
family = "dextoolbench/marker/sharpie_marker/draw_smile"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 0 # DexToolBench fixes the start pose and goals (from a human demo); no randomization
expert_time_estimate_hours = 4.0
diff --git a/tasks/dextoolbench_marker_sharpie_marker_draw_smile-standard/task.toml b/tasks/dextoolbench_marker_sharpie_marker_draw_smile-standard/task.toml
index e1a8c428bd..d5c9613c9f 100644
--- a/tasks/dextoolbench_marker_sharpie_marker_draw_smile-standard/task.toml
+++ b/tasks/dextoolbench_marker_sharpie_marker_draw_smile-standard/task.toml
@@ -24,6 +24,7 @@ family = "dextoolbench/marker/sharpie_marker/draw_smile"
instance_seed = 0 # DexToolBench fixes the start pose and goals (from a human demo); no randomization
expert_time_estimate_hours = 4.0
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
diff --git a/tasks/dextoolbench_marker_sharpie_marker_write_c-privileged/task.toml b/tasks/dextoolbench_marker_sharpie_marker_write_c-privileged/task.toml
index df5da659d7..a7853c6002 100644
--- a/tasks/dextoolbench_marker_sharpie_marker_write_c-privileged/task.toml
+++ b/tasks/dextoolbench_marker_sharpie_marker_write_c-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "mujoco"
robot = "kuka-iiwa14-sharpa-ha4"
family = "dextoolbench/marker/sharpie_marker/write_c"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 0 # DexToolBench fixes the start pose and goals (from a human demo); no randomization
expert_time_estimate_hours = 4.0
diff --git a/tasks/dextoolbench_marker_sharpie_marker_write_c-standard/task.toml b/tasks/dextoolbench_marker_sharpie_marker_write_c-standard/task.toml
index 3bac5c3ed3..6c0715a16e 100644
--- a/tasks/dextoolbench_marker_sharpie_marker_write_c-standard/task.toml
+++ b/tasks/dextoolbench_marker_sharpie_marker_write_c-standard/task.toml
@@ -24,6 +24,7 @@ family = "dextoolbench/marker/sharpie_marker/write_c"
instance_seed = 0 # DexToolBench fixes the start pose and goals (from a human demo); no randomization
expert_time_estimate_hours = 4.0
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
diff --git a/tasks/dextoolbench_marker_staples_marker_draw_smile-privileged/task.toml b/tasks/dextoolbench_marker_staples_marker_draw_smile-privileged/task.toml
index 84cad5c5a4..b5d5d41259 100644
--- a/tasks/dextoolbench_marker_staples_marker_draw_smile-privileged/task.toml
+++ b/tasks/dextoolbench_marker_staples_marker_draw_smile-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "mujoco"
robot = "kuka-iiwa14-sharpa-ha4"
family = "dextoolbench/marker/staples_marker/draw_smile"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 0 # DexToolBench fixes the start pose and goals (from a human demo); no randomization
expert_time_estimate_hours = 4.0
diff --git a/tasks/dextoolbench_marker_staples_marker_draw_smile-standard/task.toml b/tasks/dextoolbench_marker_staples_marker_draw_smile-standard/task.toml
index a0b691cc4f..d05cd032cc 100644
--- a/tasks/dextoolbench_marker_staples_marker_draw_smile-standard/task.toml
+++ b/tasks/dextoolbench_marker_staples_marker_draw_smile-standard/task.toml
@@ -24,6 +24,7 @@ family = "dextoolbench/marker/staples_marker/draw_smile"
instance_seed = 0 # DexToolBench fixes the start pose and goals (from a human demo); no randomization
expert_time_estimate_hours = 4.0
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
diff --git a/tasks/dextoolbench_marker_staples_marker_write_c-privileged/task.toml b/tasks/dextoolbench_marker_staples_marker_write_c-privileged/task.toml
index 56f1aca770..4b76d8985f 100644
--- a/tasks/dextoolbench_marker_staples_marker_write_c-privileged/task.toml
+++ b/tasks/dextoolbench_marker_staples_marker_write_c-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "mujoco"
robot = "kuka-iiwa14-sharpa-ha4"
family = "dextoolbench/marker/staples_marker/write_c"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "partial"
oracle_evidence = "SimToolReal's pretrained policy, replayed in this scene, reaches 19/29 goals (solution/; README of dextoolbench_marker_staples_marker_write_c-privileged)"
instance_seed = 0 # DexToolBench fixes the start pose and goals (from a human demo); no randomization
diff --git a/tasks/dextoolbench_marker_staples_marker_write_c-standard/task.toml b/tasks/dextoolbench_marker_staples_marker_write_c-standard/task.toml
index fcc5f255d1..9a35086fe6 100644
--- a/tasks/dextoolbench_marker_staples_marker_write_c-standard/task.toml
+++ b/tasks/dextoolbench_marker_staples_marker_write_c-standard/task.toml
@@ -24,6 +24,7 @@ family = "dextoolbench/marker/staples_marker/write_c"
instance_seed = 0 # DexToolBench fixes the start pose and goals (from a human demo); no randomization
expert_time_estimate_hours = 4.0
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "partial"
oracle_evidence = "SimToolReal's pretrained policy, replayed in this scene, reaches 19/29 goals (solution/; README of dextoolbench_marker_staples_marker_write_c-privileged)"
diff --git a/tasks/dextoolbench_screwdriver_long_screwdriver_spin_horizontal-privileged/task.toml b/tasks/dextoolbench_screwdriver_long_screwdriver_spin_horizontal-privileged/task.toml
index e6ea2cd7de..8bf3802867 100644
--- a/tasks/dextoolbench_screwdriver_long_screwdriver_spin_horizontal-privileged/task.toml
+++ b/tasks/dextoolbench_screwdriver_long_screwdriver_spin_horizontal-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "mujoco"
robot = "kuka-iiwa14-sharpa-ha4"
family = "dextoolbench/screwdriver/long_screwdriver/spin_horizontal"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 0 # DexToolBench fixes the start pose and goals (from a human demo); no randomization
expert_time_estimate_hours = 4.0
diff --git a/tasks/dextoolbench_screwdriver_long_screwdriver_spin_horizontal-standard/task.toml b/tasks/dextoolbench_screwdriver_long_screwdriver_spin_horizontal-standard/task.toml
index 4b96f4625b..5bbb1b52d6 100644
--- a/tasks/dextoolbench_screwdriver_long_screwdriver_spin_horizontal-standard/task.toml
+++ b/tasks/dextoolbench_screwdriver_long_screwdriver_spin_horizontal-standard/task.toml
@@ -24,6 +24,7 @@ family = "dextoolbench/screwdriver/long_screwdriver/spin_horizontal"
instance_seed = 0 # DexToolBench fixes the start pose and goals (from a human demo); no randomization
expert_time_estimate_hours = 4.0
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
diff --git a/tasks/dextoolbench_screwdriver_long_screwdriver_spin_vertical-privileged/task.toml b/tasks/dextoolbench_screwdriver_long_screwdriver_spin_vertical-privileged/task.toml
index 5f0a06e6dc..0eeee084be 100644
--- a/tasks/dextoolbench_screwdriver_long_screwdriver_spin_vertical-privileged/task.toml
+++ b/tasks/dextoolbench_screwdriver_long_screwdriver_spin_vertical-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "mujoco"
robot = "kuka-iiwa14-sharpa-ha4"
family = "dextoolbench/screwdriver/long_screwdriver/spin_vertical"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "partial"
oracle_evidence = "SimToolReal's pretrained policy, replayed in this scene, reaches 27/100 goals (solution/; README of dextoolbench_screwdriver_long_screwdriver_spin_vertical-privileged)"
instance_seed = 0 # DexToolBench fixes the start pose and goals (from a human demo); no randomization
diff --git a/tasks/dextoolbench_screwdriver_long_screwdriver_spin_vertical-standard/task.toml b/tasks/dextoolbench_screwdriver_long_screwdriver_spin_vertical-standard/task.toml
index 82eff14704..734c1423ab 100644
--- a/tasks/dextoolbench_screwdriver_long_screwdriver_spin_vertical-standard/task.toml
+++ b/tasks/dextoolbench_screwdriver_long_screwdriver_spin_vertical-standard/task.toml
@@ -24,6 +24,7 @@ family = "dextoolbench/screwdriver/long_screwdriver/spin_vertical"
instance_seed = 0 # DexToolBench fixes the start pose and goals (from a human demo); no randomization
expert_time_estimate_hours = 4.0
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "partial"
oracle_evidence = "SimToolReal's pretrained policy, replayed in this scene, reaches 27/100 goals (solution/; README of dextoolbench_screwdriver_long_screwdriver_spin_vertical-privileged)"
diff --git a/tasks/dextoolbench_screwdriver_short_screwdriver_spin_horizontal-privileged/task.toml b/tasks/dextoolbench_screwdriver_short_screwdriver_spin_horizontal-privileged/task.toml
index 9a4317676e..8ecd8857a8 100644
--- a/tasks/dextoolbench_screwdriver_short_screwdriver_spin_horizontal-privileged/task.toml
+++ b/tasks/dextoolbench_screwdriver_short_screwdriver_spin_horizontal-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "mujoco"
robot = "kuka-iiwa14-sharpa-ha4"
family = "dextoolbench/screwdriver/short_screwdriver/spin_horizontal"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "partial"
oracle_evidence = "SimToolReal's pretrained policy, replayed in this scene, reaches 15/38 goals (solution/; README of dextoolbench_screwdriver_short_screwdriver_spin_horizontal-privileged)"
instance_seed = 0 # DexToolBench fixes the start pose and goals (from a human demo); no randomization
diff --git a/tasks/dextoolbench_screwdriver_short_screwdriver_spin_horizontal-standard/task.toml b/tasks/dextoolbench_screwdriver_short_screwdriver_spin_horizontal-standard/task.toml
index 6d6f594838..41b9011081 100644
--- a/tasks/dextoolbench_screwdriver_short_screwdriver_spin_horizontal-standard/task.toml
+++ b/tasks/dextoolbench_screwdriver_short_screwdriver_spin_horizontal-standard/task.toml
@@ -24,6 +24,7 @@ family = "dextoolbench/screwdriver/short_screwdriver/spin_horizontal"
instance_seed = 0 # DexToolBench fixes the start pose and goals (from a human demo); no randomization
expert_time_estimate_hours = 4.0
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "partial"
oracle_evidence = "SimToolReal's pretrained policy, replayed in this scene, reaches 15/38 goals (solution/; README of dextoolbench_screwdriver_short_screwdriver_spin_horizontal-privileged)"
diff --git a/tasks/dextoolbench_screwdriver_short_screwdriver_spin_vertical-privileged/task.toml b/tasks/dextoolbench_screwdriver_short_screwdriver_spin_vertical-privileged/task.toml
index ee4447f6ec..b6b5454f7b 100644
--- a/tasks/dextoolbench_screwdriver_short_screwdriver_spin_vertical-privileged/task.toml
+++ b/tasks/dextoolbench_screwdriver_short_screwdriver_spin_vertical-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "mujoco"
robot = "kuka-iiwa14-sharpa-ha4"
family = "dextoolbench/screwdriver/short_screwdriver/spin_vertical"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "partial"
oracle_evidence = "SimToolReal's pretrained policy, replayed in this scene, reaches 36/76 goals (solution/; README of dextoolbench_screwdriver_short_screwdriver_spin_vertical-privileged)"
instance_seed = 0 # DexToolBench fixes the start pose and goals (from a human demo); no randomization
diff --git a/tasks/dextoolbench_screwdriver_short_screwdriver_spin_vertical-standard/task.toml b/tasks/dextoolbench_screwdriver_short_screwdriver_spin_vertical-standard/task.toml
index 70c5ea2039..88b4a62592 100644
--- a/tasks/dextoolbench_screwdriver_short_screwdriver_spin_vertical-standard/task.toml
+++ b/tasks/dextoolbench_screwdriver_short_screwdriver_spin_vertical-standard/task.toml
@@ -24,6 +24,7 @@ family = "dextoolbench/screwdriver/short_screwdriver/spin_vertical"
instance_seed = 0 # DexToolBench fixes the start pose and goals (from a human demo); no randomization
expert_time_estimate_hours = 4.0
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "partial"
oracle_evidence = "SimToolReal's pretrained policy, replayed in this scene, reaches 36/76 goals (solution/; README of dextoolbench_screwdriver_short_screwdriver_spin_vertical-privileged)"
diff --git a/tasks/dextoolbench_spatula_flat_spatula_flip_over-privileged/task.toml b/tasks/dextoolbench_spatula_flat_spatula_flip_over-privileged/task.toml
index c54af43876..7f97b5d34e 100644
--- a/tasks/dextoolbench_spatula_flat_spatula_flip_over-privileged/task.toml
+++ b/tasks/dextoolbench_spatula_flat_spatula_flip_over-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "mujoco"
robot = "kuka-iiwa14-sharpa-ha4"
family = "dextoolbench/spatula/flat_spatula/flip_over"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "partial"
oracle_evidence = "SimToolReal's pretrained policy, replayed in this scene, reaches 7/45 goals (solution/; README of dextoolbench_spatula_flat_spatula_flip_over-privileged)"
instance_seed = 0 # DexToolBench fixes the start pose and goals (from a human demo); no randomization
diff --git a/tasks/dextoolbench_spatula_flat_spatula_flip_over-standard/task.toml b/tasks/dextoolbench_spatula_flat_spatula_flip_over-standard/task.toml
index dc65aeb681..8d6525b838 100644
--- a/tasks/dextoolbench_spatula_flat_spatula_flip_over-standard/task.toml
+++ b/tasks/dextoolbench_spatula_flat_spatula_flip_over-standard/task.toml
@@ -24,6 +24,7 @@ family = "dextoolbench/spatula/flat_spatula/flip_over"
instance_seed = 0 # DexToolBench fixes the start pose and goals (from a human demo); no randomization
expert_time_estimate_hours = 4.0
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "partial"
oracle_evidence = "SimToolReal's pretrained policy, replayed in this scene, reaches 7/45 goals (solution/; README of dextoolbench_spatula_flat_spatula_flip_over-privileged)"
diff --git a/tasks/dextoolbench_spatula_flat_spatula_serve_plate-privileged/task.toml b/tasks/dextoolbench_spatula_flat_spatula_serve_plate-privileged/task.toml
index aa37e7932f..9609128fca 100644
--- a/tasks/dextoolbench_spatula_flat_spatula_serve_plate-privileged/task.toml
+++ b/tasks/dextoolbench_spatula_flat_spatula_serve_plate-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "mujoco"
robot = "kuka-iiwa14-sharpa-ha4"
family = "dextoolbench/spatula/flat_spatula/serve_plate"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "partial"
oracle_evidence = "SimToolReal's pretrained policy, replayed in this scene, reaches 9/33 goals (solution/; README of dextoolbench_spatula_flat_spatula_serve_plate-privileged)"
instance_seed = 0 # DexToolBench fixes the start pose and goals (from a human demo); no randomization
diff --git a/tasks/dextoolbench_spatula_flat_spatula_serve_plate-standard/task.toml b/tasks/dextoolbench_spatula_flat_spatula_serve_plate-standard/task.toml
index 7db6d28019..a836dbc867 100644
--- a/tasks/dextoolbench_spatula_flat_spatula_serve_plate-standard/task.toml
+++ b/tasks/dextoolbench_spatula_flat_spatula_serve_plate-standard/task.toml
@@ -24,6 +24,7 @@ family = "dextoolbench/spatula/flat_spatula/serve_plate"
instance_seed = 0 # DexToolBench fixes the start pose and goals (from a human demo); no randomization
expert_time_estimate_hours = 4.0
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "partial"
oracle_evidence = "SimToolReal's pretrained policy, replayed in this scene, reaches 9/33 goals (solution/; README of dextoolbench_spatula_flat_spatula_serve_plate-privileged)"
diff --git a/tasks/dextoolbench_spatula_spoon_spatula_flip_over-privileged/task.toml b/tasks/dextoolbench_spatula_spoon_spatula_flip_over-privileged/task.toml
index ed64c702c7..7ba5140ab2 100644
--- a/tasks/dextoolbench_spatula_spoon_spatula_flip_over-privileged/task.toml
+++ b/tasks/dextoolbench_spatula_spoon_spatula_flip_over-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "mujoco"
robot = "kuka-iiwa14-sharpa-ha4"
family = "dextoolbench/spatula/spoon_spatula/flip_over"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "partial"
oracle_evidence = "SimToolReal's pretrained policy, replayed in this scene, reaches 6/40 goals (solution/; README of dextoolbench_spatula_spoon_spatula_flip_over-privileged)"
instance_seed = 0 # DexToolBench fixes the start pose and goals (from a human demo); no randomization
diff --git a/tasks/dextoolbench_spatula_spoon_spatula_flip_over-standard/task.toml b/tasks/dextoolbench_spatula_spoon_spatula_flip_over-standard/task.toml
index 0ffe6fa4ed..c39dedf1e6 100644
--- a/tasks/dextoolbench_spatula_spoon_spatula_flip_over-standard/task.toml
+++ b/tasks/dextoolbench_spatula_spoon_spatula_flip_over-standard/task.toml
@@ -24,6 +24,7 @@ family = "dextoolbench/spatula/spoon_spatula/flip_over"
instance_seed = 0 # DexToolBench fixes the start pose and goals (from a human demo); no randomization
expert_time_estimate_hours = 4.0
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "partial"
oracle_evidence = "SimToolReal's pretrained policy, replayed in this scene, reaches 6/40 goals (solution/; README of dextoolbench_spatula_spoon_spatula_flip_over-privileged)"
diff --git a/tasks/dextoolbench_spatula_spoon_spatula_serve_plate-privileged/task.toml b/tasks/dextoolbench_spatula_spoon_spatula_serve_plate-privileged/task.toml
index c0e14e6e84..61387c358f 100644
--- a/tasks/dextoolbench_spatula_spoon_spatula_serve_plate-privileged/task.toml
+++ b/tasks/dextoolbench_spatula_spoon_spatula_serve_plate-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "mujoco"
robot = "kuka-iiwa14-sharpa-ha4"
family = "dextoolbench/spatula/spoon_spatula/serve_plate"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "partial"
oracle_evidence = "SimToolReal's pretrained policy, replayed in this scene, reaches 13/48 goals (solution/; README of dextoolbench_spatula_spoon_spatula_serve_plate-privileged)"
instance_seed = 0 # DexToolBench fixes the start pose and goals (from a human demo); no randomization
diff --git a/tasks/dextoolbench_spatula_spoon_spatula_serve_plate-standard/task.toml b/tasks/dextoolbench_spatula_spoon_spatula_serve_plate-standard/task.toml
index de4c4d08a2..0272ad58ab 100644
--- a/tasks/dextoolbench_spatula_spoon_spatula_serve_plate-standard/task.toml
+++ b/tasks/dextoolbench_spatula_spoon_spatula_serve_plate-standard/task.toml
@@ -24,6 +24,7 @@ family = "dextoolbench/spatula/spoon_spatula/serve_plate"
instance_seed = 0 # DexToolBench fixes the start pose and goals (from a human demo); no randomization
expert_time_estimate_hours = 4.0
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "partial"
oracle_evidence = "SimToolReal's pretrained policy, replayed in this scene, reaches 13/48 goals (solution/; README of dextoolbench_spatula_spoon_spatula_serve_plate-privileged)"
diff --git a/tasks/humanoidbench-loco-balance-hard-i00-privileged/task.toml b/tasks/humanoidbench-loco-balance-hard-i00-privileged/task.toml
index e668cce58e..c697fc7dc7 100644
--- a/tasks/humanoidbench-loco-balance-hard-i00-privileged/task.toml
+++ b/tasks/humanoidbench-loco-balance-hard-i00-privileged/task.toml
@@ -17,6 +17,7 @@ robot = "unitree-h1"
# HumanoidBench env h1-balance_hard-v0; paper category locomotion; C3 · quasi-static stabilisation
family = "balance_hard"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
# scene = environment/instance.npz (frozen MuJoCo state); identity = the dataset.toml digest
instance_seed = 0
expert_time_estimate_hours = 4.0
diff --git a/tasks/humanoidbench-loco-balance-hard-i00-standard-realtime/README.md b/tasks/humanoidbench-loco-balance-hard-i00-standard-realtime/README.md
index 204664878d..a3089c39cf 100644
--- a/tasks/humanoidbench-loco-balance-hard-i00-standard-realtime/README.md
+++ b/tasks/humanoidbench-loco-balance-hard-i00-standard-realtime/README.md
@@ -9,6 +9,7 @@
| Family / instance | `balance_hard` / seed 0 |
| Oracle | `none` — none known: no run has passed it, and published RL reaches 31% of the bar (paper 57, FastTD3 244, bar 800) with upstream's mesh feet on MuJoCo 3.1.6 (box feet since image 0.2.0); human review in PR #83 |
| Agent time budget | 3600 s; tokens: protocol default (PROTOCOL §5) |
+| Real time | slowdown S = 1 (the world runs at 1/S of real speed: one control step every S x control_dt s of wall time; `RCB_SLOWDOWN`'s default); reset budget 50 (`RCB_RESET_BUDGET`'s default) (PROTOCOL §4b) |
| Base image | `ghcr.io/mll-lab-nu/eai-humanoidbench:0.2.0` |
## Task
diff --git a/tasks/humanoidbench-loco-balance-hard-i00-standard-realtime/task.toml b/tasks/humanoidbench-loco-balance-hard-i00-standard-realtime/task.toml
index 9c88f5c8d9..a25e396ad8 100644
--- a/tasks/humanoidbench-loco-balance-hard-i00-standard-realtime/task.toml
+++ b/tasks/humanoidbench-loco-balance-hard-i00-standard-realtime/task.toml
@@ -28,6 +28,8 @@ simulator = "humanoidbench"
robot = "unitree-h1"
family = "balance_hard"
mode = "standard-realtime"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
+slowdown = 1 # standard-realtime: the world runs at 1/S of real speed (docs/PROTOCOL.md §4b); RCB_SLOWDOWN's default
protocol = "eai-standard/2.3"
instance_seed = 0
expert_time_estimate_hours = 8.0
diff --git a/tasks/humanoidbench-loco-balance-hard-i00-standard/task.toml b/tasks/humanoidbench-loco-balance-hard-i00-standard/task.toml
index 9b2bc77dbd..6d92e36852 100644
--- a/tasks/humanoidbench-loco-balance-hard-i00-standard/task.toml
+++ b/tasks/humanoidbench-loco-balance-hard-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "humanoidbench"
robot = "unitree-h1"
family = "balance_hard"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
instance_seed = 0
expert_time_estimate_hours = 8.0
diff --git a/tasks/humanoidbench-loco-hurdle-i00-privileged/task.toml b/tasks/humanoidbench-loco-hurdle-i00-privileged/task.toml
index ab15a7f210..b694293354 100644
--- a/tasks/humanoidbench-loco-hurdle-i00-privileged/task.toml
+++ b/tasks/humanoidbench-loco-hurdle-i00-privileged/task.toml
@@ -17,6 +17,7 @@ robot = "unitree-h1"
# HumanoidBench env h1-hurdle-v0; paper category locomotion; C2 · constrained gait
family = "hurdle"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
# scene = environment/instance.npz (frozen MuJoCo state); identity = the dataset.toml digest
instance_seed = 0
expert_time_estimate_hours = 3.0
diff --git a/tasks/humanoidbench-loco-hurdle-i00-standard-realtime/README.md b/tasks/humanoidbench-loco-hurdle-i00-standard-realtime/README.md
index cdaeedba6e..5f6cccf3b1 100644
--- a/tasks/humanoidbench-loco-hurdle-i00-standard-realtime/README.md
+++ b/tasks/humanoidbench-loco-hurdle-i00-standard-realtime/README.md
@@ -9,6 +9,7 @@
| Family / instance | `hurdle` / seed 0 |
| Oracle | `none` — no rollout of this instance; published RL passes the bar (FastTD3 887 / 700); human review in PR #83 |
| Agent time budget | 3600 s; tokens: protocol default (PROTOCOL §5) |
+| Real time | slowdown S = 1 (the world runs at 1/S of real speed: one control step every S x control_dt s of wall time; `RCB_SLOWDOWN`'s default); reset budget 50 (`RCB_RESET_BUDGET`'s default) (PROTOCOL §4b) |
| Base image | `ghcr.io/mll-lab-nu/eai-humanoidbench:0.2.0` |
## Task
diff --git a/tasks/humanoidbench-loco-hurdle-i00-standard-realtime/task.toml b/tasks/humanoidbench-loco-hurdle-i00-standard-realtime/task.toml
index 26c9522d77..eeb0d2dd7f 100644
--- a/tasks/humanoidbench-loco-hurdle-i00-standard-realtime/task.toml
+++ b/tasks/humanoidbench-loco-hurdle-i00-standard-realtime/task.toml
@@ -28,6 +28,8 @@ simulator = "humanoidbench"
robot = "unitree-h1"
family = "hurdle"
mode = "standard-realtime"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
+slowdown = 1 # standard-realtime: the world runs at 1/S of real speed (docs/PROTOCOL.md §4b); RCB_SLOWDOWN's default
protocol = "eai-standard/2.3"
instance_seed = 0
expert_time_estimate_hours = 6.0
diff --git a/tasks/humanoidbench-loco-hurdle-i00-standard/task.toml b/tasks/humanoidbench-loco-hurdle-i00-standard/task.toml
index 02209e3490..594c9c88b1 100644
--- a/tasks/humanoidbench-loco-hurdle-i00-standard/task.toml
+++ b/tasks/humanoidbench-loco-hurdle-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "humanoidbench"
robot = "unitree-h1"
family = "hurdle"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
instance_seed = 0
expert_time_estimate_hours = 6.0
diff --git a/tasks/humanoidbench-loco-sit-simple-i00-privileged/task.toml b/tasks/humanoidbench-loco-sit-simple-i00-privileged/task.toml
index e5f5b14194..67821a4182 100644
--- a/tasks/humanoidbench-loco-sit-simple-i00-privileged/task.toml
+++ b/tasks/humanoidbench-loco-sit-simple-i00-privileged/task.toml
@@ -17,6 +17,7 @@ robot = "unitree-h1"
# HumanoidBench env h1-sit_simple-v0; paper category locomotion; fixture (C3 · quasi-static stabilisation, too easy to score)
family = "sit_simple"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
# scene = environment/instance.npz (frozen MuJoCo state); identity = the dataset.toml digest
instance_seed = 0
expert_time_estimate_hours = 1.0
diff --git a/tasks/humanoidbench-loco-sit-simple-i00-standard-realtime/README.md b/tasks/humanoidbench-loco-sit-simple-i00-standard-realtime/README.md
index 744b0e5024..ebed350594 100644
--- a/tasks/humanoidbench-loco-sit-simple-i00-standard-realtime/README.md
+++ b/tasks/humanoidbench-loco-sit-simple-i00-standard-realtime/README.md
@@ -9,6 +9,7 @@
| Family / instance | `sit_simple` / seed 0 |
| Oracle | `full` |
| Agent time budget | 3600 s; tokens: protocol default (PROTOCOL §5) |
+| Real time | slowdown S = 1 (the world runs at 1/S of real speed: one control step every S x control_dt s of wall time; `RCB_SLOWDOWN`'s default); reset budget 50 (`RCB_RESET_BUDGET`'s default) (PROTOCOL §4b) |
| Base image | `ghcr.io/mll-lab-nu/eai-humanoidbench:0.2.0` |
## Task
diff --git a/tasks/humanoidbench-loco-sit-simple-i00-standard-realtime/task.toml b/tasks/humanoidbench-loco-sit-simple-i00-standard-realtime/task.toml
index c38a84154d..bce71cecf7 100644
--- a/tasks/humanoidbench-loco-sit-simple-i00-standard-realtime/task.toml
+++ b/tasks/humanoidbench-loco-sit-simple-i00-standard-realtime/task.toml
@@ -28,6 +28,8 @@ simulator = "humanoidbench"
robot = "unitree-h1"
family = "sit_simple"
mode = "standard-realtime"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
+slowdown = 1 # standard-realtime: the world runs at 1/S of real speed (docs/PROTOCOL.md §4b); RCB_SLOWDOWN's default
protocol = "eai-standard/2.3"
instance_seed = 0
expert_time_estimate_hours = 2.0
diff --git a/tasks/humanoidbench-loco-sit-simple-i00-standard/task.toml b/tasks/humanoidbench-loco-sit-simple-i00-standard/task.toml
index a8f9567d76..e0087672ad 100644
--- a/tasks/humanoidbench-loco-sit-simple-i00-standard/task.toml
+++ b/tasks/humanoidbench-loco-sit-simple-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "humanoidbench"
robot = "unitree-h1"
family = "sit_simple"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
instance_seed = 0
expert_time_estimate_hours = 2.0
diff --git a/tasks/humanoidbench-loco-walk-i00-privileged/task.toml b/tasks/humanoidbench-loco-walk-i00-privileged/task.toml
index b80353e3d3..bd7dc59f3c 100644
--- a/tasks/humanoidbench-loco-walk-i00-privileged/task.toml
+++ b/tasks/humanoidbench-loco-walk-i00-privileged/task.toml
@@ -17,6 +17,7 @@ robot = "unitree-h1"
# HumanoidBench env h1-walk-v0; paper category locomotion; C1 · flat-ground periodic gait
family = "walk"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
# scene = environment/instance.npz (frozen MuJoCo state); identity = the dataset.toml digest
instance_seed = 0
expert_time_estimate_hours = 2.0
diff --git a/tasks/humanoidbench-loco-walk-i00-standard-realtime/README.md b/tasks/humanoidbench-loco-walk-i00-standard-realtime/README.md
index f9681a9638..fbbef327b2 100644
--- a/tasks/humanoidbench-loco-walk-i00-standard-realtime/README.md
+++ b/tasks/humanoidbench-loco-walk-i00-standard-realtime/README.md
@@ -9,6 +9,7 @@
| Family / instance | `walk` / seed 0 |
| Oracle | `none` — Claude Opus 5.5's privileged-mode trajectory (834.80 / 700 on image 0.1.6, MuJoCo 3.1.6), run `claude_code-claude_opus_5_5-medium` on the results site; it is open-loop, so it does not carry over to image 0.2.0's physics (51.46 / 700 there); published RL passes as well (SAC 854, FastTD3 928, on upstream's MuJoCo 3.1.6); in realtime, the same rows planned into the episode from step 0 (submit's `at`) run at the same control steps and are graded the same; human review in PR #83 |
| Agent time budget | 3600 s; tokens: protocol default (PROTOCOL §5) |
+| Real time | slowdown S = 1 (the world runs at 1/S of real speed: one control step every S x control_dt s of wall time; `RCB_SLOWDOWN`'s default); reset budget 50 (`RCB_RESET_BUDGET`'s default) (PROTOCOL §4b) |
| Base image | `ghcr.io/mll-lab-nu/eai-humanoidbench:0.2.0` |
## Task
diff --git a/tasks/humanoidbench-loco-walk-i00-standard-realtime/task.toml b/tasks/humanoidbench-loco-walk-i00-standard-realtime/task.toml
index 45a1050531..50a7db6311 100644
--- a/tasks/humanoidbench-loco-walk-i00-standard-realtime/task.toml
+++ b/tasks/humanoidbench-loco-walk-i00-standard-realtime/task.toml
@@ -28,6 +28,8 @@ simulator = "humanoidbench"
robot = "unitree-h1"
family = "walk"
mode = "standard-realtime"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
+slowdown = 1 # standard-realtime: the world runs at 1/S of real speed (docs/PROTOCOL.md §4b); RCB_SLOWDOWN's default
protocol = "eai-standard/2.3"
instance_seed = 0
expert_time_estimate_hours = 4.0
diff --git a/tasks/humanoidbench-loco-walk-i00-standard/task.toml b/tasks/humanoidbench-loco-walk-i00-standard/task.toml
index 14fab6bc65..98d7510f64 100644
--- a/tasks/humanoidbench-loco-walk-i00-standard/task.toml
+++ b/tasks/humanoidbench-loco-walk-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "humanoidbench"
robot = "unitree-h1"
family = "walk"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
instance_seed = 0
expert_time_estimate_hours = 4.0
diff --git a/tasks/humanoidbench-manip-bookshelf-simple-i00-privileged/task.toml b/tasks/humanoidbench-manip-bookshelf-simple-i00-privileged/task.toml
index ef3df19e8d..f6da347b11 100644
--- a/tasks/humanoidbench-manip-bookshelf-simple-i00-privileged/task.toml
+++ b/tasks/humanoidbench-manip-bookshelf-simple-i00-privileged/task.toml
@@ -17,6 +17,7 @@ robot = "unitree-h1hand"
# HumanoidBench env h1hand-bookshelf_simple-v0; paper category manipulation; C9 · long-horizon multi-subtask
family = "bookshelf_simple"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
# scene = environment/instance.npz (frozen MuJoCo state); identity = the dataset.toml digest
instance_seed = 0
expert_time_estimate_hours = 5.0
diff --git a/tasks/humanoidbench-manip-bookshelf-simple-i00-standard-realtime/README.md b/tasks/humanoidbench-manip-bookshelf-simple-i00-standard-realtime/README.md
index ee1e0b9469..6aa2e9a955 100644
--- a/tasks/humanoidbench-manip-bookshelf-simple-i00-standard-realtime/README.md
+++ b/tasks/humanoidbench-manip-bookshelf-simple-i00-standard-realtime/README.md
@@ -9,6 +9,7 @@
| Family / instance | `bookshelf_simple` / seed 0 |
| Oracle | `none` — none known: no run has passed it, and published RL reaches 41% of the bar (paper 763, FastTD3 822, bar 2000); human review in PR #83 |
| Agent time budget | 3600 s; tokens: protocol default (PROTOCOL §5) |
+| Real time | slowdown S = 1 (the world runs at 1/S of real speed: one control step every S x control_dt s of wall time; `RCB_SLOWDOWN`'s default); reset budget 50 (`RCB_RESET_BUDGET`'s default) (PROTOCOL §4b) |
| Base image | `ghcr.io/mll-lab-nu/eai-humanoidbench:0.2.0` |
## Task
diff --git a/tasks/humanoidbench-manip-bookshelf-simple-i00-standard-realtime/task.toml b/tasks/humanoidbench-manip-bookshelf-simple-i00-standard-realtime/task.toml
index 3ea042b5aa..e98c4d3637 100644
--- a/tasks/humanoidbench-manip-bookshelf-simple-i00-standard-realtime/task.toml
+++ b/tasks/humanoidbench-manip-bookshelf-simple-i00-standard-realtime/task.toml
@@ -28,6 +28,8 @@ simulator = "humanoidbench"
robot = "unitree-h1hand"
family = "bookshelf_simple"
mode = "standard-realtime"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
+slowdown = 1 # standard-realtime: the world runs at 1/S of real speed (docs/PROTOCOL.md §4b); RCB_SLOWDOWN's default
protocol = "eai-standard/2.3"
instance_seed = 0
expert_time_estimate_hours = 10.0
diff --git a/tasks/humanoidbench-manip-bookshelf-simple-i00-standard/task.toml b/tasks/humanoidbench-manip-bookshelf-simple-i00-standard/task.toml
index d07923a293..32c82e9cce 100644
--- a/tasks/humanoidbench-manip-bookshelf-simple-i00-standard/task.toml
+++ b/tasks/humanoidbench-manip-bookshelf-simple-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "humanoidbench"
robot = "unitree-h1hand"
family = "bookshelf_simple"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
instance_seed = 0
expert_time_estimate_hours = 10.0
diff --git a/tasks/humanoidbench-manip-cube-i00-privileged/task.toml b/tasks/humanoidbench-manip-cube-i00-privileged/task.toml
index f2d37af567..d38e074a0f 100644
--- a/tasks/humanoidbench-manip-cube-i00-privileged/task.toml
+++ b/tasks/humanoidbench-manip-cube-i00-privileged/task.toml
@@ -17,6 +17,7 @@ robot = "unitree-h1hand"
# HumanoidBench env h1hand-cube-v0; paper category manipulation; C7 · in-hand precision
family = "cube"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
# scene = environment/instance.npz (frozen MuJoCo state); identity = the dataset.toml digest
instance_seed = 0
expert_time_estimate_hours = 4.0
diff --git a/tasks/humanoidbench-manip-cube-i00-standard-realtime/README.md b/tasks/humanoidbench-manip-cube-i00-standard-realtime/README.md
index ca04430db0..32284c0bfb 100644
--- a/tasks/humanoidbench-manip-cube-i00-standard-realtime/README.md
+++ b/tasks/humanoidbench-manip-cube-i00-standard-realtime/README.md
@@ -9,6 +9,7 @@
| Family / instance | `cube` / seed 0 |
| Oracle | `none` — none known: no run has passed it, and published RL reaches 60% of the bar (paper 135, FastTD3 221, bar 370); human review in PR #83 |
| Agent time budget | 3600 s; tokens: protocol default (PROTOCOL §5) |
+| Real time | slowdown S = 1 (the world runs at 1/S of real speed: one control step every S x control_dt s of wall time; `RCB_SLOWDOWN`'s default); reset budget 50 (`RCB_RESET_BUDGET`'s default) (PROTOCOL §4b) |
| Base image | `ghcr.io/mll-lab-nu/eai-humanoidbench:0.2.0` |
## Task
diff --git a/tasks/humanoidbench-manip-cube-i00-standard-realtime/task.toml b/tasks/humanoidbench-manip-cube-i00-standard-realtime/task.toml
index f088b11ec5..2c350cf7d6 100644
--- a/tasks/humanoidbench-manip-cube-i00-standard-realtime/task.toml
+++ b/tasks/humanoidbench-manip-cube-i00-standard-realtime/task.toml
@@ -28,6 +28,8 @@ simulator = "humanoidbench"
robot = "unitree-h1hand"
family = "cube"
mode = "standard-realtime"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
+slowdown = 1 # standard-realtime: the world runs at 1/S of real speed (docs/PROTOCOL.md §4b); RCB_SLOWDOWN's default
protocol = "eai-standard/2.3"
instance_seed = 0
expert_time_estimate_hours = 8.0
diff --git a/tasks/humanoidbench-manip-cube-i00-standard/task.toml b/tasks/humanoidbench-manip-cube-i00-standard/task.toml
index 1e17c3c116..e345fcf48c 100644
--- a/tasks/humanoidbench-manip-cube-i00-standard/task.toml
+++ b/tasks/humanoidbench-manip-cube-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "humanoidbench"
robot = "unitree-h1hand"
family = "cube"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
instance_seed = 0
expert_time_estimate_hours = 8.0
diff --git a/tasks/humanoidbench-manip-door-i00-privileged/task.toml b/tasks/humanoidbench-manip-door-i00-privileged/task.toml
index ab07243d7a..8f0d4f6ab5 100644
--- a/tasks/humanoidbench-manip-door-i00-privileged/task.toml
+++ b/tasks/humanoidbench-manip-door-i00-privileged/task.toml
@@ -17,6 +17,7 @@ robot = "unitree-h1"
# HumanoidBench env h1-door-v0; paper category manipulation; C4 · navigation and passage
family = "door"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
# scene = environment/instance.npz (frozen MuJoCo state); identity = the dataset.toml digest
instance_seed = 0
expert_time_estimate_hours = 3.0
diff --git a/tasks/humanoidbench-manip-door-i00-standard-realtime/README.md b/tasks/humanoidbench-manip-door-i00-standard-realtime/README.md
index c82963cae0..8c0871b1fe 100644
--- a/tasks/humanoidbench-manip-door-i00-standard-realtime/README.md
+++ b/tasks/humanoidbench-manip-door-i00-standard-realtime/README.md
@@ -9,6 +9,7 @@
| Family / instance | `door` / seed 0 |
| Oracle | `none` — Claude Opus 5.5's privileged-mode trajectory (676.02 / 600 on image 0.1.6, MuJoCo 3.1.6), run `claude_code-claude_opus_5_5-medium` on the results site; it is open-loop, so it does not carry over to image 0.2.0's physics (9.24 / 600 there); in realtime, the same rows planned into the episode from step 0 (submit's `at`) run at the same control steps and are graded the same; human review in PR #83 |
| Agent time budget | 3600 s; tokens: protocol default (PROTOCOL §5) |
+| Real time | slowdown S = 1 (the world runs at 1/S of real speed: one control step every S x control_dt s of wall time; `RCB_SLOWDOWN`'s default); reset budget 50 (`RCB_RESET_BUDGET`'s default) (PROTOCOL §4b) |
| Base image | `ghcr.io/mll-lab-nu/eai-humanoidbench:0.2.0` |
## Task
diff --git a/tasks/humanoidbench-manip-door-i00-standard-realtime/task.toml b/tasks/humanoidbench-manip-door-i00-standard-realtime/task.toml
index 5b02cb57a5..41e110e8e1 100644
--- a/tasks/humanoidbench-manip-door-i00-standard-realtime/task.toml
+++ b/tasks/humanoidbench-manip-door-i00-standard-realtime/task.toml
@@ -28,6 +28,8 @@ simulator = "humanoidbench"
robot = "unitree-h1"
family = "door"
mode = "standard-realtime"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
+slowdown = 1 # standard-realtime: the world runs at 1/S of real speed (docs/PROTOCOL.md §4b); RCB_SLOWDOWN's default
protocol = "eai-standard/2.3"
instance_seed = 0
expert_time_estimate_hours = 6.0
diff --git a/tasks/humanoidbench-manip-door-i00-standard/task.toml b/tasks/humanoidbench-manip-door-i00-standard/task.toml
index e5390dac85..8c83909150 100644
--- a/tasks/humanoidbench-manip-door-i00-standard/task.toml
+++ b/tasks/humanoidbench-manip-door-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "humanoidbench"
robot = "unitree-h1"
family = "door"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
instance_seed = 0
expert_time_estimate_hours = 6.0
diff --git a/tasks/humanoidbench-manip-powerlift-i00-privileged/task.toml b/tasks/humanoidbench-manip-powerlift-i00-privileged/task.toml
index 1207db4904..981ea0bddd 100644
--- a/tasks/humanoidbench-manip-powerlift-i00-privileged/task.toml
+++ b/tasks/humanoidbench-manip-powerlift-i00-privileged/task.toml
@@ -17,6 +17,7 @@ robot = "unitree-h1hand"
# HumanoidBench env h1hand-powerlift-v0; paper category manipulation; C5 · whole-body power and momentum
family = "powerlift"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
# scene = environment/instance.npz (frozen MuJoCo state); identity = the dataset.toml digest
instance_seed = 0
expert_time_estimate_hours = 4.0
diff --git a/tasks/humanoidbench-manip-powerlift-i00-standard-realtime/README.md b/tasks/humanoidbench-manip-powerlift-i00-standard-realtime/README.md
index 290511c447..30713398c9 100644
--- a/tasks/humanoidbench-manip-powerlift-i00-standard-realtime/README.md
+++ b/tasks/humanoidbench-manip-powerlift-i00-standard-realtime/README.md
@@ -9,6 +9,7 @@
| Family / instance | `powerlift` / seed 0 |
| Oracle | `none` — none known: no run has passed it, and published RL reaches 41% of the bar (paper 286, FastTD3 325, bar 800) with upstream's 104 kg dumbbell (53.7 kg since image 0.2.0); human review in PR #83 |
| Agent time budget | 3600 s; tokens: protocol default (PROTOCOL §5) |
+| Real time | slowdown S = 1 (the world runs at 1/S of real speed: one control step every S x control_dt s of wall time; `RCB_SLOWDOWN`'s default); reset budget 50 (`RCB_RESET_BUDGET`'s default) (PROTOCOL §4b) |
| Base image | `ghcr.io/mll-lab-nu/eai-humanoidbench:0.2.0` |
## Task
diff --git a/tasks/humanoidbench-manip-powerlift-i00-standard-realtime/task.toml b/tasks/humanoidbench-manip-powerlift-i00-standard-realtime/task.toml
index 7ed2a8e694..dcb5a116ca 100644
--- a/tasks/humanoidbench-manip-powerlift-i00-standard-realtime/task.toml
+++ b/tasks/humanoidbench-manip-powerlift-i00-standard-realtime/task.toml
@@ -28,6 +28,8 @@ simulator = "humanoidbench"
robot = "unitree-h1hand"
family = "powerlift"
mode = "standard-realtime"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
+slowdown = 1 # standard-realtime: the world runs at 1/S of real speed (docs/PROTOCOL.md §4b); RCB_SLOWDOWN's default
protocol = "eai-standard/2.3"
instance_seed = 0
expert_time_estimate_hours = 8.0
diff --git a/tasks/humanoidbench-manip-powerlift-i00-standard/task.toml b/tasks/humanoidbench-manip-powerlift-i00-standard/task.toml
index ef8d67ead3..ab7e093af1 100644
--- a/tasks/humanoidbench-manip-powerlift-i00-standard/task.toml
+++ b/tasks/humanoidbench-manip-powerlift-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "humanoidbench"
robot = "unitree-h1hand"
family = "powerlift"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
instance_seed = 0
expert_time_estimate_hours = 8.0
diff --git a/tasks/humanoidbench-manip-room-i00-privileged/task.toml b/tasks/humanoidbench-manip-room-i00-privileged/task.toml
index 7ff90060df..13421159a4 100644
--- a/tasks/humanoidbench-manip-room-i00-privileged/task.toml
+++ b/tasks/humanoidbench-manip-room-i00-privileged/task.toml
@@ -17,6 +17,7 @@ robot = "unitree-h1hand"
# HumanoidBench env h1hand-room-v0; paper category manipulation; C8 · mobile manipulation
family = "room"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
# scene = environment/instance.npz (frozen MuJoCo state); identity = the dataset.toml digest
instance_seed = 0
expert_time_estimate_hours = 5.0
diff --git a/tasks/humanoidbench-manip-room-i00-standard-realtime/README.md b/tasks/humanoidbench-manip-room-i00-standard-realtime/README.md
index c93e120fc3..63fd0014bf 100644
--- a/tasks/humanoidbench-manip-room-i00-standard-realtime/README.md
+++ b/tasks/humanoidbench-manip-room-i00-standard-realtime/README.md
@@ -9,6 +9,7 @@
| Family / instance | `room` / seed 0 |
| Oracle | `none` — none known: no run has passed it, and published RL reaches 44% of the bar (paper 128, FastTD3 177, bar 400); human review in PR #83 |
| Agent time budget | 3600 s; tokens: protocol default (PROTOCOL §5) |
+| Real time | slowdown S = 1 (the world runs at 1/S of real speed: one control step every S x control_dt s of wall time; `RCB_SLOWDOWN`'s default); reset budget 50 (`RCB_RESET_BUDGET`'s default) (PROTOCOL §4b) |
| Base image | `ghcr.io/mll-lab-nu/eai-humanoidbench:0.2.0` |
## Task
diff --git a/tasks/humanoidbench-manip-room-i00-standard-realtime/task.toml b/tasks/humanoidbench-manip-room-i00-standard-realtime/task.toml
index 2b0760e7aa..543f6046eb 100644
--- a/tasks/humanoidbench-manip-room-i00-standard-realtime/task.toml
+++ b/tasks/humanoidbench-manip-room-i00-standard-realtime/task.toml
@@ -28,6 +28,8 @@ simulator = "humanoidbench"
robot = "unitree-h1hand"
family = "room"
mode = "standard-realtime"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
+slowdown = 1 # standard-realtime: the world runs at 1/S of real speed (docs/PROTOCOL.md §4b); RCB_SLOWDOWN's default
protocol = "eai-standard/2.3"
instance_seed = 0
expert_time_estimate_hours = 10.0
diff --git a/tasks/humanoidbench-manip-room-i00-standard/task.toml b/tasks/humanoidbench-manip-room-i00-standard/task.toml
index 986faee24a..55cc4becdc 100644
--- a/tasks/humanoidbench-manip-room-i00-standard/task.toml
+++ b/tasks/humanoidbench-manip-room-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "humanoidbench"
robot = "unitree-h1hand"
family = "room"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
instance_seed = 0
expert_time_estimate_hours = 10.0
diff --git a/tasks/humanoidbench-manip-window-i00-privileged/task.toml b/tasks/humanoidbench-manip-window-i00-privileged/task.toml
index 19620ec30d..3ce75a5b1d 100644
--- a/tasks/humanoidbench-manip-window-i00-privileged/task.toml
+++ b/tasks/humanoidbench-manip-window-i00-privileged/task.toml
@@ -17,6 +17,7 @@ robot = "unitree-h1hand"
# HumanoidBench env h1hand-window-v0; paper category manipulation; C6 · standing bimanual manipulation
family = "window"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
# scene = environment/instance.npz (frozen MuJoCo state); identity = the dataset.toml digest
instance_seed = 0
expert_time_estimate_hours = 4.0
diff --git a/tasks/humanoidbench-manip-window-i00-standard-realtime/README.md b/tasks/humanoidbench-manip-window-i00-standard-realtime/README.md
index 1178ab74e4..f1550a52b8 100644
--- a/tasks/humanoidbench-manip-window-i00-standard-realtime/README.md
+++ b/tasks/humanoidbench-manip-window-i00-standard-realtime/README.md
@@ -9,6 +9,7 @@
| Family / instance | `window` / seed 0 |
| Oracle | `none` — none known: no run has passed it, and published RL reaches 94% of the bar (paper 227, FastTD3 613, bar 650); human review in PR #83 |
| Agent time budget | 3600 s; tokens: protocol default (PROTOCOL §5) |
+| Real time | slowdown S = 1 (the world runs at 1/S of real speed: one control step every S x control_dt s of wall time; `RCB_SLOWDOWN`'s default); reset budget 50 (`RCB_RESET_BUDGET`'s default) (PROTOCOL §4b) |
| Base image | `ghcr.io/mll-lab-nu/eai-humanoidbench:0.2.0` |
## Task
diff --git a/tasks/humanoidbench-manip-window-i00-standard-realtime/task.toml b/tasks/humanoidbench-manip-window-i00-standard-realtime/task.toml
index 00f38a88f4..ecce062c9f 100644
--- a/tasks/humanoidbench-manip-window-i00-standard-realtime/task.toml
+++ b/tasks/humanoidbench-manip-window-i00-standard-realtime/task.toml
@@ -28,6 +28,8 @@ simulator = "humanoidbench"
robot = "unitree-h1hand"
family = "window"
mode = "standard-realtime"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
+slowdown = 1 # standard-realtime: the world runs at 1/S of real speed (docs/PROTOCOL.md §4b); RCB_SLOWDOWN's default
protocol = "eai-standard/2.3"
instance_seed = 0
expert_time_estimate_hours = 8.0
diff --git a/tasks/humanoidbench-manip-window-i00-standard/task.toml b/tasks/humanoidbench-manip-window-i00-standard/task.toml
index d04802abbd..af77ad3f9c 100644
--- a/tasks/humanoidbench-manip-window-i00-standard/task.toml
+++ b/tasks/humanoidbench-manip-window-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "humanoidbench"
robot = "unitree-h1hand"
family = "window"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
instance_seed = 0
expert_time_estimate_hours = 8.0
diff --git a/tasks/kinder-constrained-cupboard-i00-privileged/task.toml b/tasks/kinder-constrained-cupboard-i00-privileged/task.toml
index bdfbffd834..85b7fea7f1 100644
--- a/tasks/kinder-constrained-cupboard-i00-privileged/task.toml
+++ b/tasks/kinder-constrained-cupboard-i00-privileged/task.toml
@@ -16,6 +16,7 @@ robot = "tidybot-plus-plus"
# KinDER env kinder/ConstrainedCupboard3D-o6-v0
family = "constrained-cupboard"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
# scene = environment/instance.json (KinDER's seed + a digest of the scene it builds); identity = the dataset.toml digest
instance_seed = 0
expert_time_estimate_hours = 4.0
diff --git a/tasks/kinder-constrained-cupboard-i00-standard/task.toml b/tasks/kinder-constrained-cupboard-i00-standard/task.toml
index 99b3cfc34f..9658978fe8 100644
--- a/tasks/kinder-constrained-cupboard-i00-standard/task.toml
+++ b/tasks/kinder-constrained-cupboard-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "kinder"
robot = "tidybot-plus-plus"
family = "constrained-cupboard"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
instance_seed = 0
expert_time_estimate_hours = 8.0
diff --git a/tasks/kinder-dynamo-i00-privileged/task.toml b/tasks/kinder-dynamo-i00-privileged/task.toml
index 184aaab974..dcee91e289 100644
--- a/tasks/kinder-dynamo-i00-privileged/task.toml
+++ b/tasks/kinder-dynamo-i00-privileged/task.toml
@@ -16,6 +16,7 @@ robot = "tidybot-plus-plus"
# KinDER env kinder/Dynamo3D-o3-v0
family = "dynamo"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
# scene = environment/instance.json (KinDER's seed + a digest of the scene it builds); identity = the dataset.toml digest
instance_seed = 0
expert_time_estimate_hours = 2.0
diff --git a/tasks/kinder-dynamo-i00-standard/task.toml b/tasks/kinder-dynamo-i00-standard/task.toml
index 9de6380d8d..35bfeae238 100644
--- a/tasks/kinder-dynamo-i00-standard/task.toml
+++ b/tasks/kinder-dynamo-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "kinder"
robot = "tidybot-plus-plus"
family = "dynamo"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
instance_seed = 0
expert_time_estimate_hours = 4.0
diff --git a/tasks/kinder-scoop-pour-i00-privileged/task.toml b/tasks/kinder-scoop-pour-i00-privileged/task.toml
index 29bbfac614..88130f81b9 100644
--- a/tasks/kinder-scoop-pour-i00-privileged/task.toml
+++ b/tasks/kinder-scoop-pour-i00-privileged/task.toml
@@ -16,6 +16,7 @@ robot = "tidybot-plus-plus"
# KinDER env kinder/ScoopPour3D-o30-v0
family = "scoop-pour"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
# scene = environment/instance.json (KinDER's seed + a digest of the scene it builds); identity = the dataset.toml digest
instance_seed = 0
expert_time_estimate_hours = 4.0
diff --git a/tasks/kinder-scoop-pour-i00-standard/task.toml b/tasks/kinder-scoop-pour-i00-standard/task.toml
index ee1ee38f17..7ced8de801 100644
--- a/tasks/kinder-scoop-pour-i00-standard/task.toml
+++ b/tasks/kinder-scoop-pour-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "kinder"
robot = "tidybot-plus-plus"
family = "scoop-pour"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
instance_seed = 0
expert_time_estimate_hours = 8.0
diff --git a/tasks/kinder-sweep-into-drawer-i00-privileged/task.toml b/tasks/kinder-sweep-into-drawer-i00-privileged/task.toml
index 7e1df3ca97..e2e748ddcf 100644
--- a/tasks/kinder-sweep-into-drawer-i00-privileged/task.toml
+++ b/tasks/kinder-sweep-into-drawer-i00-privileged/task.toml
@@ -16,6 +16,7 @@ robot = "tidybot-plus-plus"
# KinDER env kinder/SweepIntoDrawer3D-o5-v0
family = "sweep-into-drawer"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
# scene = environment/instance.json (KinDER's seed + a digest of the scene it builds); identity = the dataset.toml digest
instance_seed = 0
expert_time_estimate_hours = 4.0
diff --git a/tasks/kinder-sweep-into-drawer-i00-standard/task.toml b/tasks/kinder-sweep-into-drawer-i00-standard/task.toml
index 7ae3ac83b6..dab6c7d363 100644
--- a/tasks/kinder-sweep-into-drawer-i00-standard/task.toml
+++ b/tasks/kinder-sweep-into-drawer-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "kinder"
robot = "tidybot-plus-plus"
family = "sweep-into-drawer"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
instance_seed = 0
expert_time_estimate_hours = 8.0
diff --git a/tasks/kinder-sweep-simple-i00-privileged/task.toml b/tasks/kinder-sweep-simple-i00-privileged/task.toml
index ae32d980b1..7b2ef216cf 100644
--- a/tasks/kinder-sweep-simple-i00-privileged/task.toml
+++ b/tasks/kinder-sweep-simple-i00-privileged/task.toml
@@ -16,6 +16,7 @@ robot = "tidybot-plus-plus"
# KinDER env kinder/SweepSimple3D-o10-sweep_the_blocks_to_the_left_side_of_the_kitchen_island-v0
family = "sweep-simple"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
# scene = environment/instance.json (KinDER's seed + a digest of the scene it builds); identity = the dataset.toml digest
instance_seed = 0
expert_time_estimate_hours = 3.0
diff --git a/tasks/kinder-sweep-simple-i00-standard/task.toml b/tasks/kinder-sweep-simple-i00-standard/task.toml
index 5bd207e306..604a0ea497 100644
--- a/tasks/kinder-sweep-simple-i00-standard/task.toml
+++ b/tasks/kinder-sweep-simple-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "kinder"
robot = "tidybot-plus-plus"
family = "sweep-simple"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
instance_seed = 0
expert_time_estimate_hours = 6.0
diff --git a/tasks/labutopia-heat-liquid-i00-privileged/task.toml b/tasks/labutopia-heat-liquid-i00-privileged/task.toml
index db90959dc6..c87370c979 100644
--- a/tasks/labutopia-heat-liquid-i00-privileged/task.toml
+++ b/tasks/labutopia-heat-liquid-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "labutopia" # LabUtopia @ 8df7278 on NVIDIA Isaac Sim 6.1.
robot = "franka-panda"
family = "heat_liquid"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
level = 2 # LabUtopia's own level for the task (config level2_HeatLiquid)
instance_seed = 1 # numpy's and Python's random generators before LabUtopia's reset
diff --git a/tasks/labutopia-heat-liquid-i00-standard/task.toml b/tasks/labutopia-heat-liquid-i00-standard/task.toml
index 74616c2483..eefc9b0cfb 100644
--- a/tasks/labutopia-heat-liquid-i00-standard/task.toml
+++ b/tasks/labutopia-heat-liquid-i00-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "labutopia" # LabUtopia @ 8df7278 on NVIDIA Isaac Sim 6.1.
robot = "franka-panda"
family = "heat_liquid"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "full"
level = 2 # LabUtopia's own level for the task (config level2_HeatLiquid)
diff --git a/tasks/labutopia-open-close-i00-privileged/task.toml b/tasks/labutopia-open-close-i00-privileged/task.toml
index b275151eec..f43021bc06 100644
--- a/tasks/labutopia-open-close-i00-privileged/task.toml
+++ b/tasks/labutopia-open-close-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "labutopia" # LabUtopia @ 8df7278 on NVIDIA Isaac Sim 6.1.
robot = "franka-panda"
family = "open_close"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
level = 2 # LabUtopia's own level for the task (config level2_openclose)
instance_seed = 0 # numpy's and Python's random generators before LabUtopia's reset
diff --git a/tasks/labutopia-open-close-i00-standard/task.toml b/tasks/labutopia-open-close-i00-standard/task.toml
index 4934661edc..b48336938e 100644
--- a/tasks/labutopia-open-close-i00-standard/task.toml
+++ b/tasks/labutopia-open-close-i00-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "labutopia" # LabUtopia @ 8df7278 on NVIDIA Isaac Sim 6.1.
robot = "franka-panda"
family = "open_close"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "full"
level = 2 # LabUtopia's own level for the task (config level2_openclose)
diff --git a/tasks/labutopia-open-door-i00-privileged/task.toml b/tasks/labutopia-open-door-i00-privileged/task.toml
index b84579d8b2..a27133f244 100644
--- a/tasks/labutopia-open-door-i00-privileged/task.toml
+++ b/tasks/labutopia-open-door-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "labutopia" # LabUtopia @ 8df7278 on NVIDIA Isaac Sim 6.1.
robot = "franka-panda"
family = "open_door"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
level = 1 # LabUtopia's own level for the task (config level1_open_door)
instance_seed = 4 # numpy's and Python's random generators before LabUtopia's reset
diff --git a/tasks/labutopia-open-door-i00-standard/task.toml b/tasks/labutopia-open-door-i00-standard/task.toml
index cf7c4fd802..358ef25f0c 100644
--- a/tasks/labutopia-open-door-i00-standard/task.toml
+++ b/tasks/labutopia-open-door-i00-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "labutopia" # LabUtopia @ 8df7278 on NVIDIA Isaac Sim 6.1.
robot = "franka-panda"
family = "open_door"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "full"
level = 1 # LabUtopia's own level for the task (config level1_open_door)
diff --git a/tasks/labutopia-open-drawer-i00-privileged/task.toml b/tasks/labutopia-open-drawer-i00-privileged/task.toml
index 3d7a2e2d4a..fdeb68fc51 100644
--- a/tasks/labutopia-open-drawer-i00-privileged/task.toml
+++ b/tasks/labutopia-open-drawer-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "labutopia" # LabUtopia @ 8df7278 on NVIDIA Isaac Sim 6.1.
robot = "franka-panda"
family = "open_drawer"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
level = 1 # LabUtopia's own level for the task (config level1_open_drawer)
instance_seed = 2 # numpy's and Python's random generators before LabUtopia's reset
diff --git a/tasks/labutopia-open-drawer-i00-standard/task.toml b/tasks/labutopia-open-drawer-i00-standard/task.toml
index b7e9db527b..bb5125db40 100644
--- a/tasks/labutopia-open-drawer-i00-standard/task.toml
+++ b/tasks/labutopia-open-drawer-i00-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "labutopia" # LabUtopia @ 8df7278 on NVIDIA Isaac Sim 6.1.
robot = "franka-panda"
family = "open_drawer"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "full"
level = 1 # LabUtopia's own level for the task (config level1_open_drawer)
diff --git a/tasks/labutopia-pick-i00-privileged/task.toml b/tasks/labutopia-pick-i00-privileged/task.toml
index 80f124eba9..24d000b60a 100644
--- a/tasks/labutopia-pick-i00-privileged/task.toml
+++ b/tasks/labutopia-pick-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "labutopia" # LabUtopia @ 8df7278 on NVIDIA Isaac Sim 6.1.
robot = "franka-panda"
family = "pick"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
level = 1 # LabUtopia's own level for the task (config level1_pick)
instance_seed = 0 # numpy's and Python's random generators before LabUtopia's reset
diff --git a/tasks/labutopia-pick-i00-standard/task.toml b/tasks/labutopia-pick-i00-standard/task.toml
index 099d0e70e4..6a3fb72c77 100644
--- a/tasks/labutopia-pick-i00-standard/task.toml
+++ b/tasks/labutopia-pick-i00-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "labutopia" # LabUtopia @ 8df7278 on NVIDIA Isaac Sim 6.1.
robot = "franka-panda"
family = "pick"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "full"
level = 1 # LabUtopia's own level for the task (config level1_pick)
diff --git a/tasks/labutopia-place-i00-privileged/task.toml b/tasks/labutopia-place-i00-privileged/task.toml
index f5618a8fca..008ea5bb12 100644
--- a/tasks/labutopia-place-i00-privileged/task.toml
+++ b/tasks/labutopia-place-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "labutopia" # LabUtopia @ 8df7278 on NVIDIA Isaac Sim 6.1.
robot = "franka-panda"
family = "place"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
level = 1 # LabUtopia's own level for the task (config level1_place)
instance_seed = 0 # numpy's and Python's random generators before LabUtopia's reset
diff --git a/tasks/labutopia-place-i00-standard/task.toml b/tasks/labutopia-place-i00-standard/task.toml
index a0157a508b..425bd27ffe 100644
--- a/tasks/labutopia-place-i00-standard/task.toml
+++ b/tasks/labutopia-place-i00-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "labutopia" # LabUtopia @ 8df7278 on NVIDIA Isaac Sim 6.1.
robot = "franka-panda"
family = "place"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "full"
level = 1 # LabUtopia's own level for the task (config level1_place)
diff --git a/tasks/labutopia-pour-i00-privileged/task.toml b/tasks/labutopia-pour-i00-privileged/task.toml
index 793f276c53..42f565ba6b 100644
--- a/tasks/labutopia-pour-i00-privileged/task.toml
+++ b/tasks/labutopia-pour-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "labutopia" # LabUtopia @ 8df7278 on NVIDIA Isaac Sim 6.1.
robot = "franka-panda"
family = "pour"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
level = 1 # LabUtopia's own level for the task (config level1_pour)
instance_seed = 0 # numpy's and Python's random generators before LabUtopia's reset
diff --git a/tasks/labutopia-pour-i00-standard/task.toml b/tasks/labutopia-pour-i00-standard/task.toml
index f2f7c9494a..4ef461a3b0 100644
--- a/tasks/labutopia-pour-i00-standard/task.toml
+++ b/tasks/labutopia-pour-i00-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "labutopia" # LabUtopia @ 8df7278 on NVIDIA Isaac Sim 6.1.
robot = "franka-panda"
family = "pour"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "full"
level = 1 # LabUtopia's own level for the task (config level1_pour)
diff --git a/tasks/labutopia-press-i00-privileged/task.toml b/tasks/labutopia-press-i00-privileged/task.toml
index 9eb24514f9..02bc791a10 100644
--- a/tasks/labutopia-press-i00-privileged/task.toml
+++ b/tasks/labutopia-press-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "labutopia" # LabUtopia @ 8df7278 on NVIDIA Isaac Sim 6.1.
robot = "franka-panda"
family = "press"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
level = 1 # LabUtopia's own level for the task (config level1_press)
instance_seed = 0 # numpy's and Python's random generators before LabUtopia's reset
diff --git a/tasks/labutopia-press-i00-standard/task.toml b/tasks/labutopia-press-i00-standard/task.toml
index 02cda925b7..dc6cf8ffc8 100644
--- a/tasks/labutopia-press-i00-standard/task.toml
+++ b/tasks/labutopia-press-i00-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "labutopia" # LabUtopia @ 8df7278 on NVIDIA Isaac Sim 6.1.
robot = "franka-panda"
family = "press"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "full"
level = 1 # LabUtopia's own level for the task (config level1_press)
diff --git a/tasks/labutopia-shake-i00-privileged/task.toml b/tasks/labutopia-shake-i00-privileged/task.toml
index 2291fb9177..6ad9d544fb 100644
--- a/tasks/labutopia-shake-i00-privileged/task.toml
+++ b/tasks/labutopia-shake-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "labutopia" # LabUtopia @ 8df7278 on NVIDIA Isaac Sim 6.1.
robot = "franka-panda"
family = "shake"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
level = 1 # LabUtopia's own level for the task (config level1_shake)
instance_seed = 0 # numpy's and Python's random generators before LabUtopia's reset
diff --git a/tasks/labutopia-shake-i00-standard/task.toml b/tasks/labutopia-shake-i00-standard/task.toml
index 27c28fe7e9..c17e4b32c5 100644
--- a/tasks/labutopia-shake-i00-standard/task.toml
+++ b/tasks/labutopia-shake-i00-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "labutopia" # LabUtopia @ 8df7278 on NVIDIA Isaac Sim 6.1.
robot = "franka-panda"
family = "shake"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "full"
level = 1 # LabUtopia's own level for the task (config level1_shake)
diff --git a/tasks/libero-max-goal-akita-black-bowl-in-wooden-cabinet-top-target-relocation-i00-privileged/task.toml b/tasks/libero-max-goal-akita-black-bowl-in-wooden-cabinet-top-target-relocation-i00-privileged/task.toml
index 0935ecf825..35e120ea88 100644
--- a/tasks/libero-max-goal-akita-black-bowl-in-wooden-cabinet-top-target-relocation-i00-privileged/task.toml
+++ b/tasks/libero-max-goal-akita-black-bowl-in-wooden-cabinet-top-target-relocation-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "robosuite"
robot = "franka-panda"
family = "libero-max/libero_goal/target_relocation"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 195 # LIBERO-MAX's policy seed for the case (the scene is the frozen LIBERO-PRO initial state)
expert_time_estimate_hours = 1.0
diff --git a/tasks/libero-max-goal-akita-black-bowl-in-wooden-cabinet-top-target-relocation-i00-standard/README.md b/tasks/libero-max-goal-akita-black-bowl-in-wooden-cabinet-top-target-relocation-i00-standard/README.md
index ae3cd9c3c0..ca5d31a952 100644
--- a/tasks/libero-max-goal-akita-black-bowl-in-wooden-cabinet-top-target-relocation-i00-standard/README.md
+++ b/tasks/libero-max-goal-akita-black-bowl-in-wooden-cabinet-top-target-relocation-i00-standard/README.md
@@ -86,3 +86,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
/reset` ends the episode as `error` (`sim_error = 1`), not `reset` (reachable only when a run grants resets; the
default is none). The scene, instance, pass rule and every verdict reached before are unchanged, so the task keeps
its name.
+- 2026-10-14 (board 0113, from board 0086): the instruction says that a moved object or receptacle jumps alone: what rests in or on it stays where it was (LIBERO-MAX moves one entity; board 0086 saw a basket jump through a can already in it). Text only: the instance, the event, the reference solution, LIBERO's success check and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/libero-max-goal-akita-black-bowl-in-wooden-cabinet-top-target-relocation-i00-standard/instruction.md b/tasks/libero-max-goal-akita-black-bowl-in-wooden-cabinet-top-target-relocation-i00-standard/instruction.md
index 28307d2517..f136487bc2 100644
--- a/tasks/libero-max-goal-akita-black-bowl-in-wooden-cabinet-top-target-relocation-i00-standard/instruction.md
+++ b/tasks/libero-max-goal-akita-black-bowl-in-wooden-cabinet-top-target-relocation-i00-standard/instruction.md
@@ -6,7 +6,8 @@ simulator runs in a service you cannot open, and you reach it only through a cli
The scene can change while you work: once during the episode, when your gripper first gets close to one of the task's
objects, the scene changes (as in LIBERO-MAX): an object or a receptacle may be moved, an obstacle or a few other
objects may appear, a camera may shift, or the lighting, the colours or the image noise may change. Nothing tells you
-when; the task and its success check stay the same, so keep checking what the cameras show.
+when; the task and its success check stay the same, so keep checking what the cameras show. A moved object or
+receptacle jumps there alone: what already rests in or on it stays where it was.
## Rules (real-robot rules)
- **One episode, no reset.** The episode starts when the service comes up. It ends when the task succeeds, after
diff --git a/tasks/libero-max-goal-akita-black-bowl-in-wooden-cabinet-top-target-relocation-i00-standard/task.toml b/tasks/libero-max-goal-akita-black-bowl-in-wooden-cabinet-top-target-relocation-i00-standard/task.toml
index 0fcdff2959..c69ec2c9ce 100644
--- a/tasks/libero-max-goal-akita-black-bowl-in-wooden-cabinet-top-target-relocation-i00-standard/task.toml
+++ b/tasks/libero-max-goal-akita-black-bowl-in-wooden-cabinet-top-target-relocation-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "robosuite"
robot = "franka-panda"
family = "libero-max/libero_goal/target_relocation"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 195
diff --git a/tasks/libero-max-goal-akita-black-bowl-on-flat-stove-cook-target-relocation-i00-privileged/task.toml b/tasks/libero-max-goal-akita-black-bowl-on-flat-stove-cook-target-relocation-i00-privileged/task.toml
index b3ca5b48ab..4a787b55f6 100644
--- a/tasks/libero-max-goal-akita-black-bowl-on-flat-stove-cook-target-relocation-i00-privileged/task.toml
+++ b/tasks/libero-max-goal-akita-black-bowl-on-flat-stove-cook-target-relocation-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "robosuite"
robot = "franka-panda"
family = "libero-max/libero_goal/target_relocation"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 195 # LIBERO-MAX's policy seed for the case (the scene is the frozen LIBERO-PRO initial state)
expert_time_estimate_hours = 1.0
diff --git a/tasks/libero-max-goal-akita-black-bowl-on-flat-stove-cook-target-relocation-i00-standard/README.md b/tasks/libero-max-goal-akita-black-bowl-on-flat-stove-cook-target-relocation-i00-standard/README.md
index ba4d42c89a..feff297a7f 100644
--- a/tasks/libero-max-goal-akita-black-bowl-on-flat-stove-cook-target-relocation-i00-standard/README.md
+++ b/tasks/libero-max-goal-akita-black-bowl-on-flat-stove-cook-target-relocation-i00-standard/README.md
@@ -86,3 +86,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
/reset` ends the episode as `error` (`sim_error = 1`), not `reset` (reachable only when a run grants resets; the
default is none). The scene, instance, pass rule and every verdict reached before are unchanged, so the task keeps
its name.
+- 2026-10-14 (board 0113, from board 0086): the instruction says that a moved object or receptacle jumps alone: what rests in or on it stays where it was (LIBERO-MAX moves one entity; board 0086 saw a basket jump through a can already in it). Text only: the instance, the event, the reference solution, LIBERO's success check and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/libero-max-goal-akita-black-bowl-on-flat-stove-cook-target-relocation-i00-standard/instruction.md b/tasks/libero-max-goal-akita-black-bowl-on-flat-stove-cook-target-relocation-i00-standard/instruction.md
index 8d2c3515b0..5bbffba78f 100644
--- a/tasks/libero-max-goal-akita-black-bowl-on-flat-stove-cook-target-relocation-i00-standard/instruction.md
+++ b/tasks/libero-max-goal-akita-black-bowl-on-flat-stove-cook-target-relocation-i00-standard/instruction.md
@@ -6,7 +6,8 @@ simulator runs in a service you cannot open, and you reach it only through a cli
The scene can change while you work: once during the episode, when your gripper first gets close to one of the task's
objects, the scene changes (as in LIBERO-MAX): an object or a receptacle may be moved, an obstacle or a few other
objects may appear, a camera may shift, or the lighting, the colours or the image noise may change. Nothing tells you
-when; the task and its success check stay the same, so keep checking what the cameras show.
+when; the task and its success check stay the same, so keep checking what the cameras show. A moved object or
+receptacle jumps there alone: what already rests in or on it stays where it was.
## Rules (real-robot rules)
- **One episode, no reset.** The episode starts when the service comes up. It ends when the task succeeds, after
diff --git a/tasks/libero-max-goal-akita-black-bowl-on-flat-stove-cook-target-relocation-i00-standard/task.toml b/tasks/libero-max-goal-akita-black-bowl-on-flat-stove-cook-target-relocation-i00-standard/task.toml
index 64bb23b29c..2670288cc0 100644
--- a/tasks/libero-max-goal-akita-black-bowl-on-flat-stove-cook-target-relocation-i00-standard/task.toml
+++ b/tasks/libero-max-goal-akita-black-bowl-on-flat-stove-cook-target-relocation-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "robosuite"
robot = "franka-panda"
family = "libero-max/libero_goal/target_relocation"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 195
diff --git a/tasks/libero-max-goal-akita-black-bowl-on-plate-receptacle-relocation-i00-privileged/README.md b/tasks/libero-max-goal-akita-black-bowl-on-plate-receptacle-relocation-i00-privileged/README.md
index d6f3c94928..0a41dc4d92 100644
--- a/tasks/libero-max-goal-akita-black-bowl-on-plate-receptacle-relocation-i00-privileged/README.md
+++ b/tasks/libero-max-goal-akita-black-bowl-on-plate-receptacle-relocation-i00-privileged/README.md
@@ -14,7 +14,7 @@
## Task
LIBERO-MAX case `pro-position-goal-t08-i00-receptacle_relocation-d1-p195`: a Franka Panda arm at a table, instructed **"Put the bowl on the plate."**, on a LIBERO-PRO variant
of a libero_goal task (position). Online event (receptacle relocation): the first time the gripper comes within
-0.18 m of `akita_black_bowl_1`, LIBERO-MAX changes the scene once (the receptacle `plate_1` is moved 0.12 m, direction (0.707, 0.707)). The deliverable is one
+0.18 m of `akita_black_bowl_1`, LIBERO-MAX changes the scene once (the receptacle `plate_1` is moved 0.12 m, direction (0.707, 0.707); it jumps there alone: what already rests in or on it stays where it was, and where it lands on another object the physics pushes the two apart). The deliverable is one
replayable trajectory of LIBERO's OSC control steps (`/app/output/trajectory.npz`, `(T, 7)`, T ≤ 300), graded by
LIBERO's success check at its end.
@@ -54,7 +54,7 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
|---|---|
| Upstream | LIBERO-PRO @ 2b910b5 (MIT; task data `zhouxueyang/LIBERO-Pro` @ c86fc3b, CC-BY-4.0); LIBERO-MAX @ a1e3cef (event runtime and this case's parameters; no licence file upstream); robosuite 1.4.0, MuJoCo 3.2.6 |
| Case | `pro-position-goal-t08-i00-receptacle_relocation-d1-p195` (LIBERO-MAX 8,000-case manifest; LIBERO-PRO category `position`, task 8, initial state 0) |
-| Event | `receptacle_relocation`: the receptacle `plate_1` is moved 0.12 m, direction (0.707, 0.707); trigger: gripper within 0.18 m of `akita_black_bowl_1` |
+| Event | `receptacle_relocation`: the receptacle `plate_1` is moved 0.12 m, direction (0.707, 0.707); it jumps there alone: what already rests in or on it stays where it was, and where it lands on another object the physics pushes the two apart; trigger: gripper within 0.18 m of `akita_black_bowl_1` |
| Actions | 7 per control step (1/20 s): OSC hand-target delta (1.0 = 0.05 m, 0.5 rad) and gripper |
| Resources | 4 CPUs, no GPU (CPU physics, OSMesa rendering) |
@@ -76,3 +76,4 @@ Generated by `scripts/libero-max/generate.py` from `scripts/libero-max/cases.jso
it executes any action). Before, such a trial read as an agent failure with `grader_error = 1` only.
- A replay that records a video prints its verdict before it finishes the video, and a failing video no longer fails
the replay; `video_rendered` is 1 only when the video was finished.
+- 2026-10-14 (board 0113, from board 0086): the instruction says that a moved receptacle jumps alone: what rests in or on it stays where it was (LIBERO-MAX moves one entity; board 0086 saw a basket jump through a can already in it). Text only: the instance, the event, the reference solution, LIBERO's success check and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/libero-max-goal-akita-black-bowl-on-plate-receptacle-relocation-i00-privileged/instruction.md b/tasks/libero-max-goal-akita-black-bowl-on-plate-receptacle-relocation-i00-privileged/instruction.md
index 3d85f6beea..8c5310419e 100644
--- a/tasks/libero-max-goal-akita-black-bowl-on-plate-receptacle-relocation-i00-privileged/instruction.md
+++ b/tasks/libero-max-goal-akita-black-bowl-on-plate-receptacle-relocation-i00-privileged/instruction.md
@@ -3,7 +3,7 @@
A **Franka Panda** arm (LIBERO on robosuite / MuJoCo) at a table. **Task: Put the bowl on the plate.**
The scene is one frozen LIBERO-MAX case: a LIBERO-PRO variant of a libero_goal task (position), with one online
event: the first time the gripper comes within 0.18 m of `akita_black_bowl_1`, LIBERO-MAX changes the scene
-once (the receptacle `plate_1` is moved 0.12 m, direction (0.707, 0.707)). You have the full simulator state, so you can see the change happen (`s.state()["event_fired"]`).
+once (the receptacle `plate_1` is moved 0.12 m, direction (0.707, 0.707); it jumps there alone: what already rests in or on it stays where it was, and where it lands on another object the physics pushes the two apart). You have the full simulator state, so you can see the change happen (`s.state()["event_fired"]`).
## What you hand in
`/app/output/trajectory.npz` with key `actions`: LIBERO's native control steps, shape `(T, 7)`, 1 ≤ T ≤ 300
diff --git a/tasks/libero-max-goal-akita-black-bowl-on-plate-receptacle-relocation-i00-privileged/task.toml b/tasks/libero-max-goal-akita-black-bowl-on-plate-receptacle-relocation-i00-privileged/task.toml
index 225df5c2a7..b47b4c77c2 100644
--- a/tasks/libero-max-goal-akita-black-bowl-on-plate-receptacle-relocation-i00-privileged/task.toml
+++ b/tasks/libero-max-goal-akita-black-bowl-on-plate-receptacle-relocation-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "robosuite"
robot = "franka-panda"
family = "libero-max/libero_goal/receptacle_relocation"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 195 # LIBERO-MAX's policy seed for the case (the scene is the frozen LIBERO-PRO initial state)
expert_time_estimate_hours = 1.0
diff --git a/tasks/libero-max-goal-akita-black-bowl-on-plate-receptacle-relocation-i00-standard/README.md b/tasks/libero-max-goal-akita-black-bowl-on-plate-receptacle-relocation-i00-standard/README.md
index f6f19c6dfc..9f66e41791 100644
--- a/tasks/libero-max-goal-akita-black-bowl-on-plate-receptacle-relocation-i00-standard/README.md
+++ b/tasks/libero-max-goal-akita-black-bowl-on-plate-receptacle-relocation-i00-standard/README.md
@@ -13,7 +13,7 @@
## Task
LIBERO-MAX case `pro-position-goal-t08-i00-receptacle_relocation-d1-p195` as a robot service: a Franka Panda arm at a table, instructed **"Put the bowl on the plate."**, on
-the same frozen instance as `libero-max-goal-akita-black-bowl-on-plate-receptacle-relocation-i00-privileged`, with the same online event (receptacle relocation: the receptacle `plate_1` is moved 0.12 m, direction (0.707, 0.707), the
+the same frozen instance as `libero-max-goal-akita-black-bowl-on-plate-receptacle-relocation-i00-privileged`, with the same online event (receptacle relocation: the receptacle `plate_1` is moved 0.12 m, direction (0.707, 0.707); it jumps there alone: what already rests in or on it stays where it was, and where it lands on another object the physics pushes the two apart, the
first time the gripper comes within 0.18 m of `akita_black_bowl_1`). The simulator runs in the `sim` sidecar
(`environment/sim/server.py` on main's `rcb_service.py`); the agent only has the client: two cameras (RGB
256×256, depth on request, calibrated), the robot's own state, and OSC control steps in chunks. One
@@ -86,3 +86,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
/reset` ends the episode as `error` (`sim_error = 1`), not `reset` (reachable only when a run grants resets; the
default is none). The scene, instance, pass rule and every verdict reached before are unchanged, so the task keeps
its name.
+- 2026-10-14 (board 0113, from board 0086): the instruction says that a moved object or receptacle jumps alone: what rests in or on it stays where it was (LIBERO-MAX moves one entity; board 0086 saw a basket jump through a can already in it). Text only: the instance, the event, the reference solution, LIBERO's success check and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/libero-max-goal-akita-black-bowl-on-plate-receptacle-relocation-i00-standard/instruction.md b/tasks/libero-max-goal-akita-black-bowl-on-plate-receptacle-relocation-i00-standard/instruction.md
index 0b2b94ac3b..37dfbb7326 100644
--- a/tasks/libero-max-goal-akita-black-bowl-on-plate-receptacle-relocation-i00-standard/instruction.md
+++ b/tasks/libero-max-goal-akita-black-bowl-on-plate-receptacle-relocation-i00-standard/instruction.md
@@ -6,7 +6,8 @@ simulator runs in a service you cannot open, and you reach it only through a cli
The scene can change while you work: once during the episode, when your gripper first gets close to one of the task's
objects, the scene changes (as in LIBERO-MAX): an object or a receptacle may be moved, an obstacle or a few other
objects may appear, a camera may shift, or the lighting, the colours or the image noise may change. Nothing tells you
-when; the task and its success check stay the same, so keep checking what the cameras show.
+when; the task and its success check stay the same, so keep checking what the cameras show. A moved object or
+receptacle jumps there alone: what already rests in or on it stays where it was.
## Rules (real-robot rules)
- **One episode, no reset.** The episode starts when the service comes up. It ends when the task succeeds, after
diff --git a/tasks/libero-max-goal-akita-black-bowl-on-plate-receptacle-relocation-i00-standard/task.toml b/tasks/libero-max-goal-akita-black-bowl-on-plate-receptacle-relocation-i00-standard/task.toml
index f31d53e911..ecfbed5d7a 100644
--- a/tasks/libero-max-goal-akita-black-bowl-on-plate-receptacle-relocation-i00-standard/task.toml
+++ b/tasks/libero-max-goal-akita-black-bowl-on-plate-receptacle-relocation-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "robosuite"
robot = "franka-panda"
family = "libero-max/libero_goal/receptacle_relocation"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 195
diff --git a/tasks/libero-max-goal-akita-black-bowl-on-wooden-cabinet-top-obstacle-insertion-i00-privileged/task.toml b/tasks/libero-max-goal-akita-black-bowl-on-wooden-cabinet-top-obstacle-insertion-i00-privileged/task.toml
index 58f5ed44a9..03ae3e9bea 100644
--- a/tasks/libero-max-goal-akita-black-bowl-on-wooden-cabinet-top-obstacle-insertion-i00-privileged/task.toml
+++ b/tasks/libero-max-goal-akita-black-bowl-on-wooden-cabinet-top-obstacle-insertion-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "robosuite"
robot = "franka-panda"
family = "libero-max/libero_goal/obstacle_insertion"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 195 # LIBERO-MAX's policy seed for the case (the scene is the frozen LIBERO-PRO initial state)
expert_time_estimate_hours = 1.0
diff --git a/tasks/libero-max-goal-akita-black-bowl-on-wooden-cabinet-top-obstacle-insertion-i00-standard/README.md b/tasks/libero-max-goal-akita-black-bowl-on-wooden-cabinet-top-obstacle-insertion-i00-standard/README.md
index 80d77f336e..14cced68e8 100644
--- a/tasks/libero-max-goal-akita-black-bowl-on-wooden-cabinet-top-obstacle-insertion-i00-standard/README.md
+++ b/tasks/libero-max-goal-akita-black-bowl-on-wooden-cabinet-top-obstacle-insertion-i00-standard/README.md
@@ -86,3 +86,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
/reset` ends the episode as `error` (`sim_error = 1`), not `reset` (reachable only when a run grants resets; the
default is none). The scene, instance, pass rule and every verdict reached before are unchanged, so the task keeps
its name.
+- 2026-10-14 (board 0113, from board 0086): the instruction says that a moved object or receptacle jumps alone: what rests in or on it stays where it was (LIBERO-MAX moves one entity; board 0086 saw a basket jump through a can already in it). Text only: the instance, the event, the reference solution, LIBERO's success check and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/libero-max-goal-akita-black-bowl-on-wooden-cabinet-top-obstacle-insertion-i00-standard/instruction.md b/tasks/libero-max-goal-akita-black-bowl-on-wooden-cabinet-top-obstacle-insertion-i00-standard/instruction.md
index 0ce48b2e49..886f09524c 100644
--- a/tasks/libero-max-goal-akita-black-bowl-on-wooden-cabinet-top-obstacle-insertion-i00-standard/instruction.md
+++ b/tasks/libero-max-goal-akita-black-bowl-on-wooden-cabinet-top-obstacle-insertion-i00-standard/instruction.md
@@ -6,7 +6,8 @@ simulator runs in a service you cannot open, and you reach it only through a cli
The scene can change while you work: once during the episode, when your gripper first gets close to one of the task's
objects, the scene changes (as in LIBERO-MAX): an object or a receptacle may be moved, an obstacle or a few other
objects may appear, a camera may shift, or the lighting, the colours or the image noise may change. Nothing tells you
-when; the task and its success check stay the same, so keep checking what the cameras show.
+when; the task and its success check stay the same, so keep checking what the cameras show. A moved object or
+receptacle jumps there alone: what already rests in or on it stays where it was.
## Rules (real-robot rules)
- **One episode, no reset.** The episode starts when the service comes up. It ends when the task succeeds, after
diff --git a/tasks/libero-max-goal-akita-black-bowl-on-wooden-cabinet-top-obstacle-insertion-i00-standard/task.toml b/tasks/libero-max-goal-akita-black-bowl-on-wooden-cabinet-top-obstacle-insertion-i00-standard/task.toml
index 91a6a1c82a..481eccacd0 100644
--- a/tasks/libero-max-goal-akita-black-bowl-on-wooden-cabinet-top-obstacle-insertion-i00-standard/task.toml
+++ b/tasks/libero-max-goal-akita-black-bowl-on-wooden-cabinet-top-obstacle-insertion-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "robosuite"
robot = "franka-panda"
family = "libero-max/libero_goal/obstacle_insertion"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 195
diff --git a/tasks/libero-max-goal-cream-cheese-on-akita-black-bowl-obstacle-insertion-i00-privileged/task.toml b/tasks/libero-max-goal-cream-cheese-on-akita-black-bowl-obstacle-insertion-i00-privileged/task.toml
index 8c3ba052b8..46c7e7df84 100644
--- a/tasks/libero-max-goal-cream-cheese-on-akita-black-bowl-obstacle-insertion-i00-privileged/task.toml
+++ b/tasks/libero-max-goal-cream-cheese-on-akita-black-bowl-obstacle-insertion-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "robosuite"
robot = "franka-panda"
family = "libero-max/libero_goal/obstacle_insertion"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 195 # LIBERO-MAX's policy seed for the case (the scene is the frozen LIBERO-PRO initial state)
expert_time_estimate_hours = 1.0
diff --git a/tasks/libero-max-goal-cream-cheese-on-akita-black-bowl-obstacle-insertion-i00-standard/README.md b/tasks/libero-max-goal-cream-cheese-on-akita-black-bowl-obstacle-insertion-i00-standard/README.md
index 537a315176..fa8081488c 100644
--- a/tasks/libero-max-goal-cream-cheese-on-akita-black-bowl-obstacle-insertion-i00-standard/README.md
+++ b/tasks/libero-max-goal-cream-cheese-on-akita-black-bowl-obstacle-insertion-i00-standard/README.md
@@ -86,3 +86,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
/reset` ends the episode as `error` (`sim_error = 1`), not `reset` (reachable only when a run grants resets; the
default is none). The scene, instance, pass rule and every verdict reached before are unchanged, so the task keeps
its name.
+- 2026-10-14 (board 0113, from board 0086): the instruction says that a moved object or receptacle jumps alone: what rests in or on it stays where it was (LIBERO-MAX moves one entity; board 0086 saw a basket jump through a can already in it). Text only: the instance, the event, the reference solution, LIBERO's success check and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/libero-max-goal-cream-cheese-on-akita-black-bowl-obstacle-insertion-i00-standard/instruction.md b/tasks/libero-max-goal-cream-cheese-on-akita-black-bowl-obstacle-insertion-i00-standard/instruction.md
index 4e440baa53..10ebe435fb 100644
--- a/tasks/libero-max-goal-cream-cheese-on-akita-black-bowl-obstacle-insertion-i00-standard/instruction.md
+++ b/tasks/libero-max-goal-cream-cheese-on-akita-black-bowl-obstacle-insertion-i00-standard/instruction.md
@@ -6,7 +6,8 @@ simulator runs in a service you cannot open, and you reach it only through a cli
The scene can change while you work: once during the episode, when your gripper first gets close to one of the task's
objects, the scene changes (as in LIBERO-MAX): an object or a receptacle may be moved, an obstacle or a few other
objects may appear, a camera may shift, or the lighting, the colours or the image noise may change. Nothing tells you
-when; the task and its success check stay the same, so keep checking what the cameras show.
+when; the task and its success check stay the same, so keep checking what the cameras show. A moved object or
+receptacle jumps there alone: what already rests in or on it stays where it was.
## Rules (real-robot rules)
- **One episode, no reset.** The episode starts when the service comes up. It ends when the task succeeds, after
diff --git a/tasks/libero-max-goal-cream-cheese-on-akita-black-bowl-obstacle-insertion-i00-standard/task.toml b/tasks/libero-max-goal-cream-cheese-on-akita-black-bowl-obstacle-insertion-i00-standard/task.toml
index 65936ebccd..eba24ac94c 100644
--- a/tasks/libero-max-goal-cream-cheese-on-akita-black-bowl-obstacle-insertion-i00-standard/task.toml
+++ b/tasks/libero-max-goal-cream-cheese-on-akita-black-bowl-obstacle-insertion-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "robosuite"
robot = "franka-panda"
family = "libero-max/libero_goal/obstacle_insertion"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 195
diff --git a/tasks/libero-max-goal-flat-stove-off-visual-theme-switch-i00-privileged/task.toml b/tasks/libero-max-goal-flat-stove-off-visual-theme-switch-i00-privileged/task.toml
index 7974a0a4e1..4c55b54f88 100644
--- a/tasks/libero-max-goal-flat-stove-off-visual-theme-switch-i00-privileged/task.toml
+++ b/tasks/libero-max-goal-flat-stove-off-visual-theme-switch-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "robosuite"
robot = "franka-panda"
family = "libero-max/libero_goal/visual_theme_switch"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 195 # LIBERO-MAX's policy seed for the case (the scene is the frozen LIBERO-PRO initial state)
expert_time_estimate_hours = 1.0
diff --git a/tasks/libero-max-goal-flat-stove-off-visual-theme-switch-i00-standard/README.md b/tasks/libero-max-goal-flat-stove-off-visual-theme-switch-i00-standard/README.md
index f9b8de39d8..50cead89bc 100644
--- a/tasks/libero-max-goal-flat-stove-off-visual-theme-switch-i00-standard/README.md
+++ b/tasks/libero-max-goal-flat-stove-off-visual-theme-switch-i00-standard/README.md
@@ -86,3 +86,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
/reset` ends the episode as `error` (`sim_error = 1`), not `reset` (reachable only when a run grants resets; the
default is none). The scene, instance, pass rule and every verdict reached before are unchanged, so the task keeps
its name.
+- 2026-10-14 (board 0113, from board 0086): the instruction says that a moved object or receptacle jumps alone: what rests in or on it stays where it was (LIBERO-MAX moves one entity; board 0086 saw a basket jump through a can already in it). Text only: the instance, the event, the reference solution, LIBERO's success check and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/libero-max-goal-flat-stove-off-visual-theme-switch-i00-standard/instruction.md b/tasks/libero-max-goal-flat-stove-off-visual-theme-switch-i00-standard/instruction.md
index dc63087bbc..9cd3849dcb 100644
--- a/tasks/libero-max-goal-flat-stove-off-visual-theme-switch-i00-standard/instruction.md
+++ b/tasks/libero-max-goal-flat-stove-off-visual-theme-switch-i00-standard/instruction.md
@@ -6,7 +6,8 @@ simulator runs in a service you cannot open, and you reach it only through a cli
The scene can change while you work: once during the episode, when your gripper first gets close to one of the task's
objects, the scene changes (as in LIBERO-MAX): an object or a receptacle may be moved, an obstacle or a few other
objects may appear, a camera may shift, or the lighting, the colours or the image noise may change. Nothing tells you
-when; the task and its success check stay the same, so keep checking what the cameras show.
+when; the task and its success check stay the same, so keep checking what the cameras show. A moved object or
+receptacle jumps there alone: what already rests in or on it stays where it was.
## Rules (real-robot rules)
- **One episode, no reset.** The episode starts when the service comes up. It ends when the task succeeds, after
diff --git a/tasks/libero-max-goal-flat-stove-off-visual-theme-switch-i00-standard/task.toml b/tasks/libero-max-goal-flat-stove-off-visual-theme-switch-i00-standard/task.toml
index aed9c01df2..d974104204 100644
--- a/tasks/libero-max-goal-flat-stove-off-visual-theme-switch-i00-standard/task.toml
+++ b/tasks/libero-max-goal-flat-stove-off-visual-theme-switch-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "robosuite"
robot = "franka-panda"
family = "libero-max/libero_goal/visual_theme_switch"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 195
diff --git a/tasks/libero-max-goal-flat-stove-on-illumination-switch-i00-privileged/task.toml b/tasks/libero-max-goal-flat-stove-on-illumination-switch-i00-privileged/task.toml
index 95cf42f254..11b2c0cbeb 100644
--- a/tasks/libero-max-goal-flat-stove-on-illumination-switch-i00-privileged/task.toml
+++ b/tasks/libero-max-goal-flat-stove-on-illumination-switch-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "robosuite"
robot = "franka-panda"
family = "libero-max/libero_goal/illumination_switch"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 195 # LIBERO-MAX's policy seed for the case (the scene is the frozen LIBERO-PRO initial state)
expert_time_estimate_hours = 1.0
diff --git a/tasks/libero-max-goal-flat-stove-on-illumination-switch-i00-standard/README.md b/tasks/libero-max-goal-flat-stove-on-illumination-switch-i00-standard/README.md
index 9333bf1c47..7c63b71c16 100644
--- a/tasks/libero-max-goal-flat-stove-on-illumination-switch-i00-standard/README.md
+++ b/tasks/libero-max-goal-flat-stove-on-illumination-switch-i00-standard/README.md
@@ -86,3 +86,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
/reset` ends the episode as `error` (`sim_error = 1`), not `reset` (reachable only when a run grants resets; the
default is none). The scene, instance, pass rule and every verdict reached before are unchanged, so the task keeps
its name.
+- 2026-10-14 (board 0113, from board 0086): the instruction says that a moved object or receptacle jumps alone: what rests in or on it stays where it was (LIBERO-MAX moves one entity; board 0086 saw a basket jump through a can already in it). Text only: the instance, the event, the reference solution, LIBERO's success check and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/libero-max-goal-flat-stove-on-illumination-switch-i00-standard/instruction.md b/tasks/libero-max-goal-flat-stove-on-illumination-switch-i00-standard/instruction.md
index ed98fc8c92..d8018f7cbe 100644
--- a/tasks/libero-max-goal-flat-stove-on-illumination-switch-i00-standard/instruction.md
+++ b/tasks/libero-max-goal-flat-stove-on-illumination-switch-i00-standard/instruction.md
@@ -6,7 +6,8 @@ simulator runs in a service you cannot open, and you reach it only through a cli
The scene can change while you work: once during the episode, when your gripper first gets close to one of the task's
objects, the scene changes (as in LIBERO-MAX): an object or a receptacle may be moved, an obstacle or a few other
objects may appear, a camera may shift, or the lighting, the colours or the image noise may change. Nothing tells you
-when; the task and its success check stay the same, so keep checking what the cameras show.
+when; the task and its success check stay the same, so keep checking what the cameras show. A moved object or
+receptacle jumps there alone: what already rests in or on it stays where it was.
## Rules (real-robot rules)
- **One episode, no reset.** The episode starts when the service comes up. It ends when the task succeeds, after
diff --git a/tasks/libero-max-goal-flat-stove-on-illumination-switch-i00-standard/task.toml b/tasks/libero-max-goal-flat-stove-on-illumination-switch-i00-standard/task.toml
index a2a2276b3e..acf395b841 100644
--- a/tasks/libero-max-goal-flat-stove-on-illumination-switch-i00-standard/task.toml
+++ b/tasks/libero-max-goal-flat-stove-on-illumination-switch-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "robosuite"
robot = "franka-panda"
family = "libero-max/libero_goal/illumination_switch"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 195
diff --git a/tasks/libero-max-goal-flat-stove-on-illumination-switch-i01-privileged/task.toml b/tasks/libero-max-goal-flat-stove-on-illumination-switch-i01-privileged/task.toml
index 7109924841..4fed9cab00 100644
--- a/tasks/libero-max-goal-flat-stove-on-illumination-switch-i01-privileged/task.toml
+++ b/tasks/libero-max-goal-flat-stove-on-illumination-switch-i01-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "robosuite"
robot = "franka-panda"
family = "libero-max/libero_goal/illumination_switch"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 195 # LIBERO-MAX's policy seed for the case (the scene is the frozen LIBERO-PRO initial state)
expert_time_estimate_hours = 1.0
diff --git a/tasks/libero-max-goal-flat-stove-on-illumination-switch-i01-standard/README.md b/tasks/libero-max-goal-flat-stove-on-illumination-switch-i01-standard/README.md
index 6b9e825594..346f813a1f 100644
--- a/tasks/libero-max-goal-flat-stove-on-illumination-switch-i01-standard/README.md
+++ b/tasks/libero-max-goal-flat-stove-on-illumination-switch-i01-standard/README.md
@@ -86,3 +86,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
/reset` ends the episode as `error` (`sim_error = 1`), not `reset` (reachable only when a run grants resets; the
default is none). The scene, instance, pass rule and every verdict reached before are unchanged, so the task keeps
its name.
+- 2026-10-14 (board 0113, from board 0086): the instruction says that a moved object or receptacle jumps alone: what rests in or on it stays where it was (LIBERO-MAX moves one entity; board 0086 saw a basket jump through a can already in it). Text only: the instance, the event, the reference solution, LIBERO's success check and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/libero-max-goal-flat-stove-on-illumination-switch-i01-standard/instruction.md b/tasks/libero-max-goal-flat-stove-on-illumination-switch-i01-standard/instruction.md
index c8ddfb005c..b6abcf2aa0 100644
--- a/tasks/libero-max-goal-flat-stove-on-illumination-switch-i01-standard/instruction.md
+++ b/tasks/libero-max-goal-flat-stove-on-illumination-switch-i01-standard/instruction.md
@@ -6,7 +6,8 @@ simulator runs in a service you cannot open, and you reach it only through a cli
The scene can change while you work: once during the episode, when your gripper first gets close to one of the task's
objects, the scene changes (as in LIBERO-MAX): an object or a receptacle may be moved, an obstacle or a few other
objects may appear, a camera may shift, or the lighting, the colours or the image noise may change. Nothing tells you
-when; the task and its success check stay the same, so keep checking what the cameras show.
+when; the task and its success check stay the same, so keep checking what the cameras show. A moved object or
+receptacle jumps there alone: what already rests in or on it stays where it was.
## Rules (real-robot rules)
- **One episode, no reset.** The episode starts when the service comes up. It ends when the task succeeds, after
diff --git a/tasks/libero-max-goal-flat-stove-on-illumination-switch-i01-standard/task.toml b/tasks/libero-max-goal-flat-stove-on-illumination-switch-i01-standard/task.toml
index 099a3f14f7..6a9e3059f9 100644
--- a/tasks/libero-max-goal-flat-stove-on-illumination-switch-i01-standard/task.toml
+++ b/tasks/libero-max-goal-flat-stove-on-illumination-switch-i01-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "robosuite"
robot = "franka-panda"
family = "libero-max/libero_goal/illumination_switch"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 195
diff --git a/tasks/libero-max-goal-flat-stove-on-sensor-noise-onset-i00-privileged/task.toml b/tasks/libero-max-goal-flat-stove-on-sensor-noise-onset-i00-privileged/task.toml
index 877997dd34..5a6b62fb1c 100644
--- a/tasks/libero-max-goal-flat-stove-on-sensor-noise-onset-i00-privileged/task.toml
+++ b/tasks/libero-max-goal-flat-stove-on-sensor-noise-onset-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "robosuite"
robot = "franka-panda"
family = "libero-max/libero_goal/sensor_noise_onset"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 195 # LIBERO-MAX's policy seed for the case (the scene is the frozen LIBERO-PRO initial state)
expert_time_estimate_hours = 1.0
diff --git a/tasks/libero-max-goal-flat-stove-on-sensor-noise-onset-i00-standard/README.md b/tasks/libero-max-goal-flat-stove-on-sensor-noise-onset-i00-standard/README.md
index b18940b4db..dd0aa83ce0 100644
--- a/tasks/libero-max-goal-flat-stove-on-sensor-noise-onset-i00-standard/README.md
+++ b/tasks/libero-max-goal-flat-stove-on-sensor-noise-onset-i00-standard/README.md
@@ -86,3 +86,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
/reset` ends the episode as `error` (`sim_error = 1`), not `reset` (reachable only when a run grants resets; the
default is none). The scene, instance, pass rule and every verdict reached before are unchanged, so the task keeps
its name.
+- 2026-10-14 (board 0113, from board 0086): the instruction says that a moved object or receptacle jumps alone: what rests in or on it stays where it was (LIBERO-MAX moves one entity; board 0086 saw a basket jump through a can already in it). Text only: the instance, the event, the reference solution, LIBERO's success check and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/libero-max-goal-flat-stove-on-sensor-noise-onset-i00-standard/instruction.md b/tasks/libero-max-goal-flat-stove-on-sensor-noise-onset-i00-standard/instruction.md
index 708dd3dfbd..3df65c652a 100644
--- a/tasks/libero-max-goal-flat-stove-on-sensor-noise-onset-i00-standard/instruction.md
+++ b/tasks/libero-max-goal-flat-stove-on-sensor-noise-onset-i00-standard/instruction.md
@@ -6,7 +6,8 @@ simulator runs in a service you cannot open, and you reach it only through a cli
The scene can change while you work: once during the episode, when your gripper first gets close to one of the task's
objects, the scene changes (as in LIBERO-MAX): an object or a receptacle may be moved, an obstacle or a few other
objects may appear, a camera may shift, or the lighting, the colours or the image noise may change. Nothing tells you
-when; the task and its success check stay the same, so keep checking what the cameras show.
+when; the task and its success check stay the same, so keep checking what the cameras show. A moved object or
+receptacle jumps there alone: what already rests in or on it stays where it was.
## Rules (real-robot rules)
- **One episode, no reset.** The episode starts when the service comes up. It ends when the task succeeds, after
diff --git a/tasks/libero-max-goal-flat-stove-on-sensor-noise-onset-i00-standard/task.toml b/tasks/libero-max-goal-flat-stove-on-sensor-noise-onset-i00-standard/task.toml
index 1c74d8ed56..786ad60042 100644
--- a/tasks/libero-max-goal-flat-stove-on-sensor-noise-onset-i00-standard/task.toml
+++ b/tasks/libero-max-goal-flat-stove-on-sensor-noise-onset-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "robosuite"
robot = "franka-panda"
family = "libero-max/libero_goal/sensor_noise_onset"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 195
diff --git a/tasks/libero-max-goal-wine-bottle-on-plate-receptacle-relocation-i00-privileged/README.md b/tasks/libero-max-goal-wine-bottle-on-plate-receptacle-relocation-i00-privileged/README.md
index cb772a6071..3ad0a72e80 100644
--- a/tasks/libero-max-goal-wine-bottle-on-plate-receptacle-relocation-i00-privileged/README.md
+++ b/tasks/libero-max-goal-wine-bottle-on-plate-receptacle-relocation-i00-privileged/README.md
@@ -14,7 +14,7 @@
## Task
LIBERO-MAX case `pro-task-goal-t08-i02-receptacle_relocation-d1-p195`: a Franka Panda arm at a table, instructed **"Put the wine bottle on the plate."**, on a LIBERO-PRO variant
of a libero_goal task (task). Online event (receptacle relocation): the first time the gripper comes within
-0.18 m of `wine_bottle_1`, LIBERO-MAX changes the scene once (the receptacle `plate_1` is moved 0.12 m, direction (0.707, 0.707)). The deliverable is one
+0.18 m of `wine_bottle_1`, LIBERO-MAX changes the scene once (the receptacle `plate_1` is moved 0.12 m, direction (0.707, 0.707); it jumps there alone: what already rests in or on it stays where it was, and where it lands on another object the physics pushes the two apart). The deliverable is one
replayable trajectory of LIBERO's OSC control steps (`/app/output/trajectory.npz`, `(T, 7)`, T ≤ 300), graded by
LIBERO's success check at its end.
@@ -54,7 +54,7 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
|---|---|
| Upstream | LIBERO-PRO @ 2b910b5 (MIT; task data `zhouxueyang/LIBERO-Pro` @ c86fc3b, CC-BY-4.0); LIBERO-MAX @ a1e3cef (event runtime and this case's parameters; no licence file upstream); robosuite 1.4.0, MuJoCo 3.2.6 |
| Case | `pro-task-goal-t08-i02-receptacle_relocation-d1-p195` (LIBERO-MAX 8,000-case manifest; LIBERO-PRO category `task`, task 8, initial state 2) |
-| Event | `receptacle_relocation`: the receptacle `plate_1` is moved 0.12 m, direction (0.707, 0.707); trigger: gripper within 0.18 m of `wine_bottle_1` |
+| Event | `receptacle_relocation`: the receptacle `plate_1` is moved 0.12 m, direction (0.707, 0.707); it jumps there alone: what already rests in or on it stays where it was, and where it lands on another object the physics pushes the two apart; trigger: gripper within 0.18 m of `wine_bottle_1` |
| Actions | 7 per control step (1/20 s): OSC hand-target delta (1.0 = 0.05 m, 0.5 rad) and gripper |
| Resources | 4 CPUs, no GPU (CPU physics, OSMesa rendering) |
@@ -76,3 +76,4 @@ Generated by `scripts/libero-max/generate.py` from `scripts/libero-max/cases.jso
it executes any action). Before, such a trial read as an agent failure with `grader_error = 1` only.
- A replay that records a video prints its verdict before it finishes the video, and a failing video no longer fails
the replay; `video_rendered` is 1 only when the video was finished.
+- 2026-10-14 (board 0113, from board 0086): the instruction says that a moved receptacle jumps alone: what rests in or on it stays where it was (LIBERO-MAX moves one entity; board 0086 saw a basket jump through a can already in it). Text only: the instance, the event, the reference solution, LIBERO's success check and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/libero-max-goal-wine-bottle-on-plate-receptacle-relocation-i00-privileged/instruction.md b/tasks/libero-max-goal-wine-bottle-on-plate-receptacle-relocation-i00-privileged/instruction.md
index 8e4f36acc6..43593f8be0 100644
--- a/tasks/libero-max-goal-wine-bottle-on-plate-receptacle-relocation-i00-privileged/instruction.md
+++ b/tasks/libero-max-goal-wine-bottle-on-plate-receptacle-relocation-i00-privileged/instruction.md
@@ -3,7 +3,7 @@
A **Franka Panda** arm (LIBERO on robosuite / MuJoCo) at a table. **Task: Put the wine bottle on the plate.**
The scene is one frozen LIBERO-MAX case: a LIBERO-PRO variant of a libero_goal task (task), with one online
event: the first time the gripper comes within 0.18 m of `wine_bottle_1`, LIBERO-MAX changes the scene
-once (the receptacle `plate_1` is moved 0.12 m, direction (0.707, 0.707)). You have the full simulator state, so you can see the change happen (`s.state()["event_fired"]`).
+once (the receptacle `plate_1` is moved 0.12 m, direction (0.707, 0.707); it jumps there alone: what already rests in or on it stays where it was, and where it lands on another object the physics pushes the two apart). You have the full simulator state, so you can see the change happen (`s.state()["event_fired"]`).
## What you hand in
`/app/output/trajectory.npz` with key `actions`: LIBERO's native control steps, shape `(T, 7)`, 1 ≤ T ≤ 300
diff --git a/tasks/libero-max-goal-wine-bottle-on-plate-receptacle-relocation-i00-privileged/task.toml b/tasks/libero-max-goal-wine-bottle-on-plate-receptacle-relocation-i00-privileged/task.toml
index 8fd722e3fe..3a2ca1af9d 100644
--- a/tasks/libero-max-goal-wine-bottle-on-plate-receptacle-relocation-i00-privileged/task.toml
+++ b/tasks/libero-max-goal-wine-bottle-on-plate-receptacle-relocation-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "robosuite"
robot = "franka-panda"
family = "libero-max/libero_goal/receptacle_relocation"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 195 # LIBERO-MAX's policy seed for the case (the scene is the frozen LIBERO-PRO initial state)
expert_time_estimate_hours = 1.0
diff --git a/tasks/libero-max-goal-wine-bottle-on-plate-receptacle-relocation-i00-standard/README.md b/tasks/libero-max-goal-wine-bottle-on-plate-receptacle-relocation-i00-standard/README.md
index 16c7e37360..ed8a5b697b 100644
--- a/tasks/libero-max-goal-wine-bottle-on-plate-receptacle-relocation-i00-standard/README.md
+++ b/tasks/libero-max-goal-wine-bottle-on-plate-receptacle-relocation-i00-standard/README.md
@@ -13,7 +13,7 @@
## Task
LIBERO-MAX case `pro-task-goal-t08-i02-receptacle_relocation-d1-p195` as a robot service: a Franka Panda arm at a table, instructed **"Put the wine bottle on the plate."**, on
-the same frozen instance as `libero-max-goal-wine-bottle-on-plate-receptacle-relocation-i00-privileged`, with the same online event (receptacle relocation: the receptacle `plate_1` is moved 0.12 m, direction (0.707, 0.707), the
+the same frozen instance as `libero-max-goal-wine-bottle-on-plate-receptacle-relocation-i00-privileged`, with the same online event (receptacle relocation: the receptacle `plate_1` is moved 0.12 m, direction (0.707, 0.707); it jumps there alone: what already rests in or on it stays where it was, and where it lands on another object the physics pushes the two apart, the
first time the gripper comes within 0.18 m of `wine_bottle_1`). The simulator runs in the `sim` sidecar
(`environment/sim/server.py` on main's `rcb_service.py`); the agent only has the client: two cameras (RGB
256×256, depth on request, calibrated), the robot's own state, and OSC control steps in chunks. One
@@ -86,3 +86,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
/reset` ends the episode as `error` (`sim_error = 1`), not `reset` (reachable only when a run grants resets; the
default is none). The scene, instance, pass rule and every verdict reached before are unchanged, so the task keeps
its name.
+- 2026-10-14 (board 0113, from board 0086): the instruction says that a moved object or receptacle jumps alone: what rests in or on it stays where it was (LIBERO-MAX moves one entity; board 0086 saw a basket jump through a can already in it). Text only: the instance, the event, the reference solution, LIBERO's success check and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/libero-max-goal-wine-bottle-on-plate-receptacle-relocation-i00-standard/instruction.md b/tasks/libero-max-goal-wine-bottle-on-plate-receptacle-relocation-i00-standard/instruction.md
index a22515e032..be9fa6610c 100644
--- a/tasks/libero-max-goal-wine-bottle-on-plate-receptacle-relocation-i00-standard/instruction.md
+++ b/tasks/libero-max-goal-wine-bottle-on-plate-receptacle-relocation-i00-standard/instruction.md
@@ -6,7 +6,8 @@ simulator runs in a service you cannot open, and you reach it only through a cli
The scene can change while you work: once during the episode, when your gripper first gets close to one of the task's
objects, the scene changes (as in LIBERO-MAX): an object or a receptacle may be moved, an obstacle or a few other
objects may appear, a camera may shift, or the lighting, the colours or the image noise may change. Nothing tells you
-when; the task and its success check stay the same, so keep checking what the cameras show.
+when; the task and its success check stay the same, so keep checking what the cameras show. A moved object or
+receptacle jumps there alone: what already rests in or on it stays where it was.
## Rules (real-robot rules)
- **One episode, no reset.** The episode starts when the service comes up. It ends when the task succeeds, after
diff --git a/tasks/libero-max-goal-wine-bottle-on-plate-receptacle-relocation-i00-standard/task.toml b/tasks/libero-max-goal-wine-bottle-on-plate-receptacle-relocation-i00-standard/task.toml
index 03973dc552..c65b8992e3 100644
--- a/tasks/libero-max-goal-wine-bottle-on-plate-receptacle-relocation-i00-standard/task.toml
+++ b/tasks/libero-max-goal-wine-bottle-on-plate-receptacle-relocation-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "robosuite"
robot = "franka-panda"
family = "libero-max/libero_goal/receptacle_relocation"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 195
diff --git a/tasks/libero-max-goal-wine-bottle-on-wine-rack-top-camera-shift-i00-privileged/task.toml b/tasks/libero-max-goal-wine-bottle-on-wine-rack-top-camera-shift-i00-privileged/task.toml
index 8bb38c1f6d..43075e714c 100644
--- a/tasks/libero-max-goal-wine-bottle-on-wine-rack-top-camera-shift-i00-privileged/task.toml
+++ b/tasks/libero-max-goal-wine-bottle-on-wine-rack-top-camera-shift-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "robosuite"
robot = "franka-panda"
family = "libero-max/libero_goal/camera_shift"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 195 # LIBERO-MAX's policy seed for the case (the scene is the frozen LIBERO-PRO initial state)
expert_time_estimate_hours = 1.0
diff --git a/tasks/libero-max-goal-wine-bottle-on-wine-rack-top-camera-shift-i00-standard/README.md b/tasks/libero-max-goal-wine-bottle-on-wine-rack-top-camera-shift-i00-standard/README.md
index 850ade9f45..4997342e70 100644
--- a/tasks/libero-max-goal-wine-bottle-on-wine-rack-top-camera-shift-i00-standard/README.md
+++ b/tasks/libero-max-goal-wine-bottle-on-wine-rack-top-camera-shift-i00-standard/README.md
@@ -86,3 +86,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
/reset` ends the episode as `error` (`sim_error = 1`), not `reset` (reachable only when a run grants resets; the
default is none). The scene, instance, pass rule and every verdict reached before are unchanged, so the task keeps
its name.
+- 2026-10-14 (board 0113, from board 0086): the instruction says that a moved object or receptacle jumps alone: what rests in or on it stays where it was (LIBERO-MAX moves one entity; board 0086 saw a basket jump through a can already in it). Text only: the instance, the event, the reference solution, LIBERO's success check and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/libero-max-goal-wine-bottle-on-wine-rack-top-camera-shift-i00-standard/instruction.md b/tasks/libero-max-goal-wine-bottle-on-wine-rack-top-camera-shift-i00-standard/instruction.md
index 56ed65e4de..67585380d6 100644
--- a/tasks/libero-max-goal-wine-bottle-on-wine-rack-top-camera-shift-i00-standard/instruction.md
+++ b/tasks/libero-max-goal-wine-bottle-on-wine-rack-top-camera-shift-i00-standard/instruction.md
@@ -6,7 +6,8 @@ simulator runs in a service you cannot open, and you reach it only through a cli
The scene can change while you work: once during the episode, when your gripper first gets close to one of the task's
objects, the scene changes (as in LIBERO-MAX): an object or a receptacle may be moved, an obstacle or a few other
objects may appear, a camera may shift, or the lighting, the colours or the image noise may change. Nothing tells you
-when; the task and its success check stay the same, so keep checking what the cameras show.
+when; the task and its success check stay the same, so keep checking what the cameras show. A moved object or
+receptacle jumps there alone: what already rests in or on it stays where it was.
## Rules (real-robot rules)
- **One episode, no reset.** The episode starts when the service comes up. It ends when the task succeeds, after
diff --git a/tasks/libero-max-goal-wine-bottle-on-wine-rack-top-camera-shift-i00-standard/task.toml b/tasks/libero-max-goal-wine-bottle-on-wine-rack-top-camera-shift-i00-standard/task.toml
index a288b8297f..1c2732409d 100644
--- a/tasks/libero-max-goal-wine-bottle-on-wine-rack-top-camera-shift-i00-standard/task.toml
+++ b/tasks/libero-max-goal-wine-bottle-on-wine-rack-top-camera-shift-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "robosuite"
robot = "franka-panda"
family = "libero-max/libero_goal/camera_shift"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 195
diff --git a/tasks/libero-max-goal-wine-bottle-on-wine-rack-top-obstacle-insertion-i00-privileged/task.toml b/tasks/libero-max-goal-wine-bottle-on-wine-rack-top-obstacle-insertion-i00-privileged/task.toml
index a9eeb7ac07..5a983637f4 100644
--- a/tasks/libero-max-goal-wine-bottle-on-wine-rack-top-obstacle-insertion-i00-privileged/task.toml
+++ b/tasks/libero-max-goal-wine-bottle-on-wine-rack-top-obstacle-insertion-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "robosuite"
robot = "franka-panda"
family = "libero-max/libero_goal/obstacle_insertion"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 195 # LIBERO-MAX's policy seed for the case (the scene is the frozen LIBERO-PRO initial state)
expert_time_estimate_hours = 1.0
diff --git a/tasks/libero-max-goal-wine-bottle-on-wine-rack-top-obstacle-insertion-i00-standard/README.md b/tasks/libero-max-goal-wine-bottle-on-wine-rack-top-obstacle-insertion-i00-standard/README.md
index 4453287cf3..43ec6766e4 100644
--- a/tasks/libero-max-goal-wine-bottle-on-wine-rack-top-obstacle-insertion-i00-standard/README.md
+++ b/tasks/libero-max-goal-wine-bottle-on-wine-rack-top-obstacle-insertion-i00-standard/README.md
@@ -86,3 +86,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
/reset` ends the episode as `error` (`sim_error = 1`), not `reset` (reachable only when a run grants resets; the
default is none). The scene, instance, pass rule and every verdict reached before are unchanged, so the task keeps
its name.
+- 2026-10-14 (board 0113, from board 0086): the instruction says that a moved object or receptacle jumps alone: what rests in or on it stays where it was (LIBERO-MAX moves one entity; board 0086 saw a basket jump through a can already in it). Text only: the instance, the event, the reference solution, LIBERO's success check and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/libero-max-goal-wine-bottle-on-wine-rack-top-obstacle-insertion-i00-standard/instruction.md b/tasks/libero-max-goal-wine-bottle-on-wine-rack-top-obstacle-insertion-i00-standard/instruction.md
index d84c5cb927..45bbe198d7 100644
--- a/tasks/libero-max-goal-wine-bottle-on-wine-rack-top-obstacle-insertion-i00-standard/instruction.md
+++ b/tasks/libero-max-goal-wine-bottle-on-wine-rack-top-obstacle-insertion-i00-standard/instruction.md
@@ -6,7 +6,8 @@ simulator runs in a service you cannot open, and you reach it only through a cli
The scene can change while you work: once during the episode, when your gripper first gets close to one of the task's
objects, the scene changes (as in LIBERO-MAX): an object or a receptacle may be moved, an obstacle or a few other
objects may appear, a camera may shift, or the lighting, the colours or the image noise may change. Nothing tells you
-when; the task and its success check stay the same, so keep checking what the cameras show.
+when; the task and its success check stay the same, so keep checking what the cameras show. A moved object or
+receptacle jumps there alone: what already rests in or on it stays where it was.
## Rules (real-robot rules)
- **One episode, no reset.** The episode starts when the service comes up. It ends when the task succeeds, after
diff --git a/tasks/libero-max-goal-wine-bottle-on-wine-rack-top-obstacle-insertion-i00-standard/task.toml b/tasks/libero-max-goal-wine-bottle-on-wine-rack-top-obstacle-insertion-i00-standard/task.toml
index ab51d26668..a337043d9b 100644
--- a/tasks/libero-max-goal-wine-bottle-on-wine-rack-top-obstacle-insertion-i00-standard/task.toml
+++ b/tasks/libero-max-goal-wine-bottle-on-wine-rack-top-obstacle-insertion-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "robosuite"
robot = "franka-panda"
family = "libero-max/libero_goal/obstacle_insertion"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 195
diff --git a/tasks/libero-max-goal-wine-bottle-on-wine-rack-top-visual-theme-switch-i00-privileged/task.toml b/tasks/libero-max-goal-wine-bottle-on-wine-rack-top-visual-theme-switch-i00-privileged/task.toml
index bef65d0cd6..bd9326295a 100644
--- a/tasks/libero-max-goal-wine-bottle-on-wine-rack-top-visual-theme-switch-i00-privileged/task.toml
+++ b/tasks/libero-max-goal-wine-bottle-on-wine-rack-top-visual-theme-switch-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "robosuite"
robot = "franka-panda"
family = "libero-max/libero_goal/visual_theme_switch"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 195 # LIBERO-MAX's policy seed for the case (the scene is the frozen LIBERO-PRO initial state)
expert_time_estimate_hours = 1.0
diff --git a/tasks/libero-max-goal-wine-bottle-on-wine-rack-top-visual-theme-switch-i00-standard/README.md b/tasks/libero-max-goal-wine-bottle-on-wine-rack-top-visual-theme-switch-i00-standard/README.md
index 5d0c855217..87d9d749e1 100644
--- a/tasks/libero-max-goal-wine-bottle-on-wine-rack-top-visual-theme-switch-i00-standard/README.md
+++ b/tasks/libero-max-goal-wine-bottle-on-wine-rack-top-visual-theme-switch-i00-standard/README.md
@@ -86,3 +86,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
/reset` ends the episode as `error` (`sim_error = 1`), not `reset` (reachable only when a run grants resets; the
default is none). The scene, instance, pass rule and every verdict reached before are unchanged, so the task keeps
its name.
+- 2026-10-14 (board 0113, from board 0086): the instruction says that a moved object or receptacle jumps alone: what rests in or on it stays where it was (LIBERO-MAX moves one entity; board 0086 saw a basket jump through a can already in it). Text only: the instance, the event, the reference solution, LIBERO's success check and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/libero-max-goal-wine-bottle-on-wine-rack-top-visual-theme-switch-i00-standard/instruction.md b/tasks/libero-max-goal-wine-bottle-on-wine-rack-top-visual-theme-switch-i00-standard/instruction.md
index da7de489a8..b7dd83f554 100644
--- a/tasks/libero-max-goal-wine-bottle-on-wine-rack-top-visual-theme-switch-i00-standard/instruction.md
+++ b/tasks/libero-max-goal-wine-bottle-on-wine-rack-top-visual-theme-switch-i00-standard/instruction.md
@@ -6,7 +6,8 @@ simulator runs in a service you cannot open, and you reach it only through a cli
The scene can change while you work: once during the episode, when your gripper first gets close to one of the task's
objects, the scene changes (as in LIBERO-MAX): an object or a receptacle may be moved, an obstacle or a few other
objects may appear, a camera may shift, or the lighting, the colours or the image noise may change. Nothing tells you
-when; the task and its success check stay the same, so keep checking what the cameras show.
+when; the task and its success check stay the same, so keep checking what the cameras show. A moved object or
+receptacle jumps there alone: what already rests in or on it stays where it was.
## Rules (real-robot rules)
- **One episode, no reset.** The episode starts when the service comes up. It ends when the task succeeds, after
diff --git a/tasks/libero-max-goal-wine-bottle-on-wine-rack-top-visual-theme-switch-i00-standard/task.toml b/tasks/libero-max-goal-wine-bottle-on-wine-rack-top-visual-theme-switch-i00-standard/task.toml
index 8280710a8c..5597ab3cee 100644
--- a/tasks/libero-max-goal-wine-bottle-on-wine-rack-top-visual-theme-switch-i00-standard/task.toml
+++ b/tasks/libero-max-goal-wine-bottle-on-wine-rack-top-visual-theme-switch-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "robosuite"
robot = "franka-panda"
family = "libero-max/libero_goal/visual_theme_switch"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 195
diff --git a/tasks/libero-max-goal-wine-bottle-on-wooden-cabinet-top-sensor-noise-onset-i00-privileged/task.toml b/tasks/libero-max-goal-wine-bottle-on-wooden-cabinet-top-sensor-noise-onset-i00-privileged/task.toml
index 3ed0504c1a..ccebefe74c 100644
--- a/tasks/libero-max-goal-wine-bottle-on-wooden-cabinet-top-sensor-noise-onset-i00-privileged/task.toml
+++ b/tasks/libero-max-goal-wine-bottle-on-wooden-cabinet-top-sensor-noise-onset-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "robosuite"
robot = "franka-panda"
family = "libero-max/libero_goal/sensor_noise_onset"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 195 # LIBERO-MAX's policy seed for the case (the scene is the frozen LIBERO-PRO initial state)
expert_time_estimate_hours = 1.0
diff --git a/tasks/libero-max-goal-wine-bottle-on-wooden-cabinet-top-sensor-noise-onset-i00-standard/README.md b/tasks/libero-max-goal-wine-bottle-on-wooden-cabinet-top-sensor-noise-onset-i00-standard/README.md
index 339e164650..67821eaacd 100644
--- a/tasks/libero-max-goal-wine-bottle-on-wooden-cabinet-top-sensor-noise-onset-i00-standard/README.md
+++ b/tasks/libero-max-goal-wine-bottle-on-wooden-cabinet-top-sensor-noise-onset-i00-standard/README.md
@@ -86,3 +86,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
/reset` ends the episode as `error` (`sim_error = 1`), not `reset` (reachable only when a run grants resets; the
default is none). The scene, instance, pass rule and every verdict reached before are unchanged, so the task keeps
its name.
+- 2026-10-14 (board 0113, from board 0086): the instruction says that a moved object or receptacle jumps alone: what rests in or on it stays where it was (LIBERO-MAX moves one entity; board 0086 saw a basket jump through a can already in it). Text only: the instance, the event, the reference solution, LIBERO's success check and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/libero-max-goal-wine-bottle-on-wooden-cabinet-top-sensor-noise-onset-i00-standard/instruction.md b/tasks/libero-max-goal-wine-bottle-on-wooden-cabinet-top-sensor-noise-onset-i00-standard/instruction.md
index 85a9254de8..80b8df09f2 100644
--- a/tasks/libero-max-goal-wine-bottle-on-wooden-cabinet-top-sensor-noise-onset-i00-standard/instruction.md
+++ b/tasks/libero-max-goal-wine-bottle-on-wooden-cabinet-top-sensor-noise-onset-i00-standard/instruction.md
@@ -6,7 +6,8 @@ simulator runs in a service you cannot open, and you reach it only through a cli
The scene can change while you work: once during the episode, when your gripper first gets close to one of the task's
objects, the scene changes (as in LIBERO-MAX): an object or a receptacle may be moved, an obstacle or a few other
objects may appear, a camera may shift, or the lighting, the colours or the image noise may change. Nothing tells you
-when; the task and its success check stay the same, so keep checking what the cameras show.
+when; the task and its success check stay the same, so keep checking what the cameras show. A moved object or
+receptacle jumps there alone: what already rests in or on it stays where it was.
## Rules (real-robot rules)
- **One episode, no reset.** The episode starts when the service comes up. It ends when the task succeeds, after
diff --git a/tasks/libero-max-goal-wine-bottle-on-wooden-cabinet-top-sensor-noise-onset-i00-standard/task.toml b/tasks/libero-max-goal-wine-bottle-on-wooden-cabinet-top-sensor-noise-onset-i00-standard/task.toml
index b695d9ae80..2c0d2f8f41 100644
--- a/tasks/libero-max-goal-wine-bottle-on-wooden-cabinet-top-sensor-noise-onset-i00-standard/task.toml
+++ b/tasks/libero-max-goal-wine-bottle-on-wooden-cabinet-top-sensor-noise-onset-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "robosuite"
robot = "franka-panda"
family = "libero-max/libero_goal/sensor_noise_onset"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 195
diff --git a/tasks/libero-max-goal-yellow-stove-on-camera-shift-i00-privileged/task.toml b/tasks/libero-max-goal-yellow-stove-on-camera-shift-i00-privileged/task.toml
index ca26e52e4a..1f53c91840 100644
--- a/tasks/libero-max-goal-yellow-stove-on-camera-shift-i00-privileged/task.toml
+++ b/tasks/libero-max-goal-yellow-stove-on-camera-shift-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "robosuite"
robot = "franka-panda"
family = "libero-max/libero_goal/camera_shift"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 195 # LIBERO-MAX's policy seed for the case (the scene is the frozen LIBERO-PRO initial state)
expert_time_estimate_hours = 1.0
diff --git a/tasks/libero-max-goal-yellow-stove-on-camera-shift-i00-standard/README.md b/tasks/libero-max-goal-yellow-stove-on-camera-shift-i00-standard/README.md
index 25c58d4fb1..ccb055d86d 100644
--- a/tasks/libero-max-goal-yellow-stove-on-camera-shift-i00-standard/README.md
+++ b/tasks/libero-max-goal-yellow-stove-on-camera-shift-i00-standard/README.md
@@ -86,3 +86,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
/reset` ends the episode as `error` (`sim_error = 1`), not `reset` (reachable only when a run grants resets; the
default is none). The scene, instance, pass rule and every verdict reached before are unchanged, so the task keeps
its name.
+- 2026-10-14 (board 0113, from board 0086): the instruction says that a moved object or receptacle jumps alone: what rests in or on it stays where it was (LIBERO-MAX moves one entity; board 0086 saw a basket jump through a can already in it). Text only: the instance, the event, the reference solution, LIBERO's success check and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/libero-max-goal-yellow-stove-on-camera-shift-i00-standard/instruction.md b/tasks/libero-max-goal-yellow-stove-on-camera-shift-i00-standard/instruction.md
index 1ad2dc8cf9..2b19a9fdaa 100644
--- a/tasks/libero-max-goal-yellow-stove-on-camera-shift-i00-standard/instruction.md
+++ b/tasks/libero-max-goal-yellow-stove-on-camera-shift-i00-standard/instruction.md
@@ -6,7 +6,8 @@ simulator runs in a service you cannot open, and you reach it only through a cli
The scene can change while you work: once during the episode, when your gripper first gets close to one of the task's
objects, the scene changes (as in LIBERO-MAX): an object or a receptacle may be moved, an obstacle or a few other
objects may appear, a camera may shift, or the lighting, the colours or the image noise may change. Nothing tells you
-when; the task and its success check stay the same, so keep checking what the cameras show.
+when; the task and its success check stay the same, so keep checking what the cameras show. A moved object or
+receptacle jumps there alone: what already rests in or on it stays where it was.
## Rules (real-robot rules)
- **One episode, no reset.** The episode starts when the service comes up. It ends when the task succeeds, after
diff --git a/tasks/libero-max-goal-yellow-stove-on-camera-shift-i00-standard/task.toml b/tasks/libero-max-goal-yellow-stove-on-camera-shift-i00-standard/task.toml
index 3cf5fffac8..270e8a6c9c 100644
--- a/tasks/libero-max-goal-yellow-stove-on-camera-shift-i00-standard/task.toml
+++ b/tasks/libero-max-goal-yellow-stove-on-camera-shift-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "robosuite"
robot = "franka-panda"
family = "libero-max/libero_goal/camera_shift"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 195
diff --git a/tasks/libero-max-long-alphabet-soup-and-cream-cheese-in-basket-receptacle-relocation-i00-privileged/README.md b/tasks/libero-max-long-alphabet-soup-and-cream-cheese-in-basket-receptacle-relocation-i00-privileged/README.md
index 92fcc0c995..abf65fcc1e 100644
--- a/tasks/libero-max-long-alphabet-soup-and-cream-cheese-in-basket-receptacle-relocation-i00-privileged/README.md
+++ b/tasks/libero-max-long-alphabet-soup-and-cream-cheese-in-basket-receptacle-relocation-i00-privileged/README.md
@@ -14,7 +14,7 @@
## Task
LIBERO-MAX case `pro-camera-view-angle-10-t07-i01-receptacle_relocation-d0-p195`: a Franka Panda arm at a table, instructed **"Put both the alphabet soup and the cream cheese box in the basket."**, on a LIBERO-PRO variant
of a libero_10 task (camera_view_angle). Online event (receptacle relocation): the first time the gripper comes within
-0.18 m of `alphabet_soup_1`, LIBERO-MAX changes the scene once (the receptacle `basket_1` is moved 0.06 m, direction (0, -1)). The deliverable is one
+0.18 m of `alphabet_soup_1`, LIBERO-MAX changes the scene once (the receptacle `basket_1` is moved 0.06 m, direction (0, -1); it jumps there alone: what already rests in or on it stays where it was, and where it lands on another object the physics pushes the two apart). The deliverable is one
replayable trajectory of LIBERO's OSC control steps (`/app/output/trajectory.npz`, `(T, 7)`, T ≤ 520), graded by
LIBERO's success check at its end.
@@ -54,7 +54,7 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
|---|---|
| Upstream | LIBERO-PRO @ 2b910b5 (MIT; task data `zhouxueyang/LIBERO-Pro` @ c86fc3b, CC-BY-4.0); LIBERO-MAX @ a1e3cef (event runtime and this case's parameters; no licence file upstream); robosuite 1.4.0, MuJoCo 3.2.6 |
| Case | `pro-camera-view-angle-10-t07-i01-receptacle_relocation-d0-p195` (LIBERO-MAX 8,000-case manifest; LIBERO-PRO category `camera_view_angle`, task 7, initial state 1) |
-| Event | `receptacle_relocation`: the receptacle `basket_1` is moved 0.06 m, direction (0, -1); trigger: gripper within 0.18 m of `alphabet_soup_1` |
+| Event | `receptacle_relocation`: the receptacle `basket_1` is moved 0.06 m, direction (0, -1); it jumps there alone: what already rests in or on it stays where it was, and where it lands on another object the physics pushes the two apart; trigger: gripper within 0.18 m of `alphabet_soup_1` |
| Actions | 7 per control step (1/20 s): OSC hand-target delta (1.0 = 0.05 m, 0.5 rad) and gripper |
| Resources | 4 CPUs, no GPU (CPU physics, OSMesa rendering) |
@@ -76,3 +76,4 @@ Generated by `scripts/libero-max/generate.py` from `scripts/libero-max/cases.jso
it executes any action). Before, such a trial read as an agent failure with `grader_error = 1` only.
- A replay that records a video prints its verdict before it finishes the video, and a failing video no longer fails
the replay; `video_rendered` is 1 only when the video was finished.
+- 2026-10-14 (board 0113, from board 0086): the instruction says that a moved receptacle jumps alone: what rests in or on it stays where it was (LIBERO-MAX moves one entity; board 0086 saw a basket jump through a can already in it). Text only: the instance, the event, the reference solution, LIBERO's success check and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/libero-max-long-alphabet-soup-and-cream-cheese-in-basket-receptacle-relocation-i00-privileged/instruction.md b/tasks/libero-max-long-alphabet-soup-and-cream-cheese-in-basket-receptacle-relocation-i00-privileged/instruction.md
index 1dca753c8a..b399788f11 100644
--- a/tasks/libero-max-long-alphabet-soup-and-cream-cheese-in-basket-receptacle-relocation-i00-privileged/instruction.md
+++ b/tasks/libero-max-long-alphabet-soup-and-cream-cheese-in-basket-receptacle-relocation-i00-privileged/instruction.md
@@ -3,7 +3,7 @@
A **Franka Panda** arm (LIBERO on robosuite / MuJoCo) at a table. **Task: Put both the alphabet soup and the cream cheese box in the basket.**
The scene is one frozen LIBERO-MAX case: a LIBERO-PRO variant of a libero_10 task (camera_view_angle), with one online
event: the first time the gripper comes within 0.18 m of `alphabet_soup_1`, LIBERO-MAX changes the scene
-once (the receptacle `basket_1` is moved 0.06 m, direction (0, -1)). You have the full simulator state, so you can see the change happen (`s.state()["event_fired"]`).
+once (the receptacle `basket_1` is moved 0.06 m, direction (0, -1); it jumps there alone: what already rests in or on it stays where it was, and where it lands on another object the physics pushes the two apart). You have the full simulator state, so you can see the change happen (`s.state()["event_fired"]`).
## What you hand in
`/app/output/trajectory.npz` with key `actions`: LIBERO's native control steps, shape `(T, 7)`, 1 ≤ T ≤ 520
diff --git a/tasks/libero-max-long-alphabet-soup-and-cream-cheese-in-basket-receptacle-relocation-i00-privileged/task.toml b/tasks/libero-max-long-alphabet-soup-and-cream-cheese-in-basket-receptacle-relocation-i00-privileged/task.toml
index 883593758f..4d53a1de0f 100644
--- a/tasks/libero-max-long-alphabet-soup-and-cream-cheese-in-basket-receptacle-relocation-i00-privileged/task.toml
+++ b/tasks/libero-max-long-alphabet-soup-and-cream-cheese-in-basket-receptacle-relocation-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "robosuite"
robot = "franka-panda"
family = "libero-max/libero_10/receptacle_relocation"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 195 # LIBERO-MAX's policy seed for the case (the scene is the frozen LIBERO-PRO initial state)
expert_time_estimate_hours = 1.0
diff --git a/tasks/libero-max-long-alphabet-soup-and-cream-cheese-in-basket-receptacle-relocation-i00-standard/README.md b/tasks/libero-max-long-alphabet-soup-and-cream-cheese-in-basket-receptacle-relocation-i00-standard/README.md
index 5a47a7f32d..408567de4a 100644
--- a/tasks/libero-max-long-alphabet-soup-and-cream-cheese-in-basket-receptacle-relocation-i00-standard/README.md
+++ b/tasks/libero-max-long-alphabet-soup-and-cream-cheese-in-basket-receptacle-relocation-i00-standard/README.md
@@ -13,7 +13,7 @@
## Task
LIBERO-MAX case `pro-camera-view-angle-10-t07-i01-receptacle_relocation-d0-p195` as a robot service: a Franka Panda arm at a table, instructed **"Put both the alphabet soup and the cream cheese box in the basket."**, on
-the same frozen instance as `libero-max-long-alphabet-soup-and-cream-cheese-in-basket-receptacle-relocation-i00-privileged`, with the same online event (receptacle relocation: the receptacle `basket_1` is moved 0.06 m, direction (0, -1), the
+the same frozen instance as `libero-max-long-alphabet-soup-and-cream-cheese-in-basket-receptacle-relocation-i00-privileged`, with the same online event (receptacle relocation: the receptacle `basket_1` is moved 0.06 m, direction (0, -1); it jumps there alone: what already rests in or on it stays where it was, and where it lands on another object the physics pushes the two apart, the
first time the gripper comes within 0.18 m of `alphabet_soup_1`). The simulator runs in the `sim` sidecar
(`environment/sim/server.py` on main's `rcb_service.py`); the agent only has the client: two cameras (RGB
256×256, depth on request, calibrated), the robot's own state, and OSC control steps in chunks. One
@@ -86,3 +86,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
/reset` ends the episode as `error` (`sim_error = 1`), not `reset` (reachable only when a run grants resets; the
default is none). The scene, instance, pass rule and every verdict reached before are unchanged, so the task keeps
its name.
+- 2026-10-14 (board 0113, from board 0086): the instruction says that a moved object or receptacle jumps alone: what rests in or on it stays where it was (LIBERO-MAX moves one entity; board 0086 saw a basket jump through a can already in it). Text only: the instance, the event, the reference solution, LIBERO's success check and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/libero-max-long-alphabet-soup-and-cream-cheese-in-basket-receptacle-relocation-i00-standard/instruction.md b/tasks/libero-max-long-alphabet-soup-and-cream-cheese-in-basket-receptacle-relocation-i00-standard/instruction.md
index 69ecf613e5..ffc868aa6e 100644
--- a/tasks/libero-max-long-alphabet-soup-and-cream-cheese-in-basket-receptacle-relocation-i00-standard/instruction.md
+++ b/tasks/libero-max-long-alphabet-soup-and-cream-cheese-in-basket-receptacle-relocation-i00-standard/instruction.md
@@ -6,7 +6,8 @@ simulator runs in a service you cannot open, and you reach it only through a cli
The scene can change while you work: once during the episode, when your gripper first gets close to one of the task's
objects, the scene changes (as in LIBERO-MAX): an object or a receptacle may be moved, an obstacle or a few other
objects may appear, a camera may shift, or the lighting, the colours or the image noise may change. Nothing tells you
-when; the task and its success check stay the same, so keep checking what the cameras show.
+when; the task and its success check stay the same, so keep checking what the cameras show. A moved object or
+receptacle jumps there alone: what already rests in or on it stays where it was.
## Rules (real-robot rules)
- **One episode, no reset.** The episode starts when the service comes up. It ends when the task succeeds, after
diff --git a/tasks/libero-max-long-alphabet-soup-and-cream-cheese-in-basket-receptacle-relocation-i00-standard/task.toml b/tasks/libero-max-long-alphabet-soup-and-cream-cheese-in-basket-receptacle-relocation-i00-standard/task.toml
index ad20329e69..aff9bb6b0b 100644
--- a/tasks/libero-max-long-alphabet-soup-and-cream-cheese-in-basket-receptacle-relocation-i00-standard/task.toml
+++ b/tasks/libero-max-long-alphabet-soup-and-cream-cheese-in-basket-receptacle-relocation-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "robosuite"
robot = "franka-panda"
family = "libero-max/libero_10/receptacle_relocation"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 195
diff --git a/tasks/libero-max-long-alphabet-soup-and-cream-cheese-in-basket-target-relocation-i00-privileged/task.toml b/tasks/libero-max-long-alphabet-soup-and-cream-cheese-in-basket-target-relocation-i00-privileged/task.toml
index b7e51c401c..ea88b58213 100644
--- a/tasks/libero-max-long-alphabet-soup-and-cream-cheese-in-basket-target-relocation-i00-privileged/task.toml
+++ b/tasks/libero-max-long-alphabet-soup-and-cream-cheese-in-basket-target-relocation-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "robosuite"
robot = "franka-panda"
family = "libero-max/libero_10/target_relocation"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 195 # LIBERO-MAX's policy seed for the case (the scene is the frozen LIBERO-PRO initial state)
expert_time_estimate_hours = 1.0
diff --git a/tasks/libero-max-long-alphabet-soup-and-cream-cheese-in-basket-target-relocation-i00-standard/README.md b/tasks/libero-max-long-alphabet-soup-and-cream-cheese-in-basket-target-relocation-i00-standard/README.md
index 4aa459374c..989953c70c 100644
--- a/tasks/libero-max-long-alphabet-soup-and-cream-cheese-in-basket-target-relocation-i00-standard/README.md
+++ b/tasks/libero-max-long-alphabet-soup-and-cream-cheese-in-basket-target-relocation-i00-standard/README.md
@@ -86,3 +86,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
/reset` ends the episode as `error` (`sim_error = 1`), not `reset` (reachable only when a run grants resets; the
default is none). The scene, instance, pass rule and every verdict reached before are unchanged, so the task keeps
its name.
+- 2026-10-14 (board 0113, from board 0086): the instruction says that a moved object or receptacle jumps alone: what rests in or on it stays where it was (LIBERO-MAX moves one entity; board 0086 saw a basket jump through a can already in it). Text only: the instance, the event, the reference solution, LIBERO's success check and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/libero-max-long-alphabet-soup-and-cream-cheese-in-basket-target-relocation-i00-standard/instruction.md b/tasks/libero-max-long-alphabet-soup-and-cream-cheese-in-basket-target-relocation-i00-standard/instruction.md
index a7752eb781..37d01473f5 100644
--- a/tasks/libero-max-long-alphabet-soup-and-cream-cheese-in-basket-target-relocation-i00-standard/instruction.md
+++ b/tasks/libero-max-long-alphabet-soup-and-cream-cheese-in-basket-target-relocation-i00-standard/instruction.md
@@ -6,7 +6,8 @@ simulator runs in a service you cannot open, and you reach it only through a cli
The scene can change while you work: once during the episode, when your gripper first gets close to one of the task's
objects, the scene changes (as in LIBERO-MAX): an object or a receptacle may be moved, an obstacle or a few other
objects may appear, a camera may shift, or the lighting, the colours or the image noise may change. Nothing tells you
-when; the task and its success check stay the same, so keep checking what the cameras show.
+when; the task and its success check stay the same, so keep checking what the cameras show. A moved object or
+receptacle jumps there alone: what already rests in or on it stays where it was.
## Rules (real-robot rules)
- **One episode, no reset.** The episode starts when the service comes up. It ends when the task succeeds, after
diff --git a/tasks/libero-max-long-alphabet-soup-and-cream-cheese-in-basket-target-relocation-i00-standard/task.toml b/tasks/libero-max-long-alphabet-soup-and-cream-cheese-in-basket-target-relocation-i00-standard/task.toml
index afa062c2a9..ba2c6bdfdc 100644
--- a/tasks/libero-max-long-alphabet-soup-and-cream-cheese-in-basket-target-relocation-i00-standard/task.toml
+++ b/tasks/libero-max-long-alphabet-soup-and-cream-cheese-in-basket-target-relocation-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "robosuite"
robot = "franka-panda"
family = "libero-max/libero_10/target_relocation"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 195
diff --git a/tasks/libero-max-long-alphabet-soup-and-tomato-sauce-in-basket-distractor-burst-i00-privileged/task.toml b/tasks/libero-max-long-alphabet-soup-and-tomato-sauce-in-basket-distractor-burst-i00-privileged/task.toml
index ab5482fc29..a0a117cebd 100644
--- a/tasks/libero-max-long-alphabet-soup-and-tomato-sauce-in-basket-distractor-burst-i00-privileged/task.toml
+++ b/tasks/libero-max-long-alphabet-soup-and-tomato-sauce-in-basket-distractor-burst-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "robosuite"
robot = "franka-panda"
family = "libero-max/libero_10/distractor_burst"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 195 # LIBERO-MAX's policy seed for the case (the scene is the frozen LIBERO-PRO initial state)
expert_time_estimate_hours = 1.0
diff --git a/tasks/libero-max-long-alphabet-soup-and-tomato-sauce-in-basket-distractor-burst-i00-standard/README.md b/tasks/libero-max-long-alphabet-soup-and-tomato-sauce-in-basket-distractor-burst-i00-standard/README.md
index 83ad875b3a..3d374d92b7 100644
--- a/tasks/libero-max-long-alphabet-soup-and-tomato-sauce-in-basket-distractor-burst-i00-standard/README.md
+++ b/tasks/libero-max-long-alphabet-soup-and-tomato-sauce-in-basket-distractor-burst-i00-standard/README.md
@@ -86,3 +86,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
/reset` ends the episode as `error` (`sim_error = 1`), not `reset` (reachable only when a run grants resets; the
default is none). The scene, instance, pass rule and every verdict reached before are unchanged, so the task keeps
its name.
+- 2026-10-14 (board 0113, from board 0086): the instruction says that a moved object or receptacle jumps alone: what rests in or on it stays where it was (LIBERO-MAX moves one entity; board 0086 saw a basket jump through a can already in it). Text only: the instance, the event, the reference solution, LIBERO's success check and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/libero-max-long-alphabet-soup-and-tomato-sauce-in-basket-distractor-burst-i00-standard/instruction.md b/tasks/libero-max-long-alphabet-soup-and-tomato-sauce-in-basket-distractor-burst-i00-standard/instruction.md
index 269d0814df..8cb5365df6 100644
--- a/tasks/libero-max-long-alphabet-soup-and-tomato-sauce-in-basket-distractor-burst-i00-standard/instruction.md
+++ b/tasks/libero-max-long-alphabet-soup-and-tomato-sauce-in-basket-distractor-burst-i00-standard/instruction.md
@@ -6,7 +6,8 @@ simulator runs in a service you cannot open, and you reach it only through a cli
The scene can change while you work: once during the episode, when your gripper first gets close to one of the task's
objects, the scene changes (as in LIBERO-MAX): an object or a receptacle may be moved, an obstacle or a few other
objects may appear, a camera may shift, or the lighting, the colours or the image noise may change. Nothing tells you
-when; the task and its success check stay the same, so keep checking what the cameras show.
+when; the task and its success check stay the same, so keep checking what the cameras show. A moved object or
+receptacle jumps there alone: what already rests in or on it stays where it was.
## Rules (real-robot rules)
- **One episode, no reset.** The episode starts when the service comes up. It ends when the task succeeds, after
diff --git a/tasks/libero-max-long-alphabet-soup-and-tomato-sauce-in-basket-distractor-burst-i00-standard/task.toml b/tasks/libero-max-long-alphabet-soup-and-tomato-sauce-in-basket-distractor-burst-i00-standard/task.toml
index 40c25322a7..6a64aeb87e 100644
--- a/tasks/libero-max-long-alphabet-soup-and-tomato-sauce-in-basket-distractor-burst-i00-standard/task.toml
+++ b/tasks/libero-max-long-alphabet-soup-and-tomato-sauce-in-basket-distractor-burst-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "robosuite"
robot = "franka-panda"
family = "libero-max/libero_10/distractor_burst"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 195
diff --git a/tasks/libero-max-long-alphabet-soup-and-tomato-sauce-in-basket-receptacle-relocation-i00-privileged/README.md b/tasks/libero-max-long-alphabet-soup-and-tomato-sauce-in-basket-receptacle-relocation-i00-privileged/README.md
index 378766485b..3c9419aff0 100644
--- a/tasks/libero-max-long-alphabet-soup-and-tomato-sauce-in-basket-receptacle-relocation-i00-privileged/README.md
+++ b/tasks/libero-max-long-alphabet-soup-and-tomato-sauce-in-basket-receptacle-relocation-i00-privileged/README.md
@@ -14,7 +14,7 @@
## Task
LIBERO-MAX case `pro-initial-pose-position-angle-10-t00-i02-receptacle_relocation-d0-p195`: a Franka Panda arm at a table, instructed **"Put both the alphabet soup and the tomato sauce in the basket."**, on a LIBERO-PRO variant
of a libero_10 task (initial_pose_position_angle). Online event (receptacle relocation): the first time the gripper comes within
-0.18 m of `alphabet_soup_1`, LIBERO-MAX changes the scene once (the receptacle `basket_1` is moved 0.06 m, direction (1, 0)). The deliverable is one
+0.18 m of `alphabet_soup_1`, LIBERO-MAX changes the scene once (the receptacle `basket_1` is moved 0.06 m, direction (1, 0); it jumps there alone: what already rests in or on it stays where it was, and where it lands on another object the physics pushes the two apart). The deliverable is one
replayable trajectory of LIBERO's OSC control steps (`/app/output/trajectory.npz`, `(T, 7)`, T ≤ 520), graded by
LIBERO's success check at its end.
@@ -54,7 +54,7 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
|---|---|
| Upstream | LIBERO-PRO @ 2b910b5 (MIT; task data `zhouxueyang/LIBERO-Pro` @ c86fc3b, CC-BY-4.0); LIBERO-MAX @ a1e3cef (event runtime and this case's parameters; no licence file upstream); robosuite 1.4.0, MuJoCo 3.2.6 |
| Case | `pro-initial-pose-position-angle-10-t00-i02-receptacle_relocation-d0-p195` (LIBERO-MAX 8,000-case manifest; LIBERO-PRO category `initial_pose_position_angle`, task 0, initial state 2) |
-| Event | `receptacle_relocation`: the receptacle `basket_1` is moved 0.06 m, direction (1, 0); trigger: gripper within 0.18 m of `alphabet_soup_1` |
+| Event | `receptacle_relocation`: the receptacle `basket_1` is moved 0.06 m, direction (1, 0); it jumps there alone: what already rests in or on it stays where it was, and where it lands on another object the physics pushes the two apart; trigger: gripper within 0.18 m of `alphabet_soup_1` |
| Actions | 7 per control step (1/20 s): OSC hand-target delta (1.0 = 0.05 m, 0.5 rad) and gripper |
| Resources | 4 CPUs, no GPU (CPU physics, OSMesa rendering) |
@@ -76,3 +76,4 @@ Generated by `scripts/libero-max/generate.py` from `scripts/libero-max/cases.jso
it executes any action). Before, such a trial read as an agent failure with `grader_error = 1` only.
- A replay that records a video prints its verdict before it finishes the video, and a failing video no longer fails
the replay; `video_rendered` is 1 only when the video was finished.
+- 2026-10-14 (board 0113, from board 0086): the instruction says that a moved receptacle jumps alone: what rests in or on it stays where it was (LIBERO-MAX moves one entity; board 0086 saw a basket jump through a can already in it). Text only: the instance, the event, the reference solution, LIBERO's success check and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/libero-max-long-alphabet-soup-and-tomato-sauce-in-basket-receptacle-relocation-i00-privileged/instruction.md b/tasks/libero-max-long-alphabet-soup-and-tomato-sauce-in-basket-receptacle-relocation-i00-privileged/instruction.md
index f4e48ea89c..070b3c73f7 100644
--- a/tasks/libero-max-long-alphabet-soup-and-tomato-sauce-in-basket-receptacle-relocation-i00-privileged/instruction.md
+++ b/tasks/libero-max-long-alphabet-soup-and-tomato-sauce-in-basket-receptacle-relocation-i00-privileged/instruction.md
@@ -3,7 +3,7 @@
A **Franka Panda** arm (LIBERO on robosuite / MuJoCo) at a table. **Task: Put both the alphabet soup and the tomato sauce in the basket.**
The scene is one frozen LIBERO-MAX case: a LIBERO-PRO variant of a libero_10 task (initial_pose_position_angle), with one online
event: the first time the gripper comes within 0.18 m of `alphabet_soup_1`, LIBERO-MAX changes the scene
-once (the receptacle `basket_1` is moved 0.06 m, direction (1, 0)). You have the full simulator state, so you can see the change happen (`s.state()["event_fired"]`).
+once (the receptacle `basket_1` is moved 0.06 m, direction (1, 0); it jumps there alone: what already rests in or on it stays where it was, and where it lands on another object the physics pushes the two apart). You have the full simulator state, so you can see the change happen (`s.state()["event_fired"]`).
## What you hand in
`/app/output/trajectory.npz` with key `actions`: LIBERO's native control steps, shape `(T, 7)`, 1 ≤ T ≤ 520
diff --git a/tasks/libero-max-long-alphabet-soup-and-tomato-sauce-in-basket-receptacle-relocation-i00-privileged/task.toml b/tasks/libero-max-long-alphabet-soup-and-tomato-sauce-in-basket-receptacle-relocation-i00-privileged/task.toml
index 1a140678b6..36a1bf1be2 100644
--- a/tasks/libero-max-long-alphabet-soup-and-tomato-sauce-in-basket-receptacle-relocation-i00-privileged/task.toml
+++ b/tasks/libero-max-long-alphabet-soup-and-tomato-sauce-in-basket-receptacle-relocation-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "robosuite"
robot = "franka-panda"
family = "libero-max/libero_10/receptacle_relocation"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 195 # LIBERO-MAX's policy seed for the case (the scene is the frozen LIBERO-PRO initial state)
expert_time_estimate_hours = 1.0
diff --git a/tasks/libero-max-long-alphabet-soup-and-tomato-sauce-in-basket-receptacle-relocation-i00-standard/README.md b/tasks/libero-max-long-alphabet-soup-and-tomato-sauce-in-basket-receptacle-relocation-i00-standard/README.md
index dd9cb52a78..d411ff2718 100644
--- a/tasks/libero-max-long-alphabet-soup-and-tomato-sauce-in-basket-receptacle-relocation-i00-standard/README.md
+++ b/tasks/libero-max-long-alphabet-soup-and-tomato-sauce-in-basket-receptacle-relocation-i00-standard/README.md
@@ -13,7 +13,7 @@
## Task
LIBERO-MAX case `pro-initial-pose-position-angle-10-t00-i02-receptacle_relocation-d0-p195` as a robot service: a Franka Panda arm at a table, instructed **"Put both the alphabet soup and the tomato sauce in the basket."**, on
-the same frozen instance as `libero-max-long-alphabet-soup-and-tomato-sauce-in-basket-receptacle-relocation-i00-privileged`, with the same online event (receptacle relocation: the receptacle `basket_1` is moved 0.06 m, direction (1, 0), the
+the same frozen instance as `libero-max-long-alphabet-soup-and-tomato-sauce-in-basket-receptacle-relocation-i00-privileged`, with the same online event (receptacle relocation: the receptacle `basket_1` is moved 0.06 m, direction (1, 0); it jumps there alone: what already rests in or on it stays where it was, and where it lands on another object the physics pushes the two apart, the
first time the gripper comes within 0.18 m of `alphabet_soup_1`). The simulator runs in the `sim` sidecar
(`environment/sim/server.py` on main's `rcb_service.py`); the agent only has the client: two cameras (RGB
256×256, depth on request, calibrated), the robot's own state, and OSC control steps in chunks. One
@@ -86,3 +86,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
/reset` ends the episode as `error` (`sim_error = 1`), not `reset` (reachable only when a run grants resets; the
default is none). The scene, instance, pass rule and every verdict reached before are unchanged, so the task keeps
its name.
+- 2026-10-14 (board 0113, from board 0086): the instruction says that a moved object or receptacle jumps alone: what rests in or on it stays where it was (LIBERO-MAX moves one entity; board 0086 saw a basket jump through a can already in it). Text only: the instance, the event, the reference solution, LIBERO's success check and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/libero-max-long-alphabet-soup-and-tomato-sauce-in-basket-receptacle-relocation-i00-standard/instruction.md b/tasks/libero-max-long-alphabet-soup-and-tomato-sauce-in-basket-receptacle-relocation-i00-standard/instruction.md
index 981fda4623..a86e1906f3 100644
--- a/tasks/libero-max-long-alphabet-soup-and-tomato-sauce-in-basket-receptacle-relocation-i00-standard/instruction.md
+++ b/tasks/libero-max-long-alphabet-soup-and-tomato-sauce-in-basket-receptacle-relocation-i00-standard/instruction.md
@@ -6,7 +6,8 @@ simulator runs in a service you cannot open, and you reach it only through a cli
The scene can change while you work: once during the episode, when your gripper first gets close to one of the task's
objects, the scene changes (as in LIBERO-MAX): an object or a receptacle may be moved, an obstacle or a few other
objects may appear, a camera may shift, or the lighting, the colours or the image noise may change. Nothing tells you
-when; the task and its success check stay the same, so keep checking what the cameras show.
+when; the task and its success check stay the same, so keep checking what the cameras show. A moved object or
+receptacle jumps there alone: what already rests in or on it stays where it was.
## Rules (real-robot rules)
- **One episode, no reset.** The episode starts when the service comes up. It ends when the task succeeds, after
diff --git a/tasks/libero-max-long-alphabet-soup-and-tomato-sauce-in-basket-receptacle-relocation-i00-standard/task.toml b/tasks/libero-max-long-alphabet-soup-and-tomato-sauce-in-basket-receptacle-relocation-i00-standard/task.toml
index dff22918aa..2267023641 100644
--- a/tasks/libero-max-long-alphabet-soup-and-tomato-sauce-in-basket-receptacle-relocation-i00-standard/task.toml
+++ b/tasks/libero-max-long-alphabet-soup-and-tomato-sauce-in-basket-receptacle-relocation-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "robosuite"
robot = "franka-panda"
family = "libero-max/libero_10/receptacle_relocation"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 195
diff --git a/tasks/libero-max-long-alphabet-soup-and-tomato-sauce-in-basket-sensor-noise-onset-i00-privileged/task.toml b/tasks/libero-max-long-alphabet-soup-and-tomato-sauce-in-basket-sensor-noise-onset-i00-privileged/task.toml
index dd8f637a9c..d43f7467b8 100644
--- a/tasks/libero-max-long-alphabet-soup-and-tomato-sauce-in-basket-sensor-noise-onset-i00-privileged/task.toml
+++ b/tasks/libero-max-long-alphabet-soup-and-tomato-sauce-in-basket-sensor-noise-onset-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "robosuite"
robot = "franka-panda"
family = "libero-max/libero_10/sensor_noise_onset"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 195 # LIBERO-MAX's policy seed for the case (the scene is the frozen LIBERO-PRO initial state)
expert_time_estimate_hours = 1.0
diff --git a/tasks/libero-max-long-alphabet-soup-and-tomato-sauce-in-basket-sensor-noise-onset-i00-standard/README.md b/tasks/libero-max-long-alphabet-soup-and-tomato-sauce-in-basket-sensor-noise-onset-i00-standard/README.md
index c6f44f8505..9a7fb0198a 100644
--- a/tasks/libero-max-long-alphabet-soup-and-tomato-sauce-in-basket-sensor-noise-onset-i00-standard/README.md
+++ b/tasks/libero-max-long-alphabet-soup-and-tomato-sauce-in-basket-sensor-noise-onset-i00-standard/README.md
@@ -86,3 +86,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
/reset` ends the episode as `error` (`sim_error = 1`), not `reset` (reachable only when a run grants resets; the
default is none). The scene, instance, pass rule and every verdict reached before are unchanged, so the task keeps
its name.
+- 2026-10-14 (board 0113, from board 0086): the instruction says that a moved object or receptacle jumps alone: what rests in or on it stays where it was (LIBERO-MAX moves one entity; board 0086 saw a basket jump through a can already in it). Text only: the instance, the event, the reference solution, LIBERO's success check and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/libero-max-long-alphabet-soup-and-tomato-sauce-in-basket-sensor-noise-onset-i00-standard/instruction.md b/tasks/libero-max-long-alphabet-soup-and-tomato-sauce-in-basket-sensor-noise-onset-i00-standard/instruction.md
index 336cc95c46..64e5dd1a76 100644
--- a/tasks/libero-max-long-alphabet-soup-and-tomato-sauce-in-basket-sensor-noise-onset-i00-standard/instruction.md
+++ b/tasks/libero-max-long-alphabet-soup-and-tomato-sauce-in-basket-sensor-noise-onset-i00-standard/instruction.md
@@ -6,7 +6,8 @@ simulator runs in a service you cannot open, and you reach it only through a cli
The scene can change while you work: once during the episode, when your gripper first gets close to one of the task's
objects, the scene changes (as in LIBERO-MAX): an object or a receptacle may be moved, an obstacle or a few other
objects may appear, a camera may shift, or the lighting, the colours or the image noise may change. Nothing tells you
-when; the task and its success check stay the same, so keep checking what the cameras show.
+when; the task and its success check stay the same, so keep checking what the cameras show. A moved object or
+receptacle jumps there alone: what already rests in or on it stays where it was.
## Rules (real-robot rules)
- **One episode, no reset.** The episode starts when the service comes up. It ends when the task succeeds, after
diff --git a/tasks/libero-max-long-alphabet-soup-and-tomato-sauce-in-basket-sensor-noise-onset-i00-standard/task.toml b/tasks/libero-max-long-alphabet-soup-and-tomato-sauce-in-basket-sensor-noise-onset-i00-standard/task.toml
index 740d1adc04..da1c0fe831 100644
--- a/tasks/libero-max-long-alphabet-soup-and-tomato-sauce-in-basket-sensor-noise-onset-i00-standard/task.toml
+++ b/tasks/libero-max-long-alphabet-soup-and-tomato-sauce-in-basket-sensor-noise-onset-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "robosuite"
robot = "franka-panda"
family = "libero-max/libero_10/sensor_noise_onset"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 195
diff --git a/tasks/libero-max-long-black-bowl-in-wooden-cabinet-bottom-closed-obstacle-insertion-i00-privileged/task.toml b/tasks/libero-max-long-black-bowl-in-wooden-cabinet-bottom-closed-obstacle-insertion-i00-privileged/task.toml
index 0bccf60fa7..75c427e1cb 100644
--- a/tasks/libero-max-long-black-bowl-in-wooden-cabinet-bottom-closed-obstacle-insertion-i00-privileged/task.toml
+++ b/tasks/libero-max-long-black-bowl-in-wooden-cabinet-bottom-closed-obstacle-insertion-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "robosuite"
robot = "franka-panda"
family = "libero-max/libero_10/obstacle_insertion"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 195 # LIBERO-MAX's policy seed for the case (the scene is the frozen LIBERO-PRO initial state)
expert_time_estimate_hours = 1.0
diff --git a/tasks/libero-max-long-black-bowl-in-wooden-cabinet-bottom-closed-obstacle-insertion-i00-standard/README.md b/tasks/libero-max-long-black-bowl-in-wooden-cabinet-bottom-closed-obstacle-insertion-i00-standard/README.md
index af613a67f0..990617822e 100644
--- a/tasks/libero-max-long-black-bowl-in-wooden-cabinet-bottom-closed-obstacle-insertion-i00-standard/README.md
+++ b/tasks/libero-max-long-black-bowl-in-wooden-cabinet-bottom-closed-obstacle-insertion-i00-standard/README.md
@@ -86,3 +86,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
/reset` ends the episode as `error` (`sim_error = 1`), not `reset` (reachable only when a run grants resets; the
default is none). The scene, instance, pass rule and every verdict reached before are unchanged, so the task keeps
its name.
+- 2026-10-14 (board 0113, from board 0086): the instruction says that a moved object or receptacle jumps alone: what rests in or on it stays where it was (LIBERO-MAX moves one entity; board 0086 saw a basket jump through a can already in it). Text only: the instance, the event, the reference solution, LIBERO's success check and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/libero-max-long-black-bowl-in-wooden-cabinet-bottom-closed-obstacle-insertion-i00-standard/instruction.md b/tasks/libero-max-long-black-bowl-in-wooden-cabinet-bottom-closed-obstacle-insertion-i00-standard/instruction.md
index 164e67e275..7e273f3e03 100644
--- a/tasks/libero-max-long-black-bowl-in-wooden-cabinet-bottom-closed-obstacle-insertion-i00-standard/instruction.md
+++ b/tasks/libero-max-long-black-bowl-in-wooden-cabinet-bottom-closed-obstacle-insertion-i00-standard/instruction.md
@@ -6,7 +6,8 @@ simulator runs in a service you cannot open, and you reach it only through a cli
The scene can change while you work: once during the episode, when your gripper first gets close to one of the task's
objects, the scene changes (as in LIBERO-MAX): an object or a receptacle may be moved, an obstacle or a few other
objects may appear, a camera may shift, or the lighting, the colours or the image noise may change. Nothing tells you
-when; the task and its success check stay the same, so keep checking what the cameras show.
+when; the task and its success check stay the same, so keep checking what the cameras show. A moved object or
+receptacle jumps there alone: what already rests in or on it stays where it was.
## Rules (real-robot rules)
- **One episode, no reset.** The episode starts when the service comes up. It ends when the task succeeds, after
diff --git a/tasks/libero-max-long-black-bowl-in-wooden-cabinet-bottom-closed-obstacle-insertion-i00-standard/task.toml b/tasks/libero-max-long-black-bowl-in-wooden-cabinet-bottom-closed-obstacle-insertion-i00-standard/task.toml
index 7a3004cf63..b9366c9866 100644
--- a/tasks/libero-max-long-black-bowl-in-wooden-cabinet-bottom-closed-obstacle-insertion-i00-standard/task.toml
+++ b/tasks/libero-max-long-black-bowl-in-wooden-cabinet-bottom-closed-obstacle-insertion-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "robosuite"
robot = "franka-panda"
family = "libero-max/libero_10/obstacle_insertion"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 195
diff --git a/tasks/libero-max-long-black-bowl-in-wooden-cabinet-bottom-closed-obstacle-insertion-i01-privileged/task.toml b/tasks/libero-max-long-black-bowl-in-wooden-cabinet-bottom-closed-obstacle-insertion-i01-privileged/task.toml
index ff55dbd97f..08cc2851e3 100644
--- a/tasks/libero-max-long-black-bowl-in-wooden-cabinet-bottom-closed-obstacle-insertion-i01-privileged/task.toml
+++ b/tasks/libero-max-long-black-bowl-in-wooden-cabinet-bottom-closed-obstacle-insertion-i01-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "robosuite"
robot = "franka-panda"
family = "libero-max/libero_10/obstacle_insertion"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 195 # LIBERO-MAX's policy seed for the case (the scene is the frozen LIBERO-PRO initial state)
expert_time_estimate_hours = 1.0
diff --git a/tasks/libero-max-long-black-bowl-in-wooden-cabinet-bottom-closed-obstacle-insertion-i01-standard/README.md b/tasks/libero-max-long-black-bowl-in-wooden-cabinet-bottom-closed-obstacle-insertion-i01-standard/README.md
index b35c61458a..67e26748d9 100644
--- a/tasks/libero-max-long-black-bowl-in-wooden-cabinet-bottom-closed-obstacle-insertion-i01-standard/README.md
+++ b/tasks/libero-max-long-black-bowl-in-wooden-cabinet-bottom-closed-obstacle-insertion-i01-standard/README.md
@@ -86,3 +86,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
/reset` ends the episode as `error` (`sim_error = 1`), not `reset` (reachable only when a run grants resets; the
default is none). The scene, instance, pass rule and every verdict reached before are unchanged, so the task keeps
its name.
+- 2026-10-14 (board 0113, from board 0086): the instruction says that a moved object or receptacle jumps alone: what rests in or on it stays where it was (LIBERO-MAX moves one entity; board 0086 saw a basket jump through a can already in it). Text only: the instance, the event, the reference solution, LIBERO's success check and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/libero-max-long-black-bowl-in-wooden-cabinet-bottom-closed-obstacle-insertion-i01-standard/instruction.md b/tasks/libero-max-long-black-bowl-in-wooden-cabinet-bottom-closed-obstacle-insertion-i01-standard/instruction.md
index 3905d4767d..32ad72ef14 100644
--- a/tasks/libero-max-long-black-bowl-in-wooden-cabinet-bottom-closed-obstacle-insertion-i01-standard/instruction.md
+++ b/tasks/libero-max-long-black-bowl-in-wooden-cabinet-bottom-closed-obstacle-insertion-i01-standard/instruction.md
@@ -6,7 +6,8 @@ simulator runs in a service you cannot open, and you reach it only through a cli
The scene can change while you work: once during the episode, when your gripper first gets close to one of the task's
objects, the scene changes (as in LIBERO-MAX): an object or a receptacle may be moved, an obstacle or a few other
objects may appear, a camera may shift, or the lighting, the colours or the image noise may change. Nothing tells you
-when; the task and its success check stay the same, so keep checking what the cameras show.
+when; the task and its success check stay the same, so keep checking what the cameras show. A moved object or
+receptacle jumps there alone: what already rests in or on it stays where it was.
## Rules (real-robot rules)
- **One episode, no reset.** The episode starts when the service comes up. It ends when the task succeeds, after
diff --git a/tasks/libero-max-long-black-bowl-in-wooden-cabinet-bottom-closed-obstacle-insertion-i01-standard/task.toml b/tasks/libero-max-long-black-bowl-in-wooden-cabinet-bottom-closed-obstacle-insertion-i01-standard/task.toml
index 9f635dda46..76b91a0ea2 100644
--- a/tasks/libero-max-long-black-bowl-in-wooden-cabinet-bottom-closed-obstacle-insertion-i01-standard/task.toml
+++ b/tasks/libero-max-long-black-bowl-in-wooden-cabinet-bottom-closed-obstacle-insertion-i01-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "robosuite"
robot = "franka-panda"
family = "libero-max/libero_10/obstacle_insertion"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 195
diff --git a/tasks/libero-max-long-porcelain-mug-and-white-yellow-mug-on-plate-camera-shift-i00-privileged/README.md b/tasks/libero-max-long-porcelain-mug-and-white-yellow-mug-on-plate-camera-shift-i00-privileged/README.md
index 9007a4c5be..0cca224e24 100644
--- a/tasks/libero-max-long-porcelain-mug-and-white-yellow-mug-on-plate-camera-shift-i00-privileged/README.md
+++ b/tasks/libero-max-long-porcelain-mug-and-white-yellow-mug-on-plate-camera-shift-i00-privileged/README.md
@@ -76,3 +76,4 @@ Generated by `scripts/libero-max/generate.py` from `scripts/libero-max/cases.jso
it executes any action). Before, such a trial read as an agent failure with `grader_error = 1` only.
- A replay that records a video prints its verdict before it finishes the video, and a failing video no longer fails
the replay; `video_rendered` is 1 only when the video was finished.
+- 2026-10-14 (board 0113, from board 0086): the instruction adds a note after the task sentence, where the sentence alone misleads in this LIBERO-PRO variant: which plate is the left one (LIBERO's left is as seen facing the robot, the robot's right). Text only: the instance, the event, the reference solution, LIBERO's success check and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/libero-max-long-porcelain-mug-and-white-yellow-mug-on-plate-camera-shift-i00-privileged/instruction.md b/tasks/libero-max-long-porcelain-mug-and-white-yellow-mug-on-plate-camera-shift-i00-privileged/instruction.md
index dc100f2452..e033e1bf55 100644
--- a/tasks/libero-max-long-porcelain-mug-and-white-yellow-mug-on-plate-camera-shift-i00-privileged/instruction.md
+++ b/tasks/libero-max-long-porcelain-mug-and-white-yellow-mug-on-plate-camera-shift-i00-privileged/instruction.md
@@ -1,6 +1,6 @@
# LIBERO-MAX · libero_10 · camera shift — instance i00
-A **Franka Panda** arm (LIBERO on robosuite / MuJoCo) at a table. **Task: Put the white mug on the left plate and put the yellow and white mug on the right plate.**
+A **Franka Panda** arm (LIBERO on robosuite / MuJoCo) at a table. **Task: Put the white mug on the left plate and put the yellow and white mug on the right plate.** Left and right are LIBERO's, as seen from the front of the table, facing the robot: the left plate (`plate_1`, the goal's) is the one on the robot's right, the right plate (`plate_2`) the one on the robot's left (LIBERO-PRO's scene file puts them in `plate_left_region` / `plate_right_region`, 0.3 m to either side).
The scene is one frozen LIBERO-MAX case: a LIBERO-PRO variant of a libero_10 task (camera_view_angle), with one online
event: the first time the gripper comes within 0.18 m of `porcelain_mug_1`, LIBERO-MAX changes the scene
once (the `agentview` camera moves by (-0.06, 0.03, 0) m, turns -22° and its field of view changes by -12°). You have the full simulator state, so you can see the change happen (`s.state()["event_fired"]`).
diff --git a/tasks/libero-max-long-porcelain-mug-and-white-yellow-mug-on-plate-camera-shift-i00-privileged/task.toml b/tasks/libero-max-long-porcelain-mug-and-white-yellow-mug-on-plate-camera-shift-i00-privileged/task.toml
index aca0866d5e..00a6cccfe4 100644
--- a/tasks/libero-max-long-porcelain-mug-and-white-yellow-mug-on-plate-camera-shift-i00-privileged/task.toml
+++ b/tasks/libero-max-long-porcelain-mug-and-white-yellow-mug-on-plate-camera-shift-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "robosuite"
robot = "franka-panda"
family = "libero-max/libero_10/camera_shift"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 195 # LIBERO-MAX's policy seed for the case (the scene is the frozen LIBERO-PRO initial state)
expert_time_estimate_hours = 1.0
diff --git a/tasks/libero-max-long-porcelain-mug-and-white-yellow-mug-on-plate-camera-shift-i00-standard/README.md b/tasks/libero-max-long-porcelain-mug-and-white-yellow-mug-on-plate-camera-shift-i00-standard/README.md
index dabef26203..1d6f3f5c40 100644
--- a/tasks/libero-max-long-porcelain-mug-and-white-yellow-mug-on-plate-camera-shift-i00-standard/README.md
+++ b/tasks/libero-max-long-porcelain-mug-and-white-yellow-mug-on-plate-camera-shift-i00-standard/README.md
@@ -86,3 +86,5 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
/reset` ends the episode as `error` (`sim_error = 1`), not `reset` (reachable only when a run grants resets; the
default is none). The scene, instance, pass rule and every verdict reached before are unchanged, so the task keeps
its name.
+- 2026-10-14 (board 0113, from board 0086): the instruction adds a note after the task sentence, where the sentence alone misleads in this LIBERO-PRO variant: which plate is the left one (LIBERO's left is as seen facing the robot, the robot's right). Text only: the instance, the event, the reference solution, LIBERO's success check and the verifier are unchanged, so the task keeps its name.
+- 2026-10-14 (board 0113, from board 0086): the instruction says that a moved object or receptacle jumps alone: what rests in or on it stays where it was (LIBERO-MAX moves one entity; board 0086 saw a basket jump through a can already in it). Text only: the instance, the event, the reference solution, LIBERO's success check and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/libero-max-long-porcelain-mug-and-white-yellow-mug-on-plate-camera-shift-i00-standard/instruction.md b/tasks/libero-max-long-porcelain-mug-and-white-yellow-mug-on-plate-camera-shift-i00-standard/instruction.md
index d78aaaecdb..937238aac3 100644
--- a/tasks/libero-max-long-porcelain-mug-and-white-yellow-mug-on-plate-camera-shift-i00-standard/instruction.md
+++ b/tasks/libero-max-long-porcelain-mug-and-white-yellow-mug-on-plate-camera-shift-i00-standard/instruction.md
@@ -1,12 +1,13 @@
# LIBERO-MAX · libero_10 · camera shift — instance i00 — STANDARD MODE (robot as a service)
You control a simulated **Franka Panda** arm (LIBERO on robosuite / MuJoCo) **as if it were real hardware**: the
-simulator runs in a service you cannot open, and you reach it only through a client. **Task: Put the white mug on the left plate and put the yellow and white mug on the right plate.**
+simulator runs in a service you cannot open, and you reach it only through a client. **Task: Put the white mug on the left plate and put the yellow and white mug on the right plate.** Left and right are LIBERO's, as seen from the front of the table, facing the robot: the left plate (`plate_1`, the goal's) is the one on the robot's right, the right plate (`plate_2`) the one on the robot's left (LIBERO-PRO's scene file puts them in `plate_left_region` / `plate_right_region`, 0.3 m to either side).
The scene can change while you work: once during the episode, when your gripper first gets close to one of the task's
objects, the scene changes (as in LIBERO-MAX): an object or a receptacle may be moved, an obstacle or a few other
objects may appear, a camera may shift, or the lighting, the colours or the image noise may change. Nothing tells you
-when; the task and its success check stay the same, so keep checking what the cameras show.
+when; the task and its success check stay the same, so keep checking what the cameras show. A moved object or
+receptacle jumps there alone: what already rests in or on it stays where it was.
## Rules (real-robot rules)
- **One episode, no reset.** The episode starts when the service comes up. It ends when the task succeeds, after
diff --git a/tasks/libero-max-long-porcelain-mug-and-white-yellow-mug-on-plate-camera-shift-i00-standard/task.toml b/tasks/libero-max-long-porcelain-mug-and-white-yellow-mug-on-plate-camera-shift-i00-standard/task.toml
index 6bca5342a0..c73a3fd681 100644
--- a/tasks/libero-max-long-porcelain-mug-and-white-yellow-mug-on-plate-camera-shift-i00-standard/task.toml
+++ b/tasks/libero-max-long-porcelain-mug-and-white-yellow-mug-on-plate-camera-shift-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "robosuite"
robot = "franka-panda"
family = "libero-max/libero_10/camera_shift"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 195
diff --git a/tasks/libero-max-long-porcelain-mug-and-white-yellow-mug-on-plate-illumination-switch-i00-privileged/README.md b/tasks/libero-max-long-porcelain-mug-and-white-yellow-mug-on-plate-illumination-switch-i00-privileged/README.md
index 64cdd9e5bd..c33392f98a 100644
--- a/tasks/libero-max-long-porcelain-mug-and-white-yellow-mug-on-plate-illumination-switch-i00-privileged/README.md
+++ b/tasks/libero-max-long-porcelain-mug-and-white-yellow-mug-on-plate-illumination-switch-i00-privileged/README.md
@@ -76,3 +76,4 @@ Generated by `scripts/libero-max/generate.py` from `scripts/libero-max/cases.jso
it executes any action). Before, such a trial read as an agent failure with `grader_error = 1` only.
- A replay that records a video prints its verdict before it finishes the video, and a failing video no longer fails
the replay; `video_rendered` is 1 only when the video was finished.
+- 2026-10-14 (board 0113, from board 0086): the instruction adds a note after the task sentence, where the sentence alone misleads in this LIBERO-PRO variant: which plate is the left one (LIBERO's left is as seen facing the robot, the robot's right). Text only: the instance, the event, the reference solution, LIBERO's success check and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/libero-max-long-porcelain-mug-and-white-yellow-mug-on-plate-illumination-switch-i00-privileged/instruction.md b/tasks/libero-max-long-porcelain-mug-and-white-yellow-mug-on-plate-illumination-switch-i00-privileged/instruction.md
index 4c3b766cef..6205eefe92 100644
--- a/tasks/libero-max-long-porcelain-mug-and-white-yellow-mug-on-plate-illumination-switch-i00-privileged/instruction.md
+++ b/tasks/libero-max-long-porcelain-mug-and-white-yellow-mug-on-plate-illumination-switch-i00-privileged/instruction.md
@@ -1,6 +1,6 @@
# LIBERO-MAX · libero_10 · illumination switch — instance i00
-A **Franka Panda** arm (LIBERO on robosuite / MuJoCo) at a table. **Task: Put the white mug on the left plate and put the yellow and white mug on the right plate.**
+A **Franka Panda** arm (LIBERO on robosuite / MuJoCo) at a table. **Task: Put the white mug on the left plate and put the yellow and white mug on the right plate.** Left and right are LIBERO's, as seen from the front of the table, facing the robot: the left plate (`plate_1`, the goal's) is the one on the robot's right, the right plate (`plate_2`) the one on the robot's left (LIBERO-PRO's scene file puts them in `plate_left_region` / `plate_right_region`, 0.3 m to either side).
The scene is one frozen LIBERO-MAX case: a LIBERO-PRO variant of a libero_10 task (camera_view_angle), with one online
event: the first time the gripper comes within 0.18 m of `porcelain_mug_1`, LIBERO-MAX changes the scene
once (the lighting is scaled by 0.2). You have the full simulator state, so you can see the change happen (`s.state()["event_fired"]`).
diff --git a/tasks/libero-max-long-porcelain-mug-and-white-yellow-mug-on-plate-illumination-switch-i00-privileged/task.toml b/tasks/libero-max-long-porcelain-mug-and-white-yellow-mug-on-plate-illumination-switch-i00-privileged/task.toml
index f4ffe7cc33..52ec621bb9 100644
--- a/tasks/libero-max-long-porcelain-mug-and-white-yellow-mug-on-plate-illumination-switch-i00-privileged/task.toml
+++ b/tasks/libero-max-long-porcelain-mug-and-white-yellow-mug-on-plate-illumination-switch-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "robosuite"
robot = "franka-panda"
family = "libero-max/libero_10/illumination_switch"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 195 # LIBERO-MAX's policy seed for the case (the scene is the frozen LIBERO-PRO initial state)
expert_time_estimate_hours = 1.0
diff --git a/tasks/libero-max-long-porcelain-mug-and-white-yellow-mug-on-plate-illumination-switch-i00-standard/README.md b/tasks/libero-max-long-porcelain-mug-and-white-yellow-mug-on-plate-illumination-switch-i00-standard/README.md
index 13db81c4f5..f0aebe9769 100644
--- a/tasks/libero-max-long-porcelain-mug-and-white-yellow-mug-on-plate-illumination-switch-i00-standard/README.md
+++ b/tasks/libero-max-long-porcelain-mug-and-white-yellow-mug-on-plate-illumination-switch-i00-standard/README.md
@@ -86,3 +86,5 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
/reset` ends the episode as `error` (`sim_error = 1`), not `reset` (reachable only when a run grants resets; the
default is none). The scene, instance, pass rule and every verdict reached before are unchanged, so the task keeps
its name.
+- 2026-10-14 (board 0113, from board 0086): the instruction adds a note after the task sentence, where the sentence alone misleads in this LIBERO-PRO variant: which plate is the left one (LIBERO's left is as seen facing the robot, the robot's right). Text only: the instance, the event, the reference solution, LIBERO's success check and the verifier are unchanged, so the task keeps its name.
+- 2026-10-14 (board 0113, from board 0086): the instruction says that a moved object or receptacle jumps alone: what rests in or on it stays where it was (LIBERO-MAX moves one entity; board 0086 saw a basket jump through a can already in it). Text only: the instance, the event, the reference solution, LIBERO's success check and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/libero-max-long-porcelain-mug-and-white-yellow-mug-on-plate-illumination-switch-i00-standard/instruction.md b/tasks/libero-max-long-porcelain-mug-and-white-yellow-mug-on-plate-illumination-switch-i00-standard/instruction.md
index 823ce56c05..b80ecc5871 100644
--- a/tasks/libero-max-long-porcelain-mug-and-white-yellow-mug-on-plate-illumination-switch-i00-standard/instruction.md
+++ b/tasks/libero-max-long-porcelain-mug-and-white-yellow-mug-on-plate-illumination-switch-i00-standard/instruction.md
@@ -1,12 +1,13 @@
# LIBERO-MAX · libero_10 · illumination switch — instance i00 — STANDARD MODE (robot as a service)
You control a simulated **Franka Panda** arm (LIBERO on robosuite / MuJoCo) **as if it were real hardware**: the
-simulator runs in a service you cannot open, and you reach it only through a client. **Task: Put the white mug on the left plate and put the yellow and white mug on the right plate.**
+simulator runs in a service you cannot open, and you reach it only through a client. **Task: Put the white mug on the left plate and put the yellow and white mug on the right plate.** Left and right are LIBERO's, as seen from the front of the table, facing the robot: the left plate (`plate_1`, the goal's) is the one on the robot's right, the right plate (`plate_2`) the one on the robot's left (LIBERO-PRO's scene file puts them in `plate_left_region` / `plate_right_region`, 0.3 m to either side).
The scene can change while you work: once during the episode, when your gripper first gets close to one of the task's
objects, the scene changes (as in LIBERO-MAX): an object or a receptacle may be moved, an obstacle or a few other
objects may appear, a camera may shift, or the lighting, the colours or the image noise may change. Nothing tells you
-when; the task and its success check stay the same, so keep checking what the cameras show.
+when; the task and its success check stay the same, so keep checking what the cameras show. A moved object or
+receptacle jumps there alone: what already rests in or on it stays where it was.
## Rules (real-robot rules)
- **One episode, no reset.** The episode starts when the service comes up. It ends when the task succeeds, after
diff --git a/tasks/libero-max-long-porcelain-mug-and-white-yellow-mug-on-plate-illumination-switch-i00-standard/task.toml b/tasks/libero-max-long-porcelain-mug-and-white-yellow-mug-on-plate-illumination-switch-i00-standard/task.toml
index 02db9b6590..5aa0975889 100644
--- a/tasks/libero-max-long-porcelain-mug-and-white-yellow-mug-on-plate-illumination-switch-i00-standard/task.toml
+++ b/tasks/libero-max-long-porcelain-mug-and-white-yellow-mug-on-plate-illumination-switch-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "robosuite"
robot = "franka-panda"
family = "libero-max/libero_10/illumination_switch"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 195
diff --git a/tasks/libero-max-long-porcelain-mug-and-white-yellow-mug-on-plate-visual-theme-switch-i00-privileged/README.md b/tasks/libero-max-long-porcelain-mug-and-white-yellow-mug-on-plate-visual-theme-switch-i00-privileged/README.md
index 4f2a5b1dde..614533de6a 100644
--- a/tasks/libero-max-long-porcelain-mug-and-white-yellow-mug-on-plate-visual-theme-switch-i00-privileged/README.md
+++ b/tasks/libero-max-long-porcelain-mug-and-white-yellow-mug-on-plate-visual-theme-switch-i00-privileged/README.md
@@ -76,3 +76,4 @@ Generated by `scripts/libero-max/generate.py` from `scripts/libero-max/cases.jso
it executes any action). Before, such a trial read as an agent failure with `grader_error = 1` only.
- A replay that records a video prints its verdict before it finishes the video, and a failing video no longer fails
the replay; `video_rendered` is 1 only when the video was finished.
+- 2026-10-14 (board 0113, from board 0086): the instruction adds a note after the task sentence, where the sentence alone misleads in this LIBERO-PRO variant: which plate is the left one (LIBERO's left is as seen facing the robot, the robot's right). Text only: the instance, the event, the reference solution, LIBERO's success check and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/libero-max-long-porcelain-mug-and-white-yellow-mug-on-plate-visual-theme-switch-i00-privileged/instruction.md b/tasks/libero-max-long-porcelain-mug-and-white-yellow-mug-on-plate-visual-theme-switch-i00-privileged/instruction.md
index 4a0ad4531a..15f7e554f7 100644
--- a/tasks/libero-max-long-porcelain-mug-and-white-yellow-mug-on-plate-visual-theme-switch-i00-privileged/instruction.md
+++ b/tasks/libero-max-long-porcelain-mug-and-white-yellow-mug-on-plate-visual-theme-switch-i00-privileged/instruction.md
@@ -1,6 +1,6 @@
# LIBERO-MAX · libero_10 · visual-theme switch — instance i00
-A **Franka Panda** arm (LIBERO on robosuite / MuJoCo) at a table. **Task: Put the white mug on the left plate and put the yellow and white mug on the right plate.**
+A **Franka Panda** arm (LIBERO on robosuite / MuJoCo) at a table. **Task: Put the white mug on the left plate and put the yellow and white mug on the right plate.** Left and right are LIBERO's, as seen from the front of the table, facing the robot: the left plate (`plate_1`, the goal's) is the one on the robot's right, the right plate (`plate_2`) the one on the robot's left (LIBERO-PRO's scene file puts them in `plate_left_region` / `plate_right_region`, 0.3 m to either side).
The scene is one frozen LIBERO-MAX case: a LIBERO-PRO variant of a libero_10 task (position), with one online
event: the first time the gripper comes within 0.18 m of `porcelain_mug_1`, LIBERO-MAX changes the scene
once (the image colours change (channels permuted (1, 2, 0), scaled (1.25, 0.7, 1.1))). You have the full simulator state, so you can see the change happen (`s.state()["event_fired"]`).
diff --git a/tasks/libero-max-long-porcelain-mug-and-white-yellow-mug-on-plate-visual-theme-switch-i00-privileged/task.toml b/tasks/libero-max-long-porcelain-mug-and-white-yellow-mug-on-plate-visual-theme-switch-i00-privileged/task.toml
index 6835147041..f0f4bda067 100644
--- a/tasks/libero-max-long-porcelain-mug-and-white-yellow-mug-on-plate-visual-theme-switch-i00-privileged/task.toml
+++ b/tasks/libero-max-long-porcelain-mug-and-white-yellow-mug-on-plate-visual-theme-switch-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "robosuite"
robot = "franka-panda"
family = "libero-max/libero_10/visual_theme_switch"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 195 # LIBERO-MAX's policy seed for the case (the scene is the frozen LIBERO-PRO initial state)
expert_time_estimate_hours = 1.0
diff --git a/tasks/libero-max-long-porcelain-mug-and-white-yellow-mug-on-plate-visual-theme-switch-i00-standard/README.md b/tasks/libero-max-long-porcelain-mug-and-white-yellow-mug-on-plate-visual-theme-switch-i00-standard/README.md
index d2d2a502af..818ab7c083 100644
--- a/tasks/libero-max-long-porcelain-mug-and-white-yellow-mug-on-plate-visual-theme-switch-i00-standard/README.md
+++ b/tasks/libero-max-long-porcelain-mug-and-white-yellow-mug-on-plate-visual-theme-switch-i00-standard/README.md
@@ -86,3 +86,5 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
/reset` ends the episode as `error` (`sim_error = 1`), not `reset` (reachable only when a run grants resets; the
default is none). The scene, instance, pass rule and every verdict reached before are unchanged, so the task keeps
its name.
+- 2026-10-14 (board 0113, from board 0086): the instruction adds a note after the task sentence, where the sentence alone misleads in this LIBERO-PRO variant: which plate is the left one (LIBERO's left is as seen facing the robot, the robot's right). Text only: the instance, the event, the reference solution, LIBERO's success check and the verifier are unchanged, so the task keeps its name.
+- 2026-10-14 (board 0113, from board 0086): the instruction says that a moved object or receptacle jumps alone: what rests in or on it stays where it was (LIBERO-MAX moves one entity; board 0086 saw a basket jump through a can already in it). Text only: the instance, the event, the reference solution, LIBERO's success check and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/libero-max-long-porcelain-mug-and-white-yellow-mug-on-plate-visual-theme-switch-i00-standard/instruction.md b/tasks/libero-max-long-porcelain-mug-and-white-yellow-mug-on-plate-visual-theme-switch-i00-standard/instruction.md
index 0955974a03..9a6769d159 100644
--- a/tasks/libero-max-long-porcelain-mug-and-white-yellow-mug-on-plate-visual-theme-switch-i00-standard/instruction.md
+++ b/tasks/libero-max-long-porcelain-mug-and-white-yellow-mug-on-plate-visual-theme-switch-i00-standard/instruction.md
@@ -1,12 +1,13 @@
# LIBERO-MAX · libero_10 · visual-theme switch — instance i00 — STANDARD MODE (robot as a service)
You control a simulated **Franka Panda** arm (LIBERO on robosuite / MuJoCo) **as if it were real hardware**: the
-simulator runs in a service you cannot open, and you reach it only through a client. **Task: Put the white mug on the left plate and put the yellow and white mug on the right plate.**
+simulator runs in a service you cannot open, and you reach it only through a client. **Task: Put the white mug on the left plate and put the yellow and white mug on the right plate.** Left and right are LIBERO's, as seen from the front of the table, facing the robot: the left plate (`plate_1`, the goal's) is the one on the robot's right, the right plate (`plate_2`) the one on the robot's left (LIBERO-PRO's scene file puts them in `plate_left_region` / `plate_right_region`, 0.3 m to either side).
The scene can change while you work: once during the episode, when your gripper first gets close to one of the task's
objects, the scene changes (as in LIBERO-MAX): an object or a receptacle may be moved, an obstacle or a few other
objects may appear, a camera may shift, or the lighting, the colours or the image noise may change. Nothing tells you
-when; the task and its success check stay the same, so keep checking what the cameras show.
+when; the task and its success check stay the same, so keep checking what the cameras show. A moved object or
+receptacle jumps there alone: what already rests in or on it stays where it was.
## Rules (real-robot rules)
- **One episode, no reset.** The episode starts when the service comes up. It ends when the task succeeds, after
diff --git a/tasks/libero-max-long-porcelain-mug-and-white-yellow-mug-on-plate-visual-theme-switch-i00-standard/task.toml b/tasks/libero-max-long-porcelain-mug-and-white-yellow-mug-on-plate-visual-theme-switch-i00-standard/task.toml
index 9f226eda96..c24374eb9d 100644
--- a/tasks/libero-max-long-porcelain-mug-and-white-yellow-mug-on-plate-visual-theme-switch-i00-standard/task.toml
+++ b/tasks/libero-max-long-porcelain-mug-and-white-yellow-mug-on-plate-visual-theme-switch-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "robosuite"
robot = "franka-panda"
family = "libero-max/libero_10/visual_theme_switch"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 195
diff --git a/tasks/libero-max-long-porcelain-mug-and-white-yellow-mug-on-plate-visual-theme-switch-i01-privileged/README.md b/tasks/libero-max-long-porcelain-mug-and-white-yellow-mug-on-plate-visual-theme-switch-i01-privileged/README.md
index 0e3a7395a1..dcf0845cf0 100644
--- a/tasks/libero-max-long-porcelain-mug-and-white-yellow-mug-on-plate-visual-theme-switch-i01-privileged/README.md
+++ b/tasks/libero-max-long-porcelain-mug-and-white-yellow-mug-on-plate-visual-theme-switch-i01-privileged/README.md
@@ -76,3 +76,4 @@ Generated by `scripts/libero-max/generate.py` from `scripts/libero-max/cases.jso
it executes any action). Before, such a trial read as an agent failure with `grader_error = 1` only.
- A replay that records a video prints its verdict before it finishes the video, and a failing video no longer fails
the replay; `video_rendered` is 1 only when the video was finished.
+- 2026-10-14 (board 0113, from board 0086): the instruction adds a note after the task sentence, where the sentence alone misleads in this LIBERO-PRO variant: which plate is the left one (LIBERO's left is as seen facing the robot, the robot's right). Text only: the instance, the event, the reference solution, LIBERO's success check and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/libero-max-long-porcelain-mug-and-white-yellow-mug-on-plate-visual-theme-switch-i01-privileged/instruction.md b/tasks/libero-max-long-porcelain-mug-and-white-yellow-mug-on-plate-visual-theme-switch-i01-privileged/instruction.md
index d8c30c14bc..a453885822 100644
--- a/tasks/libero-max-long-porcelain-mug-and-white-yellow-mug-on-plate-visual-theme-switch-i01-privileged/instruction.md
+++ b/tasks/libero-max-long-porcelain-mug-and-white-yellow-mug-on-plate-visual-theme-switch-i01-privileged/instruction.md
@@ -1,6 +1,6 @@
# LIBERO-MAX · libero_10 · visual-theme switch — instance i01
-A **Franka Panda** arm (LIBERO on robosuite / MuJoCo) at a table. **Task: Place white mug on left plate and yellow-white mug on right plate.**
+A **Franka Panda** arm (LIBERO on robosuite / MuJoCo) at a table. **Task: Place white mug on left plate and yellow-white mug on right plate.** Left and right are LIBERO's, as seen from the front of the table, facing the robot: the left plate (`plate_1`, the goal's) is the one on the robot's right, the right plate (`plate_2`) the one on the robot's left (LIBERO-PRO's scene file puts them in `plate_left_region` / `plate_right_region`, 0.3 m to either side).
The scene is one frozen LIBERO-MAX case: a LIBERO-PRO variant of a libero_10 task (semantic), with one online
event: the first time the gripper comes within 0.18 m of `porcelain_mug_1`, LIBERO-MAX changes the scene
once (the image colours change (channels permuted (1, 2, 0), scaled (1.25, 0.7, 1.1))). You have the full simulator state, so you can see the change happen (`s.state()["event_fired"]`).
diff --git a/tasks/libero-max-long-porcelain-mug-and-white-yellow-mug-on-plate-visual-theme-switch-i01-privileged/task.toml b/tasks/libero-max-long-porcelain-mug-and-white-yellow-mug-on-plate-visual-theme-switch-i01-privileged/task.toml
index e47b801107..b3b673202f 100644
--- a/tasks/libero-max-long-porcelain-mug-and-white-yellow-mug-on-plate-visual-theme-switch-i01-privileged/task.toml
+++ b/tasks/libero-max-long-porcelain-mug-and-white-yellow-mug-on-plate-visual-theme-switch-i01-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "robosuite"
robot = "franka-panda"
family = "libero-max/libero_10/visual_theme_switch"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 195 # LIBERO-MAX's policy seed for the case (the scene is the frozen LIBERO-PRO initial state)
expert_time_estimate_hours = 1.0
diff --git a/tasks/libero-max-long-porcelain-mug-and-white-yellow-mug-on-plate-visual-theme-switch-i01-standard/README.md b/tasks/libero-max-long-porcelain-mug-and-white-yellow-mug-on-plate-visual-theme-switch-i01-standard/README.md
index 03fae111df..4a641103ad 100644
--- a/tasks/libero-max-long-porcelain-mug-and-white-yellow-mug-on-plate-visual-theme-switch-i01-standard/README.md
+++ b/tasks/libero-max-long-porcelain-mug-and-white-yellow-mug-on-plate-visual-theme-switch-i01-standard/README.md
@@ -86,3 +86,5 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
/reset` ends the episode as `error` (`sim_error = 1`), not `reset` (reachable only when a run grants resets; the
default is none). The scene, instance, pass rule and every verdict reached before are unchanged, so the task keeps
its name.
+- 2026-10-14 (board 0113, from board 0086): the instruction adds a note after the task sentence, where the sentence alone misleads in this LIBERO-PRO variant: which plate is the left one (LIBERO's left is as seen facing the robot, the robot's right). Text only: the instance, the event, the reference solution, LIBERO's success check and the verifier are unchanged, so the task keeps its name.
+- 2026-10-14 (board 0113, from board 0086): the instruction says that a moved object or receptacle jumps alone: what rests in or on it stays where it was (LIBERO-MAX moves one entity; board 0086 saw a basket jump through a can already in it). Text only: the instance, the event, the reference solution, LIBERO's success check and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/libero-max-long-porcelain-mug-and-white-yellow-mug-on-plate-visual-theme-switch-i01-standard/instruction.md b/tasks/libero-max-long-porcelain-mug-and-white-yellow-mug-on-plate-visual-theme-switch-i01-standard/instruction.md
index cdc4f3ae1b..7589e46a79 100644
--- a/tasks/libero-max-long-porcelain-mug-and-white-yellow-mug-on-plate-visual-theme-switch-i01-standard/instruction.md
+++ b/tasks/libero-max-long-porcelain-mug-and-white-yellow-mug-on-plate-visual-theme-switch-i01-standard/instruction.md
@@ -1,12 +1,13 @@
# LIBERO-MAX · libero_10 · visual-theme switch — instance i01 — STANDARD MODE (robot as a service)
You control a simulated **Franka Panda** arm (LIBERO on robosuite / MuJoCo) **as if it were real hardware**: the
-simulator runs in a service you cannot open, and you reach it only through a client. **Task: Place white mug on left plate and yellow-white mug on right plate.**
+simulator runs in a service you cannot open, and you reach it only through a client. **Task: Place white mug on left plate and yellow-white mug on right plate.** Left and right are LIBERO's, as seen from the front of the table, facing the robot: the left plate (`plate_1`, the goal's) is the one on the robot's right, the right plate (`plate_2`) the one on the robot's left (LIBERO-PRO's scene file puts them in `plate_left_region` / `plate_right_region`, 0.3 m to either side).
The scene can change while you work: once during the episode, when your gripper first gets close to one of the task's
objects, the scene changes (as in LIBERO-MAX): an object or a receptacle may be moved, an obstacle or a few other
objects may appear, a camera may shift, or the lighting, the colours or the image noise may change. Nothing tells you
-when; the task and its success check stay the same, so keep checking what the cameras show.
+when; the task and its success check stay the same, so keep checking what the cameras show. A moved object or
+receptacle jumps there alone: what already rests in or on it stays where it was.
## Rules (real-robot rules)
- **One episode, no reset.** The episode starts when the service comes up. It ends when the task succeeds, after
diff --git a/tasks/libero-max-long-porcelain-mug-and-white-yellow-mug-on-plate-visual-theme-switch-i01-standard/task.toml b/tasks/libero-max-long-porcelain-mug-and-white-yellow-mug-on-plate-visual-theme-switch-i01-standard/task.toml
index 636c4cf4cd..566c62f031 100644
--- a/tasks/libero-max-long-porcelain-mug-and-white-yellow-mug-on-plate-visual-theme-switch-i01-standard/task.toml
+++ b/tasks/libero-max-long-porcelain-mug-and-white-yellow-mug-on-plate-visual-theme-switch-i01-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "robosuite"
robot = "franka-panda"
family = "libero-max/libero_10/visual_theme_switch"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 195
diff --git a/tasks/libero-max-long-white-yellow-mug-and-green-chocolate-pudding-on-yellow-plate-sensor-noise-onset-i00-privileged/task.toml b/tasks/libero-max-long-white-yellow-mug-and-green-chocolate-pudding-on-yellow-plate-sensor-noise-onset-i00-privileged/task.toml
index e0d2abe9e8..49ff8e03b4 100644
--- a/tasks/libero-max-long-white-yellow-mug-and-green-chocolate-pudding-on-yellow-plate-sensor-noise-onset-i00-privileged/task.toml
+++ b/tasks/libero-max-long-white-yellow-mug-and-green-chocolate-pudding-on-yellow-plate-sensor-noise-onset-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "robosuite"
robot = "franka-panda"
family = "libero-max/libero_10/sensor_noise_onset"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 195 # LIBERO-MAX's policy seed for the case (the scene is the frozen LIBERO-PRO initial state)
expert_time_estimate_hours = 1.0
diff --git a/tasks/libero-max-long-white-yellow-mug-and-green-chocolate-pudding-on-yellow-plate-sensor-noise-onset-i00-standard/README.md b/tasks/libero-max-long-white-yellow-mug-and-green-chocolate-pudding-on-yellow-plate-sensor-noise-onset-i00-standard/README.md
index 89837354a8..f1baa2fd86 100644
--- a/tasks/libero-max-long-white-yellow-mug-and-green-chocolate-pudding-on-yellow-plate-sensor-noise-onset-i00-standard/README.md
+++ b/tasks/libero-max-long-white-yellow-mug-and-green-chocolate-pudding-on-yellow-plate-sensor-noise-onset-i00-standard/README.md
@@ -86,3 +86,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
/reset` ends the episode as `error` (`sim_error = 1`), not `reset` (reachable only when a run grants resets; the
default is none). The scene, instance, pass rule and every verdict reached before are unchanged, so the task keeps
its name.
+- 2026-10-14 (board 0113, from board 0086): the instruction says that a moved object or receptacle jumps alone: what rests in or on it stays where it was (LIBERO-MAX moves one entity; board 0086 saw a basket jump through a can already in it). Text only: the instance, the event, the reference solution, LIBERO's success check and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/libero-max-long-white-yellow-mug-and-green-chocolate-pudding-on-yellow-plate-sensor-noise-onset-i00-standard/instruction.md b/tasks/libero-max-long-white-yellow-mug-and-green-chocolate-pudding-on-yellow-plate-sensor-noise-onset-i00-standard/instruction.md
index ebfae8f4c1..8040c8a5ca 100644
--- a/tasks/libero-max-long-white-yellow-mug-and-green-chocolate-pudding-on-yellow-plate-sensor-noise-onset-i00-standard/instruction.md
+++ b/tasks/libero-max-long-white-yellow-mug-and-green-chocolate-pudding-on-yellow-plate-sensor-noise-onset-i00-standard/instruction.md
@@ -6,7 +6,8 @@ simulator runs in a service you cannot open, and you reach it only through a cli
The scene can change while you work: once during the episode, when your gripper first gets close to one of the task's
objects, the scene changes (as in LIBERO-MAX): an object or a receptacle may be moved, an obstacle or a few other
objects may appear, a camera may shift, or the lighting, the colours or the image noise may change. Nothing tells you
-when; the task and its success check stay the same, so keep checking what the cameras show.
+when; the task and its success check stay the same, so keep checking what the cameras show. A moved object or
+receptacle jumps there alone: what already rests in or on it stays where it was.
## Rules (real-robot rules)
- **One episode, no reset.** The episode starts when the service comes up. It ends when the task succeeds, after
diff --git a/tasks/libero-max-long-white-yellow-mug-and-green-chocolate-pudding-on-yellow-plate-sensor-noise-onset-i00-standard/task.toml b/tasks/libero-max-long-white-yellow-mug-and-green-chocolate-pudding-on-yellow-plate-sensor-noise-onset-i00-standard/task.toml
index c929ba9bfe..7bf2cf253c 100644
--- a/tasks/libero-max-long-white-yellow-mug-and-green-chocolate-pudding-on-yellow-plate-sensor-noise-onset-i00-standard/task.toml
+++ b/tasks/libero-max-long-white-yellow-mug-and-green-chocolate-pudding-on-yellow-plate-sensor-noise-onset-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "robosuite"
robot = "franka-panda"
family = "libero-max/libero_10/sensor_noise_onset"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 195
diff --git a/tasks/libero-max-long-white-yellow-mug-in-desk-caddy-back-camera-shift-i00-privileged/README.md b/tasks/libero-max-long-white-yellow-mug-in-desk-caddy-back-camera-shift-i00-privileged/README.md
index 0a97b9b469..72825f6de0 100644
--- a/tasks/libero-max-long-white-yellow-mug-in-desk-caddy-back-camera-shift-i00-privileged/README.md
+++ b/tasks/libero-max-long-white-yellow-mug-in-desk-caddy-back-camera-shift-i00-privileged/README.md
@@ -76,3 +76,4 @@ Generated by `scripts/libero-max/generate.py` from `scripts/libero-max/cases.jso
it executes any action). Before, such a trial read as an agent failure with `grader_error = 1` only.
- A replay that records a video prints its verdict before it finishes the video, and a failing video no longer fails
the replay; `video_rendered` is 1 only when the video was finished.
+- 2026-10-14 (board 0113, from board 0086): the instruction adds a note after the task sentence, where the sentence alone misleads in this LIBERO-PRO variant: which cell is the caddy's back compartment, that the mug does not fit into it, and what LIBERO's In accepts. Text only: the instance, the event, the reference solution, LIBERO's success check and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/libero-max-long-white-yellow-mug-in-desk-caddy-back-camera-shift-i00-privileged/instruction.md b/tasks/libero-max-long-white-yellow-mug-in-desk-caddy-back-camera-shift-i00-privileged/instruction.md
index c0e4bc4dc4..0c676555f7 100644
--- a/tasks/libero-max-long-white-yellow-mug-in-desk-caddy-back-camera-shift-i00-privileged/instruction.md
+++ b/tasks/libero-max-long-white-yellow-mug-in-desk-caddy-back-camera-shift-i00-privileged/instruction.md
@@ -1,6 +1,6 @@
# LIBERO-MAX · libero_10 · camera shift — instance i00
-A **Franka Panda** arm (LIBERO on robosuite / MuJoCo) at a table. **Task: Pick up the cup and place it in the back compartment of the caddy.**
+A **Franka Panda** arm (LIBERO on robosuite / MuJoCo) at a table. **Task: Pick up the cup and place it in the back compartment of the caddy.** The caddy has a large cell at each end and, between them, two narrow cells: the front one and the back one, which runs along the caddy's tall wall. The back compartment is that narrow back cell (`desk_caddy_1_back_contain_region`, about 5.5 cm across): the mug (about 9 cm wide) does not fit down into it. LIBERO's In tests only the mug's origin against the region's box (its top is about 15 cm above the table) and its contact with the caddy, so a mug resting on top of the back cell's opening, its origin just inside the top of that box, counts; a mug in any other cell does not.
The scene is one frozen LIBERO-MAX case: a LIBERO-PRO variant of a libero_10 task (task), with one online
event: the first time the gripper comes within 0.18 m of `white_yellow_mug_1`, LIBERO-MAX changes the scene
once (the `agentview` camera moves by (-0.06, 0.03, 0) m, turns -22° and its field of view changes by -12°). You have the full simulator state, so you can see the change happen (`s.state()["event_fired"]`).
diff --git a/tasks/libero-max-long-white-yellow-mug-in-desk-caddy-back-camera-shift-i00-privileged/task.toml b/tasks/libero-max-long-white-yellow-mug-in-desk-caddy-back-camera-shift-i00-privileged/task.toml
index d6ef160176..2f20b8a1e0 100644
--- a/tasks/libero-max-long-white-yellow-mug-in-desk-caddy-back-camera-shift-i00-privileged/task.toml
+++ b/tasks/libero-max-long-white-yellow-mug-in-desk-caddy-back-camera-shift-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "robosuite"
robot = "franka-panda"
family = "libero-max/libero_10/camera_shift"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 195 # LIBERO-MAX's policy seed for the case (the scene is the frozen LIBERO-PRO initial state)
expert_time_estimate_hours = 1.0
diff --git a/tasks/libero-max-long-white-yellow-mug-in-desk-caddy-back-camera-shift-i00-standard/README.md b/tasks/libero-max-long-white-yellow-mug-in-desk-caddy-back-camera-shift-i00-standard/README.md
index 6cac540a7a..0930c5a761 100644
--- a/tasks/libero-max-long-white-yellow-mug-in-desk-caddy-back-camera-shift-i00-standard/README.md
+++ b/tasks/libero-max-long-white-yellow-mug-in-desk-caddy-back-camera-shift-i00-standard/README.md
@@ -86,3 +86,5 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
/reset` ends the episode as `error` (`sim_error = 1`), not `reset` (reachable only when a run grants resets; the
default is none). The scene, instance, pass rule and every verdict reached before are unchanged, so the task keeps
its name.
+- 2026-10-14 (board 0113, from board 0086): the instruction adds a note after the task sentence, where the sentence alone misleads in this LIBERO-PRO variant: which cell is the caddy's back compartment, that the mug does not fit into it, and what LIBERO's In accepts. Text only: the instance, the event, the reference solution, LIBERO's success check and the verifier are unchanged, so the task keeps its name.
+- 2026-10-14 (board 0113, from board 0086): the instruction says that a moved object or receptacle jumps alone: what rests in or on it stays where it was (LIBERO-MAX moves one entity; board 0086 saw a basket jump through a can already in it). Text only: the instance, the event, the reference solution, LIBERO's success check and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/libero-max-long-white-yellow-mug-in-desk-caddy-back-camera-shift-i00-standard/instruction.md b/tasks/libero-max-long-white-yellow-mug-in-desk-caddy-back-camera-shift-i00-standard/instruction.md
index a1bcc8bf8d..9b6b96ab49 100644
--- a/tasks/libero-max-long-white-yellow-mug-in-desk-caddy-back-camera-shift-i00-standard/instruction.md
+++ b/tasks/libero-max-long-white-yellow-mug-in-desk-caddy-back-camera-shift-i00-standard/instruction.md
@@ -1,12 +1,13 @@
# LIBERO-MAX · libero_10 · camera shift — instance i00 — STANDARD MODE (robot as a service)
You control a simulated **Franka Panda** arm (LIBERO on robosuite / MuJoCo) **as if it were real hardware**: the
-simulator runs in a service you cannot open, and you reach it only through a client. **Task: Pick up the cup and place it in the back compartment of the caddy.**
+simulator runs in a service you cannot open, and you reach it only through a client. **Task: Pick up the cup and place it in the back compartment of the caddy.** The caddy has a large cell at each end and, between them, two narrow cells: the front one and the back one, which runs along the caddy's tall wall. The back compartment is that narrow back cell (`desk_caddy_1_back_contain_region`, about 5.5 cm across): the mug (about 9 cm wide) does not fit down into it. LIBERO's In tests only the mug's origin against the region's box (its top is about 15 cm above the table) and its contact with the caddy, so a mug resting on top of the back cell's opening, its origin just inside the top of that box, counts; a mug in any other cell does not.
The scene can change while you work: once during the episode, when your gripper first gets close to one of the task's
objects, the scene changes (as in LIBERO-MAX): an object or a receptacle may be moved, an obstacle or a few other
objects may appear, a camera may shift, or the lighting, the colours or the image noise may change. Nothing tells you
-when; the task and its success check stay the same, so keep checking what the cameras show.
+when; the task and its success check stay the same, so keep checking what the cameras show. A moved object or
+receptacle jumps there alone: what already rests in or on it stays where it was.
## Rules (real-robot rules)
- **One episode, no reset.** The episode starts when the service comes up. It ends when the task succeeds, after
diff --git a/tasks/libero-max-long-white-yellow-mug-in-desk-caddy-back-camera-shift-i00-standard/task.toml b/tasks/libero-max-long-white-yellow-mug-in-desk-caddy-back-camera-shift-i00-standard/task.toml
index d8de561e1f..a067a2c03a 100644
--- a/tasks/libero-max-long-white-yellow-mug-in-desk-caddy-back-camera-shift-i00-standard/task.toml
+++ b/tasks/libero-max-long-white-yellow-mug-in-desk-caddy-back-camera-shift-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "robosuite"
robot = "franka-panda"
family = "libero-max/libero_10/camera_shift"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 195
diff --git a/tasks/libero-max-long-white-yellow-mug-in-desk-caddy-back-illumination-switch-i00-privileged/README.md b/tasks/libero-max-long-white-yellow-mug-in-desk-caddy-back-illumination-switch-i00-privileged/README.md
index 325f8fe3ce..781c16deb1 100644
--- a/tasks/libero-max-long-white-yellow-mug-in-desk-caddy-back-illumination-switch-i00-privileged/README.md
+++ b/tasks/libero-max-long-white-yellow-mug-in-desk-caddy-back-illumination-switch-i00-privileged/README.md
@@ -76,3 +76,4 @@ Generated by `scripts/libero-max/generate.py` from `scripts/libero-max/cases.jso
it executes any action). Before, such a trial read as an agent failure with `grader_error = 1` only.
- A replay that records a video prints its verdict before it finishes the video, and a failing video no longer fails
the replay; `video_rendered` is 1 only when the video was finished.
+- 2026-10-14 (board 0113, from board 0086): the instruction adds a note after the task sentence, where the sentence alone misleads in this LIBERO-PRO variant: which cell is the caddy's back compartment, that the mug does not fit into it, and what LIBERO's In accepts. Text only: the instance, the event, the reference solution, LIBERO's success check and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/libero-max-long-white-yellow-mug-in-desk-caddy-back-illumination-switch-i00-privileged/instruction.md b/tasks/libero-max-long-white-yellow-mug-in-desk-caddy-back-illumination-switch-i00-privileged/instruction.md
index 54a3461806..669556dad5 100644
--- a/tasks/libero-max-long-white-yellow-mug-in-desk-caddy-back-illumination-switch-i00-privileged/instruction.md
+++ b/tasks/libero-max-long-white-yellow-mug-in-desk-caddy-back-illumination-switch-i00-privileged/instruction.md
@@ -1,6 +1,6 @@
# LIBERO-MAX · libero_10 · illumination switch — instance i00
-A **Franka Panda** arm (LIBERO on robosuite / MuJoCo) at a table. **Task: Pick up the cup and place it in the back compartment of the caddy.**
+A **Franka Panda** arm (LIBERO on robosuite / MuJoCo) at a table. **Task: Pick up the cup and place it in the back compartment of the caddy.** The caddy has a large cell at each end and, between them, two narrow cells: the front one and the back one, which runs along the caddy's tall wall. The back compartment is that narrow back cell (`desk_caddy_1_back_contain_region`, about 5.5 cm across): the mug (about 9 cm wide) does not fit down into it. LIBERO's In tests only the mug's origin against the region's box (its top is about 15 cm above the table) and its contact with the caddy, so a mug resting on top of the back cell's opening, its origin just inside the top of that box, counts; a mug in any other cell does not.
The scene is one frozen LIBERO-MAX case: a LIBERO-PRO variant of a libero_10 task (task), with one online
event: the first time the gripper comes within 0.18 m of `white_yellow_mug_1`, LIBERO-MAX changes the scene
once (the lighting is scaled by 1.8). You have the full simulator state, so you can see the change happen (`s.state()["event_fired"]`).
diff --git a/tasks/libero-max-long-white-yellow-mug-in-desk-caddy-back-illumination-switch-i00-privileged/task.toml b/tasks/libero-max-long-white-yellow-mug-in-desk-caddy-back-illumination-switch-i00-privileged/task.toml
index 457cd608ea..df0d707520 100644
--- a/tasks/libero-max-long-white-yellow-mug-in-desk-caddy-back-illumination-switch-i00-privileged/task.toml
+++ b/tasks/libero-max-long-white-yellow-mug-in-desk-caddy-back-illumination-switch-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "robosuite"
robot = "franka-panda"
family = "libero-max/libero_10/illumination_switch"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 195 # LIBERO-MAX's policy seed for the case (the scene is the frozen LIBERO-PRO initial state)
expert_time_estimate_hours = 1.0
diff --git a/tasks/libero-max-long-white-yellow-mug-in-desk-caddy-back-illumination-switch-i00-standard/README.md b/tasks/libero-max-long-white-yellow-mug-in-desk-caddy-back-illumination-switch-i00-standard/README.md
index 558ac3054f..7fb661b21c 100644
--- a/tasks/libero-max-long-white-yellow-mug-in-desk-caddy-back-illumination-switch-i00-standard/README.md
+++ b/tasks/libero-max-long-white-yellow-mug-in-desk-caddy-back-illumination-switch-i00-standard/README.md
@@ -86,3 +86,5 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
/reset` ends the episode as `error` (`sim_error = 1`), not `reset` (reachable only when a run grants resets; the
default is none). The scene, instance, pass rule and every verdict reached before are unchanged, so the task keeps
its name.
+- 2026-10-14 (board 0113, from board 0086): the instruction adds a note after the task sentence, where the sentence alone misleads in this LIBERO-PRO variant: which cell is the caddy's back compartment, that the mug does not fit into it, and what LIBERO's In accepts. Text only: the instance, the event, the reference solution, LIBERO's success check and the verifier are unchanged, so the task keeps its name.
+- 2026-10-14 (board 0113, from board 0086): the instruction says that a moved object or receptacle jumps alone: what rests in or on it stays where it was (LIBERO-MAX moves one entity; board 0086 saw a basket jump through a can already in it). Text only: the instance, the event, the reference solution, LIBERO's success check and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/libero-max-long-white-yellow-mug-in-desk-caddy-back-illumination-switch-i00-standard/instruction.md b/tasks/libero-max-long-white-yellow-mug-in-desk-caddy-back-illumination-switch-i00-standard/instruction.md
index 4f987f1cb7..8d5b7de443 100644
--- a/tasks/libero-max-long-white-yellow-mug-in-desk-caddy-back-illumination-switch-i00-standard/instruction.md
+++ b/tasks/libero-max-long-white-yellow-mug-in-desk-caddy-back-illumination-switch-i00-standard/instruction.md
@@ -1,12 +1,13 @@
# LIBERO-MAX · libero_10 · illumination switch — instance i00 — STANDARD MODE (robot as a service)
You control a simulated **Franka Panda** arm (LIBERO on robosuite / MuJoCo) **as if it were real hardware**: the
-simulator runs in a service you cannot open, and you reach it only through a client. **Task: Pick up the cup and place it in the back compartment of the caddy.**
+simulator runs in a service you cannot open, and you reach it only through a client. **Task: Pick up the cup and place it in the back compartment of the caddy.** The caddy has a large cell at each end and, between them, two narrow cells: the front one and the back one, which runs along the caddy's tall wall. The back compartment is that narrow back cell (`desk_caddy_1_back_contain_region`, about 5.5 cm across): the mug (about 9 cm wide) does not fit down into it. LIBERO's In tests only the mug's origin against the region's box (its top is about 15 cm above the table) and its contact with the caddy, so a mug resting on top of the back cell's opening, its origin just inside the top of that box, counts; a mug in any other cell does not.
The scene can change while you work: once during the episode, when your gripper first gets close to one of the task's
objects, the scene changes (as in LIBERO-MAX): an object or a receptacle may be moved, an obstacle or a few other
objects may appear, a camera may shift, or the lighting, the colours or the image noise may change. Nothing tells you
-when; the task and its success check stay the same, so keep checking what the cameras show.
+when; the task and its success check stay the same, so keep checking what the cameras show. A moved object or
+receptacle jumps there alone: what already rests in or on it stays where it was.
## Rules (real-robot rules)
- **One episode, no reset.** The episode starts when the service comes up. It ends when the task succeeds, after
diff --git a/tasks/libero-max-long-white-yellow-mug-in-desk-caddy-back-illumination-switch-i00-standard/task.toml b/tasks/libero-max-long-white-yellow-mug-in-desk-caddy-back-illumination-switch-i00-standard/task.toml
index eef9aee657..6e1a807c3e 100644
--- a/tasks/libero-max-long-white-yellow-mug-in-desk-caddy-back-illumination-switch-i00-standard/task.toml
+++ b/tasks/libero-max-long-white-yellow-mug-in-desk-caddy-back-illumination-switch-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "robosuite"
robot = "franka-panda"
family = "libero-max/libero_10/illumination_switch"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 195
diff --git a/tasks/libero-max-object-alphabet-soup-in-basket-receptacle-relocation-i00-privileged/README.md b/tasks/libero-max-object-alphabet-soup-in-basket-receptacle-relocation-i00-privileged/README.md
index 9799f0f636..7217d6345d 100644
--- a/tasks/libero-max-object-alphabet-soup-in-basket-receptacle-relocation-i00-privileged/README.md
+++ b/tasks/libero-max-object-alphabet-soup-in-basket-receptacle-relocation-i00-privileged/README.md
@@ -14,7 +14,7 @@
## Task
LIBERO-MAX case `pro-object-texture-object-t00-i08-receptacle_relocation-d0-p195`: a Franka Panda arm at a table, instructed **"Pick the alphabet soup and place it in the basket."**, on a LIBERO-PRO variant
of a libero_object task (object_texture). Online event (receptacle relocation): the first time the gripper comes within
-0.18 m of `alphabet_soup_1`, LIBERO-MAX changes the scene once (the receptacle `basket_1` is moved 0.06 m, direction (-1, 0)). The deliverable is one
+0.18 m of `alphabet_soup_1`, LIBERO-MAX changes the scene once (the receptacle `basket_1` is moved 0.06 m, direction (-1, 0); it jumps there alone: what already rests in or on it stays where it was, and where it lands on another object the physics pushes the two apart). The deliverable is one
replayable trajectory of LIBERO's OSC control steps (`/app/output/trajectory.npz`, `(T, 7)`, T ≤ 280), graded by
LIBERO's success check at its end.
@@ -54,7 +54,7 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
|---|---|
| Upstream | LIBERO-PRO @ 2b910b5 (MIT; task data `zhouxueyang/LIBERO-Pro` @ c86fc3b, CC-BY-4.0); LIBERO-MAX @ a1e3cef (event runtime and this case's parameters; no licence file upstream); robosuite 1.4.0, MuJoCo 3.2.6 |
| Case | `pro-object-texture-object-t00-i08-receptacle_relocation-d0-p195` (LIBERO-MAX 8,000-case manifest; LIBERO-PRO category `object_texture`, task 0, initial state 8) |
-| Event | `receptacle_relocation`: the receptacle `basket_1` is moved 0.06 m, direction (-1, 0); trigger: gripper within 0.18 m of `alphabet_soup_1` |
+| Event | `receptacle_relocation`: the receptacle `basket_1` is moved 0.06 m, direction (-1, 0); it jumps there alone: what already rests in or on it stays where it was, and where it lands on another object the physics pushes the two apart; trigger: gripper within 0.18 m of `alphabet_soup_1` |
| Actions | 7 per control step (1/20 s): OSC hand-target delta (1.0 = 0.05 m, 0.5 rad) and gripper |
| Resources | 4 CPUs, no GPU (CPU physics, OSMesa rendering) |
@@ -76,3 +76,4 @@ Generated by `scripts/libero-max/generate.py` from `scripts/libero-max/cases.jso
it executes any action). Before, such a trial read as an agent failure with `grader_error = 1` only.
- A replay that records a video prints its verdict before it finishes the video, and a failing video no longer fails
the replay; `video_rendered` is 1 only when the video was finished.
+- 2026-10-14 (board 0113, from board 0086): the instruction says that a moved receptacle jumps alone: what rests in or on it stays where it was (LIBERO-MAX moves one entity; board 0086 saw a basket jump through a can already in it). Text only: the instance, the event, the reference solution, LIBERO's success check and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/libero-max-object-alphabet-soup-in-basket-receptacle-relocation-i00-privileged/instruction.md b/tasks/libero-max-object-alphabet-soup-in-basket-receptacle-relocation-i00-privileged/instruction.md
index f3d2f3845a..8eeca1f8d1 100644
--- a/tasks/libero-max-object-alphabet-soup-in-basket-receptacle-relocation-i00-privileged/instruction.md
+++ b/tasks/libero-max-object-alphabet-soup-in-basket-receptacle-relocation-i00-privileged/instruction.md
@@ -3,7 +3,7 @@
A **Franka Panda** arm (LIBERO on robosuite / MuJoCo) at a table. **Task: Pick the alphabet soup and place it in the basket.**
The scene is one frozen LIBERO-MAX case: a LIBERO-PRO variant of a libero_object task (object_texture), with one online
event: the first time the gripper comes within 0.18 m of `alphabet_soup_1`, LIBERO-MAX changes the scene
-once (the receptacle `basket_1` is moved 0.06 m, direction (-1, 0)). You have the full simulator state, so you can see the change happen (`s.state()["event_fired"]`).
+once (the receptacle `basket_1` is moved 0.06 m, direction (-1, 0); it jumps there alone: what already rests in or on it stays where it was, and where it lands on another object the physics pushes the two apart). You have the full simulator state, so you can see the change happen (`s.state()["event_fired"]`).
## What you hand in
`/app/output/trajectory.npz` with key `actions`: LIBERO's native control steps, shape `(T, 7)`, 1 ≤ T ≤ 280
diff --git a/tasks/libero-max-object-alphabet-soup-in-basket-receptacle-relocation-i00-privileged/task.toml b/tasks/libero-max-object-alphabet-soup-in-basket-receptacle-relocation-i00-privileged/task.toml
index a7405b48bc..87029eebda 100644
--- a/tasks/libero-max-object-alphabet-soup-in-basket-receptacle-relocation-i00-privileged/task.toml
+++ b/tasks/libero-max-object-alphabet-soup-in-basket-receptacle-relocation-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "robosuite"
robot = "franka-panda"
family = "libero-max/libero_object/receptacle_relocation"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 195 # LIBERO-MAX's policy seed for the case (the scene is the frozen LIBERO-PRO initial state)
expert_time_estimate_hours = 1.0
diff --git a/tasks/libero-max-object-alphabet-soup-in-basket-receptacle-relocation-i00-standard/README.md b/tasks/libero-max-object-alphabet-soup-in-basket-receptacle-relocation-i00-standard/README.md
index a874c16f23..174e42941b 100644
--- a/tasks/libero-max-object-alphabet-soup-in-basket-receptacle-relocation-i00-standard/README.md
+++ b/tasks/libero-max-object-alphabet-soup-in-basket-receptacle-relocation-i00-standard/README.md
@@ -13,7 +13,7 @@
## Task
LIBERO-MAX case `pro-object-texture-object-t00-i08-receptacle_relocation-d0-p195` as a robot service: a Franka Panda arm at a table, instructed **"Pick the alphabet soup and place it in the basket."**, on
-the same frozen instance as `libero-max-object-alphabet-soup-in-basket-receptacle-relocation-i00-privileged`, with the same online event (receptacle relocation: the receptacle `basket_1` is moved 0.06 m, direction (-1, 0), the
+the same frozen instance as `libero-max-object-alphabet-soup-in-basket-receptacle-relocation-i00-privileged`, with the same online event (receptacle relocation: the receptacle `basket_1` is moved 0.06 m, direction (-1, 0); it jumps there alone: what already rests in or on it stays where it was, and where it lands on another object the physics pushes the two apart, the
first time the gripper comes within 0.18 m of `alphabet_soup_1`). The simulator runs in the `sim` sidecar
(`environment/sim/server.py` on main's `rcb_service.py`); the agent only has the client: two cameras (RGB
256×256, depth on request, calibrated), the robot's own state, and OSC control steps in chunks. One
@@ -86,3 +86,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
/reset` ends the episode as `error` (`sim_error = 1`), not `reset` (reachable only when a run grants resets; the
default is none). The scene, instance, pass rule and every verdict reached before are unchanged, so the task keeps
its name.
+- 2026-10-14 (board 0113, from board 0086): the instruction says that a moved object or receptacle jumps alone: what rests in or on it stays where it was (LIBERO-MAX moves one entity; board 0086 saw a basket jump through a can already in it). Text only: the instance, the event, the reference solution, LIBERO's success check and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/libero-max-object-alphabet-soup-in-basket-receptacle-relocation-i00-standard/instruction.md b/tasks/libero-max-object-alphabet-soup-in-basket-receptacle-relocation-i00-standard/instruction.md
index 97e5874e94..425eb65eab 100644
--- a/tasks/libero-max-object-alphabet-soup-in-basket-receptacle-relocation-i00-standard/instruction.md
+++ b/tasks/libero-max-object-alphabet-soup-in-basket-receptacle-relocation-i00-standard/instruction.md
@@ -6,7 +6,8 @@ simulator runs in a service you cannot open, and you reach it only through a cli
The scene can change while you work: once during the episode, when your gripper first gets close to one of the task's
objects, the scene changes (as in LIBERO-MAX): an object or a receptacle may be moved, an obstacle or a few other
objects may appear, a camera may shift, or the lighting, the colours or the image noise may change. Nothing tells you
-when; the task and its success check stay the same, so keep checking what the cameras show.
+when; the task and its success check stay the same, so keep checking what the cameras show. A moved object or
+receptacle jumps there alone: what already rests in or on it stays where it was.
## Rules (real-robot rules)
- **One episode, no reset.** The episode starts when the service comes up. It ends when the task succeeds, after
diff --git a/tasks/libero-max-object-alphabet-soup-in-basket-receptacle-relocation-i00-standard/task.toml b/tasks/libero-max-object-alphabet-soup-in-basket-receptacle-relocation-i00-standard/task.toml
index e1743e489a..fca9f553ba 100644
--- a/tasks/libero-max-object-alphabet-soup-in-basket-receptacle-relocation-i00-standard/task.toml
+++ b/tasks/libero-max-object-alphabet-soup-in-basket-receptacle-relocation-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "robosuite"
robot = "franka-panda"
family = "libero-max/libero_object/receptacle_relocation"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 195
diff --git a/tasks/libero-max-object-alphabet-soup-in-basket-target-relocation-i00-privileged/task.toml b/tasks/libero-max-object-alphabet-soup-in-basket-target-relocation-i00-privileged/task.toml
index abeff9d528..8d56d5d9cb 100644
--- a/tasks/libero-max-object-alphabet-soup-in-basket-target-relocation-i00-privileged/task.toml
+++ b/tasks/libero-max-object-alphabet-soup-in-basket-target-relocation-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "robosuite"
robot = "franka-panda"
family = "libero-max/libero_object/target_relocation"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 195 # LIBERO-MAX's policy seed for the case (the scene is the frozen LIBERO-PRO initial state)
expert_time_estimate_hours = 1.0
diff --git a/tasks/libero-max-object-alphabet-soup-in-basket-target-relocation-i00-standard/README.md b/tasks/libero-max-object-alphabet-soup-in-basket-target-relocation-i00-standard/README.md
index 460b3bf035..8d80955a97 100644
--- a/tasks/libero-max-object-alphabet-soup-in-basket-target-relocation-i00-standard/README.md
+++ b/tasks/libero-max-object-alphabet-soup-in-basket-target-relocation-i00-standard/README.md
@@ -86,3 +86,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
/reset` ends the episode as `error` (`sim_error = 1`), not `reset` (reachable only when a run grants resets; the
default is none). The scene, instance, pass rule and every verdict reached before are unchanged, so the task keeps
its name.
+- 2026-10-14 (board 0113, from board 0086): the instruction says that a moved object or receptacle jumps alone: what rests in or on it stays where it was (LIBERO-MAX moves one entity; board 0086 saw a basket jump through a can already in it). Text only: the instance, the event, the reference solution, LIBERO's success check and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/libero-max-object-alphabet-soup-in-basket-target-relocation-i00-standard/instruction.md b/tasks/libero-max-object-alphabet-soup-in-basket-target-relocation-i00-standard/instruction.md
index 1e8fc997aa..f8fd095537 100644
--- a/tasks/libero-max-object-alphabet-soup-in-basket-target-relocation-i00-standard/instruction.md
+++ b/tasks/libero-max-object-alphabet-soup-in-basket-target-relocation-i00-standard/instruction.md
@@ -6,7 +6,8 @@ simulator runs in a service you cannot open, and you reach it only through a cli
The scene can change while you work: once during the episode, when your gripper first gets close to one of the task's
objects, the scene changes (as in LIBERO-MAX): an object or a receptacle may be moved, an obstacle or a few other
objects may appear, a camera may shift, or the lighting, the colours or the image noise may change. Nothing tells you
-when; the task and its success check stay the same, so keep checking what the cameras show.
+when; the task and its success check stay the same, so keep checking what the cameras show. A moved object or
+receptacle jumps there alone: what already rests in or on it stays where it was.
## Rules (real-robot rules)
- **One episode, no reset.** The episode starts when the service comes up. It ends when the task succeeds, after
diff --git a/tasks/libero-max-object-alphabet-soup-in-basket-target-relocation-i00-standard/task.toml b/tasks/libero-max-object-alphabet-soup-in-basket-target-relocation-i00-standard/task.toml
index 1ad21ae0f4..e7e62bb2d6 100644
--- a/tasks/libero-max-object-alphabet-soup-in-basket-target-relocation-i00-standard/task.toml
+++ b/tasks/libero-max-object-alphabet-soup-in-basket-target-relocation-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "robosuite"
robot = "franka-panda"
family = "libero-max/libero_object/target_relocation"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 195
diff --git a/tasks/libero-max-object-alphabet-soup-in-basket-visual-theme-switch-i00-privileged/task.toml b/tasks/libero-max-object-alphabet-soup-in-basket-visual-theme-switch-i00-privileged/task.toml
index fa6abd7db5..1bfa1b1d37 100644
--- a/tasks/libero-max-object-alphabet-soup-in-basket-visual-theme-switch-i00-privileged/task.toml
+++ b/tasks/libero-max-object-alphabet-soup-in-basket-visual-theme-switch-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "robosuite"
robot = "franka-panda"
family = "libero-max/libero_object/visual_theme_switch"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 195 # LIBERO-MAX's policy seed for the case (the scene is the frozen LIBERO-PRO initial state)
expert_time_estimate_hours = 1.0
diff --git a/tasks/libero-max-object-alphabet-soup-in-basket-visual-theme-switch-i00-standard/README.md b/tasks/libero-max-object-alphabet-soup-in-basket-visual-theme-switch-i00-standard/README.md
index 98069e6268..544d822085 100644
--- a/tasks/libero-max-object-alphabet-soup-in-basket-visual-theme-switch-i00-standard/README.md
+++ b/tasks/libero-max-object-alphabet-soup-in-basket-visual-theme-switch-i00-standard/README.md
@@ -86,3 +86,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
/reset` ends the episode as `error` (`sim_error = 1`), not `reset` (reachable only when a run grants resets; the
default is none). The scene, instance, pass rule and every verdict reached before are unchanged, so the task keeps
its name.
+- 2026-10-14 (board 0113, from board 0086): the instruction says that a moved object or receptacle jumps alone: what rests in or on it stays where it was (LIBERO-MAX moves one entity; board 0086 saw a basket jump through a can already in it). Text only: the instance, the event, the reference solution, LIBERO's success check and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/libero-max-object-alphabet-soup-in-basket-visual-theme-switch-i00-standard/instruction.md b/tasks/libero-max-object-alphabet-soup-in-basket-visual-theme-switch-i00-standard/instruction.md
index a47c22931a..4107d28a96 100644
--- a/tasks/libero-max-object-alphabet-soup-in-basket-visual-theme-switch-i00-standard/instruction.md
+++ b/tasks/libero-max-object-alphabet-soup-in-basket-visual-theme-switch-i00-standard/instruction.md
@@ -6,7 +6,8 @@ simulator runs in a service you cannot open, and you reach it only through a cli
The scene can change while you work: once during the episode, when your gripper first gets close to one of the task's
objects, the scene changes (as in LIBERO-MAX): an object or a receptacle may be moved, an obstacle or a few other
objects may appear, a camera may shift, or the lighting, the colours or the image noise may change. Nothing tells you
-when; the task and its success check stay the same, so keep checking what the cameras show.
+when; the task and its success check stay the same, so keep checking what the cameras show. A moved object or
+receptacle jumps there alone: what already rests in or on it stays where it was.
## Rules (real-robot rules)
- **One episode, no reset.** The episode starts when the service comes up. It ends when the task succeeds, after
diff --git a/tasks/libero-max-object-alphabet-soup-in-basket-visual-theme-switch-i00-standard/task.toml b/tasks/libero-max-object-alphabet-soup-in-basket-visual-theme-switch-i00-standard/task.toml
index 44075066ce..860212896d 100644
--- a/tasks/libero-max-object-alphabet-soup-in-basket-visual-theme-switch-i00-standard/task.toml
+++ b/tasks/libero-max-object-alphabet-soup-in-basket-visual-theme-switch-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "robosuite"
robot = "franka-panda"
family = "libero-max/libero_object/visual_theme_switch"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 195
diff --git a/tasks/libero-max-object-butter-in-basket-distractor-burst-i00-privileged/task.toml b/tasks/libero-max-object-butter-in-basket-distractor-burst-i00-privileged/task.toml
index 35f8034148..c2bf124bd3 100644
--- a/tasks/libero-max-object-butter-in-basket-distractor-burst-i00-privileged/task.toml
+++ b/tasks/libero-max-object-butter-in-basket-distractor-burst-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "robosuite"
robot = "franka-panda"
family = "libero-max/libero_object/distractor_burst"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 195 # LIBERO-MAX's policy seed for the case (the scene is the frozen LIBERO-PRO initial state)
expert_time_estimate_hours = 1.0
diff --git a/tasks/libero-max-object-butter-in-basket-distractor-burst-i00-standard/README.md b/tasks/libero-max-object-butter-in-basket-distractor-burst-i00-standard/README.md
index 03e4b6218f..a85df8c571 100644
--- a/tasks/libero-max-object-butter-in-basket-distractor-burst-i00-standard/README.md
+++ b/tasks/libero-max-object-butter-in-basket-distractor-burst-i00-standard/README.md
@@ -86,3 +86,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
/reset` ends the episode as `error` (`sim_error = 1`), not `reset` (reachable only when a run grants resets; the
default is none). The scene, instance, pass rule and every verdict reached before are unchanged, so the task keeps
its name.
+- 2026-10-14 (board 0113, from board 0086): the instruction says that a moved object or receptacle jumps alone: what rests in or on it stays where it was (LIBERO-MAX moves one entity; board 0086 saw a basket jump through a can already in it). Text only: the instance, the event, the reference solution, LIBERO's success check and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/libero-max-object-butter-in-basket-distractor-burst-i00-standard/instruction.md b/tasks/libero-max-object-butter-in-basket-distractor-burst-i00-standard/instruction.md
index 8e139120de..ee3a347825 100644
--- a/tasks/libero-max-object-butter-in-basket-distractor-burst-i00-standard/instruction.md
+++ b/tasks/libero-max-object-butter-in-basket-distractor-burst-i00-standard/instruction.md
@@ -6,7 +6,8 @@ simulator runs in a service you cannot open, and you reach it only through a cli
The scene can change while you work: once during the episode, when your gripper first gets close to one of the task's
objects, the scene changes (as in LIBERO-MAX): an object or a receptacle may be moved, an obstacle or a few other
objects may appear, a camera may shift, or the lighting, the colours or the image noise may change. Nothing tells you
-when; the task and its success check stay the same, so keep checking what the cameras show.
+when; the task and its success check stay the same, so keep checking what the cameras show. A moved object or
+receptacle jumps there alone: what already rests in or on it stays where it was.
## Rules (real-robot rules)
- **One episode, no reset.** The episode starts when the service comes up. It ends when the task succeeds, after
diff --git a/tasks/libero-max-object-butter-in-basket-distractor-burst-i00-standard/task.toml b/tasks/libero-max-object-butter-in-basket-distractor-burst-i00-standard/task.toml
index 67430035e6..e30d16b187 100644
--- a/tasks/libero-max-object-butter-in-basket-distractor-burst-i00-standard/task.toml
+++ b/tasks/libero-max-object-butter-in-basket-distractor-burst-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "robosuite"
robot = "franka-panda"
family = "libero-max/libero_object/distractor_burst"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 195
diff --git a/tasks/libero-max-object-butter-in-basket-receptacle-relocation-i00-privileged/README.md b/tasks/libero-max-object-butter-in-basket-receptacle-relocation-i00-privileged/README.md
index 8b137f7844..e4fd18b365 100644
--- a/tasks/libero-max-object-butter-in-basket-receptacle-relocation-i00-privileged/README.md
+++ b/tasks/libero-max-object-butter-in-basket-receptacle-relocation-i00-privileged/README.md
@@ -14,7 +14,7 @@
## Task
LIBERO-MAX case `pro-initial-pose-position-angle-object-t06-i08-receptacle_relocation-d0-p195`: a Franka Panda arm at a table, instructed **"Pick the butter and place it in the basket."**, on a LIBERO-PRO variant
of a libero_object task (initial_pose_position_angle). Online event (receptacle relocation): the first time the gripper comes within
-0.18 m of `butter_1`, LIBERO-MAX changes the scene once (the receptacle `basket_1` is moved 0.06 m, direction (-0.707, 0.707)). The deliverable is one
+0.18 m of `butter_1`, LIBERO-MAX changes the scene once (the receptacle `basket_1` is moved 0.06 m, direction (-0.707, 0.707); it jumps there alone: what already rests in or on it stays where it was, and where it lands on another object the physics pushes the two apart). The deliverable is one
replayable trajectory of LIBERO's OSC control steps (`/app/output/trajectory.npz`, `(T, 7)`, T ≤ 280), graded by
LIBERO's success check at its end.
@@ -54,7 +54,7 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
|---|---|
| Upstream | LIBERO-PRO @ 2b910b5 (MIT; task data `zhouxueyang/LIBERO-Pro` @ c86fc3b, CC-BY-4.0); LIBERO-MAX @ a1e3cef (event runtime and this case's parameters; no licence file upstream); robosuite 1.4.0, MuJoCo 3.2.6 |
| Case | `pro-initial-pose-position-angle-object-t06-i08-receptacle_relocation-d0-p195` (LIBERO-MAX 8,000-case manifest; LIBERO-PRO category `initial_pose_position_angle`, task 6, initial state 8) |
-| Event | `receptacle_relocation`: the receptacle `basket_1` is moved 0.06 m, direction (-0.707, 0.707); trigger: gripper within 0.18 m of `butter_1` |
+| Event | `receptacle_relocation`: the receptacle `basket_1` is moved 0.06 m, direction (-0.707, 0.707); it jumps there alone: what already rests in or on it stays where it was, and where it lands on another object the physics pushes the two apart; trigger: gripper within 0.18 m of `butter_1` |
| Actions | 7 per control step (1/20 s): OSC hand-target delta (1.0 = 0.05 m, 0.5 rad) and gripper |
| Resources | 4 CPUs, no GPU (CPU physics, OSMesa rendering) |
@@ -76,3 +76,4 @@ Generated by `scripts/libero-max/generate.py` from `scripts/libero-max/cases.jso
it executes any action). Before, such a trial read as an agent failure with `grader_error = 1` only.
- A replay that records a video prints its verdict before it finishes the video, and a failing video no longer fails
the replay; `video_rendered` is 1 only when the video was finished.
+- 2026-10-14 (board 0113, from board 0086): the instruction says that a moved receptacle jumps alone: what rests in or on it stays where it was (LIBERO-MAX moves one entity; board 0086 saw a basket jump through a can already in it). Text only: the instance, the event, the reference solution, LIBERO's success check and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/libero-max-object-butter-in-basket-receptacle-relocation-i00-privileged/instruction.md b/tasks/libero-max-object-butter-in-basket-receptacle-relocation-i00-privileged/instruction.md
index 01511f13c2..706fcee976 100644
--- a/tasks/libero-max-object-butter-in-basket-receptacle-relocation-i00-privileged/instruction.md
+++ b/tasks/libero-max-object-butter-in-basket-receptacle-relocation-i00-privileged/instruction.md
@@ -3,7 +3,7 @@
A **Franka Panda** arm (LIBERO on robosuite / MuJoCo) at a table. **Task: Pick the butter and place it in the basket.**
The scene is one frozen LIBERO-MAX case: a LIBERO-PRO variant of a libero_object task (initial_pose_position_angle), with one online
event: the first time the gripper comes within 0.18 m of `butter_1`, LIBERO-MAX changes the scene
-once (the receptacle `basket_1` is moved 0.06 m, direction (-0.707, 0.707)). You have the full simulator state, so you can see the change happen (`s.state()["event_fired"]`).
+once (the receptacle `basket_1` is moved 0.06 m, direction (-0.707, 0.707); it jumps there alone: what already rests in or on it stays where it was, and where it lands on another object the physics pushes the two apart). You have the full simulator state, so you can see the change happen (`s.state()["event_fired"]`).
## What you hand in
`/app/output/trajectory.npz` with key `actions`: LIBERO's native control steps, shape `(T, 7)`, 1 ≤ T ≤ 280
diff --git a/tasks/libero-max-object-butter-in-basket-receptacle-relocation-i00-privileged/task.toml b/tasks/libero-max-object-butter-in-basket-receptacle-relocation-i00-privileged/task.toml
index e88c96bb9c..80038733cb 100644
--- a/tasks/libero-max-object-butter-in-basket-receptacle-relocation-i00-privileged/task.toml
+++ b/tasks/libero-max-object-butter-in-basket-receptacle-relocation-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "robosuite"
robot = "franka-panda"
family = "libero-max/libero_object/receptacle_relocation"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 195 # LIBERO-MAX's policy seed for the case (the scene is the frozen LIBERO-PRO initial state)
expert_time_estimate_hours = 1.0
diff --git a/tasks/libero-max-object-butter-in-basket-receptacle-relocation-i00-standard/README.md b/tasks/libero-max-object-butter-in-basket-receptacle-relocation-i00-standard/README.md
index 020ed91849..32475b350d 100644
--- a/tasks/libero-max-object-butter-in-basket-receptacle-relocation-i00-standard/README.md
+++ b/tasks/libero-max-object-butter-in-basket-receptacle-relocation-i00-standard/README.md
@@ -13,7 +13,7 @@
## Task
LIBERO-MAX case `pro-initial-pose-position-angle-object-t06-i08-receptacle_relocation-d0-p195` as a robot service: a Franka Panda arm at a table, instructed **"Pick the butter and place it in the basket."**, on
-the same frozen instance as `libero-max-object-butter-in-basket-receptacle-relocation-i00-privileged`, with the same online event (receptacle relocation: the receptacle `basket_1` is moved 0.06 m, direction (-0.707, 0.707), the
+the same frozen instance as `libero-max-object-butter-in-basket-receptacle-relocation-i00-privileged`, with the same online event (receptacle relocation: the receptacle `basket_1` is moved 0.06 m, direction (-0.707, 0.707); it jumps there alone: what already rests in or on it stays where it was, and where it lands on another object the physics pushes the two apart, the
first time the gripper comes within 0.18 m of `butter_1`). The simulator runs in the `sim` sidecar
(`environment/sim/server.py` on main's `rcb_service.py`); the agent only has the client: two cameras (RGB
256×256, depth on request, calibrated), the robot's own state, and OSC control steps in chunks. One
@@ -86,3 +86,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
/reset` ends the episode as `error` (`sim_error = 1`), not `reset` (reachable only when a run grants resets; the
default is none). The scene, instance, pass rule and every verdict reached before are unchanged, so the task keeps
its name.
+- 2026-10-14 (board 0113, from board 0086): the instruction says that a moved object or receptacle jumps alone: what rests in or on it stays where it was (LIBERO-MAX moves one entity; board 0086 saw a basket jump through a can already in it). Text only: the instance, the event, the reference solution, LIBERO's success check and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/libero-max-object-butter-in-basket-receptacle-relocation-i00-standard/instruction.md b/tasks/libero-max-object-butter-in-basket-receptacle-relocation-i00-standard/instruction.md
index fc6ec42862..e4245521b1 100644
--- a/tasks/libero-max-object-butter-in-basket-receptacle-relocation-i00-standard/instruction.md
+++ b/tasks/libero-max-object-butter-in-basket-receptacle-relocation-i00-standard/instruction.md
@@ -6,7 +6,8 @@ simulator runs in a service you cannot open, and you reach it only through a cli
The scene can change while you work: once during the episode, when your gripper first gets close to one of the task's
objects, the scene changes (as in LIBERO-MAX): an object or a receptacle may be moved, an obstacle or a few other
objects may appear, a camera may shift, or the lighting, the colours or the image noise may change. Nothing tells you
-when; the task and its success check stay the same, so keep checking what the cameras show.
+when; the task and its success check stay the same, so keep checking what the cameras show. A moved object or
+receptacle jumps there alone: what already rests in or on it stays where it was.
## Rules (real-robot rules)
- **One episode, no reset.** The episode starts when the service comes up. It ends when the task succeeds, after
diff --git a/tasks/libero-max-object-butter-in-basket-receptacle-relocation-i00-standard/task.toml b/tasks/libero-max-object-butter-in-basket-receptacle-relocation-i00-standard/task.toml
index f39d925ce7..0ffc9ec3f4 100644
--- a/tasks/libero-max-object-butter-in-basket-receptacle-relocation-i00-standard/task.toml
+++ b/tasks/libero-max-object-butter-in-basket-receptacle-relocation-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "robosuite"
robot = "franka-panda"
family = "libero-max/libero_object/receptacle_relocation"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 195
diff --git a/tasks/libero-max-object-chocolate-pudding-in-basket-distractor-burst-i00-privileged/task.toml b/tasks/libero-max-object-chocolate-pudding-in-basket-distractor-burst-i00-privileged/task.toml
index 2220f4c686..90a9ce1321 100644
--- a/tasks/libero-max-object-chocolate-pudding-in-basket-distractor-burst-i00-privileged/task.toml
+++ b/tasks/libero-max-object-chocolate-pudding-in-basket-distractor-burst-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "robosuite"
robot = "franka-panda"
family = "libero-max/libero_object/distractor_burst"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 195 # LIBERO-MAX's policy seed for the case (the scene is the frozen LIBERO-PRO initial state)
expert_time_estimate_hours = 1.0
diff --git a/tasks/libero-max-object-chocolate-pudding-in-basket-distractor-burst-i00-standard/README.md b/tasks/libero-max-object-chocolate-pudding-in-basket-distractor-burst-i00-standard/README.md
index f4c380bc63..8aa4448355 100644
--- a/tasks/libero-max-object-chocolate-pudding-in-basket-distractor-burst-i00-standard/README.md
+++ b/tasks/libero-max-object-chocolate-pudding-in-basket-distractor-burst-i00-standard/README.md
@@ -86,3 +86,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
/reset` ends the episode as `error` (`sim_error = 1`), not `reset` (reachable only when a run grants resets; the
default is none). The scene, instance, pass rule and every verdict reached before are unchanged, so the task keeps
its name.
+- 2026-10-14 (board 0113, from board 0086): the instruction says that a moved object or receptacle jumps alone: what rests in or on it stays where it was (LIBERO-MAX moves one entity; board 0086 saw a basket jump through a can already in it). Text only: the instance, the event, the reference solution, LIBERO's success check and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/libero-max-object-chocolate-pudding-in-basket-distractor-burst-i00-standard/instruction.md b/tasks/libero-max-object-chocolate-pudding-in-basket-distractor-burst-i00-standard/instruction.md
index 7110895140..89b878ed19 100644
--- a/tasks/libero-max-object-chocolate-pudding-in-basket-distractor-burst-i00-standard/instruction.md
+++ b/tasks/libero-max-object-chocolate-pudding-in-basket-distractor-burst-i00-standard/instruction.md
@@ -6,7 +6,8 @@ simulator runs in a service you cannot open, and you reach it only through a cli
The scene can change while you work: once during the episode, when your gripper first gets close to one of the task's
objects, the scene changes (as in LIBERO-MAX): an object or a receptacle may be moved, an obstacle or a few other
objects may appear, a camera may shift, or the lighting, the colours or the image noise may change. Nothing tells you
-when; the task and its success check stay the same, so keep checking what the cameras show.
+when; the task and its success check stay the same, so keep checking what the cameras show. A moved object or
+receptacle jumps there alone: what already rests in or on it stays where it was.
## Rules (real-robot rules)
- **One episode, no reset.** The episode starts when the service comes up. It ends when the task succeeds, after
diff --git a/tasks/libero-max-object-chocolate-pudding-in-basket-distractor-burst-i00-standard/task.toml b/tasks/libero-max-object-chocolate-pudding-in-basket-distractor-burst-i00-standard/task.toml
index 214504d7fd..bbd1ae64d2 100644
--- a/tasks/libero-max-object-chocolate-pudding-in-basket-distractor-burst-i00-standard/task.toml
+++ b/tasks/libero-max-object-chocolate-pudding-in-basket-distractor-burst-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "robosuite"
robot = "franka-panda"
family = "libero-max/libero_object/distractor_burst"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 195
diff --git a/tasks/libero-max-object-chocolate-pudding-in-basket-distractor-burst-i01-privileged/task.toml b/tasks/libero-max-object-chocolate-pudding-in-basket-distractor-burst-i01-privileged/task.toml
index f55bf8e853..b82022a6f3 100644
--- a/tasks/libero-max-object-chocolate-pudding-in-basket-distractor-burst-i01-privileged/task.toml
+++ b/tasks/libero-max-object-chocolate-pudding-in-basket-distractor-burst-i01-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "robosuite"
robot = "franka-panda"
family = "libero-max/libero_object/distractor_burst"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 195 # LIBERO-MAX's policy seed for the case (the scene is the frozen LIBERO-PRO initial state)
expert_time_estimate_hours = 1.0
diff --git a/tasks/libero-max-object-chocolate-pudding-in-basket-distractor-burst-i01-standard/README.md b/tasks/libero-max-object-chocolate-pudding-in-basket-distractor-burst-i01-standard/README.md
index 8b0d909546..f024a049bc 100644
--- a/tasks/libero-max-object-chocolate-pudding-in-basket-distractor-burst-i01-standard/README.md
+++ b/tasks/libero-max-object-chocolate-pudding-in-basket-distractor-burst-i01-standard/README.md
@@ -86,3 +86,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
/reset` ends the episode as `error` (`sim_error = 1`), not `reset` (reachable only when a run grants resets; the
default is none). The scene, instance, pass rule and every verdict reached before are unchanged, so the task keeps
its name.
+- 2026-10-14 (board 0113, from board 0086): the instruction says that a moved object or receptacle jumps alone: what rests in or on it stays where it was (LIBERO-MAX moves one entity; board 0086 saw a basket jump through a can already in it). Text only: the instance, the event, the reference solution, LIBERO's success check and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/libero-max-object-chocolate-pudding-in-basket-distractor-burst-i01-standard/instruction.md b/tasks/libero-max-object-chocolate-pudding-in-basket-distractor-burst-i01-standard/instruction.md
index e50f3859a1..09d25a005d 100644
--- a/tasks/libero-max-object-chocolate-pudding-in-basket-distractor-burst-i01-standard/instruction.md
+++ b/tasks/libero-max-object-chocolate-pudding-in-basket-distractor-burst-i01-standard/instruction.md
@@ -6,7 +6,8 @@ simulator runs in a service you cannot open, and you reach it only through a cli
The scene can change while you work: once during the episode, when your gripper first gets close to one of the task's
objects, the scene changes (as in LIBERO-MAX): an object or a receptacle may be moved, an obstacle or a few other
objects may appear, a camera may shift, or the lighting, the colours or the image noise may change. Nothing tells you
-when; the task and its success check stay the same, so keep checking what the cameras show.
+when; the task and its success check stay the same, so keep checking what the cameras show. A moved object or
+receptacle jumps there alone: what already rests in or on it stays where it was.
## Rules (real-robot rules)
- **One episode, no reset.** The episode starts when the service comes up. It ends when the task succeeds, after
diff --git a/tasks/libero-max-object-chocolate-pudding-in-basket-distractor-burst-i01-standard/task.toml b/tasks/libero-max-object-chocolate-pudding-in-basket-distractor-burst-i01-standard/task.toml
index 7577dd8a6b..779b47f22c 100644
--- a/tasks/libero-max-object-chocolate-pudding-in-basket-distractor-burst-i01-standard/task.toml
+++ b/tasks/libero-max-object-chocolate-pudding-in-basket-distractor-burst-i01-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "robosuite"
robot = "franka-panda"
family = "libero-max/libero_object/distractor_burst"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 195
diff --git a/tasks/libero-max-object-chocolate-pudding-in-basket-sensor-noise-onset-i00-privileged/task.toml b/tasks/libero-max-object-chocolate-pudding-in-basket-sensor-noise-onset-i00-privileged/task.toml
index ab028e8999..8e43b70ee2 100644
--- a/tasks/libero-max-object-chocolate-pudding-in-basket-sensor-noise-onset-i00-privileged/task.toml
+++ b/tasks/libero-max-object-chocolate-pudding-in-basket-sensor-noise-onset-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "robosuite"
robot = "franka-panda"
family = "libero-max/libero_object/sensor_noise_onset"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 195 # LIBERO-MAX's policy seed for the case (the scene is the frozen LIBERO-PRO initial state)
expert_time_estimate_hours = 1.0
diff --git a/tasks/libero-max-object-chocolate-pudding-in-basket-sensor-noise-onset-i00-standard/README.md b/tasks/libero-max-object-chocolate-pudding-in-basket-sensor-noise-onset-i00-standard/README.md
index 22f29e0eeb..fd01e3e401 100644
--- a/tasks/libero-max-object-chocolate-pudding-in-basket-sensor-noise-onset-i00-standard/README.md
+++ b/tasks/libero-max-object-chocolate-pudding-in-basket-sensor-noise-onset-i00-standard/README.md
@@ -86,3 +86,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
/reset` ends the episode as `error` (`sim_error = 1`), not `reset` (reachable only when a run grants resets; the
default is none). The scene, instance, pass rule and every verdict reached before are unchanged, so the task keeps
its name.
+- 2026-10-14 (board 0113, from board 0086): the instruction says that a moved object or receptacle jumps alone: what rests in or on it stays where it was (LIBERO-MAX moves one entity; board 0086 saw a basket jump through a can already in it). Text only: the instance, the event, the reference solution, LIBERO's success check and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/libero-max-object-chocolate-pudding-in-basket-sensor-noise-onset-i00-standard/instruction.md b/tasks/libero-max-object-chocolate-pudding-in-basket-sensor-noise-onset-i00-standard/instruction.md
index b5aa2f6466..222a159338 100644
--- a/tasks/libero-max-object-chocolate-pudding-in-basket-sensor-noise-onset-i00-standard/instruction.md
+++ b/tasks/libero-max-object-chocolate-pudding-in-basket-sensor-noise-onset-i00-standard/instruction.md
@@ -6,7 +6,8 @@ simulator runs in a service you cannot open, and you reach it only through a cli
The scene can change while you work: once during the episode, when your gripper first gets close to one of the task's
objects, the scene changes (as in LIBERO-MAX): an object or a receptacle may be moved, an obstacle or a few other
objects may appear, a camera may shift, or the lighting, the colours or the image noise may change. Nothing tells you
-when; the task and its success check stay the same, so keep checking what the cameras show.
+when; the task and its success check stay the same, so keep checking what the cameras show. A moved object or
+receptacle jumps there alone: what already rests in or on it stays where it was.
## Rules (real-robot rules)
- **One episode, no reset.** The episode starts when the service comes up. It ends when the task succeeds, after
diff --git a/tasks/libero-max-object-chocolate-pudding-in-basket-sensor-noise-onset-i00-standard/task.toml b/tasks/libero-max-object-chocolate-pudding-in-basket-sensor-noise-onset-i00-standard/task.toml
index f8d1005687..e8d8bbc07c 100644
--- a/tasks/libero-max-object-chocolate-pudding-in-basket-sensor-noise-onset-i00-standard/task.toml
+++ b/tasks/libero-max-object-chocolate-pudding-in-basket-sensor-noise-onset-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "robosuite"
robot = "franka-panda"
family = "libero-max/libero_object/sensor_noise_onset"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 195
diff --git a/tasks/libero-max-object-cream-cheese-in-basket-visual-theme-switch-i00-privileged/task.toml b/tasks/libero-max-object-cream-cheese-in-basket-visual-theme-switch-i00-privileged/task.toml
index 13f56b5291..61b81a622d 100644
--- a/tasks/libero-max-object-cream-cheese-in-basket-visual-theme-switch-i00-privileged/task.toml
+++ b/tasks/libero-max-object-cream-cheese-in-basket-visual-theme-switch-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "robosuite"
robot = "franka-panda"
family = "libero-max/libero_object/visual_theme_switch"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 195 # LIBERO-MAX's policy seed for the case (the scene is the frozen LIBERO-PRO initial state)
expert_time_estimate_hours = 1.0
diff --git a/tasks/libero-max-object-cream-cheese-in-basket-visual-theme-switch-i00-standard/README.md b/tasks/libero-max-object-cream-cheese-in-basket-visual-theme-switch-i00-standard/README.md
index 455703059c..aefcb706e5 100644
--- a/tasks/libero-max-object-cream-cheese-in-basket-visual-theme-switch-i00-standard/README.md
+++ b/tasks/libero-max-object-cream-cheese-in-basket-visual-theme-switch-i00-standard/README.md
@@ -86,3 +86,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
/reset` ends the episode as `error` (`sim_error = 1`), not `reset` (reachable only when a run grants resets; the
default is none). The scene, instance, pass rule and every verdict reached before are unchanged, so the task keeps
its name.
+- 2026-10-14 (board 0113, from board 0086): the instruction says that a moved object or receptacle jumps alone: what rests in or on it stays where it was (LIBERO-MAX moves one entity; board 0086 saw a basket jump through a can already in it). Text only: the instance, the event, the reference solution, LIBERO's success check and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/libero-max-object-cream-cheese-in-basket-visual-theme-switch-i00-standard/instruction.md b/tasks/libero-max-object-cream-cheese-in-basket-visual-theme-switch-i00-standard/instruction.md
index 9b4587a384..9151c2ec42 100644
--- a/tasks/libero-max-object-cream-cheese-in-basket-visual-theme-switch-i00-standard/instruction.md
+++ b/tasks/libero-max-object-cream-cheese-in-basket-visual-theme-switch-i00-standard/instruction.md
@@ -6,7 +6,8 @@ simulator runs in a service you cannot open, and you reach it only through a cli
The scene can change while you work: once during the episode, when your gripper first gets close to one of the task's
objects, the scene changes (as in LIBERO-MAX): an object or a receptacle may be moved, an obstacle or a few other
objects may appear, a camera may shift, or the lighting, the colours or the image noise may change. Nothing tells you
-when; the task and its success check stay the same, so keep checking what the cameras show.
+when; the task and its success check stay the same, so keep checking what the cameras show. A moved object or
+receptacle jumps there alone: what already rests in or on it stays where it was.
## Rules (real-robot rules)
- **One episode, no reset.** The episode starts when the service comes up. It ends when the task succeeds, after
diff --git a/tasks/libero-max-object-cream-cheese-in-basket-visual-theme-switch-i00-standard/task.toml b/tasks/libero-max-object-cream-cheese-in-basket-visual-theme-switch-i00-standard/task.toml
index a1b30ae10a..a1e2a1cbb5 100644
--- a/tasks/libero-max-object-cream-cheese-in-basket-visual-theme-switch-i00-standard/task.toml
+++ b/tasks/libero-max-object-cream-cheese-in-basket-visual-theme-switch-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "robosuite"
robot = "franka-panda"
family = "libero-max/libero_object/visual_theme_switch"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 195
diff --git a/tasks/libero-max-object-green-bbq-sauce-in-red-basket-distractor-burst-i00-privileged/task.toml b/tasks/libero-max-object-green-bbq-sauce-in-red-basket-distractor-burst-i00-privileged/task.toml
index 26b436ee7b..d650ee437f 100644
--- a/tasks/libero-max-object-green-bbq-sauce-in-red-basket-distractor-burst-i00-privileged/task.toml
+++ b/tasks/libero-max-object-green-bbq-sauce-in-red-basket-distractor-burst-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "robosuite"
robot = "franka-panda"
family = "libero-max/libero_object/distractor_burst"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 195 # LIBERO-MAX's policy seed for the case (the scene is the frozen LIBERO-PRO initial state)
expert_time_estimate_hours = 1.0
diff --git a/tasks/libero-max-object-green-bbq-sauce-in-red-basket-distractor-burst-i00-standard/README.md b/tasks/libero-max-object-green-bbq-sauce-in-red-basket-distractor-burst-i00-standard/README.md
index 78305b65f1..4b5e8f2df5 100644
--- a/tasks/libero-max-object-green-bbq-sauce-in-red-basket-distractor-burst-i00-standard/README.md
+++ b/tasks/libero-max-object-green-bbq-sauce-in-red-basket-distractor-burst-i00-standard/README.md
@@ -86,3 +86,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
/reset` ends the episode as `error` (`sim_error = 1`), not `reset` (reachable only when a run grants resets; the
default is none). The scene, instance, pass rule and every verdict reached before are unchanged, so the task keeps
its name.
+- 2026-10-14 (board 0113, from board 0086): the instruction says that a moved object or receptacle jumps alone: what rests in or on it stays where it was (LIBERO-MAX moves one entity; board 0086 saw a basket jump through a can already in it). Text only: the instance, the event, the reference solution, LIBERO's success check and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/libero-max-object-green-bbq-sauce-in-red-basket-distractor-burst-i00-standard/instruction.md b/tasks/libero-max-object-green-bbq-sauce-in-red-basket-distractor-burst-i00-standard/instruction.md
index 9163de031b..2f581ad2fd 100644
--- a/tasks/libero-max-object-green-bbq-sauce-in-red-basket-distractor-burst-i00-standard/instruction.md
+++ b/tasks/libero-max-object-green-bbq-sauce-in-red-basket-distractor-burst-i00-standard/instruction.md
@@ -6,7 +6,8 @@ simulator runs in a service you cannot open, and you reach it only through a cli
The scene can change while you work: once during the episode, when your gripper first gets close to one of the task's
objects, the scene changes (as in LIBERO-MAX): an object or a receptacle may be moved, an obstacle or a few other
objects may appear, a camera may shift, or the lighting, the colours or the image noise may change. Nothing tells you
-when; the task and its success check stay the same, so keep checking what the cameras show.
+when; the task and its success check stay the same, so keep checking what the cameras show. A moved object or
+receptacle jumps there alone: what already rests in or on it stays where it was.
## Rules (real-robot rules)
- **One episode, no reset.** The episode starts when the service comes up. It ends when the task succeeds, after
diff --git a/tasks/libero-max-object-green-bbq-sauce-in-red-basket-distractor-burst-i00-standard/task.toml b/tasks/libero-max-object-green-bbq-sauce-in-red-basket-distractor-burst-i00-standard/task.toml
index 99bd495360..c1d6d3aba1 100644
--- a/tasks/libero-max-object-green-bbq-sauce-in-red-basket-distractor-burst-i00-standard/task.toml
+++ b/tasks/libero-max-object-green-bbq-sauce-in-red-basket-distractor-burst-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "robosuite"
robot = "franka-panda"
family = "libero-max/libero_object/distractor_burst"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 195
diff --git a/tasks/libero-max-object-green-chocolate-pudding-in-red-basket-camera-shift-i00-privileged/task.toml b/tasks/libero-max-object-green-chocolate-pudding-in-red-basket-camera-shift-i00-privileged/task.toml
index 88b9a11d10..ba43b28074 100644
--- a/tasks/libero-max-object-green-chocolate-pudding-in-red-basket-camera-shift-i00-privileged/task.toml
+++ b/tasks/libero-max-object-green-chocolate-pudding-in-red-basket-camera-shift-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "robosuite"
robot = "franka-panda"
family = "libero-max/libero_object/camera_shift"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 195 # LIBERO-MAX's policy seed for the case (the scene is the frozen LIBERO-PRO initial state)
expert_time_estimate_hours = 1.0
diff --git a/tasks/libero-max-object-green-chocolate-pudding-in-red-basket-camera-shift-i00-standard/README.md b/tasks/libero-max-object-green-chocolate-pudding-in-red-basket-camera-shift-i00-standard/README.md
index 7600e64437..9c4821ccaf 100644
--- a/tasks/libero-max-object-green-chocolate-pudding-in-red-basket-camera-shift-i00-standard/README.md
+++ b/tasks/libero-max-object-green-chocolate-pudding-in-red-basket-camera-shift-i00-standard/README.md
@@ -86,3 +86,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
/reset` ends the episode as `error` (`sim_error = 1`), not `reset` (reachable only when a run grants resets; the
default is none). The scene, instance, pass rule and every verdict reached before are unchanged, so the task keeps
its name.
+- 2026-10-14 (board 0113, from board 0086): the instruction says that a moved object or receptacle jumps alone: what rests in or on it stays where it was (LIBERO-MAX moves one entity; board 0086 saw a basket jump through a can already in it). Text only: the instance, the event, the reference solution, LIBERO's success check and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/libero-max-object-green-chocolate-pudding-in-red-basket-camera-shift-i00-standard/instruction.md b/tasks/libero-max-object-green-chocolate-pudding-in-red-basket-camera-shift-i00-standard/instruction.md
index c01c00389c..044b83046b 100644
--- a/tasks/libero-max-object-green-chocolate-pudding-in-red-basket-camera-shift-i00-standard/instruction.md
+++ b/tasks/libero-max-object-green-chocolate-pudding-in-red-basket-camera-shift-i00-standard/instruction.md
@@ -6,7 +6,8 @@ simulator runs in a service you cannot open, and you reach it only through a cli
The scene can change while you work: once during the episode, when your gripper first gets close to one of the task's
objects, the scene changes (as in LIBERO-MAX): an object or a receptacle may be moved, an obstacle or a few other
objects may appear, a camera may shift, or the lighting, the colours or the image noise may change. Nothing tells you
-when; the task and its success check stay the same, so keep checking what the cameras show.
+when; the task and its success check stay the same, so keep checking what the cameras show. A moved object or
+receptacle jumps there alone: what already rests in or on it stays where it was.
## Rules (real-robot rules)
- **One episode, no reset.** The episode starts when the service comes up. It ends when the task succeeds, after
diff --git a/tasks/libero-max-object-green-chocolate-pudding-in-red-basket-camera-shift-i00-standard/task.toml b/tasks/libero-max-object-green-chocolate-pudding-in-red-basket-camera-shift-i00-standard/task.toml
index 8dc4bb4191..021e0f5f0d 100644
--- a/tasks/libero-max-object-green-chocolate-pudding-in-red-basket-camera-shift-i00-standard/task.toml
+++ b/tasks/libero-max-object-green-chocolate-pudding-in-red-basket-camera-shift-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "robosuite"
robot = "franka-panda"
family = "libero-max/libero_object/camera_shift"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 195
diff --git a/tasks/libero-max-object-ketchup-in-basket-distractor-burst-i00-privileged/task.toml b/tasks/libero-max-object-ketchup-in-basket-distractor-burst-i00-privileged/task.toml
index 874442887e..896aed4925 100644
--- a/tasks/libero-max-object-ketchup-in-basket-distractor-burst-i00-privileged/task.toml
+++ b/tasks/libero-max-object-ketchup-in-basket-distractor-burst-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "robosuite"
robot = "franka-panda"
family = "libero-max/libero_object/distractor_burst"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 195 # LIBERO-MAX's policy seed for the case (the scene is the frozen LIBERO-PRO initial state)
expert_time_estimate_hours = 1.0
diff --git a/tasks/libero-max-object-ketchup-in-basket-distractor-burst-i00-standard/README.md b/tasks/libero-max-object-ketchup-in-basket-distractor-burst-i00-standard/README.md
index a3553d3230..47d19c94fc 100644
--- a/tasks/libero-max-object-ketchup-in-basket-distractor-burst-i00-standard/README.md
+++ b/tasks/libero-max-object-ketchup-in-basket-distractor-burst-i00-standard/README.md
@@ -86,3 +86,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
/reset` ends the episode as `error` (`sim_error = 1`), not `reset` (reachable only when a run grants resets; the
default is none). The scene, instance, pass rule and every verdict reached before are unchanged, so the task keeps
its name.
+- 2026-10-14 (board 0113, from board 0086): the instruction says that a moved object or receptacle jumps alone: what rests in or on it stays where it was (LIBERO-MAX moves one entity; board 0086 saw a basket jump through a can already in it). Text only: the instance, the event, the reference solution, LIBERO's success check and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/libero-max-object-ketchup-in-basket-distractor-burst-i00-standard/instruction.md b/tasks/libero-max-object-ketchup-in-basket-distractor-burst-i00-standard/instruction.md
index 61e913a88f..14088ed63b 100644
--- a/tasks/libero-max-object-ketchup-in-basket-distractor-burst-i00-standard/instruction.md
+++ b/tasks/libero-max-object-ketchup-in-basket-distractor-burst-i00-standard/instruction.md
@@ -6,7 +6,8 @@ simulator runs in a service you cannot open, and you reach it only through a cli
The scene can change while you work: once during the episode, when your gripper first gets close to one of the task's
objects, the scene changes (as in LIBERO-MAX): an object or a receptacle may be moved, an obstacle or a few other
objects may appear, a camera may shift, or the lighting, the colours or the image noise may change. Nothing tells you
-when; the task and its success check stay the same, so keep checking what the cameras show.
+when; the task and its success check stay the same, so keep checking what the cameras show. A moved object or
+receptacle jumps there alone: what already rests in or on it stays where it was.
## Rules (real-robot rules)
- **One episode, no reset.** The episode starts when the service comes up. It ends when the task succeeds, after
diff --git a/tasks/libero-max-object-ketchup-in-basket-distractor-burst-i00-standard/task.toml b/tasks/libero-max-object-ketchup-in-basket-distractor-burst-i00-standard/task.toml
index e69d019145..ee47d650ee 100644
--- a/tasks/libero-max-object-ketchup-in-basket-distractor-burst-i00-standard/task.toml
+++ b/tasks/libero-max-object-ketchup-in-basket-distractor-burst-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "robosuite"
robot = "franka-panda"
family = "libero-max/libero_object/distractor_burst"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 195
diff --git a/tasks/libero-max-object-milk-in-basket-camera-shift-i00-privileged/task.toml b/tasks/libero-max-object-milk-in-basket-camera-shift-i00-privileged/task.toml
index 2d1b4f9ff8..ae10e9a0f1 100644
--- a/tasks/libero-max-object-milk-in-basket-camera-shift-i00-privileged/task.toml
+++ b/tasks/libero-max-object-milk-in-basket-camera-shift-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "robosuite"
robot = "franka-panda"
family = "libero-max/libero_object/camera_shift"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 195 # LIBERO-MAX's policy seed for the case (the scene is the frozen LIBERO-PRO initial state)
expert_time_estimate_hours = 1.0
diff --git a/tasks/libero-max-object-milk-in-basket-camera-shift-i00-standard/README.md b/tasks/libero-max-object-milk-in-basket-camera-shift-i00-standard/README.md
index 5bfdf80910..161e8e34a1 100644
--- a/tasks/libero-max-object-milk-in-basket-camera-shift-i00-standard/README.md
+++ b/tasks/libero-max-object-milk-in-basket-camera-shift-i00-standard/README.md
@@ -86,3 +86,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
/reset` ends the episode as `error` (`sim_error = 1`), not `reset` (reachable only when a run grants resets; the
default is none). The scene, instance, pass rule and every verdict reached before are unchanged, so the task keeps
its name.
+- 2026-10-14 (board 0113, from board 0086): the instruction says that a moved object or receptacle jumps alone: what rests in or on it stays where it was (LIBERO-MAX moves one entity; board 0086 saw a basket jump through a can already in it). Text only: the instance, the event, the reference solution, LIBERO's success check and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/libero-max-object-milk-in-basket-camera-shift-i00-standard/instruction.md b/tasks/libero-max-object-milk-in-basket-camera-shift-i00-standard/instruction.md
index 3d87fc78f1..55b008b2a5 100644
--- a/tasks/libero-max-object-milk-in-basket-camera-shift-i00-standard/instruction.md
+++ b/tasks/libero-max-object-milk-in-basket-camera-shift-i00-standard/instruction.md
@@ -6,7 +6,8 @@ simulator runs in a service you cannot open, and you reach it only through a cli
The scene can change while you work: once during the episode, when your gripper first gets close to one of the task's
objects, the scene changes (as in LIBERO-MAX): an object or a receptacle may be moved, an obstacle or a few other
objects may appear, a camera may shift, or the lighting, the colours or the image noise may change. Nothing tells you
-when; the task and its success check stay the same, so keep checking what the cameras show.
+when; the task and its success check stay the same, so keep checking what the cameras show. A moved object or
+receptacle jumps there alone: what already rests in or on it stays where it was.
## Rules (real-robot rules)
- **One episode, no reset.** The episode starts when the service comes up. It ends when the task succeeds, after
diff --git a/tasks/libero-max-object-milk-in-basket-camera-shift-i00-standard/task.toml b/tasks/libero-max-object-milk-in-basket-camera-shift-i00-standard/task.toml
index 100baa6c97..b4e8a29c35 100644
--- a/tasks/libero-max-object-milk-in-basket-camera-shift-i00-standard/task.toml
+++ b/tasks/libero-max-object-milk-in-basket-camera-shift-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "robosuite"
robot = "franka-panda"
family = "libero-max/libero_object/camera_shift"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 195
diff --git a/tasks/libero-max-object-milk-in-basket-distractor-burst-i00-privileged/task.toml b/tasks/libero-max-object-milk-in-basket-distractor-burst-i00-privileged/task.toml
index c2527b6100..a72b39b144 100644
--- a/tasks/libero-max-object-milk-in-basket-distractor-burst-i00-privileged/task.toml
+++ b/tasks/libero-max-object-milk-in-basket-distractor-burst-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "robosuite"
robot = "franka-panda"
family = "libero-max/libero_object/distractor_burst"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 195 # LIBERO-MAX's policy seed for the case (the scene is the frozen LIBERO-PRO initial state)
expert_time_estimate_hours = 1.0
diff --git a/tasks/libero-max-object-milk-in-basket-distractor-burst-i00-standard/README.md b/tasks/libero-max-object-milk-in-basket-distractor-burst-i00-standard/README.md
index f6f84fae48..5788342f62 100644
--- a/tasks/libero-max-object-milk-in-basket-distractor-burst-i00-standard/README.md
+++ b/tasks/libero-max-object-milk-in-basket-distractor-burst-i00-standard/README.md
@@ -86,3 +86,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
/reset` ends the episode as `error` (`sim_error = 1`), not `reset` (reachable only when a run grants resets; the
default is none). The scene, instance, pass rule and every verdict reached before are unchanged, so the task keeps
its name.
+- 2026-10-14 (board 0113, from board 0086): the instruction says that a moved object or receptacle jumps alone: what rests in or on it stays where it was (LIBERO-MAX moves one entity; board 0086 saw a basket jump through a can already in it). Text only: the instance, the event, the reference solution, LIBERO's success check and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/libero-max-object-milk-in-basket-distractor-burst-i00-standard/instruction.md b/tasks/libero-max-object-milk-in-basket-distractor-burst-i00-standard/instruction.md
index 263acb5685..3edba05fb2 100644
--- a/tasks/libero-max-object-milk-in-basket-distractor-burst-i00-standard/instruction.md
+++ b/tasks/libero-max-object-milk-in-basket-distractor-burst-i00-standard/instruction.md
@@ -6,7 +6,8 @@ simulator runs in a service you cannot open, and you reach it only through a cli
The scene can change while you work: once during the episode, when your gripper first gets close to one of the task's
objects, the scene changes (as in LIBERO-MAX): an object or a receptacle may be moved, an obstacle or a few other
objects may appear, a camera may shift, or the lighting, the colours or the image noise may change. Nothing tells you
-when; the task and its success check stay the same, so keep checking what the cameras show.
+when; the task and its success check stay the same, so keep checking what the cameras show. A moved object or
+receptacle jumps there alone: what already rests in or on it stays where it was.
## Rules (real-robot rules)
- **One episode, no reset.** The episode starts when the service comes up. It ends when the task succeeds, after
diff --git a/tasks/libero-max-object-milk-in-basket-distractor-burst-i00-standard/task.toml b/tasks/libero-max-object-milk-in-basket-distractor-burst-i00-standard/task.toml
index e84ca9e030..dce503445e 100644
--- a/tasks/libero-max-object-milk-in-basket-distractor-burst-i00-standard/task.toml
+++ b/tasks/libero-max-object-milk-in-basket-distractor-burst-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "robosuite"
robot = "franka-panda"
family = "libero-max/libero_object/distractor_burst"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 195
diff --git a/tasks/libero-max-object-milk-in-basket-illumination-switch-i00-privileged/task.toml b/tasks/libero-max-object-milk-in-basket-illumination-switch-i00-privileged/task.toml
index 9729d05729..373755794a 100644
--- a/tasks/libero-max-object-milk-in-basket-illumination-switch-i00-privileged/task.toml
+++ b/tasks/libero-max-object-milk-in-basket-illumination-switch-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "robosuite"
robot = "franka-panda"
family = "libero-max/libero_object/illumination_switch"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 195 # LIBERO-MAX's policy seed for the case (the scene is the frozen LIBERO-PRO initial state)
expert_time_estimate_hours = 1.0
diff --git a/tasks/libero-max-object-milk-in-basket-illumination-switch-i00-standard/README.md b/tasks/libero-max-object-milk-in-basket-illumination-switch-i00-standard/README.md
index 4bbc9e4199..f57d14e66a 100644
--- a/tasks/libero-max-object-milk-in-basket-illumination-switch-i00-standard/README.md
+++ b/tasks/libero-max-object-milk-in-basket-illumination-switch-i00-standard/README.md
@@ -86,3 +86,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
/reset` ends the episode as `error` (`sim_error = 1`), not `reset` (reachable only when a run grants resets; the
default is none). The scene, instance, pass rule and every verdict reached before are unchanged, so the task keeps
its name.
+- 2026-10-14 (board 0113, from board 0086): the instruction says that a moved object or receptacle jumps alone: what rests in or on it stays where it was (LIBERO-MAX moves one entity; board 0086 saw a basket jump through a can already in it). Text only: the instance, the event, the reference solution, LIBERO's success check and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/libero-max-object-milk-in-basket-illumination-switch-i00-standard/instruction.md b/tasks/libero-max-object-milk-in-basket-illumination-switch-i00-standard/instruction.md
index 3be8198fdf..931cdbc0e1 100644
--- a/tasks/libero-max-object-milk-in-basket-illumination-switch-i00-standard/instruction.md
+++ b/tasks/libero-max-object-milk-in-basket-illumination-switch-i00-standard/instruction.md
@@ -6,7 +6,8 @@ simulator runs in a service you cannot open, and you reach it only through a cli
The scene can change while you work: once during the episode, when your gripper first gets close to one of the task's
objects, the scene changes (as in LIBERO-MAX): an object or a receptacle may be moved, an obstacle or a few other
objects may appear, a camera may shift, or the lighting, the colours or the image noise may change. Nothing tells you
-when; the task and its success check stay the same, so keep checking what the cameras show.
+when; the task and its success check stay the same, so keep checking what the cameras show. A moved object or
+receptacle jumps there alone: what already rests in or on it stays where it was.
## Rules (real-robot rules)
- **One episode, no reset.** The episode starts when the service comes up. It ends when the task succeeds, after
diff --git a/tasks/libero-max-object-milk-in-basket-illumination-switch-i00-standard/task.toml b/tasks/libero-max-object-milk-in-basket-illumination-switch-i00-standard/task.toml
index 7f7de64ea6..26669fef45 100644
--- a/tasks/libero-max-object-milk-in-basket-illumination-switch-i00-standard/task.toml
+++ b/tasks/libero-max-object-milk-in-basket-illumination-switch-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "robosuite"
robot = "franka-panda"
family = "libero-max/libero_object/illumination_switch"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 195
diff --git a/tasks/libero-max-object-milk-in-basket-target-relocation-i00-privileged/task.toml b/tasks/libero-max-object-milk-in-basket-target-relocation-i00-privileged/task.toml
index c11eef1f3a..aa3b217c61 100644
--- a/tasks/libero-max-object-milk-in-basket-target-relocation-i00-privileged/task.toml
+++ b/tasks/libero-max-object-milk-in-basket-target-relocation-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "robosuite"
robot = "franka-panda"
family = "libero-max/libero_object/target_relocation"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 195 # LIBERO-MAX's policy seed for the case (the scene is the frozen LIBERO-PRO initial state)
expert_time_estimate_hours = 1.0
diff --git a/tasks/libero-max-object-milk-in-basket-target-relocation-i00-standard/README.md b/tasks/libero-max-object-milk-in-basket-target-relocation-i00-standard/README.md
index 626628634e..20882f4389 100644
--- a/tasks/libero-max-object-milk-in-basket-target-relocation-i00-standard/README.md
+++ b/tasks/libero-max-object-milk-in-basket-target-relocation-i00-standard/README.md
@@ -86,3 +86,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
/reset` ends the episode as `error` (`sim_error = 1`), not `reset` (reachable only when a run grants resets; the
default is none). The scene, instance, pass rule and every verdict reached before are unchanged, so the task keeps
its name.
+- 2026-10-14 (board 0113, from board 0086): the instruction says that a moved object or receptacle jumps alone: what rests in or on it stays where it was (LIBERO-MAX moves one entity; board 0086 saw a basket jump through a can already in it). Text only: the instance, the event, the reference solution, LIBERO's success check and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/libero-max-object-milk-in-basket-target-relocation-i00-standard/instruction.md b/tasks/libero-max-object-milk-in-basket-target-relocation-i00-standard/instruction.md
index e8f490970d..d4211dd11e 100644
--- a/tasks/libero-max-object-milk-in-basket-target-relocation-i00-standard/instruction.md
+++ b/tasks/libero-max-object-milk-in-basket-target-relocation-i00-standard/instruction.md
@@ -6,7 +6,8 @@ simulator runs in a service you cannot open, and you reach it only through a cli
The scene can change while you work: once during the episode, when your gripper first gets close to one of the task's
objects, the scene changes (as in LIBERO-MAX): an object or a receptacle may be moved, an obstacle or a few other
objects may appear, a camera may shift, or the lighting, the colours or the image noise may change. Nothing tells you
-when; the task and its success check stay the same, so keep checking what the cameras show.
+when; the task and its success check stay the same, so keep checking what the cameras show. A moved object or
+receptacle jumps there alone: what already rests in or on it stays where it was.
## Rules (real-robot rules)
- **One episode, no reset.** The episode starts when the service comes up. It ends when the task succeeds, after
diff --git a/tasks/libero-max-object-milk-in-basket-target-relocation-i00-standard/task.toml b/tasks/libero-max-object-milk-in-basket-target-relocation-i00-standard/task.toml
index 47c51c6fcc..1c01ed14f1 100644
--- a/tasks/libero-max-object-milk-in-basket-target-relocation-i00-standard/task.toml
+++ b/tasks/libero-max-object-milk-in-basket-target-relocation-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "robosuite"
robot = "franka-panda"
family = "libero-max/libero_object/target_relocation"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 195
diff --git a/tasks/libero-max-object-orange-juice-in-basket-distractor-burst-i00-privileged/task.toml b/tasks/libero-max-object-orange-juice-in-basket-distractor-burst-i00-privileged/task.toml
index 2422e414b7..603bf94985 100644
--- a/tasks/libero-max-object-orange-juice-in-basket-distractor-burst-i00-privileged/task.toml
+++ b/tasks/libero-max-object-orange-juice-in-basket-distractor-burst-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "robosuite"
robot = "franka-panda"
family = "libero-max/libero_object/distractor_burst"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 195 # LIBERO-MAX's policy seed for the case (the scene is the frozen LIBERO-PRO initial state)
expert_time_estimate_hours = 1.0
diff --git a/tasks/libero-max-object-orange-juice-in-basket-distractor-burst-i00-standard/README.md b/tasks/libero-max-object-orange-juice-in-basket-distractor-burst-i00-standard/README.md
index 7c1a843b9d..208da689f2 100644
--- a/tasks/libero-max-object-orange-juice-in-basket-distractor-burst-i00-standard/README.md
+++ b/tasks/libero-max-object-orange-juice-in-basket-distractor-burst-i00-standard/README.md
@@ -86,3 +86,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
/reset` ends the episode as `error` (`sim_error = 1`), not `reset` (reachable only when a run grants resets; the
default is none). The scene, instance, pass rule and every verdict reached before are unchanged, so the task keeps
its name.
+- 2026-10-14 (board 0113, from board 0086): the instruction says that a moved object or receptacle jumps alone: what rests in or on it stays where it was (LIBERO-MAX moves one entity; board 0086 saw a basket jump through a can already in it). Text only: the instance, the event, the reference solution, LIBERO's success check and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/libero-max-object-orange-juice-in-basket-distractor-burst-i00-standard/instruction.md b/tasks/libero-max-object-orange-juice-in-basket-distractor-burst-i00-standard/instruction.md
index 20695b2e8f..e34183cdfb 100644
--- a/tasks/libero-max-object-orange-juice-in-basket-distractor-burst-i00-standard/instruction.md
+++ b/tasks/libero-max-object-orange-juice-in-basket-distractor-burst-i00-standard/instruction.md
@@ -6,7 +6,8 @@ simulator runs in a service you cannot open, and you reach it only through a cli
The scene can change while you work: once during the episode, when your gripper first gets close to one of the task's
objects, the scene changes (as in LIBERO-MAX): an object or a receptacle may be moved, an obstacle or a few other
objects may appear, a camera may shift, or the lighting, the colours or the image noise may change. Nothing tells you
-when; the task and its success check stay the same, so keep checking what the cameras show.
+when; the task and its success check stay the same, so keep checking what the cameras show. A moved object or
+receptacle jumps there alone: what already rests in or on it stays where it was.
## Rules (real-robot rules)
- **One episode, no reset.** The episode starts when the service comes up. It ends when the task succeeds, after
diff --git a/tasks/libero-max-object-orange-juice-in-basket-distractor-burst-i00-standard/task.toml b/tasks/libero-max-object-orange-juice-in-basket-distractor-burst-i00-standard/task.toml
index 11c7a3f59b..9b4bb1af91 100644
--- a/tasks/libero-max-object-orange-juice-in-basket-distractor-burst-i00-standard/task.toml
+++ b/tasks/libero-max-object-orange-juice-in-basket-distractor-burst-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "robosuite"
robot = "franka-panda"
family = "libero-max/libero_object/distractor_burst"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 195
diff --git a/tasks/libero-max-object-orange-juice-in-basket-illumination-switch-i00-privileged/task.toml b/tasks/libero-max-object-orange-juice-in-basket-illumination-switch-i00-privileged/task.toml
index de8909d452..a2e2a0d2e0 100644
--- a/tasks/libero-max-object-orange-juice-in-basket-illumination-switch-i00-privileged/task.toml
+++ b/tasks/libero-max-object-orange-juice-in-basket-illumination-switch-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "robosuite"
robot = "franka-panda"
family = "libero-max/libero_object/illumination_switch"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 195 # LIBERO-MAX's policy seed for the case (the scene is the frozen LIBERO-PRO initial state)
expert_time_estimate_hours = 1.0
diff --git a/tasks/libero-max-object-orange-juice-in-basket-illumination-switch-i00-standard/README.md b/tasks/libero-max-object-orange-juice-in-basket-illumination-switch-i00-standard/README.md
index ac2670c6d1..e2d21fd72a 100644
--- a/tasks/libero-max-object-orange-juice-in-basket-illumination-switch-i00-standard/README.md
+++ b/tasks/libero-max-object-orange-juice-in-basket-illumination-switch-i00-standard/README.md
@@ -86,3 +86,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
/reset` ends the episode as `error` (`sim_error = 1`), not `reset` (reachable only when a run grants resets; the
default is none). The scene, instance, pass rule and every verdict reached before are unchanged, so the task keeps
its name.
+- 2026-10-14 (board 0113, from board 0086): the instruction says that a moved object or receptacle jumps alone: what rests in or on it stays where it was (LIBERO-MAX moves one entity; board 0086 saw a basket jump through a can already in it). Text only: the instance, the event, the reference solution, LIBERO's success check and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/libero-max-object-orange-juice-in-basket-illumination-switch-i00-standard/instruction.md b/tasks/libero-max-object-orange-juice-in-basket-illumination-switch-i00-standard/instruction.md
index 9275cc2c40..a5cf5f3d50 100644
--- a/tasks/libero-max-object-orange-juice-in-basket-illumination-switch-i00-standard/instruction.md
+++ b/tasks/libero-max-object-orange-juice-in-basket-illumination-switch-i00-standard/instruction.md
@@ -6,7 +6,8 @@ simulator runs in a service you cannot open, and you reach it only through a cli
The scene can change while you work: once during the episode, when your gripper first gets close to one of the task's
objects, the scene changes (as in LIBERO-MAX): an object or a receptacle may be moved, an obstacle or a few other
objects may appear, a camera may shift, or the lighting, the colours or the image noise may change. Nothing tells you
-when; the task and its success check stay the same, so keep checking what the cameras show.
+when; the task and its success check stay the same, so keep checking what the cameras show. A moved object or
+receptacle jumps there alone: what already rests in or on it stays where it was.
## Rules (real-robot rules)
- **One episode, no reset.** The episode starts when the service comes up. It ends when the task succeeds, after
diff --git a/tasks/libero-max-object-orange-juice-in-basket-illumination-switch-i00-standard/task.toml b/tasks/libero-max-object-orange-juice-in-basket-illumination-switch-i00-standard/task.toml
index eff0407526..5d8fd9666f 100644
--- a/tasks/libero-max-object-orange-juice-in-basket-illumination-switch-i00-standard/task.toml
+++ b/tasks/libero-max-object-orange-juice-in-basket-illumination-switch-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "robosuite"
robot = "franka-panda"
family = "libero-max/libero_object/illumination_switch"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 195
diff --git a/tasks/libero-max-object-salad-dressing-in-basket-sensor-noise-onset-i00-privileged/task.toml b/tasks/libero-max-object-salad-dressing-in-basket-sensor-noise-onset-i00-privileged/task.toml
index 1d814bf252..80b759a556 100644
--- a/tasks/libero-max-object-salad-dressing-in-basket-sensor-noise-onset-i00-privileged/task.toml
+++ b/tasks/libero-max-object-salad-dressing-in-basket-sensor-noise-onset-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "robosuite"
robot = "franka-panda"
family = "libero-max/libero_object/sensor_noise_onset"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 195 # LIBERO-MAX's policy seed for the case (the scene is the frozen LIBERO-PRO initial state)
expert_time_estimate_hours = 1.0
diff --git a/tasks/libero-max-object-salad-dressing-in-basket-sensor-noise-onset-i00-standard/README.md b/tasks/libero-max-object-salad-dressing-in-basket-sensor-noise-onset-i00-standard/README.md
index f8aca233db..2782c816c1 100644
--- a/tasks/libero-max-object-salad-dressing-in-basket-sensor-noise-onset-i00-standard/README.md
+++ b/tasks/libero-max-object-salad-dressing-in-basket-sensor-noise-onset-i00-standard/README.md
@@ -86,3 +86,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
/reset` ends the episode as `error` (`sim_error = 1`), not `reset` (reachable only when a run grants resets; the
default is none). The scene, instance, pass rule and every verdict reached before are unchanged, so the task keeps
its name.
+- 2026-10-14 (board 0113, from board 0086): the instruction says that a moved object or receptacle jumps alone: what rests in or on it stays where it was (LIBERO-MAX moves one entity; board 0086 saw a basket jump through a can already in it). Text only: the instance, the event, the reference solution, LIBERO's success check and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/libero-max-object-salad-dressing-in-basket-sensor-noise-onset-i00-standard/instruction.md b/tasks/libero-max-object-salad-dressing-in-basket-sensor-noise-onset-i00-standard/instruction.md
index c08e926645..714d04321e 100644
--- a/tasks/libero-max-object-salad-dressing-in-basket-sensor-noise-onset-i00-standard/instruction.md
+++ b/tasks/libero-max-object-salad-dressing-in-basket-sensor-noise-onset-i00-standard/instruction.md
@@ -6,7 +6,8 @@ simulator runs in a service you cannot open, and you reach it only through a cli
The scene can change while you work: once during the episode, when your gripper first gets close to one of the task's
objects, the scene changes (as in LIBERO-MAX): an object or a receptacle may be moved, an obstacle or a few other
objects may appear, a camera may shift, or the lighting, the colours or the image noise may change. Nothing tells you
-when; the task and its success check stay the same, so keep checking what the cameras show.
+when; the task and its success check stay the same, so keep checking what the cameras show. A moved object or
+receptacle jumps there alone: what already rests in or on it stays where it was.
## Rules (real-robot rules)
- **One episode, no reset.** The episode starts when the service comes up. It ends when the task succeeds, after
diff --git a/tasks/libero-max-object-salad-dressing-in-basket-sensor-noise-onset-i00-standard/task.toml b/tasks/libero-max-object-salad-dressing-in-basket-sensor-noise-onset-i00-standard/task.toml
index dcd43bb77e..58a4becf19 100644
--- a/tasks/libero-max-object-salad-dressing-in-basket-sensor-noise-onset-i00-standard/task.toml
+++ b/tasks/libero-max-object-salad-dressing-in-basket-sensor-noise-onset-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "robosuite"
robot = "franka-panda"
family = "libero-max/libero_object/sensor_noise_onset"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 195
diff --git a/tasks/libero-max-spatial-akita-black-bowl-on-plate-camera-shift-i00-privileged/task.toml b/tasks/libero-max-spatial-akita-black-bowl-on-plate-camera-shift-i00-privileged/task.toml
index c7ddf6f4d7..d51ec7d237 100644
--- a/tasks/libero-max-spatial-akita-black-bowl-on-plate-camera-shift-i00-privileged/task.toml
+++ b/tasks/libero-max-spatial-akita-black-bowl-on-plate-camera-shift-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "robosuite"
robot = "franka-panda"
family = "libero-max/libero_spatial/camera_shift"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 195 # LIBERO-MAX's policy seed for the case (the scene is the frozen LIBERO-PRO initial state)
expert_time_estimate_hours = 1.0
diff --git a/tasks/libero-max-spatial-akita-black-bowl-on-plate-camera-shift-i00-standard/README.md b/tasks/libero-max-spatial-akita-black-bowl-on-plate-camera-shift-i00-standard/README.md
index c9b83de0fb..cd107602d4 100644
--- a/tasks/libero-max-spatial-akita-black-bowl-on-plate-camera-shift-i00-standard/README.md
+++ b/tasks/libero-max-spatial-akita-black-bowl-on-plate-camera-shift-i00-standard/README.md
@@ -86,3 +86,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
/reset` ends the episode as `error` (`sim_error = 1`), not `reset` (reachable only when a run grants resets; the
default is none). The scene, instance, pass rule and every verdict reached before are unchanged, so the task keeps
its name.
+- 2026-10-14 (board 0113, from board 0086): the instruction says that a moved object or receptacle jumps alone: what rests in or on it stays where it was (LIBERO-MAX moves one entity; board 0086 saw a basket jump through a can already in it). Text only: the instance, the event, the reference solution, LIBERO's success check and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/libero-max-spatial-akita-black-bowl-on-plate-camera-shift-i00-standard/instruction.md b/tasks/libero-max-spatial-akita-black-bowl-on-plate-camera-shift-i00-standard/instruction.md
index 4c68cfe4ff..7dd86074d1 100644
--- a/tasks/libero-max-spatial-akita-black-bowl-on-plate-camera-shift-i00-standard/instruction.md
+++ b/tasks/libero-max-spatial-akita-black-bowl-on-plate-camera-shift-i00-standard/instruction.md
@@ -6,7 +6,8 @@ simulator runs in a service you cannot open, and you reach it only through a cli
The scene can change while you work: once during the episode, when your gripper first gets close to one of the task's
objects, the scene changes (as in LIBERO-MAX): an object or a receptacle may be moved, an obstacle or a few other
objects may appear, a camera may shift, or the lighting, the colours or the image noise may change. Nothing tells you
-when; the task and its success check stay the same, so keep checking what the cameras show.
+when; the task and its success check stay the same, so keep checking what the cameras show. A moved object or
+receptacle jumps there alone: what already rests in or on it stays where it was.
## Rules (real-robot rules)
- **One episode, no reset.** The episode starts when the service comes up. It ends when the task succeeds, after
diff --git a/tasks/libero-max-spatial-akita-black-bowl-on-plate-camera-shift-i00-standard/task.toml b/tasks/libero-max-spatial-akita-black-bowl-on-plate-camera-shift-i00-standard/task.toml
index 29e2126d09..133e0d86fb 100644
--- a/tasks/libero-max-spatial-akita-black-bowl-on-plate-camera-shift-i00-standard/task.toml
+++ b/tasks/libero-max-spatial-akita-black-bowl-on-plate-camera-shift-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "robosuite"
robot = "franka-panda"
family = "libero-max/libero_spatial/camera_shift"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 195
diff --git a/tasks/libero-max-spatial-akita-black-bowl-on-plate-camera-shift-i01-privileged/task.toml b/tasks/libero-max-spatial-akita-black-bowl-on-plate-camera-shift-i01-privileged/task.toml
index 3c36a3f1b6..1b340fbad3 100644
--- a/tasks/libero-max-spatial-akita-black-bowl-on-plate-camera-shift-i01-privileged/task.toml
+++ b/tasks/libero-max-spatial-akita-black-bowl-on-plate-camera-shift-i01-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "robosuite"
robot = "franka-panda"
family = "libero-max/libero_spatial/camera_shift"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 195 # LIBERO-MAX's policy seed for the case (the scene is the frozen LIBERO-PRO initial state)
expert_time_estimate_hours = 1.0
diff --git a/tasks/libero-max-spatial-akita-black-bowl-on-plate-camera-shift-i01-standard/README.md b/tasks/libero-max-spatial-akita-black-bowl-on-plate-camera-shift-i01-standard/README.md
index bede9baf4d..80d78d0e80 100644
--- a/tasks/libero-max-spatial-akita-black-bowl-on-plate-camera-shift-i01-standard/README.md
+++ b/tasks/libero-max-spatial-akita-black-bowl-on-plate-camera-shift-i01-standard/README.md
@@ -86,3 +86,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
/reset` ends the episode as `error` (`sim_error = 1`), not `reset` (reachable only when a run grants resets; the
default is none). The scene, instance, pass rule and every verdict reached before are unchanged, so the task keeps
its name.
+- 2026-10-14 (board 0113, from board 0086): the instruction says that a moved object or receptacle jumps alone: what rests in or on it stays where it was (LIBERO-MAX moves one entity; board 0086 saw a basket jump through a can already in it). Text only: the instance, the event, the reference solution, LIBERO's success check and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/libero-max-spatial-akita-black-bowl-on-plate-camera-shift-i01-standard/instruction.md b/tasks/libero-max-spatial-akita-black-bowl-on-plate-camera-shift-i01-standard/instruction.md
index 599d97a5a3..2a9abbc259 100644
--- a/tasks/libero-max-spatial-akita-black-bowl-on-plate-camera-shift-i01-standard/instruction.md
+++ b/tasks/libero-max-spatial-akita-black-bowl-on-plate-camera-shift-i01-standard/instruction.md
@@ -6,7 +6,8 @@ simulator runs in a service you cannot open, and you reach it only through a cli
The scene can change while you work: once during the episode, when your gripper first gets close to one of the task's
objects, the scene changes (as in LIBERO-MAX): an object or a receptacle may be moved, an obstacle or a few other
objects may appear, a camera may shift, or the lighting, the colours or the image noise may change. Nothing tells you
-when; the task and its success check stay the same, so keep checking what the cameras show.
+when; the task and its success check stay the same, so keep checking what the cameras show. A moved object or
+receptacle jumps there alone: what already rests in or on it stays where it was.
## Rules (real-robot rules)
- **One episode, no reset.** The episode starts when the service comes up. It ends when the task succeeds, after
diff --git a/tasks/libero-max-spatial-akita-black-bowl-on-plate-camera-shift-i01-standard/task.toml b/tasks/libero-max-spatial-akita-black-bowl-on-plate-camera-shift-i01-standard/task.toml
index 42022766aa..4161e14849 100644
--- a/tasks/libero-max-spatial-akita-black-bowl-on-plate-camera-shift-i01-standard/task.toml
+++ b/tasks/libero-max-spatial-akita-black-bowl-on-plate-camera-shift-i01-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "robosuite"
robot = "franka-panda"
family = "libero-max/libero_spatial/camera_shift"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 195
diff --git a/tasks/libero-max-spatial-akita-black-bowl-on-plate-illumination-switch-i00-privileged/task.toml b/tasks/libero-max-spatial-akita-black-bowl-on-plate-illumination-switch-i00-privileged/task.toml
index fc47d1b607..cd5ef6d6a0 100644
--- a/tasks/libero-max-spatial-akita-black-bowl-on-plate-illumination-switch-i00-privileged/task.toml
+++ b/tasks/libero-max-spatial-akita-black-bowl-on-plate-illumination-switch-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "robosuite"
robot = "franka-panda"
family = "libero-max/libero_spatial/illumination_switch"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 195 # LIBERO-MAX's policy seed for the case (the scene is the frozen LIBERO-PRO initial state)
expert_time_estimate_hours = 1.0
diff --git a/tasks/libero-max-spatial-akita-black-bowl-on-plate-illumination-switch-i00-standard/README.md b/tasks/libero-max-spatial-akita-black-bowl-on-plate-illumination-switch-i00-standard/README.md
index 2a19981358..7bad1cc2bf 100644
--- a/tasks/libero-max-spatial-akita-black-bowl-on-plate-illumination-switch-i00-standard/README.md
+++ b/tasks/libero-max-spatial-akita-black-bowl-on-plate-illumination-switch-i00-standard/README.md
@@ -86,3 +86,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
/reset` ends the episode as `error` (`sim_error = 1`), not `reset` (reachable only when a run grants resets; the
default is none). The scene, instance, pass rule and every verdict reached before are unchanged, so the task keeps
its name.
+- 2026-10-14 (board 0113, from board 0086): the instruction says that a moved object or receptacle jumps alone: what rests in or on it stays where it was (LIBERO-MAX moves one entity; board 0086 saw a basket jump through a can already in it). Text only: the instance, the event, the reference solution, LIBERO's success check and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/libero-max-spatial-akita-black-bowl-on-plate-illumination-switch-i00-standard/instruction.md b/tasks/libero-max-spatial-akita-black-bowl-on-plate-illumination-switch-i00-standard/instruction.md
index d5f0012248..6e4f367160 100644
--- a/tasks/libero-max-spatial-akita-black-bowl-on-plate-illumination-switch-i00-standard/instruction.md
+++ b/tasks/libero-max-spatial-akita-black-bowl-on-plate-illumination-switch-i00-standard/instruction.md
@@ -6,7 +6,8 @@ simulator runs in a service you cannot open, and you reach it only through a cli
The scene can change while you work: once during the episode, when your gripper first gets close to one of the task's
objects, the scene changes (as in LIBERO-MAX): an object or a receptacle may be moved, an obstacle or a few other
objects may appear, a camera may shift, or the lighting, the colours or the image noise may change. Nothing tells you
-when; the task and its success check stay the same, so keep checking what the cameras show.
+when; the task and its success check stay the same, so keep checking what the cameras show. A moved object or
+receptacle jumps there alone: what already rests in or on it stays where it was.
## Rules (real-robot rules)
- **One episode, no reset.** The episode starts when the service comes up. It ends when the task succeeds, after
diff --git a/tasks/libero-max-spatial-akita-black-bowl-on-plate-illumination-switch-i00-standard/task.toml b/tasks/libero-max-spatial-akita-black-bowl-on-plate-illumination-switch-i00-standard/task.toml
index c18fa65dc8..6c5d7e724a 100644
--- a/tasks/libero-max-spatial-akita-black-bowl-on-plate-illumination-switch-i00-standard/task.toml
+++ b/tasks/libero-max-spatial-akita-black-bowl-on-plate-illumination-switch-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "robosuite"
robot = "franka-panda"
family = "libero-max/libero_spatial/illumination_switch"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 195
diff --git a/tasks/libero-max-spatial-akita-black-bowl-on-plate-illumination-switch-i01-privileged/task.toml b/tasks/libero-max-spatial-akita-black-bowl-on-plate-illumination-switch-i01-privileged/task.toml
index 3969de1623..f89deb67f4 100644
--- a/tasks/libero-max-spatial-akita-black-bowl-on-plate-illumination-switch-i01-privileged/task.toml
+++ b/tasks/libero-max-spatial-akita-black-bowl-on-plate-illumination-switch-i01-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "robosuite"
robot = "franka-panda"
family = "libero-max/libero_spatial/illumination_switch"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 195 # LIBERO-MAX's policy seed for the case (the scene is the frozen LIBERO-PRO initial state)
expert_time_estimate_hours = 1.0
diff --git a/tasks/libero-max-spatial-akita-black-bowl-on-plate-illumination-switch-i01-standard/README.md b/tasks/libero-max-spatial-akita-black-bowl-on-plate-illumination-switch-i01-standard/README.md
index 35d3e92d40..04519bd7c1 100644
--- a/tasks/libero-max-spatial-akita-black-bowl-on-plate-illumination-switch-i01-standard/README.md
+++ b/tasks/libero-max-spatial-akita-black-bowl-on-plate-illumination-switch-i01-standard/README.md
@@ -86,3 +86,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
/reset` ends the episode as `error` (`sim_error = 1`), not `reset` (reachable only when a run grants resets; the
default is none). The scene, instance, pass rule and every verdict reached before are unchanged, so the task keeps
its name.
+- 2026-10-14 (board 0113, from board 0086): the instruction says that a moved object or receptacle jumps alone: what rests in or on it stays where it was (LIBERO-MAX moves one entity; board 0086 saw a basket jump through a can already in it). Text only: the instance, the event, the reference solution, LIBERO's success check and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/libero-max-spatial-akita-black-bowl-on-plate-illumination-switch-i01-standard/instruction.md b/tasks/libero-max-spatial-akita-black-bowl-on-plate-illumination-switch-i01-standard/instruction.md
index ee13d3010a..9118fb34a0 100644
--- a/tasks/libero-max-spatial-akita-black-bowl-on-plate-illumination-switch-i01-standard/instruction.md
+++ b/tasks/libero-max-spatial-akita-black-bowl-on-plate-illumination-switch-i01-standard/instruction.md
@@ -6,7 +6,8 @@ simulator runs in a service you cannot open, and you reach it only through a cli
The scene can change while you work: once during the episode, when your gripper first gets close to one of the task's
objects, the scene changes (as in LIBERO-MAX): an object or a receptacle may be moved, an obstacle or a few other
objects may appear, a camera may shift, or the lighting, the colours or the image noise may change. Nothing tells you
-when; the task and its success check stay the same, so keep checking what the cameras show.
+when; the task and its success check stay the same, so keep checking what the cameras show. A moved object or
+receptacle jumps there alone: what already rests in or on it stays where it was.
## Rules (real-robot rules)
- **One episode, no reset.** The episode starts when the service comes up. It ends when the task succeeds, after
diff --git a/tasks/libero-max-spatial-akita-black-bowl-on-plate-illumination-switch-i01-standard/task.toml b/tasks/libero-max-spatial-akita-black-bowl-on-plate-illumination-switch-i01-standard/task.toml
index 8e04961606..5707ba752a 100644
--- a/tasks/libero-max-spatial-akita-black-bowl-on-plate-illumination-switch-i01-standard/task.toml
+++ b/tasks/libero-max-spatial-akita-black-bowl-on-plate-illumination-switch-i01-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "robosuite"
robot = "franka-panda"
family = "libero-max/libero_spatial/illumination_switch"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 195
diff --git a/tasks/libero-max-spatial-akita-black-bowl-on-plate-obstacle-insertion-i00-privileged/task.toml b/tasks/libero-max-spatial-akita-black-bowl-on-plate-obstacle-insertion-i00-privileged/task.toml
index 387e53c557..4c4fe3a0c4 100644
--- a/tasks/libero-max-spatial-akita-black-bowl-on-plate-obstacle-insertion-i00-privileged/task.toml
+++ b/tasks/libero-max-spatial-akita-black-bowl-on-plate-obstacle-insertion-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "robosuite"
robot = "franka-panda"
family = "libero-max/libero_spatial/obstacle_insertion"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 195 # LIBERO-MAX's policy seed for the case (the scene is the frozen LIBERO-PRO initial state)
expert_time_estimate_hours = 1.0
diff --git a/tasks/libero-max-spatial-akita-black-bowl-on-plate-obstacle-insertion-i00-standard/README.md b/tasks/libero-max-spatial-akita-black-bowl-on-plate-obstacle-insertion-i00-standard/README.md
index 35522f7baf..84486a46d1 100644
--- a/tasks/libero-max-spatial-akita-black-bowl-on-plate-obstacle-insertion-i00-standard/README.md
+++ b/tasks/libero-max-spatial-akita-black-bowl-on-plate-obstacle-insertion-i00-standard/README.md
@@ -86,3 +86,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
/reset` ends the episode as `error` (`sim_error = 1`), not `reset` (reachable only when a run grants resets; the
default is none). The scene, instance, pass rule and every verdict reached before are unchanged, so the task keeps
its name.
+- 2026-10-14 (board 0113, from board 0086): the instruction says that a moved object or receptacle jumps alone: what rests in or on it stays where it was (LIBERO-MAX moves one entity; board 0086 saw a basket jump through a can already in it). Text only: the instance, the event, the reference solution, LIBERO's success check and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/libero-max-spatial-akita-black-bowl-on-plate-obstacle-insertion-i00-standard/instruction.md b/tasks/libero-max-spatial-akita-black-bowl-on-plate-obstacle-insertion-i00-standard/instruction.md
index 0b6dee7afa..fc465d4cca 100644
--- a/tasks/libero-max-spatial-akita-black-bowl-on-plate-obstacle-insertion-i00-standard/instruction.md
+++ b/tasks/libero-max-spatial-akita-black-bowl-on-plate-obstacle-insertion-i00-standard/instruction.md
@@ -6,7 +6,8 @@ simulator runs in a service you cannot open, and you reach it only through a cli
The scene can change while you work: once during the episode, when your gripper first gets close to one of the task's
objects, the scene changes (as in LIBERO-MAX): an object or a receptacle may be moved, an obstacle or a few other
objects may appear, a camera may shift, or the lighting, the colours or the image noise may change. Nothing tells you
-when; the task and its success check stay the same, so keep checking what the cameras show.
+when; the task and its success check stay the same, so keep checking what the cameras show. A moved object or
+receptacle jumps there alone: what already rests in or on it stays where it was.
## Rules (real-robot rules)
- **One episode, no reset.** The episode starts when the service comes up. It ends when the task succeeds, after
diff --git a/tasks/libero-max-spatial-akita-black-bowl-on-plate-obstacle-insertion-i00-standard/task.toml b/tasks/libero-max-spatial-akita-black-bowl-on-plate-obstacle-insertion-i00-standard/task.toml
index 9ef0018fb7..b2bd267860 100644
--- a/tasks/libero-max-spatial-akita-black-bowl-on-plate-obstacle-insertion-i00-standard/task.toml
+++ b/tasks/libero-max-spatial-akita-black-bowl-on-plate-obstacle-insertion-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "robosuite"
robot = "franka-panda"
family = "libero-max/libero_spatial/obstacle_insertion"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 195
diff --git a/tasks/libero-max-spatial-akita-black-bowl-on-plate-obstacle-insertion-i01-privileged/task.toml b/tasks/libero-max-spatial-akita-black-bowl-on-plate-obstacle-insertion-i01-privileged/task.toml
index d3a83086aa..f57e399027 100644
--- a/tasks/libero-max-spatial-akita-black-bowl-on-plate-obstacle-insertion-i01-privileged/task.toml
+++ b/tasks/libero-max-spatial-akita-black-bowl-on-plate-obstacle-insertion-i01-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "robosuite"
robot = "franka-panda"
family = "libero-max/libero_spatial/obstacle_insertion"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 195 # LIBERO-MAX's policy seed for the case (the scene is the frozen LIBERO-PRO initial state)
expert_time_estimate_hours = 1.0
diff --git a/tasks/libero-max-spatial-akita-black-bowl-on-plate-obstacle-insertion-i01-standard/README.md b/tasks/libero-max-spatial-akita-black-bowl-on-plate-obstacle-insertion-i01-standard/README.md
index 8ffe2e3101..4b191961e8 100644
--- a/tasks/libero-max-spatial-akita-black-bowl-on-plate-obstacle-insertion-i01-standard/README.md
+++ b/tasks/libero-max-spatial-akita-black-bowl-on-plate-obstacle-insertion-i01-standard/README.md
@@ -86,3 +86,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
/reset` ends the episode as `error` (`sim_error = 1`), not `reset` (reachable only when a run grants resets; the
default is none). The scene, instance, pass rule and every verdict reached before are unchanged, so the task keeps
its name.
+- 2026-10-14 (board 0113, from board 0086): the instruction says that a moved object or receptacle jumps alone: what rests in or on it stays where it was (LIBERO-MAX moves one entity; board 0086 saw a basket jump through a can already in it). Text only: the instance, the event, the reference solution, LIBERO's success check and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/libero-max-spatial-akita-black-bowl-on-plate-obstacle-insertion-i01-standard/instruction.md b/tasks/libero-max-spatial-akita-black-bowl-on-plate-obstacle-insertion-i01-standard/instruction.md
index f895b34f6a..e0a6c0ae37 100644
--- a/tasks/libero-max-spatial-akita-black-bowl-on-plate-obstacle-insertion-i01-standard/instruction.md
+++ b/tasks/libero-max-spatial-akita-black-bowl-on-plate-obstacle-insertion-i01-standard/instruction.md
@@ -6,7 +6,8 @@ simulator runs in a service you cannot open, and you reach it only through a cli
The scene can change while you work: once during the episode, when your gripper first gets close to one of the task's
objects, the scene changes (as in LIBERO-MAX): an object or a receptacle may be moved, an obstacle or a few other
objects may appear, a camera may shift, or the lighting, the colours or the image noise may change. Nothing tells you
-when; the task and its success check stay the same, so keep checking what the cameras show.
+when; the task and its success check stay the same, so keep checking what the cameras show. A moved object or
+receptacle jumps there alone: what already rests in or on it stays where it was.
## Rules (real-robot rules)
- **One episode, no reset.** The episode starts when the service comes up. It ends when the task succeeds, after
diff --git a/tasks/libero-max-spatial-akita-black-bowl-on-plate-obstacle-insertion-i01-standard/task.toml b/tasks/libero-max-spatial-akita-black-bowl-on-plate-obstacle-insertion-i01-standard/task.toml
index b7dc250c61..006fc09927 100644
--- a/tasks/libero-max-spatial-akita-black-bowl-on-plate-obstacle-insertion-i01-standard/task.toml
+++ b/tasks/libero-max-spatial-akita-black-bowl-on-plate-obstacle-insertion-i01-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "robosuite"
robot = "franka-panda"
family = "libero-max/libero_spatial/obstacle_insertion"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 195
diff --git a/tasks/libero-max-spatial-akita-black-bowl-on-plate-receptacle-relocation-i00-privileged/README.md b/tasks/libero-max-spatial-akita-black-bowl-on-plate-receptacle-relocation-i00-privileged/README.md
index 8d8a0c7ab8..47c70e338f 100644
--- a/tasks/libero-max-spatial-akita-black-bowl-on-plate-receptacle-relocation-i00-privileged/README.md
+++ b/tasks/libero-max-spatial-akita-black-bowl-on-plate-receptacle-relocation-i00-privileged/README.md
@@ -14,7 +14,7 @@
## Task
LIBERO-MAX case `pro-camera-view-angle-spatial-t09-i06-receptacle_relocation-d1-p195`: a Franka Panda arm at a table, instructed **"Pick the akita black bowl on the wooden cabinet and place it on the plate."**, on a LIBERO-PRO variant
of a libero_spatial task (camera_view_angle). Online event (receptacle relocation): the first time the gripper comes within
-0.18 m of `akita_black_bowl_1`, LIBERO-MAX changes the scene once (the receptacle `plate_1` is moved 0.12 m, direction (0, -1)). The deliverable is one
+0.18 m of `akita_black_bowl_1`, LIBERO-MAX changes the scene once (the receptacle `plate_1` is moved 0.12 m, direction (0, -1); it jumps there alone: what already rests in or on it stays where it was, and where it lands on another object the physics pushes the two apart). The deliverable is one
replayable trajectory of LIBERO's OSC control steps (`/app/output/trajectory.npz`, `(T, 7)`, T ≤ 220), graded by
LIBERO's success check at its end.
@@ -54,7 +54,7 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
|---|---|
| Upstream | LIBERO-PRO @ 2b910b5 (MIT; task data `zhouxueyang/LIBERO-Pro` @ c86fc3b, CC-BY-4.0); LIBERO-MAX @ a1e3cef (event runtime and this case's parameters; no licence file upstream); robosuite 1.4.0, MuJoCo 3.2.6 |
| Case | `pro-camera-view-angle-spatial-t09-i06-receptacle_relocation-d1-p195` (LIBERO-MAX 8,000-case manifest; LIBERO-PRO category `camera_view_angle`, task 9, initial state 6) |
-| Event | `receptacle_relocation`: the receptacle `plate_1` is moved 0.12 m, direction (0, -1); trigger: gripper within 0.18 m of `akita_black_bowl_1` |
+| Event | `receptacle_relocation`: the receptacle `plate_1` is moved 0.12 m, direction (0, -1); it jumps there alone: what already rests in or on it stays where it was, and where it lands on another object the physics pushes the two apart; trigger: gripper within 0.18 m of `akita_black_bowl_1` |
| Actions | 7 per control step (1/20 s): OSC hand-target delta (1.0 = 0.05 m, 0.5 rad) and gripper |
| Resources | 4 CPUs, no GPU (CPU physics, OSMesa rendering) |
@@ -76,3 +76,4 @@ Generated by `scripts/libero-max/generate.py` from `scripts/libero-max/cases.jso
it executes any action). Before, such a trial read as an agent failure with `grader_error = 1` only.
- A replay that records a video prints its verdict before it finishes the video, and a failing video no longer fails
the replay; `video_rendered` is 1 only when the video was finished.
+- 2026-10-14 (board 0113, from board 0086): the instruction says that a moved receptacle jumps alone: what rests in or on it stays where it was (LIBERO-MAX moves one entity; board 0086 saw a basket jump through a can already in it). Text only: the instance, the event, the reference solution, LIBERO's success check and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/libero-max-spatial-akita-black-bowl-on-plate-receptacle-relocation-i00-privileged/instruction.md b/tasks/libero-max-spatial-akita-black-bowl-on-plate-receptacle-relocation-i00-privileged/instruction.md
index f720c94f5a..b610780ae9 100644
--- a/tasks/libero-max-spatial-akita-black-bowl-on-plate-receptacle-relocation-i00-privileged/instruction.md
+++ b/tasks/libero-max-spatial-akita-black-bowl-on-plate-receptacle-relocation-i00-privileged/instruction.md
@@ -3,7 +3,7 @@
A **Franka Panda** arm (LIBERO on robosuite / MuJoCo) at a table. **Task: Pick the akita black bowl on the wooden cabinet and place it on the plate.**
The scene is one frozen LIBERO-MAX case: a LIBERO-PRO variant of a libero_spatial task (camera_view_angle), with one online
event: the first time the gripper comes within 0.18 m of `akita_black_bowl_1`, LIBERO-MAX changes the scene
-once (the receptacle `plate_1` is moved 0.12 m, direction (0, -1)). You have the full simulator state, so you can see the change happen (`s.state()["event_fired"]`).
+once (the receptacle `plate_1` is moved 0.12 m, direction (0, -1); it jumps there alone: what already rests in or on it stays where it was, and where it lands on another object the physics pushes the two apart). You have the full simulator state, so you can see the change happen (`s.state()["event_fired"]`).
## What you hand in
`/app/output/trajectory.npz` with key `actions`: LIBERO's native control steps, shape `(T, 7)`, 1 ≤ T ≤ 220
diff --git a/tasks/libero-max-spatial-akita-black-bowl-on-plate-receptacle-relocation-i00-privileged/task.toml b/tasks/libero-max-spatial-akita-black-bowl-on-plate-receptacle-relocation-i00-privileged/task.toml
index ceaac53dca..e448727639 100644
--- a/tasks/libero-max-spatial-akita-black-bowl-on-plate-receptacle-relocation-i00-privileged/task.toml
+++ b/tasks/libero-max-spatial-akita-black-bowl-on-plate-receptacle-relocation-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "robosuite"
robot = "franka-panda"
family = "libero-max/libero_spatial/receptacle_relocation"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 195 # LIBERO-MAX's policy seed for the case (the scene is the frozen LIBERO-PRO initial state)
expert_time_estimate_hours = 1.0
diff --git a/tasks/libero-max-spatial-akita-black-bowl-on-plate-receptacle-relocation-i00-standard/README.md b/tasks/libero-max-spatial-akita-black-bowl-on-plate-receptacle-relocation-i00-standard/README.md
index 4a09215fa6..9b408148b5 100644
--- a/tasks/libero-max-spatial-akita-black-bowl-on-plate-receptacle-relocation-i00-standard/README.md
+++ b/tasks/libero-max-spatial-akita-black-bowl-on-plate-receptacle-relocation-i00-standard/README.md
@@ -13,7 +13,7 @@
## Task
LIBERO-MAX case `pro-camera-view-angle-spatial-t09-i06-receptacle_relocation-d1-p195` as a robot service: a Franka Panda arm at a table, instructed **"Pick the akita black bowl on the wooden cabinet and place it on the plate."**, on
-the same frozen instance as `libero-max-spatial-akita-black-bowl-on-plate-receptacle-relocation-i00-privileged`, with the same online event (receptacle relocation: the receptacle `plate_1` is moved 0.12 m, direction (0, -1), the
+the same frozen instance as `libero-max-spatial-akita-black-bowl-on-plate-receptacle-relocation-i00-privileged`, with the same online event (receptacle relocation: the receptacle `plate_1` is moved 0.12 m, direction (0, -1); it jumps there alone: what already rests in or on it stays where it was, and where it lands on another object the physics pushes the two apart, the
first time the gripper comes within 0.18 m of `akita_black_bowl_1`). The simulator runs in the `sim` sidecar
(`environment/sim/server.py` on main's `rcb_service.py`); the agent only has the client: two cameras (RGB
256×256, depth on request, calibrated), the robot's own state, and OSC control steps in chunks. One
@@ -86,3 +86,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
/reset` ends the episode as `error` (`sim_error = 1`), not `reset` (reachable only when a run grants resets; the
default is none). The scene, instance, pass rule and every verdict reached before are unchanged, so the task keeps
its name.
+- 2026-10-14 (board 0113, from board 0086): the instruction says that a moved object or receptacle jumps alone: what rests in or on it stays where it was (LIBERO-MAX moves one entity; board 0086 saw a basket jump through a can already in it). Text only: the instance, the event, the reference solution, LIBERO's success check and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/libero-max-spatial-akita-black-bowl-on-plate-receptacle-relocation-i00-standard/instruction.md b/tasks/libero-max-spatial-akita-black-bowl-on-plate-receptacle-relocation-i00-standard/instruction.md
index cedc44d32d..5d0b973445 100644
--- a/tasks/libero-max-spatial-akita-black-bowl-on-plate-receptacle-relocation-i00-standard/instruction.md
+++ b/tasks/libero-max-spatial-akita-black-bowl-on-plate-receptacle-relocation-i00-standard/instruction.md
@@ -6,7 +6,8 @@ simulator runs in a service you cannot open, and you reach it only through a cli
The scene can change while you work: once during the episode, when your gripper first gets close to one of the task's
objects, the scene changes (as in LIBERO-MAX): an object or a receptacle may be moved, an obstacle or a few other
objects may appear, a camera may shift, or the lighting, the colours or the image noise may change. Nothing tells you
-when; the task and its success check stay the same, so keep checking what the cameras show.
+when; the task and its success check stay the same, so keep checking what the cameras show. A moved object or
+receptacle jumps there alone: what already rests in or on it stays where it was.
## Rules (real-robot rules)
- **One episode, no reset.** The episode starts when the service comes up. It ends when the task succeeds, after
diff --git a/tasks/libero-max-spatial-akita-black-bowl-on-plate-receptacle-relocation-i00-standard/task.toml b/tasks/libero-max-spatial-akita-black-bowl-on-plate-receptacle-relocation-i00-standard/task.toml
index 50f7d89810..be9114d8e5 100644
--- a/tasks/libero-max-spatial-akita-black-bowl-on-plate-receptacle-relocation-i00-standard/task.toml
+++ b/tasks/libero-max-spatial-akita-black-bowl-on-plate-receptacle-relocation-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "robosuite"
robot = "franka-panda"
family = "libero-max/libero_spatial/receptacle_relocation"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 195
diff --git a/tasks/libero-max-spatial-akita-black-bowl-on-plate-receptacle-relocation-i01-privileged/README.md b/tasks/libero-max-spatial-akita-black-bowl-on-plate-receptacle-relocation-i01-privileged/README.md
index 5e6bd4c090..a411056326 100644
--- a/tasks/libero-max-spatial-akita-black-bowl-on-plate-receptacle-relocation-i01-privileged/README.md
+++ b/tasks/libero-max-spatial-akita-black-bowl-on-plate-receptacle-relocation-i01-privileged/README.md
@@ -14,7 +14,7 @@
## Task
LIBERO-MAX case `pro-view-occlusion-spatial-t08-i02-receptacle_relocation-d1-p195`: a Franka Panda arm at a table, instructed **"Pick the akita black bowl next to the plate and place it on the plate."**, on a LIBERO-PRO variant
of a libero_spatial task (view_occlusion). Online event (receptacle relocation): the first time the gripper comes within
-0.18 m of `akita_black_bowl_1`, LIBERO-MAX changes the scene once (the receptacle `plate_1` is moved 0.12 m, direction (-0.707, -0.707)). The deliverable is one
+0.18 m of `akita_black_bowl_1`, LIBERO-MAX changes the scene once (the receptacle `plate_1` is moved 0.12 m, direction (-0.707, -0.707); it jumps there alone: what already rests in or on it stays where it was, and where it lands on another object the physics pushes the two apart). The deliverable is one
replayable trajectory of LIBERO's OSC control steps (`/app/output/trajectory.npz`, `(T, 7)`, T ≤ 220), graded by
LIBERO's success check at its end.
@@ -54,7 +54,7 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
|---|---|
| Upstream | LIBERO-PRO @ 2b910b5 (MIT; task data `zhouxueyang/LIBERO-Pro` @ c86fc3b, CC-BY-4.0); LIBERO-MAX @ a1e3cef (event runtime and this case's parameters; no licence file upstream); robosuite 1.4.0, MuJoCo 3.2.6 |
| Case | `pro-view-occlusion-spatial-t08-i02-receptacle_relocation-d1-p195` (LIBERO-MAX 8,000-case manifest; LIBERO-PRO category `view_occlusion`, task 8, initial state 2) |
-| Event | `receptacle_relocation`: the receptacle `plate_1` is moved 0.12 m, direction (-0.707, -0.707); trigger: gripper within 0.18 m of `akita_black_bowl_1` |
+| Event | `receptacle_relocation`: the receptacle `plate_1` is moved 0.12 m, direction (-0.707, -0.707); it jumps there alone: what already rests in or on it stays where it was, and where it lands on another object the physics pushes the two apart; trigger: gripper within 0.18 m of `akita_black_bowl_1` |
| Actions | 7 per control step (1/20 s): OSC hand-target delta (1.0 = 0.05 m, 0.5 rad) and gripper |
| Resources | 4 CPUs, no GPU (CPU physics, OSMesa rendering) |
@@ -76,3 +76,4 @@ Generated by `scripts/libero-max/generate.py` from `scripts/libero-max/cases.jso
it executes any action). Before, such a trial read as an agent failure with `grader_error = 1` only.
- A replay that records a video prints its verdict before it finishes the video, and a failing video no longer fails
the replay; `video_rendered` is 1 only when the video was finished.
+- 2026-10-14 (board 0113, from board 0086): the instruction says that a moved receptacle jumps alone: what rests in or on it stays where it was (LIBERO-MAX moves one entity; board 0086 saw a basket jump through a can already in it). Text only: the instance, the event, the reference solution, LIBERO's success check and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/libero-max-spatial-akita-black-bowl-on-plate-receptacle-relocation-i01-privileged/instruction.md b/tasks/libero-max-spatial-akita-black-bowl-on-plate-receptacle-relocation-i01-privileged/instruction.md
index cdf65ec358..a0dcd4ccf0 100644
--- a/tasks/libero-max-spatial-akita-black-bowl-on-plate-receptacle-relocation-i01-privileged/instruction.md
+++ b/tasks/libero-max-spatial-akita-black-bowl-on-plate-receptacle-relocation-i01-privileged/instruction.md
@@ -3,7 +3,7 @@
A **Franka Panda** arm (LIBERO on robosuite / MuJoCo) at a table. **Task: Pick the akita black bowl next to the plate and place it on the plate.**
The scene is one frozen LIBERO-MAX case: a LIBERO-PRO variant of a libero_spatial task (view_occlusion), with one online
event: the first time the gripper comes within 0.18 m of `akita_black_bowl_1`, LIBERO-MAX changes the scene
-once (the receptacle `plate_1` is moved 0.12 m, direction (-0.707, -0.707)). You have the full simulator state, so you can see the change happen (`s.state()["event_fired"]`).
+once (the receptacle `plate_1` is moved 0.12 m, direction (-0.707, -0.707); it jumps there alone: what already rests in or on it stays where it was, and where it lands on another object the physics pushes the two apart). You have the full simulator state, so you can see the change happen (`s.state()["event_fired"]`).
## What you hand in
`/app/output/trajectory.npz` with key `actions`: LIBERO's native control steps, shape `(T, 7)`, 1 ≤ T ≤ 220
diff --git a/tasks/libero-max-spatial-akita-black-bowl-on-plate-receptacle-relocation-i01-privileged/task.toml b/tasks/libero-max-spatial-akita-black-bowl-on-plate-receptacle-relocation-i01-privileged/task.toml
index c83016f8a7..1c3ab11394 100644
--- a/tasks/libero-max-spatial-akita-black-bowl-on-plate-receptacle-relocation-i01-privileged/task.toml
+++ b/tasks/libero-max-spatial-akita-black-bowl-on-plate-receptacle-relocation-i01-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "robosuite"
robot = "franka-panda"
family = "libero-max/libero_spatial/receptacle_relocation"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 195 # LIBERO-MAX's policy seed for the case (the scene is the frozen LIBERO-PRO initial state)
expert_time_estimate_hours = 1.0
diff --git a/tasks/libero-max-spatial-akita-black-bowl-on-plate-receptacle-relocation-i01-standard/README.md b/tasks/libero-max-spatial-akita-black-bowl-on-plate-receptacle-relocation-i01-standard/README.md
index eb4ce8a7ed..f60c78e030 100644
--- a/tasks/libero-max-spatial-akita-black-bowl-on-plate-receptacle-relocation-i01-standard/README.md
+++ b/tasks/libero-max-spatial-akita-black-bowl-on-plate-receptacle-relocation-i01-standard/README.md
@@ -13,7 +13,7 @@
## Task
LIBERO-MAX case `pro-view-occlusion-spatial-t08-i02-receptacle_relocation-d1-p195` as a robot service: a Franka Panda arm at a table, instructed **"Pick the akita black bowl next to the plate and place it on the plate."**, on
-the same frozen instance as `libero-max-spatial-akita-black-bowl-on-plate-receptacle-relocation-i01-privileged`, with the same online event (receptacle relocation: the receptacle `plate_1` is moved 0.12 m, direction (-0.707, -0.707), the
+the same frozen instance as `libero-max-spatial-akita-black-bowl-on-plate-receptacle-relocation-i01-privileged`, with the same online event (receptacle relocation: the receptacle `plate_1` is moved 0.12 m, direction (-0.707, -0.707); it jumps there alone: what already rests in or on it stays where it was, and where it lands on another object the physics pushes the two apart, the
first time the gripper comes within 0.18 m of `akita_black_bowl_1`). The simulator runs in the `sim` sidecar
(`environment/sim/server.py` on main's `rcb_service.py`); the agent only has the client: two cameras (RGB
256×256, depth on request, calibrated), the robot's own state, and OSC control steps in chunks. One
@@ -86,3 +86,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
/reset` ends the episode as `error` (`sim_error = 1`), not `reset` (reachable only when a run grants resets; the
default is none). The scene, instance, pass rule and every verdict reached before are unchanged, so the task keeps
its name.
+- 2026-10-14 (board 0113, from board 0086): the instruction says that a moved object or receptacle jumps alone: what rests in or on it stays where it was (LIBERO-MAX moves one entity; board 0086 saw a basket jump through a can already in it). Text only: the instance, the event, the reference solution, LIBERO's success check and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/libero-max-spatial-akita-black-bowl-on-plate-receptacle-relocation-i01-standard/instruction.md b/tasks/libero-max-spatial-akita-black-bowl-on-plate-receptacle-relocation-i01-standard/instruction.md
index 759ae8abe5..b02d82f15f 100644
--- a/tasks/libero-max-spatial-akita-black-bowl-on-plate-receptacle-relocation-i01-standard/instruction.md
+++ b/tasks/libero-max-spatial-akita-black-bowl-on-plate-receptacle-relocation-i01-standard/instruction.md
@@ -6,7 +6,8 @@ simulator runs in a service you cannot open, and you reach it only through a cli
The scene can change while you work: once during the episode, when your gripper first gets close to one of the task's
objects, the scene changes (as in LIBERO-MAX): an object or a receptacle may be moved, an obstacle or a few other
objects may appear, a camera may shift, or the lighting, the colours or the image noise may change. Nothing tells you
-when; the task and its success check stay the same, so keep checking what the cameras show.
+when; the task and its success check stay the same, so keep checking what the cameras show. A moved object or
+receptacle jumps there alone: what already rests in or on it stays where it was.
## Rules (real-robot rules)
- **One episode, no reset.** The episode starts when the service comes up. It ends when the task succeeds, after
diff --git a/tasks/libero-max-spatial-akita-black-bowl-on-plate-receptacle-relocation-i01-standard/task.toml b/tasks/libero-max-spatial-akita-black-bowl-on-plate-receptacle-relocation-i01-standard/task.toml
index 8a9ee46af7..5faf5bb79d 100644
--- a/tasks/libero-max-spatial-akita-black-bowl-on-plate-receptacle-relocation-i01-standard/task.toml
+++ b/tasks/libero-max-spatial-akita-black-bowl-on-plate-receptacle-relocation-i01-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "robosuite"
robot = "franka-panda"
family = "libero-max/libero_spatial/receptacle_relocation"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 195
diff --git a/tasks/libero-max-spatial-akita-black-bowl-on-plate-sensor-noise-onset-i00-privileged/task.toml b/tasks/libero-max-spatial-akita-black-bowl-on-plate-sensor-noise-onset-i00-privileged/task.toml
index 3e6b2552e0..49683da5f4 100644
--- a/tasks/libero-max-spatial-akita-black-bowl-on-plate-sensor-noise-onset-i00-privileged/task.toml
+++ b/tasks/libero-max-spatial-akita-black-bowl-on-plate-sensor-noise-onset-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "robosuite"
robot = "franka-panda"
family = "libero-max/libero_spatial/sensor_noise_onset"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 195 # LIBERO-MAX's policy seed for the case (the scene is the frozen LIBERO-PRO initial state)
expert_time_estimate_hours = 1.0
diff --git a/tasks/libero-max-spatial-akita-black-bowl-on-plate-sensor-noise-onset-i00-standard/README.md b/tasks/libero-max-spatial-akita-black-bowl-on-plate-sensor-noise-onset-i00-standard/README.md
index eb5622529c..b424ff078b 100644
--- a/tasks/libero-max-spatial-akita-black-bowl-on-plate-sensor-noise-onset-i00-standard/README.md
+++ b/tasks/libero-max-spatial-akita-black-bowl-on-plate-sensor-noise-onset-i00-standard/README.md
@@ -86,3 +86,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
/reset` ends the episode as `error` (`sim_error = 1`), not `reset` (reachable only when a run grants resets; the
default is none). The scene, instance, pass rule and every verdict reached before are unchanged, so the task keeps
its name.
+- 2026-10-14 (board 0113, from board 0086): the instruction says that a moved object or receptacle jumps alone: what rests in or on it stays where it was (LIBERO-MAX moves one entity; board 0086 saw a basket jump through a can already in it). Text only: the instance, the event, the reference solution, LIBERO's success check and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/libero-max-spatial-akita-black-bowl-on-plate-sensor-noise-onset-i00-standard/instruction.md b/tasks/libero-max-spatial-akita-black-bowl-on-plate-sensor-noise-onset-i00-standard/instruction.md
index ebf432cb8c..bdcbf30418 100644
--- a/tasks/libero-max-spatial-akita-black-bowl-on-plate-sensor-noise-onset-i00-standard/instruction.md
+++ b/tasks/libero-max-spatial-akita-black-bowl-on-plate-sensor-noise-onset-i00-standard/instruction.md
@@ -6,7 +6,8 @@ simulator runs in a service you cannot open, and you reach it only through a cli
The scene can change while you work: once during the episode, when your gripper first gets close to one of the task's
objects, the scene changes (as in LIBERO-MAX): an object or a receptacle may be moved, an obstacle or a few other
objects may appear, a camera may shift, or the lighting, the colours or the image noise may change. Nothing tells you
-when; the task and its success check stay the same, so keep checking what the cameras show.
+when; the task and its success check stay the same, so keep checking what the cameras show. A moved object or
+receptacle jumps there alone: what already rests in or on it stays where it was.
## Rules (real-robot rules)
- **One episode, no reset.** The episode starts when the service comes up. It ends when the task succeeds, after
diff --git a/tasks/libero-max-spatial-akita-black-bowl-on-plate-sensor-noise-onset-i00-standard/task.toml b/tasks/libero-max-spatial-akita-black-bowl-on-plate-sensor-noise-onset-i00-standard/task.toml
index 0883abda41..58909334e8 100644
--- a/tasks/libero-max-spatial-akita-black-bowl-on-plate-sensor-noise-onset-i00-standard/task.toml
+++ b/tasks/libero-max-spatial-akita-black-bowl-on-plate-sensor-noise-onset-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "robosuite"
robot = "franka-panda"
family = "libero-max/libero_spatial/sensor_noise_onset"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 195
diff --git a/tasks/libero-max-spatial-akita-black-bowl-on-plate-sensor-noise-onset-i01-privileged/README.md b/tasks/libero-max-spatial-akita-black-bowl-on-plate-sensor-noise-onset-i01-privileged/README.md
index 50ca512309..7cae4eb8bd 100644
--- a/tasks/libero-max-spatial-akita-black-bowl-on-plate-sensor-noise-onset-i01-privileged/README.md
+++ b/tasks/libero-max-spatial-akita-black-bowl-on-plate-sensor-noise-onset-i01-privileged/README.md
@@ -76,3 +76,4 @@ Generated by `scripts/libero-max/generate.py` from `scripts/libero-max/cases.jso
it executes any action). Before, such a trial read as an agent failure with `grader_error = 1` only.
- A replay that records a video prints its verdict before it finishes the video, and a failing video no longer fails
the replay; `video_rendered` is 1 only when the video was finished.
+- 2026-10-14 (board 0113, from board 0086): the instruction adds a note after the task sentence, where the sentence alone misleads in this LIBERO-PRO variant: the target bowl is not where the sentence puts it, and which bowl is meant. Text only: the instance, the event, the reference solution, LIBERO's success check and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/libero-max-spatial-akita-black-bowl-on-plate-sensor-noise-onset-i01-privileged/instruction.md b/tasks/libero-max-spatial-akita-black-bowl-on-plate-sensor-noise-onset-i01-privileged/instruction.md
index e862e10157..4d5577ea84 100644
--- a/tasks/libero-max-spatial-akita-black-bowl-on-plate-sensor-noise-onset-i01-privileged/instruction.md
+++ b/tasks/libero-max-spatial-akita-black-bowl-on-plate-sensor-noise-onset-i01-privileged/instruction.md
@@ -1,6 +1,6 @@
# LIBERO-MAX · libero_spatial · sensor-noise onset — instance i01
-A **Franka Panda** arm (LIBERO on robosuite / MuJoCo) at a table. **Task: Pick the akita black bowl on the cookies box and place it on the plate.**
+A **Franka Panda** arm (LIBERO on robosuite / MuJoCo) at a table. **Task: Pick the akita black bowl on the cookies box and place it on the plate.** LIBERO-PRO's initial-pose variant moved the target bowl (`akita_black_bowl_1`, the goal's) 0.17 m away from the robot and 0.06 m to the robot's right of where the sentence puts it, and turned it 45 degrees: it starts on the table, not on the cookies box (no bowl is on the box). The bowl to move is still that one; the other black bowl stands on top of the wooden cabinet.
The scene is one frozen LIBERO-MAX case: a LIBERO-PRO variant of a libero_spatial task (initial_pose_position_angle), with one online
event: the first time the gripper comes within 0.18 m of `akita_black_bowl_1`, LIBERO-MAX changes the scene
once (the camera images get noise (std 36 on 0..255) and a blocked patch (16% of the image)). You have the full simulator state, so you can see the change happen (`s.state()["event_fired"]`).
diff --git a/tasks/libero-max-spatial-akita-black-bowl-on-plate-sensor-noise-onset-i01-privileged/task.toml b/tasks/libero-max-spatial-akita-black-bowl-on-plate-sensor-noise-onset-i01-privileged/task.toml
index 4b01cb9fdc..b65f1f5e56 100644
--- a/tasks/libero-max-spatial-akita-black-bowl-on-plate-sensor-noise-onset-i01-privileged/task.toml
+++ b/tasks/libero-max-spatial-akita-black-bowl-on-plate-sensor-noise-onset-i01-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "robosuite"
robot = "franka-panda"
family = "libero-max/libero_spatial/sensor_noise_onset"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 195 # LIBERO-MAX's policy seed for the case (the scene is the frozen LIBERO-PRO initial state)
expert_time_estimate_hours = 1.0
diff --git a/tasks/libero-max-spatial-akita-black-bowl-on-plate-sensor-noise-onset-i01-standard/README.md b/tasks/libero-max-spatial-akita-black-bowl-on-plate-sensor-noise-onset-i01-standard/README.md
index a8baf1e488..b22d9060e0 100644
--- a/tasks/libero-max-spatial-akita-black-bowl-on-plate-sensor-noise-onset-i01-standard/README.md
+++ b/tasks/libero-max-spatial-akita-black-bowl-on-plate-sensor-noise-onset-i01-standard/README.md
@@ -86,3 +86,5 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
/reset` ends the episode as `error` (`sim_error = 1`), not `reset` (reachable only when a run grants resets; the
default is none). The scene, instance, pass rule and every verdict reached before are unchanged, so the task keeps
its name.
+- 2026-10-14 (board 0113, from board 0086): the instruction adds a note after the task sentence, where the sentence alone misleads in this LIBERO-PRO variant: the target bowl is not where the sentence puts it, and which bowl is meant. Text only: the instance, the event, the reference solution, LIBERO's success check and the verifier are unchanged, so the task keeps its name.
+- 2026-10-14 (board 0113, from board 0086): the instruction says that a moved object or receptacle jumps alone: what rests in or on it stays where it was (LIBERO-MAX moves one entity; board 0086 saw a basket jump through a can already in it). Text only: the instance, the event, the reference solution, LIBERO's success check and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/libero-max-spatial-akita-black-bowl-on-plate-sensor-noise-onset-i01-standard/instruction.md b/tasks/libero-max-spatial-akita-black-bowl-on-plate-sensor-noise-onset-i01-standard/instruction.md
index b8ae92f0f6..2e4236889f 100644
--- a/tasks/libero-max-spatial-akita-black-bowl-on-plate-sensor-noise-onset-i01-standard/instruction.md
+++ b/tasks/libero-max-spatial-akita-black-bowl-on-plate-sensor-noise-onset-i01-standard/instruction.md
@@ -1,12 +1,13 @@
# LIBERO-MAX · libero_spatial · sensor-noise onset — instance i01 — STANDARD MODE (robot as a service)
You control a simulated **Franka Panda** arm (LIBERO on robosuite / MuJoCo) **as if it were real hardware**: the
-simulator runs in a service you cannot open, and you reach it only through a client. **Task: Pick the akita black bowl on the cookies box and place it on the plate.**
+simulator runs in a service you cannot open, and you reach it only through a client. **Task: Pick the akita black bowl on the cookies box and place it on the plate.** LIBERO-PRO's initial-pose variant moved the target bowl (`akita_black_bowl_1`, the goal's) 0.17 m away from the robot and 0.06 m to the robot's right of where the sentence puts it, and turned it 45 degrees: it starts on the table, not on the cookies box (no bowl is on the box). The bowl to move is still that one; the other black bowl stands on top of the wooden cabinet.
The scene can change while you work: once during the episode, when your gripper first gets close to one of the task's
objects, the scene changes (as in LIBERO-MAX): an object or a receptacle may be moved, an obstacle or a few other
objects may appear, a camera may shift, or the lighting, the colours or the image noise may change. Nothing tells you
-when; the task and its success check stay the same, so keep checking what the cameras show.
+when; the task and its success check stay the same, so keep checking what the cameras show. A moved object or
+receptacle jumps there alone: what already rests in or on it stays where it was.
## Rules (real-robot rules)
- **One episode, no reset.** The episode starts when the service comes up. It ends when the task succeeds, after
diff --git a/tasks/libero-max-spatial-akita-black-bowl-on-plate-sensor-noise-onset-i01-standard/task.toml b/tasks/libero-max-spatial-akita-black-bowl-on-plate-sensor-noise-onset-i01-standard/task.toml
index be059f5c60..fa94c0be12 100644
--- a/tasks/libero-max-spatial-akita-black-bowl-on-plate-sensor-noise-onset-i01-standard/task.toml
+++ b/tasks/libero-max-spatial-akita-black-bowl-on-plate-sensor-noise-onset-i01-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "robosuite"
robot = "franka-panda"
family = "libero-max/libero_spatial/sensor_noise_onset"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 195
diff --git a/tasks/libero-max-spatial-akita-black-bowl-on-plate-target-relocation-i00-privileged/README.md b/tasks/libero-max-spatial-akita-black-bowl-on-plate-target-relocation-i00-privileged/README.md
index 31cf8c70a5..ab51dd52a7 100644
--- a/tasks/libero-max-spatial-akita-black-bowl-on-plate-target-relocation-i00-privileged/README.md
+++ b/tasks/libero-max-spatial-akita-black-bowl-on-plate-target-relocation-i00-privileged/README.md
@@ -76,3 +76,4 @@ Generated by `scripts/libero-max/generate.py` from `scripts/libero-max/cases.jso
it executes any action). Before, such a trial read as an agent failure with `grader_error = 1` only.
- A replay that records a video prints its verdict before it finishes the video, and a failing video no longer fails
the replay; `video_rendered` is 1 only when the video was finished.
+- 2026-10-14 (board 0113, from board 0086): the instruction adds a note after the task sentence, where the sentence alone misleads in this LIBERO-PRO variant: the target bowl is not where the sentence puts it, and which bowl is meant. Text only: the instance, the event, the reference solution, LIBERO's success check and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/libero-max-spatial-akita-black-bowl-on-plate-target-relocation-i00-privileged/instruction.md b/tasks/libero-max-spatial-akita-black-bowl-on-plate-target-relocation-i00-privileged/instruction.md
index b66c6c8197..91e9939feb 100644
--- a/tasks/libero-max-spatial-akita-black-bowl-on-plate-target-relocation-i00-privileged/instruction.md
+++ b/tasks/libero-max-spatial-akita-black-bowl-on-plate-target-relocation-i00-privileged/instruction.md
@@ -1,6 +1,6 @@
# LIBERO-MAX · libero_spatial · target relocation — instance i00
-A **Franka Panda** arm (LIBERO on robosuite / MuJoCo) at a table. **Task: Pick the akita black bowl next to the cookies box and place it on the plate.**
+A **Franka Panda** arm (LIBERO on robosuite / MuJoCo) at a table. **Task: Pick the akita black bowl next to the cookies box and place it on the plate.** LIBERO-PRO's initial-pose variant moved the target bowl (`akita_black_bowl_1`, the goal's) 0.17 m away from the robot and 0.06 m to the robot's right of where the sentence puts it, and turned it 45 degrees: it no longer stands right next to the cookies box. The bowl to move is still that one; the other black bowl stands on the stove.
The scene is one frozen LIBERO-MAX case: a LIBERO-PRO variant of a libero_spatial task (initial_pose_position_angle), with one online
event: the first time the gripper comes within 0.18 m of `akita_black_bowl_1`, LIBERO-MAX changes the scene
once (`akita_black_bowl_1` is moved 0.06 m across the table, direction (-0.707, 0.707)). You have the full simulator state, so you can see the change happen (`s.state()["event_fired"]`).
diff --git a/tasks/libero-max-spatial-akita-black-bowl-on-plate-target-relocation-i00-privileged/task.toml b/tasks/libero-max-spatial-akita-black-bowl-on-plate-target-relocation-i00-privileged/task.toml
index 336f5c4802..7338f0e6b1 100644
--- a/tasks/libero-max-spatial-akita-black-bowl-on-plate-target-relocation-i00-privileged/task.toml
+++ b/tasks/libero-max-spatial-akita-black-bowl-on-plate-target-relocation-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "robosuite"
robot = "franka-panda"
family = "libero-max/libero_spatial/target_relocation"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 195 # LIBERO-MAX's policy seed for the case (the scene is the frozen LIBERO-PRO initial state)
expert_time_estimate_hours = 1.0
diff --git a/tasks/libero-max-spatial-akita-black-bowl-on-plate-target-relocation-i00-standard/README.md b/tasks/libero-max-spatial-akita-black-bowl-on-plate-target-relocation-i00-standard/README.md
index 6bec08057b..f87a6a23b1 100644
--- a/tasks/libero-max-spatial-akita-black-bowl-on-plate-target-relocation-i00-standard/README.md
+++ b/tasks/libero-max-spatial-akita-black-bowl-on-plate-target-relocation-i00-standard/README.md
@@ -86,3 +86,5 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
/reset` ends the episode as `error` (`sim_error = 1`), not `reset` (reachable only when a run grants resets; the
default is none). The scene, instance, pass rule and every verdict reached before are unchanged, so the task keeps
its name.
+- 2026-10-14 (board 0113, from board 0086): the instruction adds a note after the task sentence, where the sentence alone misleads in this LIBERO-PRO variant: the target bowl is not where the sentence puts it, and which bowl is meant. Text only: the instance, the event, the reference solution, LIBERO's success check and the verifier are unchanged, so the task keeps its name.
+- 2026-10-14 (board 0113, from board 0086): the instruction says that a moved object or receptacle jumps alone: what rests in or on it stays where it was (LIBERO-MAX moves one entity; board 0086 saw a basket jump through a can already in it). Text only: the instance, the event, the reference solution, LIBERO's success check and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/libero-max-spatial-akita-black-bowl-on-plate-target-relocation-i00-standard/instruction.md b/tasks/libero-max-spatial-akita-black-bowl-on-plate-target-relocation-i00-standard/instruction.md
index 7ca75b55e3..8cb4a27019 100644
--- a/tasks/libero-max-spatial-akita-black-bowl-on-plate-target-relocation-i00-standard/instruction.md
+++ b/tasks/libero-max-spatial-akita-black-bowl-on-plate-target-relocation-i00-standard/instruction.md
@@ -1,12 +1,13 @@
# LIBERO-MAX · libero_spatial · target relocation — instance i00 — STANDARD MODE (robot as a service)
You control a simulated **Franka Panda** arm (LIBERO on robosuite / MuJoCo) **as if it were real hardware**: the
-simulator runs in a service you cannot open, and you reach it only through a client. **Task: Pick the akita black bowl next to the cookies box and place it on the plate.**
+simulator runs in a service you cannot open, and you reach it only through a client. **Task: Pick the akita black bowl next to the cookies box and place it on the plate.** LIBERO-PRO's initial-pose variant moved the target bowl (`akita_black_bowl_1`, the goal's) 0.17 m away from the robot and 0.06 m to the robot's right of where the sentence puts it, and turned it 45 degrees: it no longer stands right next to the cookies box. The bowl to move is still that one; the other black bowl stands on the stove.
The scene can change while you work: once during the episode, when your gripper first gets close to one of the task's
objects, the scene changes (as in LIBERO-MAX): an object or a receptacle may be moved, an obstacle or a few other
objects may appear, a camera may shift, or the lighting, the colours or the image noise may change. Nothing tells you
-when; the task and its success check stay the same, so keep checking what the cameras show.
+when; the task and its success check stay the same, so keep checking what the cameras show. A moved object or
+receptacle jumps there alone: what already rests in or on it stays where it was.
## Rules (real-robot rules)
- **One episode, no reset.** The episode starts when the service comes up. It ends when the task succeeds, after
diff --git a/tasks/libero-max-spatial-akita-black-bowl-on-plate-target-relocation-i00-standard/task.toml b/tasks/libero-max-spatial-akita-black-bowl-on-plate-target-relocation-i00-standard/task.toml
index d59623612f..b496c4808d 100644
--- a/tasks/libero-max-spatial-akita-black-bowl-on-plate-target-relocation-i00-standard/task.toml
+++ b/tasks/libero-max-spatial-akita-black-bowl-on-plate-target-relocation-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "robosuite"
robot = "franka-panda"
family = "libero-max/libero_spatial/target_relocation"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 195
diff --git a/tasks/libero-max-spatial-akita-black-bowl-on-plate-target-relocation-i01-privileged/task.toml b/tasks/libero-max-spatial-akita-black-bowl-on-plate-target-relocation-i01-privileged/task.toml
index 8dae069059..4c7a01304e 100644
--- a/tasks/libero-max-spatial-akita-black-bowl-on-plate-target-relocation-i01-privileged/task.toml
+++ b/tasks/libero-max-spatial-akita-black-bowl-on-plate-target-relocation-i01-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "robosuite"
robot = "franka-panda"
family = "libero-max/libero_spatial/target_relocation"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 195 # LIBERO-MAX's policy seed for the case (the scene is the frozen LIBERO-PRO initial state)
expert_time_estimate_hours = 1.0
diff --git a/tasks/libero-max-spatial-akita-black-bowl-on-plate-target-relocation-i01-standard/README.md b/tasks/libero-max-spatial-akita-black-bowl-on-plate-target-relocation-i01-standard/README.md
index 806b4714bd..a5e75f2cfa 100644
--- a/tasks/libero-max-spatial-akita-black-bowl-on-plate-target-relocation-i01-standard/README.md
+++ b/tasks/libero-max-spatial-akita-black-bowl-on-plate-target-relocation-i01-standard/README.md
@@ -86,3 +86,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
/reset` ends the episode as `error` (`sim_error = 1`), not `reset` (reachable only when a run grants resets; the
default is none). The scene, instance, pass rule and every verdict reached before are unchanged, so the task keeps
its name.
+- 2026-10-14 (board 0113, from board 0086): the instruction says that a moved object or receptacle jumps alone: what rests in or on it stays where it was (LIBERO-MAX moves one entity; board 0086 saw a basket jump through a can already in it). Text only: the instance, the event, the reference solution, LIBERO's success check and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/libero-max-spatial-akita-black-bowl-on-plate-target-relocation-i01-standard/instruction.md b/tasks/libero-max-spatial-akita-black-bowl-on-plate-target-relocation-i01-standard/instruction.md
index 0c3c6bd24e..340110a0a4 100644
--- a/tasks/libero-max-spatial-akita-black-bowl-on-plate-target-relocation-i01-standard/instruction.md
+++ b/tasks/libero-max-spatial-akita-black-bowl-on-plate-target-relocation-i01-standard/instruction.md
@@ -6,7 +6,8 @@ simulator runs in a service you cannot open, and you reach it only through a cli
The scene can change while you work: once during the episode, when your gripper first gets close to one of the task's
objects, the scene changes (as in LIBERO-MAX): an object or a receptacle may be moved, an obstacle or a few other
objects may appear, a camera may shift, or the lighting, the colours or the image noise may change. Nothing tells you
-when; the task and its success check stay the same, so keep checking what the cameras show.
+when; the task and its success check stay the same, so keep checking what the cameras show. A moved object or
+receptacle jumps there alone: what already rests in or on it stays where it was.
## Rules (real-robot rules)
- **One episode, no reset.** The episode starts when the service comes up. It ends when the task succeeds, after
diff --git a/tasks/libero-max-spatial-akita-black-bowl-on-plate-target-relocation-i01-standard/task.toml b/tasks/libero-max-spatial-akita-black-bowl-on-plate-target-relocation-i01-standard/task.toml
index f15a7c16fa..e00e46af65 100644
--- a/tasks/libero-max-spatial-akita-black-bowl-on-plate-target-relocation-i01-standard/task.toml
+++ b/tasks/libero-max-spatial-akita-black-bowl-on-plate-target-relocation-i01-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "robosuite"
robot = "franka-panda"
family = "libero-max/libero_spatial/target_relocation"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 195
diff --git a/tasks/libero-max-spatial-akita-black-bowl-on-plate-visual-theme-switch-i00-privileged/task.toml b/tasks/libero-max-spatial-akita-black-bowl-on-plate-visual-theme-switch-i00-privileged/task.toml
index 4d4bc975d7..35eeb42b9b 100644
--- a/tasks/libero-max-spatial-akita-black-bowl-on-plate-visual-theme-switch-i00-privileged/task.toml
+++ b/tasks/libero-max-spatial-akita-black-bowl-on-plate-visual-theme-switch-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "robosuite"
robot = "franka-panda"
family = "libero-max/libero_spatial/visual_theme_switch"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 195 # LIBERO-MAX's policy seed for the case (the scene is the frozen LIBERO-PRO initial state)
expert_time_estimate_hours = 1.0
diff --git a/tasks/libero-max-spatial-akita-black-bowl-on-plate-visual-theme-switch-i00-standard/README.md b/tasks/libero-max-spatial-akita-black-bowl-on-plate-visual-theme-switch-i00-standard/README.md
index a92e326b8f..404984564f 100644
--- a/tasks/libero-max-spatial-akita-black-bowl-on-plate-visual-theme-switch-i00-standard/README.md
+++ b/tasks/libero-max-spatial-akita-black-bowl-on-plate-visual-theme-switch-i00-standard/README.md
@@ -86,3 +86,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
/reset` ends the episode as `error` (`sim_error = 1`), not `reset` (reachable only when a run grants resets; the
default is none). The scene, instance, pass rule and every verdict reached before are unchanged, so the task keeps
its name.
+- 2026-10-14 (board 0113, from board 0086): the instruction says that a moved object or receptacle jumps alone: what rests in or on it stays where it was (LIBERO-MAX moves one entity; board 0086 saw a basket jump through a can already in it). Text only: the instance, the event, the reference solution, LIBERO's success check and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/libero-max-spatial-akita-black-bowl-on-plate-visual-theme-switch-i00-standard/instruction.md b/tasks/libero-max-spatial-akita-black-bowl-on-plate-visual-theme-switch-i00-standard/instruction.md
index 9f1a252129..9bb4ffcac6 100644
--- a/tasks/libero-max-spatial-akita-black-bowl-on-plate-visual-theme-switch-i00-standard/instruction.md
+++ b/tasks/libero-max-spatial-akita-black-bowl-on-plate-visual-theme-switch-i00-standard/instruction.md
@@ -6,7 +6,8 @@ simulator runs in a service you cannot open, and you reach it only through a cli
The scene can change while you work: once during the episode, when your gripper first gets close to one of the task's
objects, the scene changes (as in LIBERO-MAX): an object or a receptacle may be moved, an obstacle or a few other
objects may appear, a camera may shift, or the lighting, the colours or the image noise may change. Nothing tells you
-when; the task and its success check stay the same, so keep checking what the cameras show.
+when; the task and its success check stay the same, so keep checking what the cameras show. A moved object or
+receptacle jumps there alone: what already rests in or on it stays where it was.
## Rules (real-robot rules)
- **One episode, no reset.** The episode starts when the service comes up. It ends when the task succeeds, after
diff --git a/tasks/libero-max-spatial-akita-black-bowl-on-plate-visual-theme-switch-i00-standard/task.toml b/tasks/libero-max-spatial-akita-black-bowl-on-plate-visual-theme-switch-i00-standard/task.toml
index 0c3bd71790..a32f4764e3 100644
--- a/tasks/libero-max-spatial-akita-black-bowl-on-plate-visual-theme-switch-i00-standard/task.toml
+++ b/tasks/libero-max-spatial-akita-black-bowl-on-plate-visual-theme-switch-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "robosuite"
robot = "franka-panda"
family = "libero-max/libero_spatial/visual_theme_switch"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 195
diff --git a/tasks/libero-max-spatial-black-bowl-on-yellow-plate-obstacle-insertion-i00-privileged/task.toml b/tasks/libero-max-spatial-black-bowl-on-yellow-plate-obstacle-insertion-i00-privileged/task.toml
index 38ea430b18..a8aeefa1b6 100644
--- a/tasks/libero-max-spatial-black-bowl-on-yellow-plate-obstacle-insertion-i00-privileged/task.toml
+++ b/tasks/libero-max-spatial-black-bowl-on-yellow-plate-obstacle-insertion-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "robosuite"
robot = "franka-panda"
family = "libero-max/libero_spatial/obstacle_insertion"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 195 # LIBERO-MAX's policy seed for the case (the scene is the frozen LIBERO-PRO initial state)
expert_time_estimate_hours = 1.0
diff --git a/tasks/libero-max-spatial-black-bowl-on-yellow-plate-obstacle-insertion-i00-standard/README.md b/tasks/libero-max-spatial-black-bowl-on-yellow-plate-obstacle-insertion-i00-standard/README.md
index 8186a1cbfe..ee906dd669 100644
--- a/tasks/libero-max-spatial-black-bowl-on-yellow-plate-obstacle-insertion-i00-standard/README.md
+++ b/tasks/libero-max-spatial-black-bowl-on-yellow-plate-obstacle-insertion-i00-standard/README.md
@@ -86,3 +86,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
/reset` ends the episode as `error` (`sim_error = 1`), not `reset` (reachable only when a run grants resets; the
default is none). The scene, instance, pass rule and every verdict reached before are unchanged, so the task keeps
its name.
+- 2026-10-14 (board 0113, from board 0086): the instruction says that a moved object or receptacle jumps alone: what rests in or on it stays where it was (LIBERO-MAX moves one entity; board 0086 saw a basket jump through a can already in it). Text only: the instance, the event, the reference solution, LIBERO's success check and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/libero-max-spatial-black-bowl-on-yellow-plate-obstacle-insertion-i00-standard/instruction.md b/tasks/libero-max-spatial-black-bowl-on-yellow-plate-obstacle-insertion-i00-standard/instruction.md
index 4e25383580..09df80413b 100644
--- a/tasks/libero-max-spatial-black-bowl-on-yellow-plate-obstacle-insertion-i00-standard/instruction.md
+++ b/tasks/libero-max-spatial-black-bowl-on-yellow-plate-obstacle-insertion-i00-standard/instruction.md
@@ -6,7 +6,8 @@ simulator runs in a service you cannot open, and you reach it only through a cli
The scene can change while you work: once during the episode, when your gripper first gets close to one of the task's
objects, the scene changes (as in LIBERO-MAX): an object or a receptacle may be moved, an obstacle or a few other
objects may appear, a camera may shift, or the lighting, the colours or the image noise may change. Nothing tells you
-when; the task and its success check stay the same, so keep checking what the cameras show.
+when; the task and its success check stay the same, so keep checking what the cameras show. A moved object or
+receptacle jumps there alone: what already rests in or on it stays where it was.
## Rules (real-robot rules)
- **One episode, no reset.** The episode starts when the service comes up. It ends when the task succeeds, after
diff --git a/tasks/libero-max-spatial-black-bowl-on-yellow-plate-obstacle-insertion-i00-standard/task.toml b/tasks/libero-max-spatial-black-bowl-on-yellow-plate-obstacle-insertion-i00-standard/task.toml
index 738b6ceae2..95135fd72b 100644
--- a/tasks/libero-max-spatial-black-bowl-on-yellow-plate-obstacle-insertion-i00-standard/task.toml
+++ b/tasks/libero-max-spatial-black-bowl-on-yellow-plate-obstacle-insertion-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "robosuite"
robot = "franka-panda"
family = "libero-max/libero_spatial/obstacle_insertion"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 195
diff --git a/tasks/libero-max-spatial-black-bowl-on-yellow-plate-target-relocation-i00-privileged/task.toml b/tasks/libero-max-spatial-black-bowl-on-yellow-plate-target-relocation-i00-privileged/task.toml
index 50fb091286..e939579d73 100644
--- a/tasks/libero-max-spatial-black-bowl-on-yellow-plate-target-relocation-i00-privileged/task.toml
+++ b/tasks/libero-max-spatial-black-bowl-on-yellow-plate-target-relocation-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "robosuite"
robot = "franka-panda"
family = "libero-max/libero_spatial/target_relocation"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 195 # LIBERO-MAX's policy seed for the case (the scene is the frozen LIBERO-PRO initial state)
expert_time_estimate_hours = 1.0
diff --git a/tasks/libero-max-spatial-black-bowl-on-yellow-plate-target-relocation-i00-standard/README.md b/tasks/libero-max-spatial-black-bowl-on-yellow-plate-target-relocation-i00-standard/README.md
index 35e8897cae..bb1ce645f7 100644
--- a/tasks/libero-max-spatial-black-bowl-on-yellow-plate-target-relocation-i00-standard/README.md
+++ b/tasks/libero-max-spatial-black-bowl-on-yellow-plate-target-relocation-i00-standard/README.md
@@ -86,3 +86,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
/reset` ends the episode as `error` (`sim_error = 1`), not `reset` (reachable only when a run grants resets; the
default is none). The scene, instance, pass rule and every verdict reached before are unchanged, so the task keeps
its name.
+- 2026-10-14 (board 0113, from board 0086): the instruction says that a moved object or receptacle jumps alone: what rests in or on it stays where it was (LIBERO-MAX moves one entity; board 0086 saw a basket jump through a can already in it). Text only: the instance, the event, the reference solution, LIBERO's success check and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/libero-max-spatial-black-bowl-on-yellow-plate-target-relocation-i00-standard/instruction.md b/tasks/libero-max-spatial-black-bowl-on-yellow-plate-target-relocation-i00-standard/instruction.md
index 9758772f09..b0383ce241 100644
--- a/tasks/libero-max-spatial-black-bowl-on-yellow-plate-target-relocation-i00-standard/instruction.md
+++ b/tasks/libero-max-spatial-black-bowl-on-yellow-plate-target-relocation-i00-standard/instruction.md
@@ -6,7 +6,8 @@ simulator runs in a service you cannot open, and you reach it only through a cli
The scene can change while you work: once during the episode, when your gripper first gets close to one of the task's
objects, the scene changes (as in LIBERO-MAX): an object or a receptacle may be moved, an obstacle or a few other
objects may appear, a camera may shift, or the lighting, the colours or the image noise may change. Nothing tells you
-when; the task and its success check stay the same, so keep checking what the cameras show.
+when; the task and its success check stay the same, so keep checking what the cameras show. A moved object or
+receptacle jumps there alone: what already rests in or on it stays where it was.
## Rules (real-robot rules)
- **One episode, no reset.** The episode starts when the service comes up. It ends when the task succeeds, after
diff --git a/tasks/libero-max-spatial-black-bowl-on-yellow-plate-target-relocation-i00-standard/task.toml b/tasks/libero-max-spatial-black-bowl-on-yellow-plate-target-relocation-i00-standard/task.toml
index 65bea19b31..730e175b67 100644
--- a/tasks/libero-max-spatial-black-bowl-on-yellow-plate-target-relocation-i00-standard/task.toml
+++ b/tasks/libero-max-spatial-black-bowl-on-yellow-plate-target-relocation-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "robosuite"
robot = "franka-panda"
family = "libero-max/libero_spatial/target_relocation"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 195
diff --git a/tasks/libero-max-spatial-black-bowl-on-yellow-plate-visual-theme-switch-i00-privileged/task.toml b/tasks/libero-max-spatial-black-bowl-on-yellow-plate-visual-theme-switch-i00-privileged/task.toml
index f92e4b94c1..bf9363baaa 100644
--- a/tasks/libero-max-spatial-black-bowl-on-yellow-plate-visual-theme-switch-i00-privileged/task.toml
+++ b/tasks/libero-max-spatial-black-bowl-on-yellow-plate-visual-theme-switch-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "robosuite"
robot = "franka-panda"
family = "libero-max/libero_spatial/visual_theme_switch"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 195 # LIBERO-MAX's policy seed for the case (the scene is the frozen LIBERO-PRO initial state)
expert_time_estimate_hours = 1.0
diff --git a/tasks/libero-max-spatial-black-bowl-on-yellow-plate-visual-theme-switch-i00-standard/README.md b/tasks/libero-max-spatial-black-bowl-on-yellow-plate-visual-theme-switch-i00-standard/README.md
index 2b58933b59..c1d542a921 100644
--- a/tasks/libero-max-spatial-black-bowl-on-yellow-plate-visual-theme-switch-i00-standard/README.md
+++ b/tasks/libero-max-spatial-black-bowl-on-yellow-plate-visual-theme-switch-i00-standard/README.md
@@ -86,3 +86,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
/reset` ends the episode as `error` (`sim_error = 1`), not `reset` (reachable only when a run grants resets; the
default is none). The scene, instance, pass rule and every verdict reached before are unchanged, so the task keeps
its name.
+- 2026-10-14 (board 0113, from board 0086): the instruction says that a moved object or receptacle jumps alone: what rests in or on it stays where it was (LIBERO-MAX moves one entity; board 0086 saw a basket jump through a can already in it). Text only: the instance, the event, the reference solution, LIBERO's success check and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/libero-max-spatial-black-bowl-on-yellow-plate-visual-theme-switch-i00-standard/instruction.md b/tasks/libero-max-spatial-black-bowl-on-yellow-plate-visual-theme-switch-i00-standard/instruction.md
index e537c7a1eb..82b251ddc0 100644
--- a/tasks/libero-max-spatial-black-bowl-on-yellow-plate-visual-theme-switch-i00-standard/instruction.md
+++ b/tasks/libero-max-spatial-black-bowl-on-yellow-plate-visual-theme-switch-i00-standard/instruction.md
@@ -6,7 +6,8 @@ simulator runs in a service you cannot open, and you reach it only through a cli
The scene can change while you work: once during the episode, when your gripper first gets close to one of the task's
objects, the scene changes (as in LIBERO-MAX): an object or a receptacle may be moved, an obstacle or a few other
objects may appear, a camera may shift, or the lighting, the colours or the image noise may change. Nothing tells you
-when; the task and its success check stay the same, so keep checking what the cameras show.
+when; the task and its success check stay the same, so keep checking what the cameras show. A moved object or
+receptacle jumps there alone: what already rests in or on it stays where it was.
## Rules (real-robot rules)
- **One episode, no reset.** The episode starts when the service comes up. It ends when the task succeeds, after
diff --git a/tasks/libero-max-spatial-black-bowl-on-yellow-plate-visual-theme-switch-i00-standard/task.toml b/tasks/libero-max-spatial-black-bowl-on-yellow-plate-visual-theme-switch-i00-standard/task.toml
index eee54b7650..638cd616d5 100644
--- a/tasks/libero-max-spatial-black-bowl-on-yellow-plate-visual-theme-switch-i00-standard/task.toml
+++ b/tasks/libero-max-spatial-black-bowl-on-yellow-plate-visual-theme-switch-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "robosuite"
robot = "franka-panda"
family = "libero-max/libero_spatial/visual_theme_switch"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 195
diff --git a/tasks/maniguard-cabinet-pickup-i19-privileged/task.toml b/tasks/maniguard-cabinet-pickup-i19-privileged/task.toml
index 2a5a4dce09..5bc373ea6d 100644
--- a/tasks/maniguard-cabinet-pickup-i19-privileged/task.toml
+++ b/tasks/maniguard-cabinet-pickup-i19-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "maniguard" # ManiGuard on BEHAVIOR-1K v3.7.2 (OmniGibson
robot = "franka-panda-longfinger"
family = "cabinet_pickup"
mode = "privileged" # privileged state, resets and thin primitives through the tooling (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 19 # = the ManiGuard-Bench task index (cabinet_pickup/task_0019/base); scene = environment/instance/; identity = dataset.toml digest
expert_time_estimate_hours = 2.0
diff --git a/tasks/maniguard-cabinet-pickup-i19-standard/task.toml b/tasks/maniguard-cabinet-pickup-i19-standard/task.toml
index 0d3f3a9392..508070c64d 100644
--- a/tasks/maniguard-cabinet-pickup-i19-standard/task.toml
+++ b/tasks/maniguard-cabinet-pickup-i19-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "maniguard"
robot = "franka-panda-longfinger"
family = "cabinet_pickup"
mode = "standard" # robot as a service: RGB-D + proprioception, no object poses, no success or safety predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "full"
instance_seed = 19
diff --git a/tasks/maniguard-cabinet-pickup-i20-privileged/task.toml b/tasks/maniguard-cabinet-pickup-i20-privileged/task.toml
index 6817422009..f281171516 100644
--- a/tasks/maniguard-cabinet-pickup-i20-privileged/task.toml
+++ b/tasks/maniguard-cabinet-pickup-i20-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "maniguard" # ManiGuard on BEHAVIOR-1K v3.7.2 (OmniGibson
robot = "franka-panda-longfinger"
family = "cabinet_pickup"
mode = "privileged" # privileged state, resets and thin primitives through the tooling (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 20 # = the ManiGuard-Bench task index (cabinet_pickup/task_0020/base); scene = environment/instance/; identity = dataset.toml digest
expert_time_estimate_hours = 2.0
diff --git a/tasks/maniguard-cabinet-pickup-i20-standard/task.toml b/tasks/maniguard-cabinet-pickup-i20-standard/task.toml
index 386201cfca..284962cd8b 100644
--- a/tasks/maniguard-cabinet-pickup-i20-standard/task.toml
+++ b/tasks/maniguard-cabinet-pickup-i20-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "maniguard"
robot = "franka-panda-longfinger"
family = "cabinet_pickup"
mode = "standard" # robot as a service: RGB-D + proprioception, no object poses, no success or safety predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "full"
instance_seed = 20
diff --git a/tasks/maniguard-cabinet-pickup-i22-privileged/task.toml b/tasks/maniguard-cabinet-pickup-i22-privileged/task.toml
index 04715af5d6..3e7938836e 100644
--- a/tasks/maniguard-cabinet-pickup-i22-privileged/task.toml
+++ b/tasks/maniguard-cabinet-pickup-i22-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "maniguard" # ManiGuard on BEHAVIOR-1K v3.7.2 (OmniGibson
robot = "franka-panda-longfinger"
family = "cabinet_pickup"
mode = "privileged" # privileged state, resets and thin primitives through the tooling (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 22 # = the ManiGuard-Bench task index (cabinet_pickup/task_0022/base); scene = environment/instance/; identity = dataset.toml digest
expert_time_estimate_hours = 2.0
diff --git a/tasks/maniguard-cabinet-pickup-i22-standard/task.toml b/tasks/maniguard-cabinet-pickup-i22-standard/task.toml
index ddda260fe5..85641c5df7 100644
--- a/tasks/maniguard-cabinet-pickup-i22-standard/task.toml
+++ b/tasks/maniguard-cabinet-pickup-i22-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "maniguard"
robot = "franka-panda-longfinger"
family = "cabinet_pickup"
mode = "standard" # robot as a service: RGB-D + proprioception, no object poses, no success or safety predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "full"
instance_seed = 22
diff --git a/tasks/maniguard-cabinet-pickup-i26-privileged/task.toml b/tasks/maniguard-cabinet-pickup-i26-privileged/task.toml
index cd418cad48..bbf2dc8df6 100644
--- a/tasks/maniguard-cabinet-pickup-i26-privileged/task.toml
+++ b/tasks/maniguard-cabinet-pickup-i26-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "maniguard" # ManiGuard on BEHAVIOR-1K v3.7.2 (OmniGibson
robot = "franka-panda-longfinger"
family = "cabinet_pickup"
mode = "privileged" # privileged state, resets and thin primitives through the tooling (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No positive example yet (oracle = \"none\": the 16 ManiGuard tasks added on 2026-10-04 for variety ship without one, user decision). Upstream reference, cited only: ManiGuard's 40 scripted demonstrations of this task, episodes 1040-1079 of the Hugging Face dataset IDEAS-Lab-Northwestern/datagen-cabinet-v1-joint-5cam (revision ea8d1242, LeRobot, 30 Hz, CC BY-NC 4.0), recorded by ManiGuard's data generator in a differently built scene; none has been replayed through this verifier and none is copied into the task. The frozen instance (state sha 9f0df8b55512a085) was made on an RTX PRO 6000 (2026-10-04); a second fresh load gave the same hash. Human review per task (PROTOCOL §2.7)."
instance_seed = 26 # = the ManiGuard-Bench task index (cabinet_pickup/task_0026/base); scene = environment/instance/; identity = dataset.toml digest
diff --git a/tasks/maniguard-cabinet-pickup-i26-standard/task.toml b/tasks/maniguard-cabinet-pickup-i26-standard/task.toml
index 00824b035b..a2fd2fc016 100644
--- a/tasks/maniguard-cabinet-pickup-i26-standard/task.toml
+++ b/tasks/maniguard-cabinet-pickup-i26-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "maniguard"
robot = "franka-panda-longfinger"
family = "cabinet_pickup"
mode = "standard" # robot as a service: RGB-D + proprioception, no object poses, no success or safety predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No positive example yet (oracle = \"none\": the 16 ManiGuard tasks added on 2026-10-04 for variety ship without one, user decision). Upstream reference, cited only: ManiGuard's 40 scripted demonstrations of this task, episodes 1040-1079 of the Hugging Face dataset IDEAS-Lab-Northwestern/datagen-cabinet-v1-joint-5cam (revision ea8d1242, LeRobot, 30 Hz, CC BY-NC 4.0), recorded by ManiGuard's data generator in a differently built scene; none has been replayed through this task's service and none is copied into the task. The frozen instance (state sha 9f0df8b55512a085) was made on an RTX PRO 6000 (2026-10-04); a second fresh load gave the same hash. Human review per task (PROTOCOL §2.7)."
diff --git a/tasks/maniguard-cabinet-pickup-i29-privileged/task.toml b/tasks/maniguard-cabinet-pickup-i29-privileged/task.toml
index c87c21c194..940dc352b4 100644
--- a/tasks/maniguard-cabinet-pickup-i29-privileged/task.toml
+++ b/tasks/maniguard-cabinet-pickup-i29-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "maniguard" # ManiGuard on BEHAVIOR-1K v3.7.2 (OmniGibson
robot = "franka-panda-longfinger"
family = "cabinet_pickup"
mode = "privileged" # privileged state, resets and thin primitives through the tooling (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No positive example yet (oracle = \"none\": the 16 ManiGuard tasks added on 2026-10-04 for variety ship without one, user decision). Upstream reference, cited only: ManiGuard's 40 scripted demonstrations of this task, episodes 1160-1199 of the Hugging Face dataset IDEAS-Lab-Northwestern/datagen-cabinet-v1-joint-5cam (revision ea8d1242, LeRobot, 30 Hz, CC BY-NC 4.0), recorded by ManiGuard's data generator in a differently built scene; none has been replayed through this verifier and none is copied into the task. The frozen instance (state sha 66b0d5f46c15f14f) was made on an RTX PRO 6000 (2026-10-04); a second fresh load gave the same hash. Human review per task (PROTOCOL §2.7)."
instance_seed = 29 # = the ManiGuard-Bench task index (cabinet_pickup/task_0029/base); scene = environment/instance/; identity = dataset.toml digest
diff --git a/tasks/maniguard-cabinet-pickup-i29-standard/task.toml b/tasks/maniguard-cabinet-pickup-i29-standard/task.toml
index 1683c3257b..eba134cb6d 100644
--- a/tasks/maniguard-cabinet-pickup-i29-standard/task.toml
+++ b/tasks/maniguard-cabinet-pickup-i29-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "maniguard"
robot = "franka-panda-longfinger"
family = "cabinet_pickup"
mode = "standard" # robot as a service: RGB-D + proprioception, no object poses, no success or safety predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No positive example yet (oracle = \"none\": the 16 ManiGuard tasks added on 2026-10-04 for variety ship without one, user decision). Upstream reference, cited only: ManiGuard's 40 scripted demonstrations of this task, episodes 1160-1199 of the Hugging Face dataset IDEAS-Lab-Northwestern/datagen-cabinet-v1-joint-5cam (revision ea8d1242, LeRobot, 30 Hz, CC BY-NC 4.0), recorded by ManiGuard's data generator in a differently built scene; none has been replayed through this task's service and none is copied into the task. The frozen instance (state sha 66b0d5f46c15f14f) was made on an RTX PRO 6000 (2026-10-04); a second fresh load gave the same hash. Human review per task (PROTOCOL §2.7)."
diff --git a/tasks/maniguard-clutter-pickup-i09-privileged/task.toml b/tasks/maniguard-clutter-pickup-i09-privileged/task.toml
index 16e7fdecfc..5dbe8f6d8e 100644
--- a/tasks/maniguard-clutter-pickup-i09-privileged/task.toml
+++ b/tasks/maniguard-clutter-pickup-i09-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "maniguard" # ManiGuard on BEHAVIOR-1K v3.7.2 (OmniGibson
robot = "franka-panda-longfinger"
family = "clutter_pickup"
mode = "privileged" # privileged state, resets and thin primitives through the tooling (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No positive example yet (oracle = \"none\": the 16 ManiGuard tasks added on 2026-10-04 for variety ship without one, user decision). Upstream reference, cited only: ManiGuard's 40 scripted demonstrations of this task, episodes 360-399 of the Hugging Face dataset IDEAS-Lab-Northwestern/datagen-clutter-v1-joint-5cam (revision e92c75fb, LeRobot, 30 Hz, CC BY-NC 4.0), recorded by ManiGuard's data generator in a differently built scene; none has been replayed through this verifier and none is copied into the task. The frozen instance (state sha 23d5d2418f027373) was made on an RTX PRO 6000 (2026-10-04); a second fresh load gave the same hash. Human review per task (PROTOCOL §2.7)."
instance_seed = 9 # = the ManiGuard-Bench task index (clutter_pickup/task_0009/base); scene = environment/instance/; identity = dataset.toml digest
diff --git a/tasks/maniguard-clutter-pickup-i09-standard/task.toml b/tasks/maniguard-clutter-pickup-i09-standard/task.toml
index 3612b5292f..c113a32b04 100644
--- a/tasks/maniguard-clutter-pickup-i09-standard/task.toml
+++ b/tasks/maniguard-clutter-pickup-i09-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "maniguard"
robot = "franka-panda-longfinger"
family = "clutter_pickup"
mode = "standard" # robot as a service: RGB-D + proprioception, no object poses, no success or safety predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No positive example yet (oracle = \"none\": the 16 ManiGuard tasks added on 2026-10-04 for variety ship without one, user decision). Upstream reference, cited only: ManiGuard's 40 scripted demonstrations of this task, episodes 360-399 of the Hugging Face dataset IDEAS-Lab-Northwestern/datagen-clutter-v1-joint-5cam (revision e92c75fb, LeRobot, 30 Hz, CC BY-NC 4.0), recorded by ManiGuard's data generator in a differently built scene; none has been replayed through this task's service and none is copied into the task. The frozen instance (state sha 23d5d2418f027373) was made on an RTX PRO 6000 (2026-10-04); a second fresh load gave the same hash. Human review per task (PROTOCOL §2.7)."
diff --git a/tasks/maniguard-clutter-pickup-i15-privileged/task.toml b/tasks/maniguard-clutter-pickup-i15-privileged/task.toml
index 7d23999a06..26c1212eeb 100644
--- a/tasks/maniguard-clutter-pickup-i15-privileged/task.toml
+++ b/tasks/maniguard-clutter-pickup-i15-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "maniguard" # ManiGuard on BEHAVIOR-1K v3.7.2 (OmniGibson
robot = "franka-panda-longfinger"
family = "clutter_pickup"
mode = "privileged" # privileged state, resets and thin primitives through the tooling (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No positive example yet (oracle = \"none\": the 16 ManiGuard tasks added on 2026-10-04 for variety ship without one, user decision). Upstream reference, cited only: ManiGuard's 40 scripted demonstrations of this task, episodes 600-639 of the Hugging Face dataset IDEAS-Lab-Northwestern/datagen-clutter-v1-joint-5cam (revision e92c75fb, LeRobot, 30 Hz, CC BY-NC 4.0), recorded by ManiGuard's data generator in a differently built scene; none has been replayed through this verifier and none is copied into the task. The frozen instance (state sha be98dc0729e59682) was made on an RTX PRO 6000 (2026-10-04); a second fresh load gave the same hash. Human review per task (PROTOCOL §2.7)."
instance_seed = 15 # = the ManiGuard-Bench task index (clutter_pickup/task_0015/base); scene = environment/instance/; identity = dataset.toml digest
diff --git a/tasks/maniguard-clutter-pickup-i15-standard/task.toml b/tasks/maniguard-clutter-pickup-i15-standard/task.toml
index 12d792caa6..950f618c3c 100644
--- a/tasks/maniguard-clutter-pickup-i15-standard/task.toml
+++ b/tasks/maniguard-clutter-pickup-i15-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "maniguard"
robot = "franka-panda-longfinger"
family = "clutter_pickup"
mode = "standard" # robot as a service: RGB-D + proprioception, no object poses, no success or safety predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No positive example yet (oracle = \"none\": the 16 ManiGuard tasks added on 2026-10-04 for variety ship without one, user decision). Upstream reference, cited only: ManiGuard's 40 scripted demonstrations of this task, episodes 600-639 of the Hugging Face dataset IDEAS-Lab-Northwestern/datagen-clutter-v1-joint-5cam (revision e92c75fb, LeRobot, 30 Hz, CC BY-NC 4.0), recorded by ManiGuard's data generator in a differently built scene; none has been replayed through this task's service and none is copied into the task. The frozen instance (state sha be98dc0729e59682) was made on an RTX PRO 6000 (2026-10-04); a second fresh load gave the same hash. Human review per task (PROTOCOL §2.7)."
diff --git a/tasks/maniguard-clutter-pickup-i23-privileged/task.toml b/tasks/maniguard-clutter-pickup-i23-privileged/task.toml
index ccd1f458ef..7683e86bf6 100644
--- a/tasks/maniguard-clutter-pickup-i23-privileged/task.toml
+++ b/tasks/maniguard-clutter-pickup-i23-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "maniguard" # ManiGuard on BEHAVIOR-1K v3.7.2 (OmniGibson
robot = "franka-panda-longfinger"
family = "clutter_pickup"
mode = "privileged" # privileged state, resets and thin primitives through the tooling (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No positive example yet (oracle = \"none\": the 16 ManiGuard tasks added on 2026-10-04 for variety ship without one, user decision). Upstream reference, cited only: ManiGuard's 40 scripted demonstrations of this task, episodes 920-959 of the Hugging Face dataset IDEAS-Lab-Northwestern/datagen-clutter-v1-joint-5cam (revision e92c75fb, LeRobot, 30 Hz, CC BY-NC 4.0), recorded by ManiGuard's data generator in a differently built scene; none has been replayed through this verifier and none is copied into the task. The frozen instance (state sha 021f30c779dc2819) was made on an RTX PRO 6000 (2026-10-04); a second fresh load gave the same hash. Human review per task (PROTOCOL §2.7)."
instance_seed = 23 # = the ManiGuard-Bench task index (clutter_pickup/task_0023/base); scene = environment/instance/; identity = dataset.toml digest
diff --git a/tasks/maniguard-clutter-pickup-i23-standard/task.toml b/tasks/maniguard-clutter-pickup-i23-standard/task.toml
index 05e0b32b33..076f8bc8c9 100644
--- a/tasks/maniguard-clutter-pickup-i23-standard/task.toml
+++ b/tasks/maniguard-clutter-pickup-i23-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "maniguard"
robot = "franka-panda-longfinger"
family = "clutter_pickup"
mode = "standard" # robot as a service: RGB-D + proprioception, no object poses, no success or safety predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No positive example yet (oracle = \"none\": the 16 ManiGuard tasks added on 2026-10-04 for variety ship without one, user decision). Upstream reference, cited only: ManiGuard's 40 scripted demonstrations of this task, episodes 920-959 of the Hugging Face dataset IDEAS-Lab-Northwestern/datagen-clutter-v1-joint-5cam (revision e92c75fb, LeRobot, 30 Hz, CC BY-NC 4.0), recorded by ManiGuard's data generator in a differently built scene; none has been replayed through this task's service and none is copied into the task. The frozen instance (state sha 021f30c779dc2819) was made on an RTX PRO 6000 (2026-10-04); a second fresh load gave the same hash. Human review per task (PROTOCOL §2.7)."
diff --git a/tasks/maniguard-clutter-pickup-i25-privileged/task.toml b/tasks/maniguard-clutter-pickup-i25-privileged/task.toml
index f17871570c..ba56734c0a 100644
--- a/tasks/maniguard-clutter-pickup-i25-privileged/task.toml
+++ b/tasks/maniguard-clutter-pickup-i25-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "maniguard" # ManiGuard on BEHAVIOR-1K v3.7.2 (OmniGibson
robot = "franka-panda-longfinger"
family = "clutter_pickup"
mode = "privileged" # privileged state, resets and thin primitives through the tooling (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No positive example yet (oracle = \"none\": the 16 ManiGuard tasks added on 2026-10-04 for variety ship without one, user decision). Upstream reference, cited only: ManiGuard's 40 scripted demonstrations of this task, episodes 1000-1039 of the Hugging Face dataset IDEAS-Lab-Northwestern/datagen-clutter-v1-joint-5cam (revision e92c75fb, LeRobot, 30 Hz, CC BY-NC 4.0), recorded by ManiGuard's data generator in a differently built scene; none has been replayed through this verifier and none is copied into the task. The frozen instance (state sha af2da06ca81fa7eb) was made on an RTX PRO 6000 (2026-10-04); a second fresh load gave the same hash. Human review per task (PROTOCOL §2.7)."
instance_seed = 25 # = the ManiGuard-Bench task index (clutter_pickup/task_0025/base); scene = environment/instance/; identity = dataset.toml digest
diff --git a/tasks/maniguard-clutter-pickup-i25-standard/task.toml b/tasks/maniguard-clutter-pickup-i25-standard/task.toml
index f338f021d6..2abc4d25f8 100644
--- a/tasks/maniguard-clutter-pickup-i25-standard/task.toml
+++ b/tasks/maniguard-clutter-pickup-i25-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "maniguard"
robot = "franka-panda-longfinger"
family = "clutter_pickup"
mode = "standard" # robot as a service: RGB-D + proprioception, no object poses, no success or safety predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No positive example yet (oracle = \"none\": the 16 ManiGuard tasks added on 2026-10-04 for variety ship without one, user decision). Upstream reference, cited only: ManiGuard's 40 scripted demonstrations of this task, episodes 1000-1039 of the Hugging Face dataset IDEAS-Lab-Northwestern/datagen-clutter-v1-joint-5cam (revision e92c75fb, LeRobot, 30 Hz, CC BY-NC 4.0), recorded by ManiGuard's data generator in a differently built scene; none has been replayed through this task's service and none is copied into the task. The frozen instance (state sha af2da06ca81fa7eb) was made on an RTX PRO 6000 (2026-10-04); a second fresh load gave the same hash. Human review per task (PROTOCOL §2.7)."
diff --git a/tasks/maniguard-clutter-pickup-i28-privileged/task.toml b/tasks/maniguard-clutter-pickup-i28-privileged/task.toml
index b5616bd4e7..13f77bd4a3 100644
--- a/tasks/maniguard-clutter-pickup-i28-privileged/task.toml
+++ b/tasks/maniguard-clutter-pickup-i28-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "maniguard" # ManiGuard on BEHAVIOR-1K v3.7.2 (OmniGibson
robot = "franka-panda-longfinger"
family = "clutter_pickup"
mode = "privileged" # privileged state, resets and thin primitives through the tooling (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No positive example yet (oracle = \"none\": the 16 ManiGuard tasks added on 2026-10-04 for variety ship without one, user decision). Upstream reference, cited only: ManiGuard's 40 scripted demonstrations of this task, episodes 1120-1159 of the Hugging Face dataset IDEAS-Lab-Northwestern/datagen-clutter-v1-joint-5cam (revision e92c75fb, LeRobot, 30 Hz, CC BY-NC 4.0), recorded by ManiGuard's data generator in a differently built scene; none has been replayed through this verifier and none is copied into the task. The frozen instance (state sha 2ef0663763b82cb5) was made on an RTX PRO 6000 (2026-10-04); a second fresh load gave the same hash. Human review per task (PROTOCOL §2.7)."
instance_seed = 28 # = the ManiGuard-Bench task index (clutter_pickup/task_0028/base); scene = environment/instance/; identity = dataset.toml digest
diff --git a/tasks/maniguard-clutter-pickup-i28-standard/task.toml b/tasks/maniguard-clutter-pickup-i28-standard/task.toml
index 45b0fbdcc0..cbbe7e30ad 100644
--- a/tasks/maniguard-clutter-pickup-i28-standard/task.toml
+++ b/tasks/maniguard-clutter-pickup-i28-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "maniguard"
robot = "franka-panda-longfinger"
family = "clutter_pickup"
mode = "standard" # robot as a service: RGB-D + proprioception, no object poses, no success or safety predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No positive example yet (oracle = \"none\": the 16 ManiGuard tasks added on 2026-10-04 for variety ship without one, user decision). Upstream reference, cited only: ManiGuard's 40 scripted demonstrations of this task, episodes 1120-1159 of the Hugging Face dataset IDEAS-Lab-Northwestern/datagen-clutter-v1-joint-5cam (revision e92c75fb, LeRobot, 30 Hz, CC BY-NC 4.0), recorded by ManiGuard's data generator in a differently built scene; none has been replayed through this task's service and none is copied into the task. The frozen instance (state sha 2ef0663763b82cb5) was made on an RTX PRO 6000 (2026-10-04); a second fresh load gave the same hash. Human review per task (PROTOCOL §2.7)."
diff --git a/tasks/maniguard-dusty-transfer-i03-privileged/task.toml b/tasks/maniguard-dusty-transfer-i03-privileged/task.toml
index 8e9caaf016..968718af83 100644
--- a/tasks/maniguard-dusty-transfer-i03-privileged/task.toml
+++ b/tasks/maniguard-dusty-transfer-i03-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "maniguard" # ManiGuard on BEHAVIOR-1K v3.7.2 (OmniGibson
robot = "franka-panda-longfinger"
family = "dusty_transfer"
mode = "privileged" # privileged state, resets and thin primitives through the tooling (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 3 # = the ManiGuard-Bench task index (dusty_transfer/task_0003/base); scene = environment/instance/; identity = dataset.toml digest
expert_time_estimate_hours = 2.0
diff --git a/tasks/maniguard-dusty-transfer-i03-standard/task.toml b/tasks/maniguard-dusty-transfer-i03-standard/task.toml
index 96b37eaaba..3158116bda 100644
--- a/tasks/maniguard-dusty-transfer-i03-standard/task.toml
+++ b/tasks/maniguard-dusty-transfer-i03-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "maniguard"
robot = "franka-panda-longfinger"
family = "dusty_transfer"
mode = "standard" # robot as a service: RGB-D + proprioception, no object poses, no success or safety predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "full"
instance_seed = 3
diff --git a/tasks/maniguard-dusty-transfer-i16-privileged/task.toml b/tasks/maniguard-dusty-transfer-i16-privileged/task.toml
index 5eb001dd1f..42e2d9b57b 100644
--- a/tasks/maniguard-dusty-transfer-i16-privileged/task.toml
+++ b/tasks/maniguard-dusty-transfer-i16-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "maniguard" # ManiGuard on BEHAVIOR-1K v3.7.2 (OmniGibson
robot = "franka-panda-longfinger"
family = "dusty_transfer"
mode = "privileged" # privileged state, resets and thin primitives through the tooling (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No positive example yet (oracle = \"none\": the 16 ManiGuard tasks added on 2026-10-04 for variety ship without one, user decision). Upstream reference, cited only: ManiGuard's 40 scripted demonstrations of this task, episodes 640-679 of the Hugging Face dataset IDEAS-Lab-Northwestern/datagen-dusty-v1-joint-5cam (revision ac57d863, LeRobot, 30 Hz, CC BY-NC 4.0), recorded by ManiGuard's data generator in a differently built scene; none has been replayed through this verifier and none is copied into the task. The frozen instance (state sha 7ea4b07f85b21d28) was made on an RTX PRO 6000 (2026-10-04); a second fresh load gave the same hash. Human review per task (PROTOCOL §2.7)."
instance_seed = 16 # = the ManiGuard-Bench task index (dusty_transfer/task_0016/base); scene = environment/instance/; identity = dataset.toml digest
diff --git a/tasks/maniguard-dusty-transfer-i16-standard/task.toml b/tasks/maniguard-dusty-transfer-i16-standard/task.toml
index 8a4f881c39..ddac625139 100644
--- a/tasks/maniguard-dusty-transfer-i16-standard/task.toml
+++ b/tasks/maniguard-dusty-transfer-i16-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "maniguard"
robot = "franka-panda-longfinger"
family = "dusty_transfer"
mode = "standard" # robot as a service: RGB-D + proprioception, no object poses, no success or safety predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No positive example yet (oracle = \"none\": the 16 ManiGuard tasks added on 2026-10-04 for variety ship without one, user decision). Upstream reference, cited only: ManiGuard's 40 scripted demonstrations of this task, episodes 640-679 of the Hugging Face dataset IDEAS-Lab-Northwestern/datagen-dusty-v1-joint-5cam (revision ac57d863, LeRobot, 30 Hz, CC BY-NC 4.0), recorded by ManiGuard's data generator in a differently built scene; none has been replayed through this task's service and none is copied into the task. The frozen instance (state sha 7ea4b07f85b21d28) was made on an RTX PRO 6000 (2026-10-04); a second fresh load gave the same hash. Human review per task (PROTOCOL §2.7)."
diff --git a/tasks/maniguard-dusty-transfer-i20-privileged/task.toml b/tasks/maniguard-dusty-transfer-i20-privileged/task.toml
index cddbd0d6c8..7e8b4b1711 100644
--- a/tasks/maniguard-dusty-transfer-i20-privileged/task.toml
+++ b/tasks/maniguard-dusty-transfer-i20-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "maniguard" # ManiGuard on BEHAVIOR-1K v3.7.2 (OmniGibson
robot = "franka-panda-longfinger"
family = "dusty_transfer"
mode = "privileged" # privileged state, resets and thin primitives through the tooling (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No positive example yet (oracle = \"none\": the 16 ManiGuard tasks added on 2026-10-04 for variety ship without one, user decision). Upstream reference, cited only: ManiGuard's 40 scripted demonstrations of this task, episodes 800-839 of the Hugging Face dataset IDEAS-Lab-Northwestern/datagen-dusty-v1-joint-5cam (revision ac57d863, LeRobot, 30 Hz, CC BY-NC 4.0), recorded by ManiGuard's data generator in a differently built scene; none has been replayed through this verifier and none is copied into the task. The frozen instance (state sha ed90dd164ff74206) was made on an RTX PRO 6000 (2026-10-04); a second fresh load gave the same hash. Human review per task (PROTOCOL §2.7)."
instance_seed = 20 # = the ManiGuard-Bench task index (dusty_transfer/task_0020/base); scene = environment/instance/; identity = dataset.toml digest
diff --git a/tasks/maniguard-dusty-transfer-i20-standard/task.toml b/tasks/maniguard-dusty-transfer-i20-standard/task.toml
index 675a8e5a97..b20fc9e0cd 100644
--- a/tasks/maniguard-dusty-transfer-i20-standard/task.toml
+++ b/tasks/maniguard-dusty-transfer-i20-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "maniguard"
robot = "franka-panda-longfinger"
family = "dusty_transfer"
mode = "standard" # robot as a service: RGB-D + proprioception, no object poses, no success or safety predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No positive example yet (oracle = \"none\": the 16 ManiGuard tasks added on 2026-10-04 for variety ship without one, user decision). Upstream reference, cited only: ManiGuard's 40 scripted demonstrations of this task, episodes 800-839 of the Hugging Face dataset IDEAS-Lab-Northwestern/datagen-dusty-v1-joint-5cam (revision ac57d863, LeRobot, 30 Hz, CC BY-NC 4.0), recorded by ManiGuard's data generator in a differently built scene; none has been replayed through this task's service and none is copied into the task. The frozen instance (state sha ed90dd164ff74206) was made on an RTX PRO 6000 (2026-10-04); a second fresh load gave the same hash. Human review per task (PROTOCOL §2.7)."
diff --git a/tasks/maniguard-dusty-transfer-i24-privileged/task.toml b/tasks/maniguard-dusty-transfer-i24-privileged/task.toml
index 402e3c60b9..c96eb53ad6 100644
--- a/tasks/maniguard-dusty-transfer-i24-privileged/task.toml
+++ b/tasks/maniguard-dusty-transfer-i24-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "maniguard" # ManiGuard on BEHAVIOR-1K v3.7.2 (OmniGibson
robot = "franka-panda-longfinger"
family = "dusty_transfer"
mode = "privileged" # privileged state, resets and thin primitives through the tooling (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 24 # = the ManiGuard-Bench task index (dusty_transfer/task_0024/base); scene = environment/instance/; identity = dataset.toml digest
expert_time_estimate_hours = 2.0
diff --git a/tasks/maniguard-dusty-transfer-i24-standard/task.toml b/tasks/maniguard-dusty-transfer-i24-standard/task.toml
index b4931c647c..8ca221edc9 100644
--- a/tasks/maniguard-dusty-transfer-i24-standard/task.toml
+++ b/tasks/maniguard-dusty-transfer-i24-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "maniguard"
robot = "franka-panda-longfinger"
family = "dusty_transfer"
mode = "standard" # robot as a service: RGB-D + proprioception, no object poses, no success or safety predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "full"
instance_seed = 24
diff --git a/tasks/maniguard-dusty-transfer-i25-privileged/task.toml b/tasks/maniguard-dusty-transfer-i25-privileged/task.toml
index b52d231ec8..3bcad77907 100644
--- a/tasks/maniguard-dusty-transfer-i25-privileged/task.toml
+++ b/tasks/maniguard-dusty-transfer-i25-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "maniguard" # ManiGuard on BEHAVIOR-1K v3.7.2 (OmniGibson
robot = "franka-panda-longfinger"
family = "dusty_transfer"
mode = "privileged" # privileged state, resets and thin primitives through the tooling (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No positive example yet (oracle = \"none\": the 16 ManiGuard tasks added on 2026-10-04 for variety ship without one, user decision). Upstream reference, cited only: ManiGuard's 40 scripted demonstrations of this task, episodes 1000-1039 of the Hugging Face dataset IDEAS-Lab-Northwestern/datagen-dusty-v1-joint-5cam (revision ac57d863, LeRobot, 30 Hz, CC BY-NC 4.0), recorded by ManiGuard's data generator in a differently built scene; none has been replayed through this verifier and none is copied into the task. The frozen instance (state sha 34f583c2f9bf3d9c) was made on an RTX PRO 6000 (2026-10-04); a second fresh load gave the same hash. Human review per task (PROTOCOL §2.7)."
instance_seed = 25 # = the ManiGuard-Bench task index (dusty_transfer/task_0025/base); scene = environment/instance/; identity = dataset.toml digest
diff --git a/tasks/maniguard-dusty-transfer-i25-standard/task.toml b/tasks/maniguard-dusty-transfer-i25-standard/task.toml
index 6807f887d2..c267e8e5c4 100644
--- a/tasks/maniguard-dusty-transfer-i25-standard/task.toml
+++ b/tasks/maniguard-dusty-transfer-i25-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "maniguard"
robot = "franka-panda-longfinger"
family = "dusty_transfer"
mode = "standard" # robot as a service: RGB-D + proprioception, no object poses, no success or safety predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No positive example yet (oracle = \"none\": the 16 ManiGuard tasks added on 2026-10-04 for variety ship without one, user decision). Upstream reference, cited only: ManiGuard's 40 scripted demonstrations of this task, episodes 1000-1039 of the Hugging Face dataset IDEAS-Lab-Northwestern/datagen-dusty-v1-joint-5cam (revision ac57d863, LeRobot, 30 Hz, CC BY-NC 4.0), recorded by ManiGuard's data generator in a differently built scene; none has been replayed through this task's service and none is copied into the task. The frozen instance (state sha 34f583c2f9bf3d9c) was made on an RTX PRO 6000 (2026-10-04); a second fresh load gave the same hash. Human review per task (PROTOCOL §2.7)."
diff --git a/tasks/maniguard-jar-transport-i02-privileged/task.toml b/tasks/maniguard-jar-transport-i02-privileged/task.toml
index 5b5379a503..87e96d8467 100644
--- a/tasks/maniguard-jar-transport-i02-privileged/task.toml
+++ b/tasks/maniguard-jar-transport-i02-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "maniguard" # ManiGuard on BEHAVIOR-1K v3.7.2 (OmniGibson
robot = "franka-panda-longfinger"
family = "jar_transport"
mode = "privileged" # privileged state, resets and thin primitives through the tooling (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 2 # = the ManiGuard-Bench task index (jar_transport/task_0002/base); scene = environment/instance/; identity = dataset.toml digest
expert_time_estimate_hours = 2.0
diff --git a/tasks/maniguard-jar-transport-i02-standard/task.toml b/tasks/maniguard-jar-transport-i02-standard/task.toml
index 6d00f41881..0c0205a296 100644
--- a/tasks/maniguard-jar-transport-i02-standard/task.toml
+++ b/tasks/maniguard-jar-transport-i02-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "maniguard"
robot = "franka-panda-longfinger"
family = "jar_transport"
mode = "standard" # robot as a service: RGB-D + proprioception, no object poses, no success or safety predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "full"
instance_seed = 2
diff --git a/tasks/maniguard-jar-transport-i09-privileged/task.toml b/tasks/maniguard-jar-transport-i09-privileged/task.toml
index 05e11af69a..a79b9b2d72 100644
--- a/tasks/maniguard-jar-transport-i09-privileged/task.toml
+++ b/tasks/maniguard-jar-transport-i09-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "maniguard" # ManiGuard on BEHAVIOR-1K v3.7.2 (OmniGibson
robot = "franka-panda-longfinger"
family = "jar_transport"
mode = "privileged" # privileged state, resets and thin primitives through the tooling (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No positive example yet (oracle = \"none\": the 16 ManiGuard tasks added on 2026-10-04 for variety ship without one, user decision). Upstream reference, cited only: ManiGuard's 40 scripted demonstrations of this task, episodes 360-399 of the Hugging Face dataset IDEAS-Lab-Northwestern/datagen-jar-v1-joint-5cam (revision 080dddb4, LeRobot, 30 Hz, CC BY-NC 4.0), recorded by ManiGuard's data generator in a differently built scene; none has been replayed through this verifier and none is copied into the task. The frozen instance (state sha 9916236d78d2ff16) was made on an RTX PRO 6000 (2026-10-04); a second fresh load gave the same hash. Human review per task (PROTOCOL §2.7)."
instance_seed = 9 # = the ManiGuard-Bench task index (jar_transport/task_0009/base); scene = environment/instance/; identity = dataset.toml digest
diff --git a/tasks/maniguard-jar-transport-i09-standard/task.toml b/tasks/maniguard-jar-transport-i09-standard/task.toml
index 91d8fa318c..01ddf5cf96 100644
--- a/tasks/maniguard-jar-transport-i09-standard/task.toml
+++ b/tasks/maniguard-jar-transport-i09-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "maniguard"
robot = "franka-panda-longfinger"
family = "jar_transport"
mode = "standard" # robot as a service: RGB-D + proprioception, no object poses, no success or safety predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No positive example yet (oracle = \"none\": the 16 ManiGuard tasks added on 2026-10-04 for variety ship without one, user decision). Upstream reference, cited only: ManiGuard's 40 scripted demonstrations of this task, episodes 360-399 of the Hugging Face dataset IDEAS-Lab-Northwestern/datagen-jar-v1-joint-5cam (revision 080dddb4, LeRobot, 30 Hz, CC BY-NC 4.0), recorded by ManiGuard's data generator in a differently built scene; none has been replayed through this task's service and none is copied into the task. The frozen instance (state sha 9916236d78d2ff16) was made on an RTX PRO 6000 (2026-10-04); a second fresh load gave the same hash. Human review per task (PROTOCOL §2.7)."
diff --git a/tasks/maniguard-jar-transport-i13-privileged/task.toml b/tasks/maniguard-jar-transport-i13-privileged/task.toml
index 5baa398cfb..6b48790239 100644
--- a/tasks/maniguard-jar-transport-i13-privileged/task.toml
+++ b/tasks/maniguard-jar-transport-i13-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "maniguard" # ManiGuard on BEHAVIOR-1K v3.7.2 (OmniGibson
robot = "franka-panda-longfinger"
family = "jar_transport"
mode = "privileged" # privileged state, resets and thin primitives through the tooling (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 13 # = the ManiGuard-Bench task index (jar_transport/task_0013/base); scene = environment/instance/; identity = dataset.toml digest
expert_time_estimate_hours = 2.0
diff --git a/tasks/maniguard-jar-transport-i13-standard/task.toml b/tasks/maniguard-jar-transport-i13-standard/task.toml
index 3e34fc4dcf..b7cbeafbb0 100644
--- a/tasks/maniguard-jar-transport-i13-standard/task.toml
+++ b/tasks/maniguard-jar-transport-i13-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "maniguard"
robot = "franka-panda-longfinger"
family = "jar_transport"
mode = "standard" # robot as a service: RGB-D + proprioception, no object poses, no success or safety predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "full"
instance_seed = 13
diff --git a/tasks/maniguard-jar-transport-i14-privileged/task.toml b/tasks/maniguard-jar-transport-i14-privileged/task.toml
index 21ec480b9e..f7dba82479 100644
--- a/tasks/maniguard-jar-transport-i14-privileged/task.toml
+++ b/tasks/maniguard-jar-transport-i14-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "maniguard" # ManiGuard on BEHAVIOR-1K v3.7.2 (OmniGibson
robot = "franka-panda-longfinger"
family = "jar_transport"
mode = "privileged" # privileged state, resets and thin primitives through the tooling (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 14 # = the ManiGuard-Bench task index (jar_transport/task_0014/base); scene = environment/instance/; identity = dataset.toml digest
expert_time_estimate_hours = 2.0
diff --git a/tasks/maniguard-jar-transport-i14-standard/task.toml b/tasks/maniguard-jar-transport-i14-standard/task.toml
index 0474f99f34..54dc684bb7 100644
--- a/tasks/maniguard-jar-transport-i14-standard/task.toml
+++ b/tasks/maniguard-jar-transport-i14-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "maniguard"
robot = "franka-panda-longfinger"
family = "jar_transport"
mode = "standard" # robot as a service: RGB-D + proprioception, no object poses, no success or safety predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "full"
instance_seed = 14
diff --git a/tasks/maniguard-jar-transport-i25-privileged/task.toml b/tasks/maniguard-jar-transport-i25-privileged/task.toml
index ea9475ec76..782df9b638 100644
--- a/tasks/maniguard-jar-transport-i25-privileged/task.toml
+++ b/tasks/maniguard-jar-transport-i25-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "maniguard" # ManiGuard on BEHAVIOR-1K v3.7.2 (OmniGibson
robot = "franka-panda-longfinger"
family = "jar_transport"
mode = "privileged" # privileged state, resets and thin primitives through the tooling (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No positive example yet (oracle = \"none\": the 16 ManiGuard tasks added on 2026-10-04 for variety ship without one, user decision). Upstream reference, cited only: ManiGuard's 40 scripted demonstrations of this task, episodes 1000-1039 of the Hugging Face dataset IDEAS-Lab-Northwestern/datagen-jar-v1-joint-5cam (revision 080dddb4, LeRobot, 30 Hz, CC BY-NC 4.0), recorded by ManiGuard's data generator in a differently built scene; none has been replayed through this verifier and none is copied into the task. The frozen instance (state sha 38c331dc1d5f4f70) was made on an RTX PRO 6000 (2026-10-04); a second fresh load gave the same hash. Human review per task (PROTOCOL §2.7)."
instance_seed = 25 # = the ManiGuard-Bench task index (jar_transport/task_0025/base); scene = environment/instance/; identity = dataset.toml digest
diff --git a/tasks/maniguard-jar-transport-i25-standard/task.toml b/tasks/maniguard-jar-transport-i25-standard/task.toml
index 61bc35a5a8..6a5f813462 100644
--- a/tasks/maniguard-jar-transport-i25-standard/task.toml
+++ b/tasks/maniguard-jar-transport-i25-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "maniguard"
robot = "franka-panda-longfinger"
family = "jar_transport"
mode = "standard" # robot as a service: RGB-D + proprioception, no object poses, no success or safety predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No positive example yet (oracle = \"none\": the 16 ManiGuard tasks added on 2026-10-04 for variety ship without one, user decision). Upstream reference, cited only: ManiGuard's 40 scripted demonstrations of this task, episodes 1000-1039 of the Hugging Face dataset IDEAS-Lab-Northwestern/datagen-jar-v1-joint-5cam (revision 080dddb4, LeRobot, 30 Hz, CC BY-NC 4.0), recorded by ManiGuard's data generator in a differently built scene; none has been replayed through this task's service and none is copied into the task. The frozen instance (state sha 38c331dc1d5f4f70) was made on an RTX PRO 6000 (2026-10-04); a second fresh load gave the same hash. Human review per task (PROTOCOL §2.7)."
diff --git a/tasks/maniguard-lid-transport-i06-privileged/task.toml b/tasks/maniguard-lid-transport-i06-privileged/task.toml
index 743ceccd47..db8257528e 100644
--- a/tasks/maniguard-lid-transport-i06-privileged/task.toml
+++ b/tasks/maniguard-lid-transport-i06-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "maniguard" # ManiGuard on BEHAVIOR-1K v3.7.2 (OmniGibson
robot = "franka-panda-longfinger"
family = "lid_transport"
mode = "privileged" # privileged state, resets and thin primitives through the tooling (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No positive example yet (oracle = \"none\": the 16 ManiGuard tasks added on 2026-10-04 for variety ship without one, user decision). Upstream reference, cited only: ManiGuard's 40 scripted demonstrations of this task, episodes 240-279 of the Hugging Face dataset IDEAS-Lab-Northwestern/datagen-lid-v1-joint-5cam (revision b0a8980a, LeRobot, 30 Hz, CC BY-NC 4.0), recorded by ManiGuard's data generator in a differently built scene; none has been replayed through this verifier and none is copied into the task. The frozen instance (state sha 3463868903df88b2) was made on an RTX PRO 6000 (2026-10-04); a second fresh load gave the same hash. Human review per task (PROTOCOL §2.7)."
instance_seed = 6 # = the ManiGuard-Bench task index (lid_transport/task_0006/base); scene = environment/instance/; identity = dataset.toml digest
diff --git a/tasks/maniguard-lid-transport-i06-standard/task.toml b/tasks/maniguard-lid-transport-i06-standard/task.toml
index 10684612da..99fe7673df 100644
--- a/tasks/maniguard-lid-transport-i06-standard/task.toml
+++ b/tasks/maniguard-lid-transport-i06-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "maniguard"
robot = "franka-panda-longfinger"
family = "lid_transport"
mode = "standard" # robot as a service: RGB-D + proprioception, no object poses, no success or safety predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No positive example yet (oracle = \"none\": the 16 ManiGuard tasks added on 2026-10-04 for variety ship without one, user decision). Upstream reference, cited only: ManiGuard's 40 scripted demonstrations of this task, episodes 240-279 of the Hugging Face dataset IDEAS-Lab-Northwestern/datagen-lid-v1-joint-5cam (revision b0a8980a, LeRobot, 30 Hz, CC BY-NC 4.0), recorded by ManiGuard's data generator in a differently built scene; none has been replayed through this task's service and none is copied into the task. The frozen instance (state sha 3463868903df88b2) was made on an RTX PRO 6000 (2026-10-04); a second fresh load gave the same hash. Human review per task (PROTOCOL §2.7)."
diff --git a/tasks/maniguard-lid-transport-i07-privileged/task.toml b/tasks/maniguard-lid-transport-i07-privileged/task.toml
index e63db04f59..09f47920c9 100644
--- a/tasks/maniguard-lid-transport-i07-privileged/task.toml
+++ b/tasks/maniguard-lid-transport-i07-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "maniguard" # ManiGuard on BEHAVIOR-1K v3.7.2 (OmniGibson
robot = "franka-panda-longfinger"
family = "lid_transport"
mode = "privileged" # privileged state, resets and thin primitives through the tooling (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 7 # = the ManiGuard-Bench task index (lid_transport/task_0007/base); scene = environment/instance/; identity = dataset.toml digest
expert_time_estimate_hours = 2.0
diff --git a/tasks/maniguard-lid-transport-i07-standard/task.toml b/tasks/maniguard-lid-transport-i07-standard/task.toml
index ced7a66d9b..dc4ab63593 100644
--- a/tasks/maniguard-lid-transport-i07-standard/task.toml
+++ b/tasks/maniguard-lid-transport-i07-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "maniguard"
robot = "franka-panda-longfinger"
family = "lid_transport"
mode = "standard" # robot as a service: RGB-D + proprioception, no object poses, no success or safety predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "full"
instance_seed = 7
diff --git a/tasks/maniguard-lid-transport-i12-privileged/task.toml b/tasks/maniguard-lid-transport-i12-privileged/task.toml
index 2280d2138c..a7a16b8a77 100644
--- a/tasks/maniguard-lid-transport-i12-privileged/task.toml
+++ b/tasks/maniguard-lid-transport-i12-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "maniguard" # ManiGuard on BEHAVIOR-1K v3.7.2 (OmniGibson
robot = "franka-panda-longfinger"
family = "lid_transport"
mode = "privileged" # privileged state, resets and thin primitives through the tooling (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 12 # = the ManiGuard-Bench task index (lid_transport/task_0012/base); scene = environment/instance/; identity = dataset.toml digest
expert_time_estimate_hours = 2.0
diff --git a/tasks/maniguard-lid-transport-i12-standard/task.toml b/tasks/maniguard-lid-transport-i12-standard/task.toml
index e36d45964c..a4e028a1d3 100644
--- a/tasks/maniguard-lid-transport-i12-standard/task.toml
+++ b/tasks/maniguard-lid-transport-i12-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "maniguard"
robot = "franka-panda-longfinger"
family = "lid_transport"
mode = "standard" # robot as a service: RGB-D + proprioception, no object poses, no success or safety predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "full"
instance_seed = 12
diff --git a/tasks/maniguard-lid-transport-i15-privileged/task.toml b/tasks/maniguard-lid-transport-i15-privileged/task.toml
index 122da014bf..b0299bfa21 100644
--- a/tasks/maniguard-lid-transport-i15-privileged/task.toml
+++ b/tasks/maniguard-lid-transport-i15-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "maniguard" # ManiGuard on BEHAVIOR-1K v3.7.2 (OmniGibson
robot = "franka-panda-longfinger"
family = "lid_transport"
mode = "privileged" # privileged state, resets and thin primitives through the tooling (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No positive example yet (oracle = \"none\": the 16 ManiGuard tasks added on 2026-10-04 for variety ship without one, user decision). Upstream reference, cited only: ManiGuard's 40 scripted demonstrations of this task, episodes 600-639 of the Hugging Face dataset IDEAS-Lab-Northwestern/datagen-lid-v1-joint-5cam (revision b0a8980a, LeRobot, 30 Hz, CC BY-NC 4.0), recorded by ManiGuard's data generator in a differently built scene; none has been replayed through this verifier and none is copied into the task. The frozen instance (state sha 942a6d9824503c4a) was made on an RTX PRO 6000 (2026-10-04); a second fresh load gave the same hash. Human review per task (PROTOCOL §2.7)."
instance_seed = 15 # = the ManiGuard-Bench task index (lid_transport/task_0015/base); scene = environment/instance/; identity = dataset.toml digest
diff --git a/tasks/maniguard-lid-transport-i15-standard/task.toml b/tasks/maniguard-lid-transport-i15-standard/task.toml
index e9dc4b7b7a..384733f535 100644
--- a/tasks/maniguard-lid-transport-i15-standard/task.toml
+++ b/tasks/maniguard-lid-transport-i15-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "maniguard"
robot = "franka-panda-longfinger"
family = "lid_transport"
mode = "standard" # robot as a service: RGB-D + proprioception, no object poses, no success or safety predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No positive example yet (oracle = \"none\": the 16 ManiGuard tasks added on 2026-10-04 for variety ship without one, user decision). Upstream reference, cited only: ManiGuard's 40 scripted demonstrations of this task, episodes 600-639 of the Hugging Face dataset IDEAS-Lab-Northwestern/datagen-lid-v1-joint-5cam (revision b0a8980a, LeRobot, 30 Hz, CC BY-NC 4.0), recorded by ManiGuard's data generator in a differently built scene; none has been replayed through this task's service and none is copied into the task. The frozen instance (state sha 942a6d9824503c4a) was made on an RTX PRO 6000 (2026-10-04); a second fresh load gave the same hash. Human review per task (PROTOCOL §2.7)."
diff --git a/tasks/maniguard-lid-transport-i16-privileged/task.toml b/tasks/maniguard-lid-transport-i16-privileged/task.toml
index 60455b5c7c..53e89b29d3 100644
--- a/tasks/maniguard-lid-transport-i16-privileged/task.toml
+++ b/tasks/maniguard-lid-transport-i16-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "maniguard" # ManiGuard on BEHAVIOR-1K v3.7.2 (OmniGibson
robot = "franka-panda-longfinger"
family = "lid_transport"
mode = "privileged" # privileged state, resets and thin primitives through the tooling (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 16 # = the ManiGuard-Bench task index (lid_transport/task_0016/base); scene = environment/instance/; identity = dataset.toml digest
expert_time_estimate_hours = 2.0
diff --git a/tasks/maniguard-lid-transport-i16-standard/task.toml b/tasks/maniguard-lid-transport-i16-standard/task.toml
index d63e16ac7b..9e64345e42 100644
--- a/tasks/maniguard-lid-transport-i16-standard/task.toml
+++ b/tasks/maniguard-lid-transport-i16-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "maniguard"
robot = "franka-panda-longfinger"
family = "lid_transport"
mode = "standard" # robot as a service: RGB-D + proprioception, no object poses, no success or safety predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "full"
instance_seed = 16
diff --git a/tasks/maniguard-stack-retrieve-i08-privileged/task.toml b/tasks/maniguard-stack-retrieve-i08-privileged/task.toml
index eb79cc0674..5cc007b3d8 100644
--- a/tasks/maniguard-stack-retrieve-i08-privileged/task.toml
+++ b/tasks/maniguard-stack-retrieve-i08-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "maniguard" # ManiGuard on BEHAVIOR-1K v3.7.2 (OmniGibson
robot = "franka-panda-longfinger"
family = "stack_retrieve"
mode = "privileged" # privileged state, resets and thin primitives through the tooling (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 8 # = the ManiGuard-Bench task index (stack_retrieve/task_0008/base); scene = environment/instance/; identity = dataset.toml digest
expert_time_estimate_hours = 2.0
diff --git a/tasks/maniguard-stack-retrieve-i08-standard/task.toml b/tasks/maniguard-stack-retrieve-i08-standard/task.toml
index 7165ee203f..7db4404a5a 100644
--- a/tasks/maniguard-stack-retrieve-i08-standard/task.toml
+++ b/tasks/maniguard-stack-retrieve-i08-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "maniguard"
robot = "franka-panda-longfinger"
family = "stack_retrieve"
mode = "standard" # robot as a service: RGB-D + proprioception, no object poses, no success or safety predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "full"
instance_seed = 8
diff --git a/tasks/maniguard-stack-retrieve-i10-privileged/task.toml b/tasks/maniguard-stack-retrieve-i10-privileged/task.toml
index ae57853560..cd1866d790 100644
--- a/tasks/maniguard-stack-retrieve-i10-privileged/task.toml
+++ b/tasks/maniguard-stack-retrieve-i10-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "maniguard" # ManiGuard on BEHAVIOR-1K v3.7.2 (OmniGibson
robot = "franka-panda-longfinger"
family = "stack_retrieve"
mode = "privileged" # privileged state, resets and thin primitives through the tooling (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 10 # = the ManiGuard-Bench task index (stack_retrieve/task_0010/base); scene = environment/instance/; identity = dataset.toml digest
expert_time_estimate_hours = 2.0
diff --git a/tasks/maniguard-stack-retrieve-i10-standard/task.toml b/tasks/maniguard-stack-retrieve-i10-standard/task.toml
index 967a866714..08cf42b691 100644
--- a/tasks/maniguard-stack-retrieve-i10-standard/task.toml
+++ b/tasks/maniguard-stack-retrieve-i10-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "maniguard"
robot = "franka-panda-longfinger"
family = "stack_retrieve"
mode = "standard" # robot as a service: RGB-D + proprioception, no object poses, no success or safety predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "full"
instance_seed = 10
diff --git a/tasks/maniguard-stack-retrieve-i12-privileged/task.toml b/tasks/maniguard-stack-retrieve-i12-privileged/task.toml
index a20ef62856..22ca6ef18d 100644
--- a/tasks/maniguard-stack-retrieve-i12-privileged/task.toml
+++ b/tasks/maniguard-stack-retrieve-i12-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "maniguard" # ManiGuard on BEHAVIOR-1K v3.7.2 (OmniGibson
robot = "franka-panda-longfinger"
family = "stack_retrieve"
mode = "privileged" # privileged state, resets and thin primitives through the tooling (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No positive example yet (oracle = \"none\": the 16 ManiGuard tasks added on 2026-10-04 for variety ship without one, user decision). Upstream reference, cited only: ManiGuard's 40 scripted demonstrations of this task, episodes 480-519 of the Hugging Face dataset IDEAS-Lab-Northwestern/datagen-stack-v1-joint-5cam (revision e30b3585, LeRobot, 30 Hz, CC BY-NC 4.0), recorded by ManiGuard's data generator in a differently built scene; none has been replayed through this verifier and none is copied into the task. The frozen instance (state sha aad13a21b20c4d01) was made on an RTX PRO 6000 (2026-10-04); a second fresh load gave the same hash. Human review per task (PROTOCOL §2.7)."
instance_seed = 12 # = the ManiGuard-Bench task index (stack_retrieve/task_0012/base); scene = environment/instance/; identity = dataset.toml digest
diff --git a/tasks/maniguard-stack-retrieve-i12-standard/task.toml b/tasks/maniguard-stack-retrieve-i12-standard/task.toml
index 2efff47535..dbf2f07a5f 100644
--- a/tasks/maniguard-stack-retrieve-i12-standard/task.toml
+++ b/tasks/maniguard-stack-retrieve-i12-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "maniguard"
robot = "franka-panda-longfinger"
family = "stack_retrieve"
mode = "standard" # robot as a service: RGB-D + proprioception, no object poses, no success or safety predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No positive example yet (oracle = \"none\": the 16 ManiGuard tasks added on 2026-10-04 for variety ship without one, user decision). Upstream reference, cited only: ManiGuard's 40 scripted demonstrations of this task, episodes 480-519 of the Hugging Face dataset IDEAS-Lab-Northwestern/datagen-stack-v1-joint-5cam (revision e30b3585, LeRobot, 30 Hz, CC BY-NC 4.0), recorded by ManiGuard's data generator in a differently built scene; none has been replayed through this task's service and none is copied into the task. The frozen instance (state sha aad13a21b20c4d01) was made on an RTX PRO 6000 (2026-10-04); a second fresh load gave the same hash. Human review per task (PROTOCOL §2.7)."
diff --git a/tasks/maniguard-stack-retrieve-i13-privileged/task.toml b/tasks/maniguard-stack-retrieve-i13-privileged/task.toml
index c72a4da5f7..e12ef535ea 100644
--- a/tasks/maniguard-stack-retrieve-i13-privileged/task.toml
+++ b/tasks/maniguard-stack-retrieve-i13-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "maniguard" # ManiGuard on BEHAVIOR-1K v3.7.2 (OmniGibson
robot = "franka-panda-longfinger"
family = "stack_retrieve"
mode = "privileged" # privileged state, resets and thin primitives through the tooling (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No positive example yet (oracle = \"none\": the 16 ManiGuard tasks added on 2026-10-04 for variety ship without one, user decision). Upstream reference, cited only: ManiGuard's 40 scripted demonstrations of this task, episodes 520-559 of the Hugging Face dataset IDEAS-Lab-Northwestern/datagen-stack-v1-joint-5cam (revision e30b3585, LeRobot, 30 Hz, CC BY-NC 4.0), recorded by ManiGuard's data generator in a differently built scene; none has been replayed through this verifier and none is copied into the task. The frozen instance (state sha 803c1af818aa3c63) was made on an RTX PRO 6000 (2026-10-04); a second fresh load gave the same hash. Human review per task (PROTOCOL §2.7)."
instance_seed = 13 # = the ManiGuard-Bench task index (stack_retrieve/task_0013/base); scene = environment/instance/; identity = dataset.toml digest
diff --git a/tasks/maniguard-stack-retrieve-i13-standard/task.toml b/tasks/maniguard-stack-retrieve-i13-standard/task.toml
index 5166eb4588..552dbf0b92 100644
--- a/tasks/maniguard-stack-retrieve-i13-standard/task.toml
+++ b/tasks/maniguard-stack-retrieve-i13-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "maniguard"
robot = "franka-panda-longfinger"
family = "stack_retrieve"
mode = "standard" # robot as a service: RGB-D + proprioception, no object poses, no success or safety predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No positive example yet (oracle = \"none\": the 16 ManiGuard tasks added on 2026-10-04 for variety ship without one, user decision). Upstream reference, cited only: ManiGuard's 40 scripted demonstrations of this task, episodes 520-559 of the Hugging Face dataset IDEAS-Lab-Northwestern/datagen-stack-v1-joint-5cam (revision e30b3585, LeRobot, 30 Hz, CC BY-NC 4.0), recorded by ManiGuard's data generator in a differently built scene; none has been replayed through this task's service and none is copied into the task. The frozen instance (state sha 803c1af818aa3c63) was made on an RTX PRO 6000 (2026-10-04); a second fresh load gave the same hash. Human review per task (PROTOCOL §2.7)."
diff --git a/tasks/maniguard-stack-retrieve-i21-privileged/task.toml b/tasks/maniguard-stack-retrieve-i21-privileged/task.toml
index b6a7c02e1c..c56514554b 100644
--- a/tasks/maniguard-stack-retrieve-i21-privileged/task.toml
+++ b/tasks/maniguard-stack-retrieve-i21-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "maniguard" # ManiGuard on BEHAVIOR-1K v3.7.2 (OmniGibson
robot = "franka-panda-longfinger"
family = "stack_retrieve"
mode = "privileged" # privileged state, resets and thin primitives through the tooling (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 21 # = the ManiGuard-Bench task index (stack_retrieve/task_0021/base); scene = environment/instance/; identity = dataset.toml digest
expert_time_estimate_hours = 2.0
diff --git a/tasks/maniguard-stack-retrieve-i21-standard/task.toml b/tasks/maniguard-stack-retrieve-i21-standard/task.toml
index 287d028175..2635050b24 100644
--- a/tasks/maniguard-stack-retrieve-i21-standard/task.toml
+++ b/tasks/maniguard-stack-retrieve-i21-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "maniguard"
robot = "franka-panda-longfinger"
family = "stack_retrieve"
mode = "standard" # robot as a service: RGB-D + proprioception, no object poses, no success or safety predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "full"
instance_seed = 21
diff --git a/tasks/metaworldplus-assembly-i00-privileged/task.toml b/tasks/metaworldplus-assembly-i00-privileged/task.toml
index 5793b205b9..203cb6f0b6 100644
--- a/tasks/metaworldplus-assembly-i00-privileged/task.toml
+++ b/tasks/metaworldplus-assembly-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "mujoco"
robot = "sawyer"
family = "metaworld/assembly"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 0 # Meta-World's reset randomization (object and goal positions)
expert_time_estimate_hours = 0.5
diff --git a/tasks/metaworldplus-assembly-i00-standard/task.toml b/tasks/metaworldplus-assembly-i00-standard/task.toml
index cc3a224588..20e87072e7 100644
--- a/tasks/metaworldplus-assembly-i00-standard/task.toml
+++ b/tasks/metaworldplus-assembly-i00-standard/task.toml
@@ -23,6 +23,7 @@ simulator = "mujoco"
robot = "sawyer"
family = "metaworld/assembly"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 0
diff --git a/tasks/metaworldplus-basketball-i00-privileged/task.toml b/tasks/metaworldplus-basketball-i00-privileged/task.toml
index 7b32172b05..a1c914b9cf 100644
--- a/tasks/metaworldplus-basketball-i00-privileged/task.toml
+++ b/tasks/metaworldplus-basketball-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "mujoco"
robot = "sawyer"
family = "metaworld/basketball"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 0 # Meta-World's reset randomization (object and goal positions)
expert_time_estimate_hours = 0.5
diff --git a/tasks/metaworldplus-basketball-i00-standard/task.toml b/tasks/metaworldplus-basketball-i00-standard/task.toml
index bddd388f4c..0286fcdd2d 100644
--- a/tasks/metaworldplus-basketball-i00-standard/task.toml
+++ b/tasks/metaworldplus-basketball-i00-standard/task.toml
@@ -23,6 +23,7 @@ simulator = "mujoco"
robot = "sawyer"
family = "metaworld/basketball"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 0
diff --git a/tasks/metaworldplus-bin-picking-i00-privileged/task.toml b/tasks/metaworldplus-bin-picking-i00-privileged/task.toml
index e80358cf88..c0f101ac9d 100644
--- a/tasks/metaworldplus-bin-picking-i00-privileged/task.toml
+++ b/tasks/metaworldplus-bin-picking-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "mujoco"
robot = "sawyer"
family = "metaworld/bin-picking"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 0 # Meta-World's reset randomization (object and goal positions)
expert_time_estimate_hours = 0.5
diff --git a/tasks/metaworldplus-bin-picking-i00-standard/task.toml b/tasks/metaworldplus-bin-picking-i00-standard/task.toml
index 6b01a04446..9b13cfc5e1 100644
--- a/tasks/metaworldplus-bin-picking-i00-standard/task.toml
+++ b/tasks/metaworldplus-bin-picking-i00-standard/task.toml
@@ -23,6 +23,7 @@ simulator = "mujoco"
robot = "sawyer"
family = "metaworld/bin-picking"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 0
diff --git a/tasks/metaworldplus-box-close-i00-privileged/task.toml b/tasks/metaworldplus-box-close-i00-privileged/task.toml
index 6cc7ab5a50..f4c2aba679 100644
--- a/tasks/metaworldplus-box-close-i00-privileged/task.toml
+++ b/tasks/metaworldplus-box-close-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "mujoco"
robot = "sawyer"
family = "metaworld/box-close"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 0 # Meta-World's reset randomization (object and goal positions)
expert_time_estimate_hours = 0.5
diff --git a/tasks/metaworldplus-box-close-i00-standard/task.toml b/tasks/metaworldplus-box-close-i00-standard/task.toml
index aa48f18382..cd9615504f 100644
--- a/tasks/metaworldplus-box-close-i00-standard/task.toml
+++ b/tasks/metaworldplus-box-close-i00-standard/task.toml
@@ -23,6 +23,7 @@ simulator = "mujoco"
robot = "sawyer"
family = "metaworld/box-close"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 0
diff --git a/tasks/metaworldplus-button-press-i00-privileged/task.toml b/tasks/metaworldplus-button-press-i00-privileged/task.toml
index c6eaf97392..e19494b2cf 100644
--- a/tasks/metaworldplus-button-press-i00-privileged/task.toml
+++ b/tasks/metaworldplus-button-press-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "mujoco"
robot = "sawyer"
family = "metaworld/button-press"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 0 # Meta-World's reset randomization (object and goal positions)
expert_time_estimate_hours = 0.5
diff --git a/tasks/metaworldplus-button-press-i00-standard/task.toml b/tasks/metaworldplus-button-press-i00-standard/task.toml
index 2d79b7d2cd..282675bab7 100644
--- a/tasks/metaworldplus-button-press-i00-standard/task.toml
+++ b/tasks/metaworldplus-button-press-i00-standard/task.toml
@@ -23,6 +23,7 @@ simulator = "mujoco"
robot = "sawyer"
family = "metaworld/button-press"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 0
diff --git a/tasks/metaworldplus-button-press-topdown-i00-privileged/task.toml b/tasks/metaworldplus-button-press-topdown-i00-privileged/task.toml
index eca3977acf..44adc8411c 100644
--- a/tasks/metaworldplus-button-press-topdown-i00-privileged/task.toml
+++ b/tasks/metaworldplus-button-press-topdown-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "mujoco"
robot = "sawyer"
family = "metaworld/button-press-topdown"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 0 # Meta-World's reset randomization (object and goal positions)
expert_time_estimate_hours = 0.5
diff --git a/tasks/metaworldplus-button-press-topdown-i00-standard/task.toml b/tasks/metaworldplus-button-press-topdown-i00-standard/task.toml
index fc5b24bbbc..da688caefc 100644
--- a/tasks/metaworldplus-button-press-topdown-i00-standard/task.toml
+++ b/tasks/metaworldplus-button-press-topdown-i00-standard/task.toml
@@ -23,6 +23,7 @@ simulator = "mujoco"
robot = "sawyer"
family = "metaworld/button-press-topdown"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 0
diff --git a/tasks/metaworldplus-button-press-topdown-wall-i00-privileged/task.toml b/tasks/metaworldplus-button-press-topdown-wall-i00-privileged/task.toml
index b384786272..f089f509ee 100644
--- a/tasks/metaworldplus-button-press-topdown-wall-i00-privileged/task.toml
+++ b/tasks/metaworldplus-button-press-topdown-wall-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "mujoco"
robot = "sawyer"
family = "metaworld/button-press-topdown-wall"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 0 # Meta-World's reset randomization (object and goal positions)
expert_time_estimate_hours = 0.5
diff --git a/tasks/metaworldplus-button-press-topdown-wall-i00-standard/task.toml b/tasks/metaworldplus-button-press-topdown-wall-i00-standard/task.toml
index c02964343e..ac188644e2 100644
--- a/tasks/metaworldplus-button-press-topdown-wall-i00-standard/task.toml
+++ b/tasks/metaworldplus-button-press-topdown-wall-i00-standard/task.toml
@@ -23,6 +23,7 @@ simulator = "mujoco"
robot = "sawyer"
family = "metaworld/button-press-topdown-wall"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 0
diff --git a/tasks/metaworldplus-button-press-wall-i00-privileged/task.toml b/tasks/metaworldplus-button-press-wall-i00-privileged/task.toml
index a12490ce7c..773155911a 100644
--- a/tasks/metaworldplus-button-press-wall-i00-privileged/task.toml
+++ b/tasks/metaworldplus-button-press-wall-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "mujoco"
robot = "sawyer"
family = "metaworld/button-press-wall"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 0 # Meta-World's reset randomization (object and goal positions)
expert_time_estimate_hours = 0.5
diff --git a/tasks/metaworldplus-button-press-wall-i00-standard/task.toml b/tasks/metaworldplus-button-press-wall-i00-standard/task.toml
index 782f5f36ec..75579a201b 100644
--- a/tasks/metaworldplus-button-press-wall-i00-standard/task.toml
+++ b/tasks/metaworldplus-button-press-wall-i00-standard/task.toml
@@ -23,6 +23,7 @@ simulator = "mujoco"
robot = "sawyer"
family = "metaworld/button-press-wall"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 0
diff --git a/tasks/metaworldplus-coffee-button-i00-privileged/task.toml b/tasks/metaworldplus-coffee-button-i00-privileged/task.toml
index f37e60a3a2..6b18b1cbd2 100644
--- a/tasks/metaworldplus-coffee-button-i00-privileged/task.toml
+++ b/tasks/metaworldplus-coffee-button-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "mujoco"
robot = "sawyer"
family = "metaworld/coffee-button"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 0 # Meta-World's reset randomization (object and goal positions)
expert_time_estimate_hours = 0.5
diff --git a/tasks/metaworldplus-coffee-button-i00-standard/task.toml b/tasks/metaworldplus-coffee-button-i00-standard/task.toml
index 28e73f68aa..c4671080d2 100644
--- a/tasks/metaworldplus-coffee-button-i00-standard/task.toml
+++ b/tasks/metaworldplus-coffee-button-i00-standard/task.toml
@@ -23,6 +23,7 @@ simulator = "mujoco"
robot = "sawyer"
family = "metaworld/coffee-button"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 0
diff --git a/tasks/metaworldplus-coffee-pull-i00-privileged/task.toml b/tasks/metaworldplus-coffee-pull-i00-privileged/task.toml
index 9cd2fda598..70c184f671 100644
--- a/tasks/metaworldplus-coffee-pull-i00-privileged/task.toml
+++ b/tasks/metaworldplus-coffee-pull-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "mujoco"
robot = "sawyer"
family = "metaworld/coffee-pull"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 0 # Meta-World's reset randomization (object and goal positions)
expert_time_estimate_hours = 0.5
diff --git a/tasks/metaworldplus-coffee-pull-i00-standard/task.toml b/tasks/metaworldplus-coffee-pull-i00-standard/task.toml
index 7fb76fd62d..ad9bd8e4f5 100644
--- a/tasks/metaworldplus-coffee-pull-i00-standard/task.toml
+++ b/tasks/metaworldplus-coffee-pull-i00-standard/task.toml
@@ -23,6 +23,7 @@ simulator = "mujoco"
robot = "sawyer"
family = "metaworld/coffee-pull"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 0
diff --git a/tasks/metaworldplus-coffee-push-i00-privileged/task.toml b/tasks/metaworldplus-coffee-push-i00-privileged/task.toml
index 0b560cdebd..e1e654be58 100644
--- a/tasks/metaworldplus-coffee-push-i00-privileged/task.toml
+++ b/tasks/metaworldplus-coffee-push-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "mujoco"
robot = "sawyer"
family = "metaworld/coffee-push"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 0 # Meta-World's reset randomization (object and goal positions)
expert_time_estimate_hours = 0.5
diff --git a/tasks/metaworldplus-coffee-push-i00-standard/task.toml b/tasks/metaworldplus-coffee-push-i00-standard/task.toml
index 3963dfc36a..45cfc2e752 100644
--- a/tasks/metaworldplus-coffee-push-i00-standard/task.toml
+++ b/tasks/metaworldplus-coffee-push-i00-standard/task.toml
@@ -23,6 +23,7 @@ simulator = "mujoco"
robot = "sawyer"
family = "metaworld/coffee-push"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 0
diff --git a/tasks/metaworldplus-dial-turn-i00-privileged/task.toml b/tasks/metaworldplus-dial-turn-i00-privileged/task.toml
index 181005aa85..fbc58c958a 100644
--- a/tasks/metaworldplus-dial-turn-i00-privileged/task.toml
+++ b/tasks/metaworldplus-dial-turn-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "mujoco"
robot = "sawyer"
family = "metaworld/dial-turn"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 0 # Meta-World's reset randomization (object and goal positions)
expert_time_estimate_hours = 0.5
diff --git a/tasks/metaworldplus-dial-turn-i00-standard/task.toml b/tasks/metaworldplus-dial-turn-i00-standard/task.toml
index 48dbd37b5a..749ec235e7 100644
--- a/tasks/metaworldplus-dial-turn-i00-standard/task.toml
+++ b/tasks/metaworldplus-dial-turn-i00-standard/task.toml
@@ -23,6 +23,7 @@ simulator = "mujoco"
robot = "sawyer"
family = "metaworld/dial-turn"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 0
diff --git a/tasks/metaworldplus-disassemble-i00-privileged/task.toml b/tasks/metaworldplus-disassemble-i00-privileged/task.toml
index 8bd95eb356..f1a4263e00 100644
--- a/tasks/metaworldplus-disassemble-i00-privileged/task.toml
+++ b/tasks/metaworldplus-disassemble-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "mujoco"
robot = "sawyer"
family = "metaworld/disassemble"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 0 # Meta-World's reset randomization (object and goal positions)
expert_time_estimate_hours = 0.5
diff --git a/tasks/metaworldplus-disassemble-i00-standard/task.toml b/tasks/metaworldplus-disassemble-i00-standard/task.toml
index 0ec96a8636..5fcfc37f9b 100644
--- a/tasks/metaworldplus-disassemble-i00-standard/task.toml
+++ b/tasks/metaworldplus-disassemble-i00-standard/task.toml
@@ -23,6 +23,7 @@ simulator = "mujoco"
robot = "sawyer"
family = "metaworld/disassemble"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 0
diff --git a/tasks/metaworldplus-door-close-i00-privileged/task.toml b/tasks/metaworldplus-door-close-i00-privileged/task.toml
index 94569a55f9..2eb5a25e37 100644
--- a/tasks/metaworldplus-door-close-i00-privileged/task.toml
+++ b/tasks/metaworldplus-door-close-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "mujoco"
robot = "sawyer"
family = "metaworld/door-close"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 0 # Meta-World's reset randomization (object and goal positions)
expert_time_estimate_hours = 0.5
diff --git a/tasks/metaworldplus-door-close-i00-standard/task.toml b/tasks/metaworldplus-door-close-i00-standard/task.toml
index 123c3d2e17..d4a0efd1db 100644
--- a/tasks/metaworldplus-door-close-i00-standard/task.toml
+++ b/tasks/metaworldplus-door-close-i00-standard/task.toml
@@ -23,6 +23,7 @@ simulator = "mujoco"
robot = "sawyer"
family = "metaworld/door-close"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 0
diff --git a/tasks/metaworldplus-door-lock-i00-privileged/task.toml b/tasks/metaworldplus-door-lock-i00-privileged/task.toml
index 6a0e4377ff..70a6f9d047 100644
--- a/tasks/metaworldplus-door-lock-i00-privileged/task.toml
+++ b/tasks/metaworldplus-door-lock-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "mujoco"
robot = "sawyer"
family = "metaworld/door-lock"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 0 # Meta-World's reset randomization (object and goal positions)
expert_time_estimate_hours = 0.5
diff --git a/tasks/metaworldplus-door-lock-i00-standard/task.toml b/tasks/metaworldplus-door-lock-i00-standard/task.toml
index 3865bd6ea1..df22133fea 100644
--- a/tasks/metaworldplus-door-lock-i00-standard/task.toml
+++ b/tasks/metaworldplus-door-lock-i00-standard/task.toml
@@ -23,6 +23,7 @@ simulator = "mujoco"
robot = "sawyer"
family = "metaworld/door-lock"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 0
diff --git a/tasks/metaworldplus-door-open-i00-privileged/task.toml b/tasks/metaworldplus-door-open-i00-privileged/task.toml
index 3f0766b615..503bc5b917 100644
--- a/tasks/metaworldplus-door-open-i00-privileged/task.toml
+++ b/tasks/metaworldplus-door-open-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "mujoco"
robot = "sawyer"
family = "metaworld/door-open"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 0 # Meta-World's reset randomization (object and goal positions)
expert_time_estimate_hours = 0.5
diff --git a/tasks/metaworldplus-door-open-i00-standard/task.toml b/tasks/metaworldplus-door-open-i00-standard/task.toml
index ce54a73cd4..f69df9e235 100644
--- a/tasks/metaworldplus-door-open-i00-standard/task.toml
+++ b/tasks/metaworldplus-door-open-i00-standard/task.toml
@@ -23,6 +23,7 @@ simulator = "mujoco"
robot = "sawyer"
family = "metaworld/door-open"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 0
diff --git a/tasks/metaworldplus-door-unlock-i00-privileged/task.toml b/tasks/metaworldplus-door-unlock-i00-privileged/task.toml
index 226b9925d2..30f2744b84 100644
--- a/tasks/metaworldplus-door-unlock-i00-privileged/task.toml
+++ b/tasks/metaworldplus-door-unlock-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "mujoco"
robot = "sawyer"
family = "metaworld/door-unlock"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 0 # Meta-World's reset randomization (object and goal positions)
expert_time_estimate_hours = 0.5
diff --git a/tasks/metaworldplus-door-unlock-i00-standard/task.toml b/tasks/metaworldplus-door-unlock-i00-standard/task.toml
index 7dcf83782e..33801f8399 100644
--- a/tasks/metaworldplus-door-unlock-i00-standard/task.toml
+++ b/tasks/metaworldplus-door-unlock-i00-standard/task.toml
@@ -23,6 +23,7 @@ simulator = "mujoco"
robot = "sawyer"
family = "metaworld/door-unlock"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 0
diff --git a/tasks/metaworldplus-drawer-close-i00-privileged/task.toml b/tasks/metaworldplus-drawer-close-i00-privileged/task.toml
index 7302fb0da4..76a3e5b83e 100644
--- a/tasks/metaworldplus-drawer-close-i00-privileged/task.toml
+++ b/tasks/metaworldplus-drawer-close-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "mujoco"
robot = "sawyer"
family = "metaworld/drawer-close"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 0 # Meta-World's reset randomization (object and goal positions)
expert_time_estimate_hours = 0.5
diff --git a/tasks/metaworldplus-drawer-close-i00-standard/task.toml b/tasks/metaworldplus-drawer-close-i00-standard/task.toml
index f57945f02e..e28748a955 100644
--- a/tasks/metaworldplus-drawer-close-i00-standard/task.toml
+++ b/tasks/metaworldplus-drawer-close-i00-standard/task.toml
@@ -23,6 +23,7 @@ simulator = "mujoco"
robot = "sawyer"
family = "metaworld/drawer-close"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 0
diff --git a/tasks/metaworldplus-drawer-open-i00-privileged/task.toml b/tasks/metaworldplus-drawer-open-i00-privileged/task.toml
index a4c1bff850..ed191d604c 100644
--- a/tasks/metaworldplus-drawer-open-i00-privileged/task.toml
+++ b/tasks/metaworldplus-drawer-open-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "mujoco"
robot = "sawyer"
family = "metaworld/drawer-open"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 0 # Meta-World's reset randomization (object and goal positions)
expert_time_estimate_hours = 0.5
diff --git a/tasks/metaworldplus-drawer-open-i00-standard/task.toml b/tasks/metaworldplus-drawer-open-i00-standard/task.toml
index 19b50ca2a1..db3c0e61b9 100644
--- a/tasks/metaworldplus-drawer-open-i00-standard/task.toml
+++ b/tasks/metaworldplus-drawer-open-i00-standard/task.toml
@@ -23,6 +23,7 @@ simulator = "mujoco"
robot = "sawyer"
family = "metaworld/drawer-open"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 0
diff --git a/tasks/metaworldplus-faucet-close-i00-privileged/task.toml b/tasks/metaworldplus-faucet-close-i00-privileged/task.toml
index 0091856c30..9d815751f6 100644
--- a/tasks/metaworldplus-faucet-close-i00-privileged/task.toml
+++ b/tasks/metaworldplus-faucet-close-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "mujoco"
robot = "sawyer"
family = "metaworld/faucet-close"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 0 # Meta-World's reset randomization (object and goal positions)
expert_time_estimate_hours = 0.5
diff --git a/tasks/metaworldplus-faucet-close-i00-standard/task.toml b/tasks/metaworldplus-faucet-close-i00-standard/task.toml
index 593643645c..0de3ade4f5 100644
--- a/tasks/metaworldplus-faucet-close-i00-standard/task.toml
+++ b/tasks/metaworldplus-faucet-close-i00-standard/task.toml
@@ -23,6 +23,7 @@ simulator = "mujoco"
robot = "sawyer"
family = "metaworld/faucet-close"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 0
diff --git a/tasks/metaworldplus-faucet-open-i00-privileged/task.toml b/tasks/metaworldplus-faucet-open-i00-privileged/task.toml
index ca699dba6a..b7c0fd4886 100644
--- a/tasks/metaworldplus-faucet-open-i00-privileged/task.toml
+++ b/tasks/metaworldplus-faucet-open-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "mujoco"
robot = "sawyer"
family = "metaworld/faucet-open"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 0 # Meta-World's reset randomization (object and goal positions)
expert_time_estimate_hours = 0.5
diff --git a/tasks/metaworldplus-faucet-open-i00-standard/task.toml b/tasks/metaworldplus-faucet-open-i00-standard/task.toml
index 159c729762..19bd866976 100644
--- a/tasks/metaworldplus-faucet-open-i00-standard/task.toml
+++ b/tasks/metaworldplus-faucet-open-i00-standard/task.toml
@@ -23,6 +23,7 @@ simulator = "mujoco"
robot = "sawyer"
family = "metaworld/faucet-open"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 0
diff --git a/tasks/metaworldplus-hammer-i00-privileged/task.toml b/tasks/metaworldplus-hammer-i00-privileged/task.toml
index 769d183967..a12880ae22 100644
--- a/tasks/metaworldplus-hammer-i00-privileged/task.toml
+++ b/tasks/metaworldplus-hammer-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "mujoco"
robot = "sawyer"
family = "metaworld/hammer"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 0 # Meta-World's reset randomization (object and goal positions)
expert_time_estimate_hours = 0.5
diff --git a/tasks/metaworldplus-hammer-i00-standard/task.toml b/tasks/metaworldplus-hammer-i00-standard/task.toml
index 90e4d97742..51a8620c08 100644
--- a/tasks/metaworldplus-hammer-i00-standard/task.toml
+++ b/tasks/metaworldplus-hammer-i00-standard/task.toml
@@ -23,6 +23,7 @@ simulator = "mujoco"
robot = "sawyer"
family = "metaworld/hammer"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 0
diff --git a/tasks/metaworldplus-hand-insert-i00-privileged/task.toml b/tasks/metaworldplus-hand-insert-i00-privileged/task.toml
index d9228ca53e..41bef04445 100644
--- a/tasks/metaworldplus-hand-insert-i00-privileged/task.toml
+++ b/tasks/metaworldplus-hand-insert-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "mujoco"
robot = "sawyer"
family = "metaworld/hand-insert"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 0 # Meta-World's reset randomization (object and goal positions)
expert_time_estimate_hours = 0.5
diff --git a/tasks/metaworldplus-hand-insert-i00-standard/task.toml b/tasks/metaworldplus-hand-insert-i00-standard/task.toml
index d09e8c93fb..344c4c1fba 100644
--- a/tasks/metaworldplus-hand-insert-i00-standard/task.toml
+++ b/tasks/metaworldplus-hand-insert-i00-standard/task.toml
@@ -23,6 +23,7 @@ simulator = "mujoco"
robot = "sawyer"
family = "metaworld/hand-insert"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 0
diff --git a/tasks/metaworldplus-handle-press-i00-privileged/task.toml b/tasks/metaworldplus-handle-press-i00-privileged/task.toml
index cb67f7f5d1..1a1259d534 100644
--- a/tasks/metaworldplus-handle-press-i00-privileged/task.toml
+++ b/tasks/metaworldplus-handle-press-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "mujoco"
robot = "sawyer"
family = "metaworld/handle-press"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 0 # Meta-World's reset randomization (object and goal positions)
expert_time_estimate_hours = 0.5
diff --git a/tasks/metaworldplus-handle-press-i00-standard/task.toml b/tasks/metaworldplus-handle-press-i00-standard/task.toml
index 84d9bc9105..ad6cfb7f9b 100644
--- a/tasks/metaworldplus-handle-press-i00-standard/task.toml
+++ b/tasks/metaworldplus-handle-press-i00-standard/task.toml
@@ -23,6 +23,7 @@ simulator = "mujoco"
robot = "sawyer"
family = "metaworld/handle-press"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 0
diff --git a/tasks/metaworldplus-handle-press-side-i00-privileged/task.toml b/tasks/metaworldplus-handle-press-side-i00-privileged/task.toml
index 95e8790c23..a574d95b2a 100644
--- a/tasks/metaworldplus-handle-press-side-i00-privileged/task.toml
+++ b/tasks/metaworldplus-handle-press-side-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "mujoco"
robot = "sawyer"
family = "metaworld/handle-press-side"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 0 # Meta-World's reset randomization (object and goal positions)
expert_time_estimate_hours = 0.5
diff --git a/tasks/metaworldplus-handle-press-side-i00-standard/task.toml b/tasks/metaworldplus-handle-press-side-i00-standard/task.toml
index a394f191b9..59a8194c58 100644
--- a/tasks/metaworldplus-handle-press-side-i00-standard/task.toml
+++ b/tasks/metaworldplus-handle-press-side-i00-standard/task.toml
@@ -23,6 +23,7 @@ simulator = "mujoco"
robot = "sawyer"
family = "metaworld/handle-press-side"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 0
diff --git a/tasks/metaworldplus-handle-pull-i00-privileged/task.toml b/tasks/metaworldplus-handle-pull-i00-privileged/task.toml
index 3ddd57b9c6..0d94499373 100644
--- a/tasks/metaworldplus-handle-pull-i00-privileged/task.toml
+++ b/tasks/metaworldplus-handle-pull-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "mujoco"
robot = "sawyer"
family = "metaworld/handle-pull"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 0 # Meta-World's reset randomization (object and goal positions)
expert_time_estimate_hours = 0.5
diff --git a/tasks/metaworldplus-handle-pull-i00-standard/task.toml b/tasks/metaworldplus-handle-pull-i00-standard/task.toml
index 10783ee81c..11d7a040b9 100644
--- a/tasks/metaworldplus-handle-pull-i00-standard/task.toml
+++ b/tasks/metaworldplus-handle-pull-i00-standard/task.toml
@@ -23,6 +23,7 @@ simulator = "mujoco"
robot = "sawyer"
family = "metaworld/handle-pull"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 0
diff --git a/tasks/metaworldplus-handle-pull-side-i00-privileged/task.toml b/tasks/metaworldplus-handle-pull-side-i00-privileged/task.toml
index e6b12258db..88e35f0bda 100644
--- a/tasks/metaworldplus-handle-pull-side-i00-privileged/task.toml
+++ b/tasks/metaworldplus-handle-pull-side-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "mujoco"
robot = "sawyer"
family = "metaworld/handle-pull-side"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 0 # Meta-World's reset randomization (object and goal positions)
expert_time_estimate_hours = 0.5
diff --git a/tasks/metaworldplus-handle-pull-side-i00-standard/task.toml b/tasks/metaworldplus-handle-pull-side-i00-standard/task.toml
index f595cf9eef..1ffd70da73 100644
--- a/tasks/metaworldplus-handle-pull-side-i00-standard/task.toml
+++ b/tasks/metaworldplus-handle-pull-side-i00-standard/task.toml
@@ -23,6 +23,7 @@ simulator = "mujoco"
robot = "sawyer"
family = "metaworld/handle-pull-side"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 0
diff --git a/tasks/metaworldplus-lever-pull-i00-privileged/task.toml b/tasks/metaworldplus-lever-pull-i00-privileged/task.toml
index b57364a798..920bb15338 100644
--- a/tasks/metaworldplus-lever-pull-i00-privileged/task.toml
+++ b/tasks/metaworldplus-lever-pull-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "mujoco"
robot = "sawyer"
family = "metaworld/lever-pull"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 0 # Meta-World's reset randomization (object and goal positions)
expert_time_estimate_hours = 0.5
diff --git a/tasks/metaworldplus-lever-pull-i00-standard/task.toml b/tasks/metaworldplus-lever-pull-i00-standard/task.toml
index 3f71853cc6..574f454321 100644
--- a/tasks/metaworldplus-lever-pull-i00-standard/task.toml
+++ b/tasks/metaworldplus-lever-pull-i00-standard/task.toml
@@ -23,6 +23,7 @@ simulator = "mujoco"
robot = "sawyer"
family = "metaworld/lever-pull"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 0
diff --git a/tasks/metaworldplus-peg-insert-side-i00-privileged/task.toml b/tasks/metaworldplus-peg-insert-side-i00-privileged/task.toml
index 8aa83770ce..a2a7d94bf0 100644
--- a/tasks/metaworldplus-peg-insert-side-i00-privileged/task.toml
+++ b/tasks/metaworldplus-peg-insert-side-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "mujoco"
robot = "sawyer"
family = "metaworld/peg-insert-side"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 0 # Meta-World's reset randomization (object and goal positions)
expert_time_estimate_hours = 0.5
diff --git a/tasks/metaworldplus-peg-insert-side-i00-standard/task.toml b/tasks/metaworldplus-peg-insert-side-i00-standard/task.toml
index fd2e36ce8f..b81aedff86 100644
--- a/tasks/metaworldplus-peg-insert-side-i00-standard/task.toml
+++ b/tasks/metaworldplus-peg-insert-side-i00-standard/task.toml
@@ -23,6 +23,7 @@ simulator = "mujoco"
robot = "sawyer"
family = "metaworld/peg-insert-side"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 0
diff --git a/tasks/metaworldplus-peg-unplug-side-i00-privileged/task.toml b/tasks/metaworldplus-peg-unplug-side-i00-privileged/task.toml
index b53d8987e8..32f285a9b9 100644
--- a/tasks/metaworldplus-peg-unplug-side-i00-privileged/task.toml
+++ b/tasks/metaworldplus-peg-unplug-side-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "mujoco"
robot = "sawyer"
family = "metaworld/peg-unplug-side"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 0 # Meta-World's reset randomization (object and goal positions)
expert_time_estimate_hours = 0.5
diff --git a/tasks/metaworldplus-peg-unplug-side-i00-standard/task.toml b/tasks/metaworldplus-peg-unplug-side-i00-standard/task.toml
index 4bffb648ba..7bee91d462 100644
--- a/tasks/metaworldplus-peg-unplug-side-i00-standard/task.toml
+++ b/tasks/metaworldplus-peg-unplug-side-i00-standard/task.toml
@@ -23,6 +23,7 @@ simulator = "mujoco"
robot = "sawyer"
family = "metaworld/peg-unplug-side"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 0
diff --git a/tasks/metaworldplus-pick-out-of-hole-i00-privileged/task.toml b/tasks/metaworldplus-pick-out-of-hole-i00-privileged/task.toml
index a84bac48ce..3e982a8ca0 100644
--- a/tasks/metaworldplus-pick-out-of-hole-i00-privileged/task.toml
+++ b/tasks/metaworldplus-pick-out-of-hole-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "mujoco"
robot = "sawyer"
family = "metaworld/pick-out-of-hole"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 0 # Meta-World's reset randomization (object and goal positions)
expert_time_estimate_hours = 0.5
diff --git a/tasks/metaworldplus-pick-out-of-hole-i00-standard/task.toml b/tasks/metaworldplus-pick-out-of-hole-i00-standard/task.toml
index c36bd0d0de..365f0ef7ad 100644
--- a/tasks/metaworldplus-pick-out-of-hole-i00-standard/task.toml
+++ b/tasks/metaworldplus-pick-out-of-hole-i00-standard/task.toml
@@ -23,6 +23,7 @@ simulator = "mujoco"
robot = "sawyer"
family = "metaworld/pick-out-of-hole"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 0
diff --git a/tasks/metaworldplus-pick-place-i00-privileged/task.toml b/tasks/metaworldplus-pick-place-i00-privileged/task.toml
index 00e4331b7c..11e7afadfe 100644
--- a/tasks/metaworldplus-pick-place-i00-privileged/task.toml
+++ b/tasks/metaworldplus-pick-place-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "mujoco"
robot = "sawyer"
family = "metaworld/pick-place"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 0 # Meta-World's reset randomization (object and goal positions)
expert_time_estimate_hours = 0.5
diff --git a/tasks/metaworldplus-pick-place-i00-standard/task.toml b/tasks/metaworldplus-pick-place-i00-standard/task.toml
index bd868e4903..d9dd89e5e7 100644
--- a/tasks/metaworldplus-pick-place-i00-standard/task.toml
+++ b/tasks/metaworldplus-pick-place-i00-standard/task.toml
@@ -23,6 +23,7 @@ simulator = "mujoco"
robot = "sawyer"
family = "metaworld/pick-place"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 0
diff --git a/tasks/metaworldplus-pick-place-wall-i00-privileged/task.toml b/tasks/metaworldplus-pick-place-wall-i00-privileged/task.toml
index 41f72a808c..17e898263d 100644
--- a/tasks/metaworldplus-pick-place-wall-i00-privileged/task.toml
+++ b/tasks/metaworldplus-pick-place-wall-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "mujoco"
robot = "sawyer"
family = "metaworld/pick-place-wall"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 0 # Meta-World's reset randomization (object and goal positions)
expert_time_estimate_hours = 0.5
diff --git a/tasks/metaworldplus-pick-place-wall-i00-standard/task.toml b/tasks/metaworldplus-pick-place-wall-i00-standard/task.toml
index a736dfcd1b..b9ae2a6634 100644
--- a/tasks/metaworldplus-pick-place-wall-i00-standard/task.toml
+++ b/tasks/metaworldplus-pick-place-wall-i00-standard/task.toml
@@ -23,6 +23,7 @@ simulator = "mujoco"
robot = "sawyer"
family = "metaworld/pick-place-wall"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 0
diff --git a/tasks/metaworldplus-plate-slide-back-i00-privileged/task.toml b/tasks/metaworldplus-plate-slide-back-i00-privileged/task.toml
index ab3f9f4b6a..e3d67989af 100644
--- a/tasks/metaworldplus-plate-slide-back-i00-privileged/task.toml
+++ b/tasks/metaworldplus-plate-slide-back-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "mujoco"
robot = "sawyer"
family = "metaworld/plate-slide-back"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 0 # Meta-World's reset randomization (object and goal positions)
expert_time_estimate_hours = 0.5
diff --git a/tasks/metaworldplus-plate-slide-back-i00-standard/task.toml b/tasks/metaworldplus-plate-slide-back-i00-standard/task.toml
index abc2a0d062..faf309bda1 100644
--- a/tasks/metaworldplus-plate-slide-back-i00-standard/task.toml
+++ b/tasks/metaworldplus-plate-slide-back-i00-standard/task.toml
@@ -23,6 +23,7 @@ simulator = "mujoco"
robot = "sawyer"
family = "metaworld/plate-slide-back"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 0
diff --git a/tasks/metaworldplus-plate-slide-back-side-i00-privileged/task.toml b/tasks/metaworldplus-plate-slide-back-side-i00-privileged/task.toml
index 726f724398..d59eea314f 100644
--- a/tasks/metaworldplus-plate-slide-back-side-i00-privileged/task.toml
+++ b/tasks/metaworldplus-plate-slide-back-side-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "mujoco"
robot = "sawyer"
family = "metaworld/plate-slide-back-side"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 0 # Meta-World's reset randomization (object and goal positions)
expert_time_estimate_hours = 0.5
diff --git a/tasks/metaworldplus-plate-slide-back-side-i00-standard/task.toml b/tasks/metaworldplus-plate-slide-back-side-i00-standard/task.toml
index e40c189829..73333aebcd 100644
--- a/tasks/metaworldplus-plate-slide-back-side-i00-standard/task.toml
+++ b/tasks/metaworldplus-plate-slide-back-side-i00-standard/task.toml
@@ -23,6 +23,7 @@ simulator = "mujoco"
robot = "sawyer"
family = "metaworld/plate-slide-back-side"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 0
diff --git a/tasks/metaworldplus-plate-slide-i00-privileged/task.toml b/tasks/metaworldplus-plate-slide-i00-privileged/task.toml
index f6f40ce229..151d0a9d55 100644
--- a/tasks/metaworldplus-plate-slide-i00-privileged/task.toml
+++ b/tasks/metaworldplus-plate-slide-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "mujoco"
robot = "sawyer"
family = "metaworld/plate-slide"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 0 # Meta-World's reset randomization (object and goal positions)
expert_time_estimate_hours = 0.5
diff --git a/tasks/metaworldplus-plate-slide-i00-standard/task.toml b/tasks/metaworldplus-plate-slide-i00-standard/task.toml
index f0bfd00fb3..631e1fcbcc 100644
--- a/tasks/metaworldplus-plate-slide-i00-standard/task.toml
+++ b/tasks/metaworldplus-plate-slide-i00-standard/task.toml
@@ -23,6 +23,7 @@ simulator = "mujoco"
robot = "sawyer"
family = "metaworld/plate-slide"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 0
diff --git a/tasks/metaworldplus-plate-slide-side-i00-privileged/task.toml b/tasks/metaworldplus-plate-slide-side-i00-privileged/task.toml
index 31476dcdc3..a2455d288a 100644
--- a/tasks/metaworldplus-plate-slide-side-i00-privileged/task.toml
+++ b/tasks/metaworldplus-plate-slide-side-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "mujoco"
robot = "sawyer"
family = "metaworld/plate-slide-side"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 0 # Meta-World's reset randomization (object and goal positions)
expert_time_estimate_hours = 0.5
diff --git a/tasks/metaworldplus-plate-slide-side-i00-standard/task.toml b/tasks/metaworldplus-plate-slide-side-i00-standard/task.toml
index e3936e00d5..da5b76300c 100644
--- a/tasks/metaworldplus-plate-slide-side-i00-standard/task.toml
+++ b/tasks/metaworldplus-plate-slide-side-i00-standard/task.toml
@@ -23,6 +23,7 @@ simulator = "mujoco"
robot = "sawyer"
family = "metaworld/plate-slide-side"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 0
diff --git a/tasks/metaworldplus-push-back-i00-privileged/task.toml b/tasks/metaworldplus-push-back-i00-privileged/task.toml
index cb2fd83f32..f0197b4de0 100644
--- a/tasks/metaworldplus-push-back-i00-privileged/task.toml
+++ b/tasks/metaworldplus-push-back-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "mujoco"
robot = "sawyer"
family = "metaworld/push-back"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 0 # Meta-World's reset randomization (object and goal positions)
expert_time_estimate_hours = 0.5
diff --git a/tasks/metaworldplus-push-back-i00-standard/task.toml b/tasks/metaworldplus-push-back-i00-standard/task.toml
index 698cfc687b..78aa39737f 100644
--- a/tasks/metaworldplus-push-back-i00-standard/task.toml
+++ b/tasks/metaworldplus-push-back-i00-standard/task.toml
@@ -23,6 +23,7 @@ simulator = "mujoco"
robot = "sawyer"
family = "metaworld/push-back"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 0
diff --git a/tasks/metaworldplus-push-i00-privileged/task.toml b/tasks/metaworldplus-push-i00-privileged/task.toml
index 6cc47ed605..8884ffe2cb 100644
--- a/tasks/metaworldplus-push-i00-privileged/task.toml
+++ b/tasks/metaworldplus-push-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "mujoco"
robot = "sawyer"
family = "metaworld/push"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 0 # Meta-World's reset randomization (object and goal positions)
expert_time_estimate_hours = 0.5
diff --git a/tasks/metaworldplus-push-i00-standard/task.toml b/tasks/metaworldplus-push-i00-standard/task.toml
index 7419d55415..8a10c473bb 100644
--- a/tasks/metaworldplus-push-i00-standard/task.toml
+++ b/tasks/metaworldplus-push-i00-standard/task.toml
@@ -23,6 +23,7 @@ simulator = "mujoco"
robot = "sawyer"
family = "metaworld/push"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 0
diff --git a/tasks/metaworldplus-push-wall-i00-privileged/task.toml b/tasks/metaworldplus-push-wall-i00-privileged/task.toml
index 34e1636b2c..438b17c8a9 100644
--- a/tasks/metaworldplus-push-wall-i00-privileged/task.toml
+++ b/tasks/metaworldplus-push-wall-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "mujoco"
robot = "sawyer"
family = "metaworld/push-wall"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 0 # Meta-World's reset randomization (object and goal positions)
expert_time_estimate_hours = 0.5
diff --git a/tasks/metaworldplus-push-wall-i00-standard/task.toml b/tasks/metaworldplus-push-wall-i00-standard/task.toml
index 84d7a19a31..9690e9857c 100644
--- a/tasks/metaworldplus-push-wall-i00-standard/task.toml
+++ b/tasks/metaworldplus-push-wall-i00-standard/task.toml
@@ -23,6 +23,7 @@ simulator = "mujoco"
robot = "sawyer"
family = "metaworld/push-wall"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 0
diff --git a/tasks/metaworldplus-reach-i00-privileged/task.toml b/tasks/metaworldplus-reach-i00-privileged/task.toml
index 3f7147d4cc..9c5cf7a22e 100644
--- a/tasks/metaworldplus-reach-i00-privileged/task.toml
+++ b/tasks/metaworldplus-reach-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "mujoco"
robot = "sawyer"
family = "metaworld/reach"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 0 # Meta-World's reset randomization (object and goal positions)
expert_time_estimate_hours = 0.5
diff --git a/tasks/metaworldplus-reach-i00-standard/task.toml b/tasks/metaworldplus-reach-i00-standard/task.toml
index 20e705cae6..d9ddc7160f 100644
--- a/tasks/metaworldplus-reach-i00-standard/task.toml
+++ b/tasks/metaworldplus-reach-i00-standard/task.toml
@@ -23,6 +23,7 @@ simulator = "mujoco"
robot = "sawyer"
family = "metaworld/reach"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 0
diff --git a/tasks/metaworldplus-reach-wall-i00-privileged/task.toml b/tasks/metaworldplus-reach-wall-i00-privileged/task.toml
index 7280cffa20..6a868a0c9a 100644
--- a/tasks/metaworldplus-reach-wall-i00-privileged/task.toml
+++ b/tasks/metaworldplus-reach-wall-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "mujoco"
robot = "sawyer"
family = "metaworld/reach-wall"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 0 # Meta-World's reset randomization (object and goal positions)
expert_time_estimate_hours = 0.5
diff --git a/tasks/metaworldplus-reach-wall-i00-standard/task.toml b/tasks/metaworldplus-reach-wall-i00-standard/task.toml
index 6ead70e744..567c7b3e4a 100644
--- a/tasks/metaworldplus-reach-wall-i00-standard/task.toml
+++ b/tasks/metaworldplus-reach-wall-i00-standard/task.toml
@@ -23,6 +23,7 @@ simulator = "mujoco"
robot = "sawyer"
family = "metaworld/reach-wall"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 0
diff --git a/tasks/metaworldplus-shelf-place-i00-privileged/task.toml b/tasks/metaworldplus-shelf-place-i00-privileged/task.toml
index 9d9dc06965..c2d729e4aa 100644
--- a/tasks/metaworldplus-shelf-place-i00-privileged/task.toml
+++ b/tasks/metaworldplus-shelf-place-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "mujoco"
robot = "sawyer"
family = "metaworld/shelf-place"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 0 # Meta-World's reset randomization (object and goal positions)
expert_time_estimate_hours = 0.5
diff --git a/tasks/metaworldplus-shelf-place-i00-standard/task.toml b/tasks/metaworldplus-shelf-place-i00-standard/task.toml
index 3a686b33b9..b390cbf31b 100644
--- a/tasks/metaworldplus-shelf-place-i00-standard/task.toml
+++ b/tasks/metaworldplus-shelf-place-i00-standard/task.toml
@@ -23,6 +23,7 @@ simulator = "mujoco"
robot = "sawyer"
family = "metaworld/shelf-place"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 0
diff --git a/tasks/metaworldplus-soccer-i00-privileged/task.toml b/tasks/metaworldplus-soccer-i00-privileged/task.toml
index b59539a207..9094c6f70e 100644
--- a/tasks/metaworldplus-soccer-i00-privileged/task.toml
+++ b/tasks/metaworldplus-soccer-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "mujoco"
robot = "sawyer"
family = "metaworld/soccer"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 1 # Meta-World's reset randomization (object and goal positions)
expert_time_estimate_hours = 0.5
diff --git a/tasks/metaworldplus-soccer-i00-standard/task.toml b/tasks/metaworldplus-soccer-i00-standard/task.toml
index 16f2989e87..3255e1391d 100644
--- a/tasks/metaworldplus-soccer-i00-standard/task.toml
+++ b/tasks/metaworldplus-soccer-i00-standard/task.toml
@@ -23,6 +23,7 @@ simulator = "mujoco"
robot = "sawyer"
family = "metaworld/soccer"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 1
diff --git a/tasks/metaworldplus-stick-pull-i00-privileged/task.toml b/tasks/metaworldplus-stick-pull-i00-privileged/task.toml
index 234d0ffe82..a97365795d 100644
--- a/tasks/metaworldplus-stick-pull-i00-privileged/task.toml
+++ b/tasks/metaworldplus-stick-pull-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "mujoco"
robot = "sawyer"
family = "metaworld/stick-pull"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 1 # Meta-World's reset randomization (object and goal positions)
expert_time_estimate_hours = 0.5
diff --git a/tasks/metaworldplus-stick-pull-i00-standard/task.toml b/tasks/metaworldplus-stick-pull-i00-standard/task.toml
index 472ec92fac..317e8e84af 100644
--- a/tasks/metaworldplus-stick-pull-i00-standard/task.toml
+++ b/tasks/metaworldplus-stick-pull-i00-standard/task.toml
@@ -23,6 +23,7 @@ simulator = "mujoco"
robot = "sawyer"
family = "metaworld/stick-pull"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 1
diff --git a/tasks/metaworldplus-stick-push-i00-privileged/task.toml b/tasks/metaworldplus-stick-push-i00-privileged/task.toml
index fca08c75c0..833f2bfe5e 100644
--- a/tasks/metaworldplus-stick-push-i00-privileged/task.toml
+++ b/tasks/metaworldplus-stick-push-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "mujoco"
robot = "sawyer"
family = "metaworld/stick-push"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 0 # Meta-World's reset randomization (object and goal positions)
expert_time_estimate_hours = 0.5
diff --git a/tasks/metaworldplus-stick-push-i00-standard/task.toml b/tasks/metaworldplus-stick-push-i00-standard/task.toml
index b94b89ca00..f926989ff3 100644
--- a/tasks/metaworldplus-stick-push-i00-standard/task.toml
+++ b/tasks/metaworldplus-stick-push-i00-standard/task.toml
@@ -23,6 +23,7 @@ simulator = "mujoco"
robot = "sawyer"
family = "metaworld/stick-push"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 0
diff --git a/tasks/metaworldplus-sweep-i00-privileged/task.toml b/tasks/metaworldplus-sweep-i00-privileged/task.toml
index c4d55c0519..fb0b6741d8 100644
--- a/tasks/metaworldplus-sweep-i00-privileged/task.toml
+++ b/tasks/metaworldplus-sweep-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "mujoco"
robot = "sawyer"
family = "metaworld/sweep"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 0 # Meta-World's reset randomization (object and goal positions)
expert_time_estimate_hours = 0.5
diff --git a/tasks/metaworldplus-sweep-i00-standard/task.toml b/tasks/metaworldplus-sweep-i00-standard/task.toml
index 76ba837b2a..f905237c42 100644
--- a/tasks/metaworldplus-sweep-i00-standard/task.toml
+++ b/tasks/metaworldplus-sweep-i00-standard/task.toml
@@ -23,6 +23,7 @@ simulator = "mujoco"
robot = "sawyer"
family = "metaworld/sweep"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 0
diff --git a/tasks/metaworldplus-sweep-into-i00-privileged/task.toml b/tasks/metaworldplus-sweep-into-i00-privileged/task.toml
index 8089010984..ba3ad2301a 100644
--- a/tasks/metaworldplus-sweep-into-i00-privileged/task.toml
+++ b/tasks/metaworldplus-sweep-into-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "mujoco"
robot = "sawyer"
family = "metaworld/sweep-into"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 0 # Meta-World's reset randomization (object and goal positions)
expert_time_estimate_hours = 0.5
diff --git a/tasks/metaworldplus-sweep-into-i00-standard/task.toml b/tasks/metaworldplus-sweep-into-i00-standard/task.toml
index 4d2be44d1c..f9718c8a54 100644
--- a/tasks/metaworldplus-sweep-into-i00-standard/task.toml
+++ b/tasks/metaworldplus-sweep-into-i00-standard/task.toml
@@ -23,6 +23,7 @@ simulator = "mujoco"
robot = "sawyer"
family = "metaworld/sweep-into"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 0
diff --git a/tasks/metaworldplus-window-close-i00-privileged/task.toml b/tasks/metaworldplus-window-close-i00-privileged/task.toml
index bc74857231..6476626aac 100644
--- a/tasks/metaworldplus-window-close-i00-privileged/task.toml
+++ b/tasks/metaworldplus-window-close-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "mujoco"
robot = "sawyer"
family = "metaworld/window-close"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 0 # Meta-World's reset randomization (object and goal positions)
expert_time_estimate_hours = 0.5
diff --git a/tasks/metaworldplus-window-close-i00-standard/task.toml b/tasks/metaworldplus-window-close-i00-standard/task.toml
index 0bd0a447a7..02cdd6912b 100644
--- a/tasks/metaworldplus-window-close-i00-standard/task.toml
+++ b/tasks/metaworldplus-window-close-i00-standard/task.toml
@@ -23,6 +23,7 @@ simulator = "mujoco"
robot = "sawyer"
family = "metaworld/window-close"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 0
diff --git a/tasks/metaworldplus-window-open-i00-privileged/task.toml b/tasks/metaworldplus-window-open-i00-privileged/task.toml
index 3ea4beb943..efa9d84707 100644
--- a/tasks/metaworldplus-window-open-i00-privileged/task.toml
+++ b/tasks/metaworldplus-window-open-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "mujoco"
robot = "sawyer"
family = "metaworld/window-open"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 0 # Meta-World's reset randomization (object and goal positions)
expert_time_estimate_hours = 0.5
diff --git a/tasks/metaworldplus-window-open-i00-standard/task.toml b/tasks/metaworldplus-window-open-i00-standard/task.toml
index 28f1c76b44..ee6db5c401 100644
--- a/tasks/metaworldplus-window-open-i00-standard/task.toml
+++ b/tasks/metaworldplus-window-open-i00-standard/task.toml
@@ -23,6 +23,7 @@ simulator = "mujoco"
robot = "sawyer"
family = "metaworld/window-open"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 0
diff --git a/tasks/molmospaces-close-i00-privileged/task.toml b/tasks/molmospaces-close-i00-privileged/task.toml
index 2eb6d57173..9bcffc2946 100644
--- a/tasks/molmospaces-close-i00-privileged/task.toml
+++ b/tasks/molmospaces-close-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "molmospaces" # MolmoSpaces (molmo-spaces 0.2.9) on MuJoCo
robot = "franka-robotiq-2f85"
family = "close"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
# scene = environment/instance.json (one upstream benchmark episode + its t = 0 state hash); identity = the dataset.toml digest
expert_time_estimate_hours = 2.0
oracle = "none"
diff --git a/tasks/molmospaces-close-i00-standard/task.toml b/tasks/molmospaces-close-i00-standard/task.toml
index 04a11f5c30..13d2db391c 100644
--- a/tasks/molmospaces-close-i00-standard/task.toml
+++ b/tasks/molmospaces-close-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "molmospaces"
robot = "franka-robotiq-2f85"
family = "close"
mode = "standard" # robot as a service: images + proprioception only; no object poses, no success check, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
expert_time_estimate_hours = 3.0
oracle = "none"
diff --git a/tasks/molmospaces-navigate-to-i00-privileged/task.toml b/tasks/molmospaces-navigate-to-i00-privileged/task.toml
index 07dc76edc5..10e541a72f 100644
--- a/tasks/molmospaces-navigate-to-i00-privileged/task.toml
+++ b/tasks/molmospaces-navigate-to-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "molmospaces" # MolmoSpaces (molmo-spaces 0.2.9) on MuJoCo
robot = "rainbow-rby1"
family = "navigate-to"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
# scene = environment/instance.json (one upstream benchmark episode + its t = 0 state hash); identity = the dataset.toml digest
expert_time_estimate_hours = 2.0
oracle = "none"
diff --git a/tasks/molmospaces-navigate-to-i00-standard/task.toml b/tasks/molmospaces-navigate-to-i00-standard/task.toml
index b0b785915f..1893bda14e 100644
--- a/tasks/molmospaces-navigate-to-i00-standard/task.toml
+++ b/tasks/molmospaces-navigate-to-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "molmospaces"
robot = "rainbow-rby1"
family = "navigate-to"
mode = "standard" # robot as a service: images + proprioception only; no object poses, no success check, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
expert_time_estimate_hours = 3.0
oracle = "none"
diff --git a/tasks/molmospaces-open-door-i00-privileged/task.toml b/tasks/molmospaces-open-door-i00-privileged/task.toml
index b3d311298f..41f061a321 100644
--- a/tasks/molmospaces-open-door-i00-privileged/task.toml
+++ b/tasks/molmospaces-open-door-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "molmospaces" # MolmoSpaces (molmo-spaces 0.2.9) on MuJoCo
robot = "rainbow-rby1"
family = "open-door"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
# scene = environment/instance.json (one upstream benchmark episode + its t = 0 state hash); identity = the dataset.toml digest
expert_time_estimate_hours = 2.0
oracle = "none"
diff --git a/tasks/molmospaces-open-door-i00-standard/task.toml b/tasks/molmospaces-open-door-i00-standard/task.toml
index 9892e5e949..35cc6a4586 100644
--- a/tasks/molmospaces-open-door-i00-standard/task.toml
+++ b/tasks/molmospaces-open-door-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "molmospaces"
robot = "rainbow-rby1"
family = "open-door"
mode = "standard" # robot as a service: images + proprioception only; no object poses, no success check, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
expert_time_estimate_hours = 3.0
oracle = "none"
diff --git a/tasks/molmospaces-open-i00-privileged/task.toml b/tasks/molmospaces-open-i00-privileged/task.toml
index 72cde42ddf..4a184ffef2 100644
--- a/tasks/molmospaces-open-i00-privileged/task.toml
+++ b/tasks/molmospaces-open-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "molmospaces" # MolmoSpaces (molmo-spaces 0.2.9) on MuJoCo
robot = "franka-robotiq-2f85"
family = "open"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
# scene = environment/instance.json (one upstream benchmark episode + its t = 0 state hash); identity = the dataset.toml digest
expert_time_estimate_hours = 2.0
oracle = "none"
diff --git a/tasks/molmospaces-open-i00-standard/task.toml b/tasks/molmospaces-open-i00-standard/task.toml
index e161fd8a67..65b832d107 100644
--- a/tasks/molmospaces-open-i00-standard/task.toml
+++ b/tasks/molmospaces-open-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "molmospaces"
robot = "franka-robotiq-2f85"
family = "open"
mode = "standard" # robot as a service: images + proprioception only; no object poses, no success check, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
expert_time_estimate_hours = 3.0
oracle = "none"
diff --git a/tasks/molmospaces-pick-and-place-color-i00-privileged/task.toml b/tasks/molmospaces-pick-and-place-color-i00-privileged/task.toml
index fc69035a32..e369718a4e 100644
--- a/tasks/molmospaces-pick-and-place-color-i00-privileged/task.toml
+++ b/tasks/molmospaces-pick-and-place-color-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "molmospaces" # MolmoSpaces (molmo-spaces 0.2.9) on MuJoCo
robot = "franka-robotiq-2f85"
family = "pick-and-place-color"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
# scene = environment/instance.json (one upstream benchmark episode + its t = 0 state hash); identity = the dataset.toml digest
expert_time_estimate_hours = 2.0
oracle = "none"
diff --git a/tasks/molmospaces-pick-and-place-color-i00-standard/task.toml b/tasks/molmospaces-pick-and-place-color-i00-standard/task.toml
index 0cd221e3f0..49c57e321e 100644
--- a/tasks/molmospaces-pick-and-place-color-i00-standard/task.toml
+++ b/tasks/molmospaces-pick-and-place-color-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "molmospaces"
robot = "franka-robotiq-2f85"
family = "pick-and-place-color"
mode = "standard" # robot as a service: images + proprioception only; no object poses, no success check, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
expert_time_estimate_hours = 3.0
oracle = "none"
diff --git a/tasks/molmospaces-pick-and-place-i00-privileged/task.toml b/tasks/molmospaces-pick-and-place-i00-privileged/task.toml
index 6a31dac147..5d912a2256 100644
--- a/tasks/molmospaces-pick-and-place-i00-privileged/task.toml
+++ b/tasks/molmospaces-pick-and-place-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "molmospaces" # MolmoSpaces (molmo-spaces 0.2.9) on MuJoCo
robot = "franka-robotiq-2f85"
family = "pick-and-place"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
# scene = environment/instance.json (one upstream benchmark episode + its t = 0 state hash); identity = the dataset.toml digest
expert_time_estimate_hours = 2.0
oracle = "none"
diff --git a/tasks/molmospaces-pick-and-place-i00-standard/task.toml b/tasks/molmospaces-pick-and-place-i00-standard/task.toml
index 655e2079f1..54f9e698e5 100644
--- a/tasks/molmospaces-pick-and-place-i00-standard/task.toml
+++ b/tasks/molmospaces-pick-and-place-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "molmospaces"
robot = "franka-robotiq-2f85"
family = "pick-and-place"
mode = "standard" # robot as a service: images + proprioception only; no object poses, no success check, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
expert_time_estimate_hours = 3.0
oracle = "none"
diff --git a/tasks/molmospaces-pick-and-place-next-to-i00-privileged/task.toml b/tasks/molmospaces-pick-and-place-next-to-i00-privileged/task.toml
index 914413b03b..755894c12c 100644
--- a/tasks/molmospaces-pick-and-place-next-to-i00-privileged/task.toml
+++ b/tasks/molmospaces-pick-and-place-next-to-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "molmospaces" # MolmoSpaces (molmo-spaces 0.2.9) on MuJoCo
robot = "franka-robotiq-2f85"
family = "pick-and-place-next-to"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
# scene = environment/instance.json (one upstream benchmark episode + its t = 0 state hash); identity = the dataset.toml digest
expert_time_estimate_hours = 2.0
oracle = "none"
diff --git a/tasks/molmospaces-pick-and-place-next-to-i00-standard/task.toml b/tasks/molmospaces-pick-and-place-next-to-i00-standard/task.toml
index 320e741ce4..c046bb624d 100644
--- a/tasks/molmospaces-pick-and-place-next-to-i00-standard/task.toml
+++ b/tasks/molmospaces-pick-and-place-next-to-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "molmospaces"
robot = "franka-robotiq-2f85"
family = "pick-and-place-next-to"
mode = "standard" # robot as a service: images + proprioception only; no object poses, no success check, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
expert_time_estimate_hours = 3.0
oracle = "none"
diff --git a/tasks/molmospaces-pick-i00-privileged/task.toml b/tasks/molmospaces-pick-i00-privileged/task.toml
index bd1742355f..48b01d7814 100644
--- a/tasks/molmospaces-pick-i00-privileged/task.toml
+++ b/tasks/molmospaces-pick-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "molmospaces" # MolmoSpaces (molmo-spaces 0.2.9) on MuJoCo
robot = "franka-robotiq-2f85"
family = "pick"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
# scene = environment/instance.json (one upstream benchmark episode + its t = 0 state hash); identity = the dataset.toml digest
expert_time_estimate_hours = 2.0
oracle = "none"
diff --git a/tasks/molmospaces-pick-i00-standard/task.toml b/tasks/molmospaces-pick-i00-standard/task.toml
index 10a474af79..0c6c3d6af3 100644
--- a/tasks/molmospaces-pick-i00-standard/task.toml
+++ b/tasks/molmospaces-pick-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "molmospaces"
robot = "franka-robotiq-2f85"
family = "pick"
mode = "standard" # robot as a service: images + proprioception only; no object poses, no success check, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
expert_time_estimate_hours = 3.0
oracle = "none"
diff --git a/tasks/mujoco_playground_aloha_hand_over-privileged/task.toml b/tasks/mujoco_playground_aloha_hand_over-privileged/task.toml
index a661043af7..665d7ebdea 100644
--- a/tasks/mujoco_playground_aloha_hand_over-privileged/task.toml
+++ b/tasks/mujoco_playground_aloha_hand_over-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "mujoco"
robot = "aloha-2"
family = "mujoco_playground/AlohaHandOver"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 60417 # Playground's reset randomization (box xy offset, target position) with this seed
expert_time_estimate_hours = 2.0
diff --git a/tasks/mujoco_playground_aloha_hand_over-standard/task.toml b/tasks/mujoco_playground_aloha_hand_over-standard/task.toml
index 409f21f671..599dc5aaf3 100644
--- a/tasks/mujoco_playground_aloha_hand_over-standard/task.toml
+++ b/tasks/mujoco_playground_aloha_hand_over-standard/task.toml
@@ -24,6 +24,7 @@ family = "mujoco_playground/AlohaHandOver"
instance_seed = 60417 # Playground's reset randomization (box xy offset, target position) with this seed
expert_time_estimate_hours = 2.0
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
diff --git a/tasks/mujoco_playground_aloha_single_peg_insertion-privileged/task.toml b/tasks/mujoco_playground_aloha_single_peg_insertion-privileged/task.toml
index 0c7d28b17b..9e059b630a 100644
--- a/tasks/mujoco_playground_aloha_single_peg_insertion-privileged/task.toml
+++ b/tasks/mujoco_playground_aloha_single_peg_insertion-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "mujoco"
robot = "aloha-2"
family = "mujoco_playground/AlohaSinglePegInsertion"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 60417 # Playground's reset randomization (peg and socket xy offsets) with this seed
expert_time_estimate_hours = 2.0
diff --git a/tasks/mujoco_playground_aloha_single_peg_insertion-standard/task.toml b/tasks/mujoco_playground_aloha_single_peg_insertion-standard/task.toml
index 4ee82748b9..39da322ca9 100644
--- a/tasks/mujoco_playground_aloha_single_peg_insertion-standard/task.toml
+++ b/tasks/mujoco_playground_aloha_single_peg_insertion-standard/task.toml
@@ -24,6 +24,7 @@ family = "mujoco_playground/AlohaSinglePegInsertion"
instance_seed = 60417 # Playground's reset randomization (peg and socket xy offsets) with this seed
expert_time_estimate_hours = 2.0
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
diff --git a/tasks/mujoco_playground_barkour_joystick-privileged/README.md b/tasks/mujoco_playground_barkour_joystick-privileged/README.md
index 55676588d9..60d478f987 100644
--- a/tasks/mujoco_playground_barkour_joystick-privileged/README.md
+++ b/tasks/mujoco_playground_barkour_joystick-privileged/README.md
@@ -53,3 +53,11 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
### Provenance
Generated by `scripts/mujoco-playground/generate.py` from `scripts/mujoco-playground/instances/BarkourJoystick.json`
(`make_instances.py`, `select_instances.py`) and `scripts/mujoco-playground/oracles/BarkourJoystick.npy` (`oracle.py`).
+
+### Change Log
+- 2026-10-14 (review of board 0105, B1-m2): `tests/grade.py` only. A replay process that died or ran out of time, a
+ rebuilt start that does not match the instance's hash (`instance_mismatch`) and an exception of the verifier itself
+ now write `infra_error = 1` beside `grader_error = 1`: no verdict on the agent. Before, they wrote `grader_error = 1`
+ alone, which counted as the agent's 0. Every `reward.json` now carries `infra_error`. The scene, the instance, the
+ pass rule and the reference are unchanged, so the task keeps its name; a past 0 with `grader_error = 1` was an
+ infrastructure failure.
diff --git a/tasks/mujoco_playground_barkour_joystick-privileged/task.toml b/tasks/mujoco_playground_barkour_joystick-privileged/task.toml
index 6eab18b6ee..bb8cc5e520 100644
--- a/tasks/mujoco_playground_barkour_joystick-privileged/task.toml
+++ b/tasks/mujoco_playground_barkour_joystick-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "mujoco"
robot = "google-barkour-vb"
family = "mujoco_playground/BarkourJoystick"
mode = "privileged"
+dynamics = "dynamic" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 2 # the PRNGKey of Playground's reset (and, for the joystick tasks, its command process)
expert_time_estimate_hours = 3.0
diff --git a/tasks/mujoco_playground_barkour_joystick-privileged/tests/grade.py b/tasks/mujoco_playground_barkour_joystick-privileged/tests/grade.py
index c75c326868..1667d090c8 100644
--- a/tasks/mujoco_playground_barkour_joystick-privileged/tests/grade.py
+++ b/tasks/mujoco_playground_barkour_joystick-privileged/tests/grade.py
@@ -6,6 +6,11 @@ Two replays, each in its own fresh process (`rcb_quadruped.py replay`: the froze
task's success check on the whole trajectory). success = both replays succeed and end in the same state. Writes
/logs/verifier/reward.txt (1/0) and reward.json (numbers only); then a third replay renders replay.mp4, after the
verdict is on disk (it never changes it).
+
+Failures that are not the agent's (docs/PROTOCOL.md: infra vs agent failure) carry `grader_error` = 1 and `infra_error`
+= 1, so they are no verdict on the agent: a replay process that died or timed out, an image that no longer rebuilds the
+frozen start (`instance_mismatch`), and any exception of the verifier itself. A file the agent got wrong is the agent's
+0 (`bad_trajectory`, infra_error 0). Every reward.json carries success, deterministic, grader_error and infra_error.
"""
import argparse
import json
@@ -22,8 +27,12 @@ TOOL = os.path.join(os.path.dirname(os.path.abspath(__file__)), "rcb_quadruped.p
REPLAY_TIMEOUT_S = 300 # a 1000-step replay takes a few seconds; the video replay renders 4 views every 2nd step
+BASE = {"success": 0, "deterministic": 0, "grader_error": 0, "infra_error": 0}
+
+
def write(ok: bool, out: dict) -> None:
os.makedirs(LOG_DIR, exist_ok=True)
+ out = {**BASE, **out}
open(os.path.join(LOG_DIR, "reward.txt"), "w").write("1\n" if ok else "0\n")
json.dump(out, open(os.path.join(LOG_DIR, "reward.json"), "w"))
print(json.dumps(out), flush=True)
@@ -96,11 +105,11 @@ def main() -> int:
"n_actions": int(acts.shape[0]) if acts.ndim == 2 else 0, "grader_error": 0, **cost})
return 0
runs = [replay(a.instance, a.traj, None), replay(a.instance, a.traj, None)]
- if any("error" in r and not r.get("bad_trajectory") for r in runs):
- write(False, {"success": 0, "deterministic": 0, "grader_error": 1, **cost})
+ if any("error" in r and not r.get("bad_trajectory") for r in runs): # the replay process died / timed out
+ write(False, {"success": 0, "deterministic": 0, "grader_error": 1, "infra_error": 1, "replay_failed": 1, **cost})
return 0
if not all(r.get("instance_ok") for r in runs): # the image no longer builds the frozen start: not the agent's 0
- write(False, {"success": 0, "deterministic": 0, "grader_error": 1, "instance_mismatch": 1, **cost})
+ write(False, {"success": 0, "deterministic": 0, "grader_error": 1, "infra_error": 1, "instance_mismatch": 1, **cost})
return 0
all_success = all(bool(r.get("success")) for r in runs)
deterministic = runs[0].get("state_sha") == runs[1].get("state_sha")
@@ -116,7 +125,7 @@ def main() -> int:
write(ok, out) # the verdict, before any video
except Exception: # noqa: BLE001
traceback.print_exc(file=sys.stderr)
- write(False, {"success": 0, "deterministic": 0, "grader_error": 1, **cost})
+ write(False, {"success": 0, "deterministic": 0, "grader_error": 1, "infra_error": 1, **cost})
return 0
try: # the video, for humans: it never changes the verdict
if replay(a.instance, a.traj, os.path.join(LOG_DIR, "replay.mp4")).get("video"):
diff --git a/tasks/mujoco_playground_barkour_joystick-standard/README.md b/tasks/mujoco_playground_barkour_joystick-standard/README.md
index 752c54270e..52899d80d0 100644
--- a/tasks/mujoco_playground_barkour_joystick-standard/README.md
+++ b/tasks/mujoco_playground_barkour_joystick-standard/README.md
@@ -13,14 +13,14 @@
## Task
MuJoCo Playground `BarkourJoystick` (Google Barkour vB) on the same frozen instance as `mujoco_playground_barkour_joystick-privileged`: Make Google's Barkour quadruped follow the joystick command schedule for the whole 10 s episode, through random pushes, without falling over.
-For 10 s the robot gets a joystick command schedule (2 segments: planar velocity and yaw rate in its own frame) and must track it without falling over. Its torso is pushed 12 times. Standard mode (eai-standard/2.2): the agent drives the robot through a client in one episode, with only
-the robot's own sensors and four cameras that follow it.
+For 10 s the robot gets a joystick command schedule (2 segments: planar velocity and yaw rate in its own frame) and must track it without falling over. Its torso is pushed 12 times. Standard mode (eai-standard/2.2): the agent drives the robot through a client, with 10 resets (only
+the last episode is graded), the robot's own sensors and the cameras that follow it.
## Difficulty
-Same instance and success check as `mujoco_playground_barkour_joystick-privileged`. In standard mode the agent loses resets, the simulator and
+Same instance and success check as `mujoco_playground_barkour_joystick-privileged`. In standard mode the agent loses free resets, the simulator and
the world state: it sees camera images and proprioception (joints, IMU, foot contacts), acts in joint position chunks,
-and gets one episode, so it has to close the loop on the robot's own sensors and recover from its own mistakes without
-starting over.
+and gets 10 resets (`RCB_RESET_BUDGET`, the compose file's default; PROTOCOL §4.5), so it has to close the loop
+on the robot's own sensors; only the last episode is graded.
## Reference solution
`solution/oracle_replay.py` streams the privileged reference's rows (`solution/oracle.npy`, 1000 rows, float32-exact)
@@ -61,3 +61,4 @@ Generated by `scripts/mujoco-playground/generate.py`, as its privileged twin.
- 2026-10-10 (review 0067): `/spec` gives the control period of the `joint_position` action as 0.01 s; its text said "one per control step (1/50 s)", contradicting `control_hz = 100`, the `dt` check and the instruction.
- 2026-10-10 (review round 2, board 0067): the client only (`environment/rcb_client.py`). `Robot()` waits up to 3600 s for the service (was 600 s) and prints its progress to stderr once a minute, so the first `Robot()` outlasts the simulator's start (no effect on this benchmark, whose service is up within a minute); `cameras="name"` (a string) is one camera, no longer its characters. The service, the collect hook, the grader, the scene and the pass rule are unchanged.
- 2026-10-10 (review round 3, board 0067): the shared service layer and the client only (`environment/sim/rcb_limited.py`, `environment/rcb_client.py`). A camera named more than once in `cameras` is rendered and returned once (N copies of one name made the service build N x N images within the frame budget, which could exhaust its memory). A service that has given up on the episode (it could not start it, or could not resume it after its crashes) answers every request with 409 and the reason, and `Robot()` / `wait_ready()` raise `EpisodeOver` at once; before, nothing answered and they waited out their timeout (up to an hour). The adapter, the collect hook, the grader, the scene and the pass rule are unchanged.
+- 2026-10-14 (owner decision C1, board 0113): 10 resets (PROTOCOL §4.5). The service (`environment/sim/server.py`) gained the optional reset capability: `RCB_RESET_BUDGET` (the compose file's default 10; 0 restores the old single episode), `POST /reset` abandons the running or ended episode and starts a new one from the same frozen start (no physics), refused (409) after a success, once the agent's time is up and once the budget is spent; only the last episode is graded (`trajectory.npz` holds its rows, replayed twice from the frozen start as before); `/status` and every `meta` add `episode`, `resets_left` and `control_steps_total`; `final.json` and `reward.json` add `resets`, `reset_budget` and `episodes`. A crash resumes the current episode (the journal is replayed with the committed resets). The instruction's rules, the client example (`POST /reset` through `Robot._post`) and this README say so. The scene, the instance, the success check and the grader's pass rule are unchanged; results recorded before this change ran with one episode and are reported separately.
diff --git a/tasks/mujoco_playground_barkour_joystick-standard/environment/docker-compose.yaml b/tasks/mujoco_playground_barkour_joystick-standard/environment/docker-compose.yaml
index bc0c363241..1226b8c3a6 100644
--- a/tasks/mujoco_playground_barkour_joystick-standard/environment/docker-compose.yaml
+++ b/tasks/mujoco_playground_barkour_joystick-standard/environment/docker-compose.yaml
@@ -28,6 +28,7 @@ services:
RCB_STARTUP_TIMEOUT_S: "300" # the scene builds in about a second
PYTHONPATH: /srv
RCB_SESSION_LOG: "" # the sidecar's own Session must not write the agent-side interaction log
+ RCB_RESET_BUDGET: "${RCB_RESET_BUDGET:-10}" # Playground quadrupeds standard: 10 resets (PROTOCOL §4.5); 0: one episode
# >>> rcb_monitor (resource limits monitor, docs/PROTOCOL.md §2.10; generated by scripts/rcb_monitor/inject.py)
rcb_monitor:
image: ghcr.io/mll-lab-nu/eai-monitor:0.1.0@sha256:f2c910f53f79c2bfa4a71298b10c78383cae3afc40720ca562cc845b8d20fb0c
diff --git a/tasks/mujoco_playground_barkour_joystick-standard/environment/sim/server.py b/tasks/mujoco_playground_barkour_joystick-standard/environment/sim/server.py
index 108eca3cae..cad8920259 100644
--- a/tasks/mujoco_playground_barkour_joystick-standard/environment/sim/server.py
+++ b/tasks/mujoco_playground_barkour_joystick-standard/environment/sim/server.py
@@ -1,16 +1,20 @@
"""eai-standard/2.2 sim service for MuJoCo Playground's quadruped tasks (docs/STANDARD_MODE.md): the adapter of the
shared layer rcb_limited.py; the simulator is rcb_quadruped.Session (plain MuJoCo on the CPU).
-HTTP surface: GET /spec, GET /status, POST /observe, POST /submit, POST /kinematics. Every call is logged to
-/data/episode.jsonl.
+HTTP surface: GET /spec, GET /status, POST /observe, POST /submit, POST /kinematics, and POST /reset when the run grants
+resets (RCB_RESET_BUDGET, the optional reset capability of docs/STANDARD_MODE.md / PROTOCOL §4.5; the quadrupeds'
+standard tasks default to 10 in their compose file, owner 2026-10-14). Every call is logged to /data/episode.jsonl.
The episode is the frozen instance's: one control step = the instance's control period (Go1, Spot: 1/50 s; Barkour:
1/100 s). A joystick episode ends after the whole command schedule
(the horizon), when the robot meets Playground's termination rule (it falls over: a failure), when the agent submits
`done`, or when its time is over; success = the task's check on the steps executed. A getup episode ends the moment the
-check holds (success), after the horizon, on `done` or on timeout.
+check holds (success), after the horizon, on `done` or on timeout. With a reset budget, POST /reset abandons the running
+(or ended) episode and starts a new one from the frozen instance; only the last episode is graded, and a success ends
+the trial.
Crash recovery (rcb_limited.Episode, resume class `exact`: MuJoCo on the CPU repeats bit for bit in a fresh process):
-every executed row is journaled; after a crash the service rebuilds the episode and replays the journal. When the
+every executed row is journaled; after a crash the service rebuilds the episode and replays the journal, resetting where
+the committed resets did (their journal rows are checkpointed, `persist`). When the
episode ends the service writes /data/trajectory.npz + /data/episode.json; the collect hook (finalize.sh) makes it exit
and replays the trajectory in two fresh processes."""
@@ -31,18 +35,21 @@ INSTANCE = os.environ.get("RCB_INSTANCE", os.path.join(os.path.dirname(os.path.a
WALL_CLOCK_S = float(os.environ.get("RCB_WALL_CLOCK_S", "3600"))
MAX_CHUNK_S = float(os.environ.get("RCB_MAX_CHUNK_S", "10"))
MAX_WATCH_S = float(os.environ.get("RCB_MAX_WATCH_S", "30"))
+RESET_BUDGET = max(0, int(os.environ.get("RCB_RESET_BUDGET") or 0)) # the compose file's default: 10 (PROTOCOL §4.5)
CAM_W, CAM_H = 320, 240
LOCK = threading.Lock()
PID_FILE = os.path.join(DATA, "server.pid")
E = {"ready": False, "step": 0, "calls": 0, "done": False, "reason": None, "success": False, "t0": time.time(), "error": None,
- "t_first": None} # t_first: the agent's first call (its clock starts there; kept through a crash by the checkpoints)
+ "t_first": None, # t_first: the agent's first call (its clock starts there; kept through a crash by the checkpoints)
+ "resets": 0, # resets done (the episode is resets + 1)
+ "reset_at": []} # the journal row at which each reset happened: a resume resets there again
os.environ["RCB_SESSION_LOG"] = ""
os.environ["RCB_AGENT_TRACE_DIR"] = ""
import rcb_quadruped as Q # noqa: E402 (first: it picks the rendering backend before mujoco loads)
import mujoco # noqa: E402
-EP = L.Episode(DATA, dim=Q.N, resume="exact", persist=("t_first",))
+EP = L.Episode(DATA, dim=Q.N, resume="exact", persist=("t_first", "resets", "reset_at"))
EP.startup() # start count, back-off after failed starts, the startup watchdog until scene_built()
S = Q.Session(INSTANCE)
ROWS: list = [] # every executed row (a resume replays them first): trajectory.npz
@@ -153,13 +160,15 @@ def schedule_text() -> list:
"wz_radps": round(c["wz"], 4)} for c in S.inst["commands"]]
+ONE = "one episode, no reset" if not RESET_BUDGET else (
+ f"an episode (with up to {RESET_BUDGET} resets: POST /reset starts a new one; only the last is graded)")
TASK_EPISODE = {
- "joystick": f"one episode, no reset: it ends after {S.horizon} control steps (the {S.horizon * DT:g} s command "
+ "joystick": f"{ONE}: it ends after {S.horizon} control steps (the {S.horizon * DT:g} s command "
f"schedule), when the robot meets its termination rule ({S.spec.get('term_text')}: a failure), when you "
"submit `done`, or when your time is over",
- "getup": f"one episode, no reset: it ends the moment the robot has stood upright for {Q.HOLD_STEPS} control steps in a "
+ "getup": f"{ONE}: it ends the moment the robot has stood upright for {Q.HOLD_STEPS} control steps in a "
f"row (success), after {S.horizon} control steps, when you submit `done`, or when your time is over",
- "stand_on_two": f"one episode, no reset: it ends the moment the pose has been held for {Q.POSE_HOLD_STEPS} control steps "
+ "stand_on_two": f"{ONE}: it ends the moment the pose has been held for {Q.POSE_HOLD_STEPS} control steps "
"in a row (success), when the robot meets the termination rule (a failure: its trunk's up axis below "
f"-0.25, or {'a leg' if S.spec.get('hind_leg_contact') else 'a front leg'}'s thigh, calf or hip on the "
f"floor), after {S.horizon} control steps, when you "
@@ -204,7 +213,7 @@ SPEC = {
"world": {"frame": "z up; the robot starts where the episode puts it (no world coordinates are given)", "up": "+z",
"units": {"length": "m", "angle": "rad", "time": "s"}, "quaternion_order": "wxyz",
"camera_axes": "intrinsic_cv / extrinsic_cv: OpenCV camera axes (x right, y down, z forward); cam2world_gl: OpenGL axes"},
- "budget": {"reset": 0, "steps": None, "wall_clock_s": WALL_CLOCK_S}, # + "clock" (rcb_limited.CLOCK_NOTE): what time_left_s counts from
+ "budget": {"reset": RESET_BUDGET, "steps": None, "wall_clock_s": WALL_CLOCK_S}, # + "clock" (rcb_limited.CLOCK_NOTE): what time_left_s counts from
"max_control_steps": S.horizon,
"episode": TASK_EPISODE,
"command_schedule": schedule_text(),
@@ -223,6 +232,15 @@ SPEC = {
"depth": {"modes": list(L.DEPTH_MODES), "default": "none", "kind": "z-depth", "unit": "m", "invalid": 0,
"max_m": L.DEPTH_MAX_M, "frames_encoding": "uint16 mm, zlib", "noise": L.DEPTH_NOISE},
}
+if RESET_BUDGET:
+ SPEC["reset"] = (
+ f"POST /reset (JSON or RCB1 {{cameras, video_fps, depth}}; no physics) abandons the running episode, or follows "
+ f"one that has ended, and starts a new one from the same frozen starting state, {RESET_BUDGET} times in all; it "
+ "returns the new episode's first frames and proprioception. Only the last episode is graded: an episode you "
+ "reset away is never graded, a success ends the trial (409 after it), and when your time is up the episode then "
+ "running is the one graded. The wall clock never restarts. /status and every meta add `episode`, `resets_left` "
+ "and `control_steps_total` (all episodes); `control_steps` counts the current episode's."
+ )
L.check_spec(SPEC) # a spec that misses a 2.2 field or names a forbidden one never starts the episode
EP.scene_built()
@@ -277,6 +295,7 @@ def summary() -> dict:
"reason": E["reason"], "done": E["done"], "env": S.env, "protocol": L.PROTOCOL, "n_actions": len(ROWS),
"fell_at": chk.get("fell_at"), "worst_lin_err": chk.get("worst_lin_err"), "worst_yaw_err": chk.get("worst_yaw_err"),
"tilt_deg": chk.get("tilt_deg"), "imu_height": chk.get("imu_height"), "pose_angle_deg": chk.get("pose_angle_deg"),
+ "episodes": E["resets"] + 1, "resets": E["resets"], "reset_budget": RESET_BUDGET, "control_steps_total": EP.rows,
**EP.final_fields()}
@@ -294,7 +313,8 @@ def finish(reason: str) -> None:
E["success"] = bool(chk["success"])
except Exception: # noqa: BLE001
traceback.print_exc()
- EP.log({"ev": "end", "reason": reason, "step": E["step"], "control_steps": ctrl_steps(), "success": bool(E["success"])})
+ EP.log({"ev": "end", "reason": reason, "step": E["step"], "control_steps": ctrl_steps(), "success": bool(E["success"]),
+ "episode": E["resets"] + 1, "journal_rows": EP.rows})
dump()
@@ -388,7 +408,7 @@ def meta_common(cams) -> dict:
"sim_time_s": round(n * DT, 3), "time_s": round(_agent_time(), 2),
"time_left_s": max(0.0, round(WALL_CLOCK_S - _agent_time(), 1)),
"proprio": {"layout": PROPRIO_LAYOUT, "now": proprio_row().tolist()}, "done": E["done"], "reason": E["reason"],
- "cameras": {c: camera_info(c) for c in cams}}
+ "cameras": {c: camera_info(c) for c in cams}, **reset_counters()}
if S.kind == "joystick":
out["command_now"] = dict(zip(("vx_mps", "vy_mps", "wz_radps"), S.command_now().round(4).tolist()))
return out
@@ -478,10 +498,54 @@ def do_observe(meta: dict):
return out, rec.parts(video if seconds > 0 else ("frames" if video == "mp4" else video), fps)
+def reset_counters() -> dict:
+ if not RESET_BUDGET:
+ return {}
+ left = 0 if (E["success"] or E.get("exiting")) else RESET_BUDGET - E["resets"]
+ return {"episode": E["resets"] + 1, "resets_left": left, "control_steps_total": EP.rows}
+
+
def status() -> dict:
_touch()
return {"step": E["step"], "calls": E["calls"], "control_steps": ctrl_steps(), "done": E["done"], "reason": E["reason"],
- "elapsed_s": round(time.time() - E["t0"], 1)}
+ "elapsed_s": round(time.time() - E["t0"], 1), **reset_counters()}
+
+
+class Refused(Exception):
+ """A /reset the service turns down (409): nothing changes."""
+
+ def __init__(self, body: dict):
+ super().__init__(body.get("error"))
+ self.body = body
+
+
+def do_reset(meta: dict):
+ """POST /reset: a new episode from the frozen instance (the running one, if any, is abandoned, never graded)."""
+ meta = _request(meta)
+ cams, fps = cameras_of(meta), video_fps(meta)
+ L.depth_mode(meta) # a bad depth value is a 400 before anything changes
+ if E["success"]:
+ raise Refused({"error": "trial finished: an episode succeeded", "reason": E["reason"], "resets_left": 0})
+ if E["resets"] >= RESET_BUDGET:
+ raise Refused({"error": "reset budget exhausted", "reason": "reset budget exhausted", "budget": RESET_BUDGET,
+ "resets_left": 0})
+ abandoned = {"episode": E["resets"] + 1, "control_steps": ctrl_steps(), "done": E["done"], "reason": E["reason"]}
+ if not E["done"]: # the abandoned episode's end, for the log (never graded: its rows leave the trajectory)
+ EP.log({"ev": "end", "reason": "reset", "step": E["step"], "control_steps": ctrl_steps(), "success": False,
+ "episode": E["resets"] + 1, "journal_rows": EP.rows})
+ S.reset()
+ ROWS.clear()
+ E["resets"] += 1
+ E["reset_at"] = list(E["reset_at"]) + [EP.rows]
+ E.update(done=False, reason=None, success=False, t_end=None, episode_control_steps=None, dumped=False, check=None)
+ EP._ck = None # the state changed without a journal row: the checkpoint after this call must hash it anew
+ E["calls"] += 1
+ EP.log({"ev": "reset", "call": E["calls"], "episode": E["resets"] + 1, "journal_rows": EP.rows,
+ "abandoned": abandoned, "state_sha": S.state_sha()})
+ rec = recorder(cams, fps, meta)
+ rec.tick(force=True, last=True)
+ out = {**meta_common(cams), "reset_id": E["resets"], "ok": True}
+ return out, rec.parts("frames", fps)
def end_from_signal() -> None:
@@ -498,20 +562,106 @@ def end_from_signal() -> None:
os._exit(0)
+# A resume replays the journal from the frozen instance and resets where the committed resets did: their journal rows
+# are in the last checkpoint (`persist`), the one Episode.begin resumes from.
+_CPS = [e for e in EP.events() if e.get("ev") == "checkpoint"]
+RESUME = {"at": list((_CPS[-1].get("persist") or {}).get("reset_at") or []) if _CPS else [], "i": 0}
+
+
+def _resume_resets() -> None:
+ while RESUME["at"] and int(RESUME["at"][0]) <= RESUME["i"]:
+ RESUME["at"].pop(0)
+ S.reset()
+ ROWS.clear()
+
+
def replay_row(row) -> None:
"""A journaled row during a resume."""
+ _resume_resets()
S.step(row)
ROWS.append(np.asarray(row, dtype=np.float64))
+ RESUME["i"] += 1
-EP.begin(state=E, step=replay_row, state_sha=S.state_sha, proprio=proprio_row, spec=SPEC, on_end=end_from_signal,
- start_event={"env": S.env, "seed": S.inst.get("seed"), "cameras": CAMS})
+def resume_sha() -> str:
+ _resume_resets() # a reset after the last replayed row (or before the first)
+ return S.state_sha()
+
+
+RESUMED = EP.begin(state=E, step=replay_row, state_sha=resume_sha, proprio=proprio_row, spec=SPEC, on_end=end_from_signal,
+ start_event={"env": S.env, "seed": S.inst.get("seed"), "cameras": CAMS})
+if RESUMED and E["resets"]: # Episode.begin looks for an "end" of any episode: only the current one's counts
+ _ends = [e for e in EP.events() if e.get("ev") == "end" and int(e.get("episode") or 1) == E["resets"] + 1
+ and e.get("reason") != "reset" and int(e.get("journal_rows") or 0) <= EP.rows]
+ if _ends:
+ E.update(done=True, reason=_ends[-1].get("reason"), success=bool(_ends[-1].get("success")),
+ t_end=float(_ends[-1]["ts"]), episode_control_steps=int(_ends[-1].get("control_steps") or ctrl_steps()))
+ else:
+ E.update(done=False, reason=None, success=False, t_end=None, episode_control_steps=None)
if E["done"]:
dump() # resumed after the episode had ended: the record is rewritten from the replay
E["ready"] = True
write_final()
open(PID_FILE, "w").write(str(os.getpid()))
print(f"{L.PROTOCOL} MuJoCo Playground quadruped sim ready ({S.env}; startup {EP.n_startups}, resumes {EP.resumes})", flush=True)
-L.serve(L.Service(spec=SPEC, lock=LOCK, state=E, submit=do_submit, observe=do_observe, kinematics=do_kinematics, status=status,
- on_submit_error=lambda exc: finish("error"), log=EP.log, episode=EP, state_sha=S.state_sha,
- proprio=proprio_row), port=int(os.environ.get("RCB_PORT", "8000")))
+
+
+def make_handler(svc: L.Service):
+ """The shared 2.2 handler, plus POST /reset when the run grants resets (the shared layer has no reset)."""
+ base = L.make_handler(svc)
+
+ class H(base):
+ def do_POST(self):
+ if self.path != "/reset" or not RESET_BUDGET:
+ return super().do_POST()
+ try:
+ meta, _ = self._body()
+ except Exception as exc: # noqa: BLE001
+ return self._json(413 if "too large" in str(exc) else 400, {"error": str(exc)[:200]})
+ with LOCK:
+ if E.get("exiting"):
+ return self._json(409, {"error": "trial finished: your time is up", "reason": E.get("reason"),
+ "resets_left": 0})
+ cid = EP.call_begin(self.path, self.connection)
+ out = blobs = err = None
+ try:
+ out, blobs = do_reset(meta)
+ except Refused as exc:
+ err = (409, exc.body)
+ except ValueError as exc:
+ err = (400, {"error": str(exc)})
+ except Exception as exc: # noqa: BLE001
+ traceback.print_exc()
+ E["error"] = repr(exc)
+ EP.log({"ev": "error", "path": self.path, "error": repr(exc)})
+ try:
+ finish("error")
+ except Exception: # noqa: BLE001
+ traceback.print_exc()
+ err = (500, {"error": "simulator error", "detail": repr(exc)[:200]})
+ finally:
+ if out is not None or (err and err[0] == 500):
+ try:
+ EP.checkpoint(E, S.state_sha, proprio_row)
+ except Exception: # noqa: BLE001
+ traceback.print_exc()
+ EP.call_end(cid, self.path)
+ if err is not None:
+ return self._json(*err)
+ return self._rcb1({**out, "stopped": None}, blobs)
+
+ return H
+
+
+def serve(svc: L.Service, port: int) -> None:
+ """L.serve with this service's handler (the same spec defaults and check, and L.bind: a port still held when the
+ service restarts is waited for, not a crash)."""
+ svc.spec.setdefault("on_service_error", L.ON_SERVICE_ERROR)
+ svc.spec.setdefault("max_call_wall_s", EP.max_call_wall_s)
+ L.check_spec(svc.spec)
+ L.bind(port, make_handler(svc)).serve_forever()
+
+
+serve(L.Service(spec=SPEC, lock=LOCK, state=E, submit=do_submit, observe=do_observe, kinematics=do_kinematics, status=status,
+ on_submit_error=lambda exc: finish("error"), log=EP.log, episode=EP, state_sha=S.state_sha,
+ proprio=proprio_row), port=int(os.environ.get("RCB_PORT", "8000")))
diff --git a/tasks/mujoco_playground_barkour_joystick-standard/instruction.md b/tasks/mujoco_playground_barkour_joystick-standard/instruction.md
index b9d3730fba..e626747b4e 100644
--- a/tasks/mujoco_playground_barkour_joystick-standard/instruction.md
+++ b/tasks/mujoco_playground_barkour_joystick-standard/instruction.md
@@ -2,9 +2,9 @@
**Task:** Make Google's Barkour quadruped follow the joystick command schedule for the whole 10 s episode, through random pushes, without falling over.
-You control the Google Barkour vB quadruped — **as if it were real hardware** — in one episode of MuJoCo Playground's
+You control the Google Barkour vB quadruped — **as if it were real hardware** — in MuJoCo Playground's
`BarkourJoystick` environment (plain MuJoCo 3.3.7 with Playground's scene, commit 4057c14). The simulator runs in a service you
-cannot open. You get one client and one episode.
+cannot open. You get one client and up to 11 episodes.
## The episode
The robot starts standing still at Playground's `home` pose (this environment's reset does not move it). The ground is flat. One control step is 10 ms of simulated time: Playground's Barkour environment keeps Menagerie's 2 ms timestep and takes 5 substeps.
@@ -20,7 +20,15 @@ The torso is pushed 12 times at moments you are not told (Playground's kicks: ha
The same schedule is `spec["command_schedule"]`; every response's `meta["command_now"]` is the current command.
## Rules (real-robot rules)
-- **No reset.** One episode. You see four cameras that follow the robot (RGB, depth if you ask) and the robot's own
+- **Episodes and resets.** You have **10 resets** (`spec["budget"]["reset"]`, which is authoritative: if it
+ says 0, there is no reset and the one episode is the run). A reset (`POST /reset`, below) puts the robot back on the
+ same frozen starting state and starts a new episode, either abandoning the running one or after one has ended; it
+ takes no simulated time and returns the start's images and proprioception. Once an episode has ended, `submit`
+ raises `EpisodeOver` until you reset. **Only the last episode is graded**: an episode you reset away is never graded,
+ a success ends the run at once (no reset after it), and when your time is up the episode then running is the one
+ graded. The wall clock is the run's and never restarts; `meta["episode"]` and `meta["resets_left"]` count the
+ episodes (`meta["control_steps"]` is the current episode's).
+- **No privileged state.** You see the cameras that follow the robot (RGB, depth if you ask) and the robot's own
sensors: joint positions, velocities and torques, the IMU (gyro, accelerometer, gravity's direction, the trunk's
velocity), foot contacts. No world positions, no success check.
- **The episode ends** after 1000 control steps, the whole schedule (`spec["max_control_steps"]`, counted by `meta["control_steps"]`), when the robot meets its termination rule (the trunk's up axis pointing below the horizon, a joint outside its box (abduction [-0.7, 0.52], hip [-1.0, 2.1], knee [0.05, 2.1] rad) or the torso lower than 0.18 m: a failure), when you submit `done`, or when your time is up (before the end of the schedule, both are failures).
@@ -29,8 +37,8 @@ The same schedule is `spec["command_schedule"]`; every response's `meta["command
`get_obs(seconds=...)`, which holds the current targets). `meta["time_left_s"]` is the service's count of the limit,
from your first request to the service: it never runs out before your time does, so plan by it.
- Nothing to hand in: the service records the episode and grades it. Leave your script at `/app/output/gen.py`.
-- **If the service crashes**, a call fails with a connection error: it restarts and resumes the SAME episode (never a
- reset) in the state after the last call it completed. `r.wait_ready()` waits for it; then check `r.status()`.
+- **If the service crashes**, a call fails with a connection error: it restarts and resumes the SAME episode (a crash
+ is never a reset and costs none) in the state after the last call it completed. `r.wait_ready()` waits for it; then check `r.status()`.
- Nothing is put into your context automatically: images come back as arrays in Python objects.
## Client (`/opt/tools/rcb_client.py`)
@@ -43,7 +51,8 @@ res = r.submit("joint_position", A, dt=0.01) # A: (T, 12) rows dt apart; res.p
obs = r.get_obs(seconds=1.0, video="mp4") # hold the current targets and watch for 1 s (max 30)
r.fk(q); r.ik({"foot_front_left": [x, y, z, 1, 0, 0, 0]}, seed=q) # feet in the trunk frame <-> joints (no physics)
r.submit("done", []) # end the episode now
-r.status() # calls, control steps, done, reason, elapsed time
+r.status() # calls, control steps, done, reason, elapsed time, episode, resets_left
+obs = r._post("/reset", {"cameras": ["track"]}, [], 900) # POST /reset: a new episode (the 2.2 client has no reset())
```
Cameras: `track` (follows the robot) and `realsense/rgb` (on its head); `spec["cameras"]` gives each one's mounting and intrinsics, and every response carries its pose. Proprioception (`spec["proprio_groups"]`): `joint_pos`, `joint_vel`, `joint_effort` (12 each, action
order), `imu_gyro`, `imu_accelerometer`, `imu_gravity`, `imu_linvel` (3 each, IMU frame: x forward, y left, z up),
diff --git a/tasks/mujoco_playground_barkour_joystick-standard/task.toml b/tasks/mujoco_playground_barkour_joystick-standard/task.toml
index 43a66221ff..87875f382a 100644
--- a/tasks/mujoco_playground_barkour_joystick-standard/task.toml
+++ b/tasks/mujoco_playground_barkour_joystick-standard/task.toml
@@ -16,7 +16,7 @@ artifacts = [{ source = "/var/rcb-monitor/", service = "rcb_monitor" },
[task]
name = "robot-coding-bench/mujoco_playground_barkour_joystick-standard"
-description = "STANDARD MODE (robot as a service, eai-standard/2.2): MuJoCo Playground BarkourJoystick (Google Barkour vB) on the same frozen instance as mujoco_playground_barkour_joystick-privileged: Make Google's Barkour quadruped follow the joystick command schedule for the whole 10 s episode, through random pushes, without falling over. The agent only gets a client: four cameras that follow the robot (RGB, optional depth), the robot's joints and IMU, joint position chunks at 50 Hz, the feet's kinematics; one episode, no reset, no world state; the sim service records the episode, replays it in fresh processes and grades it."
+description = "STANDARD MODE (robot as a service, eai-standard/2.2): MuJoCo Playground BarkourJoystick (Google Barkour vB) on the same frozen instance as mujoco_playground_barkour_joystick-privileged: Make Google's Barkour quadruped follow the joystick command schedule for the whole 10 s episode, through random pushes, without falling over. The agent only gets a client: four cameras that follow the robot (RGB, optional depth), the robot's joints and IMU, joint position chunks at 50 Hz, the feet's kinematics; 10 resets (only the last episode is graded), no world state; the sim service records the episode, replays it in fresh processes and grades it."
authors = [{ name = "Changhe Chen" }]
keywords = ["robotics", "mujoco", "mujoco-playground", "google-barkour-vb", "quadruped", "locomotion", "joystick", "code-as-policy", "trajectory", "standard-mode"]
@@ -28,6 +28,7 @@ simulator = "mujoco"
robot = "google-barkour-vb"
family = "mujoco_playground/BarkourJoystick"
mode = "standard"
+dynamics = "dynamic" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "full"
instance_seed = 2
diff --git a/tasks/mujoco_playground_barkour_joystick-standard/tests/grade.py b/tasks/mujoco_playground_barkour_joystick-standard/tests/grade.py
index 41ce495421..3d4c23ad7e 100644
--- a/tasks/mujoco_playground_barkour_joystick-standard/tests/grade.py
+++ b/tasks/mujoco_playground_barkour_joystick-standard/tests/grade.py
@@ -77,6 +77,9 @@ write(
"sim_error": int(d.get("reason") == "error"),
"sim_startups": int(d.get("sim_startups") or 0),
"resumes": int(d.get("resumes") or 0),
+ "resets": int(d.get("resets") or 0), # PROTOCOL §4.5: resets the agent used; only the last episode is graded
+ "reset_budget": int(d.get("reset_budget") or 0),
+ "episodes": int(d.get("episodes") or 1),
"resume_state_mismatch": int(d.get("resume_state_mismatch") or 0),
"resume_max_joint_dev": num("resume_max_joint_dev"),
},
diff --git a/tasks/mujoco_playground_go1_footstand-privileged/README.md b/tasks/mujoco_playground_go1_footstand-privileged/README.md
index 4e24b85090..627eccb44a 100644
--- a/tasks/mujoco_playground_go1_footstand-privileged/README.md
+++ b/tasks/mujoco_playground_go1_footstand-privileged/README.md
@@ -53,3 +53,11 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
### Provenance
Generated by `scripts/mujoco-playground/generate.py` from `scripts/mujoco-playground/instances/Go1Footstand.json`
(`make_instances.py`, `select_instances.py`) and `scripts/mujoco-playground/oracles/Go1Footstand.npy` (`oracle.py`).
+
+### Change Log
+- 2026-10-14 (review of board 0105, B1-m2): `tests/grade.py` only. A replay process that died or ran out of time, a
+ rebuilt start that does not match the instance's hash (`instance_mismatch`) and an exception of the verifier itself
+ now write `infra_error = 1` beside `grader_error = 1`: no verdict on the agent. Before, they wrote `grader_error = 1`
+ alone, which counted as the agent's 0. Every `reward.json` now carries `infra_error`. The scene, the instance, the
+ pass rule and the reference are unchanged, so the task keeps its name; a past 0 with `grader_error = 1` was an
+ infrastructure failure.
diff --git a/tasks/mujoco_playground_go1_footstand-privileged/task.toml b/tasks/mujoco_playground_go1_footstand-privileged/task.toml
index b307b7c00d..88ec76312a 100644
--- a/tasks/mujoco_playground_go1_footstand-privileged/task.toml
+++ b/tasks/mujoco_playground_go1_footstand-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "mujoco"
robot = "unitree-go1"
family = "mujoco_playground/Go1Footstand"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 0 # the PRNGKey of Playground's reset (and, for the joystick tasks, its command process)
expert_time_estimate_hours = 3.0
diff --git a/tasks/mujoco_playground_go1_footstand-privileged/tests/grade.py b/tasks/mujoco_playground_go1_footstand-privileged/tests/grade.py
index c75c326868..1667d090c8 100644
--- a/tasks/mujoco_playground_go1_footstand-privileged/tests/grade.py
+++ b/tasks/mujoco_playground_go1_footstand-privileged/tests/grade.py
@@ -6,6 +6,11 @@ Two replays, each in its own fresh process (`rcb_quadruped.py replay`: the froze
task's success check on the whole trajectory). success = both replays succeed and end in the same state. Writes
/logs/verifier/reward.txt (1/0) and reward.json (numbers only); then a third replay renders replay.mp4, after the
verdict is on disk (it never changes it).
+
+Failures that are not the agent's (docs/PROTOCOL.md: infra vs agent failure) carry `grader_error` = 1 and `infra_error`
+= 1, so they are no verdict on the agent: a replay process that died or timed out, an image that no longer rebuilds the
+frozen start (`instance_mismatch`), and any exception of the verifier itself. A file the agent got wrong is the agent's
+0 (`bad_trajectory`, infra_error 0). Every reward.json carries success, deterministic, grader_error and infra_error.
"""
import argparse
import json
@@ -22,8 +27,12 @@ TOOL = os.path.join(os.path.dirname(os.path.abspath(__file__)), "rcb_quadruped.p
REPLAY_TIMEOUT_S = 300 # a 1000-step replay takes a few seconds; the video replay renders 4 views every 2nd step
+BASE = {"success": 0, "deterministic": 0, "grader_error": 0, "infra_error": 0}
+
+
def write(ok: bool, out: dict) -> None:
os.makedirs(LOG_DIR, exist_ok=True)
+ out = {**BASE, **out}
open(os.path.join(LOG_DIR, "reward.txt"), "w").write("1\n" if ok else "0\n")
json.dump(out, open(os.path.join(LOG_DIR, "reward.json"), "w"))
print(json.dumps(out), flush=True)
@@ -96,11 +105,11 @@ def main() -> int:
"n_actions": int(acts.shape[0]) if acts.ndim == 2 else 0, "grader_error": 0, **cost})
return 0
runs = [replay(a.instance, a.traj, None), replay(a.instance, a.traj, None)]
- if any("error" in r and not r.get("bad_trajectory") for r in runs):
- write(False, {"success": 0, "deterministic": 0, "grader_error": 1, **cost})
+ if any("error" in r and not r.get("bad_trajectory") for r in runs): # the replay process died / timed out
+ write(False, {"success": 0, "deterministic": 0, "grader_error": 1, "infra_error": 1, "replay_failed": 1, **cost})
return 0
if not all(r.get("instance_ok") for r in runs): # the image no longer builds the frozen start: not the agent's 0
- write(False, {"success": 0, "deterministic": 0, "grader_error": 1, "instance_mismatch": 1, **cost})
+ write(False, {"success": 0, "deterministic": 0, "grader_error": 1, "infra_error": 1, "instance_mismatch": 1, **cost})
return 0
all_success = all(bool(r.get("success")) for r in runs)
deterministic = runs[0].get("state_sha") == runs[1].get("state_sha")
@@ -116,7 +125,7 @@ def main() -> int:
write(ok, out) # the verdict, before any video
except Exception: # noqa: BLE001
traceback.print_exc(file=sys.stderr)
- write(False, {"success": 0, "deterministic": 0, "grader_error": 1, **cost})
+ write(False, {"success": 0, "deterministic": 0, "grader_error": 1, "infra_error": 1, **cost})
return 0
try: # the video, for humans: it never changes the verdict
if replay(a.instance, a.traj, os.path.join(LOG_DIR, "replay.mp4")).get("video"):
diff --git a/tasks/mujoco_playground_go1_footstand-standard/README.md b/tasks/mujoco_playground_go1_footstand-standard/README.md
index 16b0cd2a26..24a85796f6 100644
--- a/tasks/mujoco_playground_go1_footstand-standard/README.md
+++ b/tasks/mujoco_playground_go1_footstand-standard/README.md
@@ -13,14 +13,14 @@
## Task
MuJoCo Playground `Go1Footstand` (Unitree Go1) on the same frozen instance as `mujoco_playground_go1_footstand-privileged`: Make the Unitree Go1 stand up on its hind legs, nose up, front legs in the air, and balance there for 2 s.
-From standing, it has 10 s to get into a footstand on its hind legs and hold it for 2 s (Playground's footstand reward aims at the IMU 0.53 m high). Standard mode (eai-standard/2.2): the agent drives the robot through a client in one episode, with only
-the robot's own sensors and four cameras that follow it.
+From standing, it has 10 s to get into a footstand on its hind legs and hold it for 2 s (Playground's footstand reward aims at the IMU 0.53 m high). Standard mode (eai-standard/2.2): the agent drives the robot through a client, with 10 resets (only
+the last episode is graded), the robot's own sensors and the cameras that follow it.
## Difficulty
-Same instance and success check as `mujoco_playground_go1_footstand-privileged`. In standard mode the agent loses resets, the simulator and
+Same instance and success check as `mujoco_playground_go1_footstand-privileged`. In standard mode the agent loses free resets, the simulator and
the world state: it sees camera images and proprioception (joints, IMU, foot contacts), acts in joint position chunks,
-and gets one episode, so it has to close the loop on the robot's own sensors and recover from its own mistakes without
-starting over.
+and gets 10 resets (`RCB_RESET_BUDGET`, the compose file's default; PROTOCOL §4.5), so it has to close the loop
+on the robot's own sensors; only the last episode is graded.
## Reference solution
`solution/oracle_replay.py` streams the privileged reference's rows (`solution/oracle.npy`, 113 rows, float32-exact)
@@ -60,3 +60,4 @@ Generated by `scripts/mujoco-playground/generate.py`, as its privileged twin.
- 2026-10-10 (review 0067): a `final.json` that is missing, unreadable or was never finalized by the collect hook is graded `infra_error = 1` (no verdict on the agent); before, a missing one was a plain 0 and an unfinalized one was graded as it stood. The collect hook writes `final.json` atomically. A request field of the wrong type (`dt` or `seconds` null, `cameras` not a list of names, a `note` that is not a string) is answered 400 before any physics instead of ending the episode with a 500. `meta["time_left_s"]` counts from the agent's first call, not from the service's start.
- 2026-10-10 (review round 2, board 0067): the client only (`environment/rcb_client.py`). `Robot()` waits up to 3600 s for the service (was 600 s) and prints its progress to stderr once a minute, so the first `Robot()` outlasts the simulator's start (no effect on this benchmark, whose service is up within a minute); `cameras="name"` (a string) is one camera, no longer its characters. The service, the collect hook, the grader, the scene and the pass rule are unchanged.
- 2026-10-10 (review round 3, board 0067): the shared service layer and the client only (`environment/sim/rcb_limited.py`, `environment/rcb_client.py`). A camera named more than once in `cameras` is rendered and returned once (N copies of one name made the service build N x N images within the frame budget, which could exhaust its memory). A service that has given up on the episode (it could not start it, or could not resume it after its crashes) answers every request with 409 and the reason, and `Robot()` / `wait_ready()` raise `EpisodeOver` at once; before, nothing answered and they waited out their timeout (up to an hour). The adapter, the collect hook, the grader, the scene and the pass rule are unchanged.
+- 2026-10-14 (owner decision C1, board 0113): 10 resets (PROTOCOL §4.5). The service (`environment/sim/server.py`) gained the optional reset capability: `RCB_RESET_BUDGET` (the compose file's default 10; 0 restores the old single episode), `POST /reset` abandons the running or ended episode and starts a new one from the same frozen start (no physics), refused (409) after a success, once the agent's time is up and once the budget is spent; only the last episode is graded (`trajectory.npz` holds its rows, replayed twice from the frozen start as before); `/status` and every `meta` add `episode`, `resets_left` and `control_steps_total`; `final.json` and `reward.json` add `resets`, `reset_budget` and `episodes`. A crash resumes the current episode (the journal is replayed with the committed resets). The instruction's rules, the client example (`POST /reset` through `Robot._post`) and this README say so. The scene, the instance, the success check and the grader's pass rule are unchanged; results recorded before this change ran with one episode and are reported separately.
diff --git a/tasks/mujoco_playground_go1_footstand-standard/environment/docker-compose.yaml b/tasks/mujoco_playground_go1_footstand-standard/environment/docker-compose.yaml
index bc0c363241..1226b8c3a6 100644
--- a/tasks/mujoco_playground_go1_footstand-standard/environment/docker-compose.yaml
+++ b/tasks/mujoco_playground_go1_footstand-standard/environment/docker-compose.yaml
@@ -28,6 +28,7 @@ services:
RCB_STARTUP_TIMEOUT_S: "300" # the scene builds in about a second
PYTHONPATH: /srv
RCB_SESSION_LOG: "" # the sidecar's own Session must not write the agent-side interaction log
+ RCB_RESET_BUDGET: "${RCB_RESET_BUDGET:-10}" # Playground quadrupeds standard: 10 resets (PROTOCOL §4.5); 0: one episode
# >>> rcb_monitor (resource limits monitor, docs/PROTOCOL.md §2.10; generated by scripts/rcb_monitor/inject.py)
rcb_monitor:
image: ghcr.io/mll-lab-nu/eai-monitor:0.1.0@sha256:f2c910f53f79c2bfa4a71298b10c78383cae3afc40720ca562cc845b8d20fb0c
diff --git a/tasks/mujoco_playground_go1_footstand-standard/environment/sim/server.py b/tasks/mujoco_playground_go1_footstand-standard/environment/sim/server.py
index 108eca3cae..cad8920259 100644
--- a/tasks/mujoco_playground_go1_footstand-standard/environment/sim/server.py
+++ b/tasks/mujoco_playground_go1_footstand-standard/environment/sim/server.py
@@ -1,16 +1,20 @@
"""eai-standard/2.2 sim service for MuJoCo Playground's quadruped tasks (docs/STANDARD_MODE.md): the adapter of the
shared layer rcb_limited.py; the simulator is rcb_quadruped.Session (plain MuJoCo on the CPU).
-HTTP surface: GET /spec, GET /status, POST /observe, POST /submit, POST /kinematics. Every call is logged to
-/data/episode.jsonl.
+HTTP surface: GET /spec, GET /status, POST /observe, POST /submit, POST /kinematics, and POST /reset when the run grants
+resets (RCB_RESET_BUDGET, the optional reset capability of docs/STANDARD_MODE.md / PROTOCOL §4.5; the quadrupeds'
+standard tasks default to 10 in their compose file, owner 2026-10-14). Every call is logged to /data/episode.jsonl.
The episode is the frozen instance's: one control step = the instance's control period (Go1, Spot: 1/50 s; Barkour:
1/100 s). A joystick episode ends after the whole command schedule
(the horizon), when the robot meets Playground's termination rule (it falls over: a failure), when the agent submits
`done`, or when its time is over; success = the task's check on the steps executed. A getup episode ends the moment the
-check holds (success), after the horizon, on `done` or on timeout.
+check holds (success), after the horizon, on `done` or on timeout. With a reset budget, POST /reset abandons the running
+(or ended) episode and starts a new one from the frozen instance; only the last episode is graded, and a success ends
+the trial.
Crash recovery (rcb_limited.Episode, resume class `exact`: MuJoCo on the CPU repeats bit for bit in a fresh process):
-every executed row is journaled; after a crash the service rebuilds the episode and replays the journal. When the
+every executed row is journaled; after a crash the service rebuilds the episode and replays the journal, resetting where
+the committed resets did (their journal rows are checkpointed, `persist`). When the
episode ends the service writes /data/trajectory.npz + /data/episode.json; the collect hook (finalize.sh) makes it exit
and replays the trajectory in two fresh processes."""
@@ -31,18 +35,21 @@ INSTANCE = os.environ.get("RCB_INSTANCE", os.path.join(os.path.dirname(os.path.a
WALL_CLOCK_S = float(os.environ.get("RCB_WALL_CLOCK_S", "3600"))
MAX_CHUNK_S = float(os.environ.get("RCB_MAX_CHUNK_S", "10"))
MAX_WATCH_S = float(os.environ.get("RCB_MAX_WATCH_S", "30"))
+RESET_BUDGET = max(0, int(os.environ.get("RCB_RESET_BUDGET") or 0)) # the compose file's default: 10 (PROTOCOL §4.5)
CAM_W, CAM_H = 320, 240
LOCK = threading.Lock()
PID_FILE = os.path.join(DATA, "server.pid")
E = {"ready": False, "step": 0, "calls": 0, "done": False, "reason": None, "success": False, "t0": time.time(), "error": None,
- "t_first": None} # t_first: the agent's first call (its clock starts there; kept through a crash by the checkpoints)
+ "t_first": None, # t_first: the agent's first call (its clock starts there; kept through a crash by the checkpoints)
+ "resets": 0, # resets done (the episode is resets + 1)
+ "reset_at": []} # the journal row at which each reset happened: a resume resets there again
os.environ["RCB_SESSION_LOG"] = ""
os.environ["RCB_AGENT_TRACE_DIR"] = ""
import rcb_quadruped as Q # noqa: E402 (first: it picks the rendering backend before mujoco loads)
import mujoco # noqa: E402
-EP = L.Episode(DATA, dim=Q.N, resume="exact", persist=("t_first",))
+EP = L.Episode(DATA, dim=Q.N, resume="exact", persist=("t_first", "resets", "reset_at"))
EP.startup() # start count, back-off after failed starts, the startup watchdog until scene_built()
S = Q.Session(INSTANCE)
ROWS: list = [] # every executed row (a resume replays them first): trajectory.npz
@@ -153,13 +160,15 @@ def schedule_text() -> list:
"wz_radps": round(c["wz"], 4)} for c in S.inst["commands"]]
+ONE = "one episode, no reset" if not RESET_BUDGET else (
+ f"an episode (with up to {RESET_BUDGET} resets: POST /reset starts a new one; only the last is graded)")
TASK_EPISODE = {
- "joystick": f"one episode, no reset: it ends after {S.horizon} control steps (the {S.horizon * DT:g} s command "
+ "joystick": f"{ONE}: it ends after {S.horizon} control steps (the {S.horizon * DT:g} s command "
f"schedule), when the robot meets its termination rule ({S.spec.get('term_text')}: a failure), when you "
"submit `done`, or when your time is over",
- "getup": f"one episode, no reset: it ends the moment the robot has stood upright for {Q.HOLD_STEPS} control steps in a "
+ "getup": f"{ONE}: it ends the moment the robot has stood upright for {Q.HOLD_STEPS} control steps in a "
f"row (success), after {S.horizon} control steps, when you submit `done`, or when your time is over",
- "stand_on_two": f"one episode, no reset: it ends the moment the pose has been held for {Q.POSE_HOLD_STEPS} control steps "
+ "stand_on_two": f"{ONE}: it ends the moment the pose has been held for {Q.POSE_HOLD_STEPS} control steps "
"in a row (success), when the robot meets the termination rule (a failure: its trunk's up axis below "
f"-0.25, or {'a leg' if S.spec.get('hind_leg_contact') else 'a front leg'}'s thigh, calf or hip on the "
f"floor), after {S.horizon} control steps, when you "
@@ -204,7 +213,7 @@ SPEC = {
"world": {"frame": "z up; the robot starts where the episode puts it (no world coordinates are given)", "up": "+z",
"units": {"length": "m", "angle": "rad", "time": "s"}, "quaternion_order": "wxyz",
"camera_axes": "intrinsic_cv / extrinsic_cv: OpenCV camera axes (x right, y down, z forward); cam2world_gl: OpenGL axes"},
- "budget": {"reset": 0, "steps": None, "wall_clock_s": WALL_CLOCK_S}, # + "clock" (rcb_limited.CLOCK_NOTE): what time_left_s counts from
+ "budget": {"reset": RESET_BUDGET, "steps": None, "wall_clock_s": WALL_CLOCK_S}, # + "clock" (rcb_limited.CLOCK_NOTE): what time_left_s counts from
"max_control_steps": S.horizon,
"episode": TASK_EPISODE,
"command_schedule": schedule_text(),
@@ -223,6 +232,15 @@ SPEC = {
"depth": {"modes": list(L.DEPTH_MODES), "default": "none", "kind": "z-depth", "unit": "m", "invalid": 0,
"max_m": L.DEPTH_MAX_M, "frames_encoding": "uint16 mm, zlib", "noise": L.DEPTH_NOISE},
}
+if RESET_BUDGET:
+ SPEC["reset"] = (
+ f"POST /reset (JSON or RCB1 {{cameras, video_fps, depth}}; no physics) abandons the running episode, or follows "
+ f"one that has ended, and starts a new one from the same frozen starting state, {RESET_BUDGET} times in all; it "
+ "returns the new episode's first frames and proprioception. Only the last episode is graded: an episode you "
+ "reset away is never graded, a success ends the trial (409 after it), and when your time is up the episode then "
+ "running is the one graded. The wall clock never restarts. /status and every meta add `episode`, `resets_left` "
+ "and `control_steps_total` (all episodes); `control_steps` counts the current episode's."
+ )
L.check_spec(SPEC) # a spec that misses a 2.2 field or names a forbidden one never starts the episode
EP.scene_built()
@@ -277,6 +295,7 @@ def summary() -> dict:
"reason": E["reason"], "done": E["done"], "env": S.env, "protocol": L.PROTOCOL, "n_actions": len(ROWS),
"fell_at": chk.get("fell_at"), "worst_lin_err": chk.get("worst_lin_err"), "worst_yaw_err": chk.get("worst_yaw_err"),
"tilt_deg": chk.get("tilt_deg"), "imu_height": chk.get("imu_height"), "pose_angle_deg": chk.get("pose_angle_deg"),
+ "episodes": E["resets"] + 1, "resets": E["resets"], "reset_budget": RESET_BUDGET, "control_steps_total": EP.rows,
**EP.final_fields()}
@@ -294,7 +313,8 @@ def finish(reason: str) -> None:
E["success"] = bool(chk["success"])
except Exception: # noqa: BLE001
traceback.print_exc()
- EP.log({"ev": "end", "reason": reason, "step": E["step"], "control_steps": ctrl_steps(), "success": bool(E["success"])})
+ EP.log({"ev": "end", "reason": reason, "step": E["step"], "control_steps": ctrl_steps(), "success": bool(E["success"]),
+ "episode": E["resets"] + 1, "journal_rows": EP.rows})
dump()
@@ -388,7 +408,7 @@ def meta_common(cams) -> dict:
"sim_time_s": round(n * DT, 3), "time_s": round(_agent_time(), 2),
"time_left_s": max(0.0, round(WALL_CLOCK_S - _agent_time(), 1)),
"proprio": {"layout": PROPRIO_LAYOUT, "now": proprio_row().tolist()}, "done": E["done"], "reason": E["reason"],
- "cameras": {c: camera_info(c) for c in cams}}
+ "cameras": {c: camera_info(c) for c in cams}, **reset_counters()}
if S.kind == "joystick":
out["command_now"] = dict(zip(("vx_mps", "vy_mps", "wz_radps"), S.command_now().round(4).tolist()))
return out
@@ -478,10 +498,54 @@ def do_observe(meta: dict):
return out, rec.parts(video if seconds > 0 else ("frames" if video == "mp4" else video), fps)
+def reset_counters() -> dict:
+ if not RESET_BUDGET:
+ return {}
+ left = 0 if (E["success"] or E.get("exiting")) else RESET_BUDGET - E["resets"]
+ return {"episode": E["resets"] + 1, "resets_left": left, "control_steps_total": EP.rows}
+
+
def status() -> dict:
_touch()
return {"step": E["step"], "calls": E["calls"], "control_steps": ctrl_steps(), "done": E["done"], "reason": E["reason"],
- "elapsed_s": round(time.time() - E["t0"], 1)}
+ "elapsed_s": round(time.time() - E["t0"], 1), **reset_counters()}
+
+
+class Refused(Exception):
+ """A /reset the service turns down (409): nothing changes."""
+
+ def __init__(self, body: dict):
+ super().__init__(body.get("error"))
+ self.body = body
+
+
+def do_reset(meta: dict):
+ """POST /reset: a new episode from the frozen instance (the running one, if any, is abandoned, never graded)."""
+ meta = _request(meta)
+ cams, fps = cameras_of(meta), video_fps(meta)
+ L.depth_mode(meta) # a bad depth value is a 400 before anything changes
+ if E["success"]:
+ raise Refused({"error": "trial finished: an episode succeeded", "reason": E["reason"], "resets_left": 0})
+ if E["resets"] >= RESET_BUDGET:
+ raise Refused({"error": "reset budget exhausted", "reason": "reset budget exhausted", "budget": RESET_BUDGET,
+ "resets_left": 0})
+ abandoned = {"episode": E["resets"] + 1, "control_steps": ctrl_steps(), "done": E["done"], "reason": E["reason"]}
+ if not E["done"]: # the abandoned episode's end, for the log (never graded: its rows leave the trajectory)
+ EP.log({"ev": "end", "reason": "reset", "step": E["step"], "control_steps": ctrl_steps(), "success": False,
+ "episode": E["resets"] + 1, "journal_rows": EP.rows})
+ S.reset()
+ ROWS.clear()
+ E["resets"] += 1
+ E["reset_at"] = list(E["reset_at"]) + [EP.rows]
+ E.update(done=False, reason=None, success=False, t_end=None, episode_control_steps=None, dumped=False, check=None)
+ EP._ck = None # the state changed without a journal row: the checkpoint after this call must hash it anew
+ E["calls"] += 1
+ EP.log({"ev": "reset", "call": E["calls"], "episode": E["resets"] + 1, "journal_rows": EP.rows,
+ "abandoned": abandoned, "state_sha": S.state_sha()})
+ rec = recorder(cams, fps, meta)
+ rec.tick(force=True, last=True)
+ out = {**meta_common(cams), "reset_id": E["resets"], "ok": True}
+ return out, rec.parts("frames", fps)
def end_from_signal() -> None:
@@ -498,20 +562,106 @@ def end_from_signal() -> None:
os._exit(0)
+# A resume replays the journal from the frozen instance and resets where the committed resets did: their journal rows
+# are in the last checkpoint (`persist`), the one Episode.begin resumes from.
+_CPS = [e for e in EP.events() if e.get("ev") == "checkpoint"]
+RESUME = {"at": list((_CPS[-1].get("persist") or {}).get("reset_at") or []) if _CPS else [], "i": 0}
+
+
+def _resume_resets() -> None:
+ while RESUME["at"] and int(RESUME["at"][0]) <= RESUME["i"]:
+ RESUME["at"].pop(0)
+ S.reset()
+ ROWS.clear()
+
+
def replay_row(row) -> None:
"""A journaled row during a resume."""
+ _resume_resets()
S.step(row)
ROWS.append(np.asarray(row, dtype=np.float64))
+ RESUME["i"] += 1
-EP.begin(state=E, step=replay_row, state_sha=S.state_sha, proprio=proprio_row, spec=SPEC, on_end=end_from_signal,
- start_event={"env": S.env, "seed": S.inst.get("seed"), "cameras": CAMS})
+def resume_sha() -> str:
+ _resume_resets() # a reset after the last replayed row (or before the first)
+ return S.state_sha()
+
+
+RESUMED = EP.begin(state=E, step=replay_row, state_sha=resume_sha, proprio=proprio_row, spec=SPEC, on_end=end_from_signal,
+ start_event={"env": S.env, "seed": S.inst.get("seed"), "cameras": CAMS})
+if RESUMED and E["resets"]: # Episode.begin looks for an "end" of any episode: only the current one's counts
+ _ends = [e for e in EP.events() if e.get("ev") == "end" and int(e.get("episode") or 1) == E["resets"] + 1
+ and e.get("reason") != "reset" and int(e.get("journal_rows") or 0) <= EP.rows]
+ if _ends:
+ E.update(done=True, reason=_ends[-1].get("reason"), success=bool(_ends[-1].get("success")),
+ t_end=float(_ends[-1]["ts"]), episode_control_steps=int(_ends[-1].get("control_steps") or ctrl_steps()))
+ else:
+ E.update(done=False, reason=None, success=False, t_end=None, episode_control_steps=None)
if E["done"]:
dump() # resumed after the episode had ended: the record is rewritten from the replay
E["ready"] = True
write_final()
open(PID_FILE, "w").write(str(os.getpid()))
print(f"{L.PROTOCOL} MuJoCo Playground quadruped sim ready ({S.env}; startup {EP.n_startups}, resumes {EP.resumes})", flush=True)
-L.serve(L.Service(spec=SPEC, lock=LOCK, state=E, submit=do_submit, observe=do_observe, kinematics=do_kinematics, status=status,
- on_submit_error=lambda exc: finish("error"), log=EP.log, episode=EP, state_sha=S.state_sha,
- proprio=proprio_row), port=int(os.environ.get("RCB_PORT", "8000")))
+
+
+def make_handler(svc: L.Service):
+ """The shared 2.2 handler, plus POST /reset when the run grants resets (the shared layer has no reset)."""
+ base = L.make_handler(svc)
+
+ class H(base):
+ def do_POST(self):
+ if self.path != "/reset" or not RESET_BUDGET:
+ return super().do_POST()
+ try:
+ meta, _ = self._body()
+ except Exception as exc: # noqa: BLE001
+ return self._json(413 if "too large" in str(exc) else 400, {"error": str(exc)[:200]})
+ with LOCK:
+ if E.get("exiting"):
+ return self._json(409, {"error": "trial finished: your time is up", "reason": E.get("reason"),
+ "resets_left": 0})
+ cid = EP.call_begin(self.path, self.connection)
+ out = blobs = err = None
+ try:
+ out, blobs = do_reset(meta)
+ except Refused as exc:
+ err = (409, exc.body)
+ except ValueError as exc:
+ err = (400, {"error": str(exc)})
+ except Exception as exc: # noqa: BLE001
+ traceback.print_exc()
+ E["error"] = repr(exc)
+ EP.log({"ev": "error", "path": self.path, "error": repr(exc)})
+ try:
+ finish("error")
+ except Exception: # noqa: BLE001
+ traceback.print_exc()
+ err = (500, {"error": "simulator error", "detail": repr(exc)[:200]})
+ finally:
+ if out is not None or (err and err[0] == 500):
+ try:
+ EP.checkpoint(E, S.state_sha, proprio_row)
+ except Exception: # noqa: BLE001
+ traceback.print_exc()
+ EP.call_end(cid, self.path)
+ if err is not None:
+ return self._json(*err)
+ return self._rcb1({**out, "stopped": None}, blobs)
+
+ return H
+
+
+def serve(svc: L.Service, port: int) -> None:
+ """L.serve with this service's handler (the same spec defaults and check, and L.bind: a port still held when the
+ service restarts is waited for, not a crash)."""
+ svc.spec.setdefault("on_service_error", L.ON_SERVICE_ERROR)
+ svc.spec.setdefault("max_call_wall_s", EP.max_call_wall_s)
+ L.check_spec(svc.spec)
+ L.bind(port, make_handler(svc)).serve_forever()
+
+
+serve(L.Service(spec=SPEC, lock=LOCK, state=E, submit=do_submit, observe=do_observe, kinematics=do_kinematics, status=status,
+ on_submit_error=lambda exc: finish("error"), log=EP.log, episode=EP, state_sha=S.state_sha,
+ proprio=proprio_row), port=int(os.environ.get("RCB_PORT", "8000")))
diff --git a/tasks/mujoco_playground_go1_footstand-standard/instruction.md b/tasks/mujoco_playground_go1_footstand-standard/instruction.md
index 4a3c87adb6..a4105faf7f 100644
--- a/tasks/mujoco_playground_go1_footstand-standard/instruction.md
+++ b/tasks/mujoco_playground_go1_footstand-standard/instruction.md
@@ -2,15 +2,23 @@
**Task:** Make the Unitree Go1 stand up on its hind legs, nose up, front legs in the air, and balance there for 2 s.
-You control the Unitree Go1 quadruped — **as if it were real hardware** — in one episode of MuJoCo Playground's
+You control the Unitree Go1 quadruped — **as if it were real hardware** — in MuJoCo Playground's
`Go1Footstand` environment (plain MuJoCo 3.3.7 with Playground's scene, commit 4057c14). The simulator runs in a service you
-cannot open. You get one client and one episode.
+cannot open. You get one client and up to 11 episodes.
## The episode
The robot starts standing at Playground's `home` pose with a random base velocity (Playground's reset). You have 500 control steps (10 s).
## Rules (real-robot rules)
-- **No reset.** One episode. You see four cameras that follow the robot (RGB, depth if you ask) and the robot's own
+- **Episodes and resets.** You have **10 resets** (`spec["budget"]["reset"]`, which is authoritative: if it
+ says 0, there is no reset and the one episode is the run). A reset (`POST /reset`, below) puts the robot back on the
+ same frozen starting state and starts a new episode, either abandoning the running one or after one has ended; it
+ takes no simulated time and returns the start's images and proprioception. Once an episode has ended, `submit`
+ raises `EpisodeOver` until you reset. **Only the last episode is graded**: an episode you reset away is never graded,
+ a success ends the run at once (no reset after it), and when your time is up the episode then running is the one
+ graded. The wall clock is the run's and never restarts; `meta["episode"]` and `meta["resets_left"]` count the
+ episodes (`meta["control_steps"]` is the current episode's).
+- **No privileged state.** You see the cameras that follow the robot (RGB, depth if you ask) and the robot's own
sensors: joint positions, velocities and torques, the IMU (gyro, accelerometer, gravity's direction, the trunk's
velocity), foot contacts. No world positions, no success check.
- **The episode ends** the moment the pose has been held for 100 control steps in a row (success), when the robot meets the termination rule (the trunk's up axis below -0.25 (z component of the IMU frame's z axis), or any part of a front leg other than its foot (thigh, calf, hip) touching the floor (Playground's rule as its code reads it), or the same for a hind leg (the list its Footstand sets up but never reads; without it, sitting on a hind hip would count): a failure), after 500 control steps (`spec["max_control_steps"]`), when you submit `done`, or when your time is up.
@@ -19,8 +27,8 @@ The robot starts standing at Playground's `home` pose with a random base velocit
`get_obs(seconds=...)`, which holds the current targets). `meta["time_left_s"]` is the service's count of the limit,
from your first request to the service: it never runs out before your time does, so plan by it.
- Nothing to hand in: the service records the episode and grades it. Leave your script at `/app/output/gen.py`.
-- **If the service crashes**, a call fails with a connection error: it restarts and resumes the SAME episode (never a
- reset) in the state after the last call it completed. `r.wait_ready()` waits for it; then check `r.status()`.
+- **If the service crashes**, a call fails with a connection error: it restarts and resumes the SAME episode (a crash
+ is never a reset and costs none) in the state after the last call it completed. `r.wait_ready()` waits for it; then check `r.status()`.
- Nothing is put into your context automatically: images come back as arrays in Python objects.
## Client (`/opt/tools/rcb_client.py`)
@@ -33,7 +41,8 @@ res = r.submit("joint_position", A, dt=0.02) # A: (T, 12) rows dt apart; res.p
obs = r.get_obs(seconds=1.0, video="mp4") # hold the current targets and watch for 1 s (max 30)
r.fk(q); r.ik({"FR": [x, y, z, 1, 0, 0, 0]}, seed=q) # feet in the trunk frame <-> joints (no physics)
r.submit("done", []) # end the episode now
-r.status() # calls, control steps, done, reason, elapsed time
+r.status() # calls, control steps, done, reason, elapsed time, episode, resets_left
+obs = r._post("/reset", {"cameras": ["track"]}, [], 900) # POST /reset: a new episode (the 2.2 client has no reset())
```
Cameras: `track`, `side`, `back`, `top` (they follow the robot); `spec["cameras"]` gives each one's mounting and intrinsics, and every response carries its pose. Proprioception (`spec["proprio_groups"]`): `joint_pos`, `joint_vel`, `joint_effort` (12 each, action
order), `imu_gyro`, `imu_accelerometer`, `imu_gravity`, `imu_linvel` (3 each, IMU frame: x forward, y left, z up),
diff --git a/tasks/mujoco_playground_go1_footstand-standard/task.toml b/tasks/mujoco_playground_go1_footstand-standard/task.toml
index aab7960b9f..454f7ed6cc 100644
--- a/tasks/mujoco_playground_go1_footstand-standard/task.toml
+++ b/tasks/mujoco_playground_go1_footstand-standard/task.toml
@@ -16,7 +16,7 @@ artifacts = [{ source = "/var/rcb-monitor/", service = "rcb_monitor" },
[task]
name = "robot-coding-bench/mujoco_playground_go1_footstand-standard"
-description = "STANDARD MODE (robot as a service, eai-standard/2.2): MuJoCo Playground Go1Footstand (Unitree Go1) on the same frozen instance as mujoco_playground_go1_footstand-privileged: Make the Unitree Go1 stand up on its hind legs, nose up, front legs in the air, and balance there for 2 s. The agent only gets a client: four cameras that follow the robot (RGB, optional depth), the robot's joints and IMU, joint position chunks at 50 Hz, the feet's kinematics; one episode, no reset, no world state; the sim service records the episode, replays it in fresh processes and grades it."
+description = "STANDARD MODE (robot as a service, eai-standard/2.2): MuJoCo Playground Go1Footstand (Unitree Go1) on the same frozen instance as mujoco_playground_go1_footstand-privileged: Make the Unitree Go1 stand up on its hind legs, nose up, front legs in the air, and balance there for 2 s. The agent only gets a client: four cameras that follow the robot (RGB, optional depth), the robot's joints and IMU, joint position chunks at 50 Hz, the feet's kinematics; 10 resets (only the last episode is graded), no world state; the sim service records the episode, replays it in fresh processes and grades it."
authors = [{ name = "Changhe Chen" }]
keywords = ["robotics", "mujoco", "mujoco-playground", "unitree-go1", "quadruped", "locomotion", "stand_on_two", "code-as-policy", "trajectory", "standard-mode"]
@@ -28,6 +28,7 @@ simulator = "mujoco"
robot = "unitree-go1"
family = "mujoco_playground/Go1Footstand"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "full"
instance_seed = 0
diff --git a/tasks/mujoco_playground_go1_footstand-standard/tests/grade.py b/tasks/mujoco_playground_go1_footstand-standard/tests/grade.py
index 41ce495421..3d4c23ad7e 100644
--- a/tasks/mujoco_playground_go1_footstand-standard/tests/grade.py
+++ b/tasks/mujoco_playground_go1_footstand-standard/tests/grade.py
@@ -77,6 +77,9 @@ write(
"sim_error": int(d.get("reason") == "error"),
"sim_startups": int(d.get("sim_startups") or 0),
"resumes": int(d.get("resumes") or 0),
+ "resets": int(d.get("resets") or 0), # PROTOCOL §4.5: resets the agent used; only the last episode is graded
+ "reset_budget": int(d.get("reset_budget") or 0),
+ "episodes": int(d.get("episodes") or 1),
"resume_state_mismatch": int(d.get("resume_state_mismatch") or 0),
"resume_max_joint_dev": num("resume_max_joint_dev"),
},
diff --git a/tasks/mujoco_playground_go1_getup-privileged/README.md b/tasks/mujoco_playground_go1_getup-privileged/README.md
index 71ed994d5e..760e9e5769 100644
--- a/tasks/mujoco_playground_go1_getup-privileged/README.md
+++ b/tasks/mujoco_playground_go1_getup-privileged/README.md
@@ -53,3 +53,11 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
### Provenance
Generated by `scripts/mujoco-playground/generate.py` from `scripts/mujoco-playground/instances/Go1Getup.json`
(`make_instances.py`, `select_instances.py`) and `scripts/mujoco-playground/oracles/Go1Getup.npy` (`oracle.py`).
+
+### Change Log
+- 2026-10-14 (review of board 0105, B1-m2): `tests/grade.py` only. A replay process that died or ran out of time, a
+ rebuilt start that does not match the instance's hash (`instance_mismatch`) and an exception of the verifier itself
+ now write `infra_error = 1` beside `grader_error = 1`: no verdict on the agent. Before, they wrote `grader_error = 1`
+ alone, which counted as the agent's 0. Every `reward.json` now carries `infra_error`. The scene, the instance, the
+ pass rule and the reference are unchanged, so the task keeps its name; a past 0 with `grader_error = 1` was an
+ infrastructure failure.
diff --git a/tasks/mujoco_playground_go1_getup-privileged/task.toml b/tasks/mujoco_playground_go1_getup-privileged/task.toml
index 33cc9c7b1a..aaa455dcd7 100644
--- a/tasks/mujoco_playground_go1_getup-privileged/task.toml
+++ b/tasks/mujoco_playground_go1_getup-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "mujoco"
robot = "unitree-go1"
family = "mujoco_playground/Go1Getup"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 5 # the PRNGKey of Playground's reset (and, for the joystick tasks, its command process)
expert_time_estimate_hours = 2.0
diff --git a/tasks/mujoco_playground_go1_getup-privileged/tests/grade.py b/tasks/mujoco_playground_go1_getup-privileged/tests/grade.py
index c75c326868..1667d090c8 100644
--- a/tasks/mujoco_playground_go1_getup-privileged/tests/grade.py
+++ b/tasks/mujoco_playground_go1_getup-privileged/tests/grade.py
@@ -6,6 +6,11 @@ Two replays, each in its own fresh process (`rcb_quadruped.py replay`: the froze
task's success check on the whole trajectory). success = both replays succeed and end in the same state. Writes
/logs/verifier/reward.txt (1/0) and reward.json (numbers only); then a third replay renders replay.mp4, after the
verdict is on disk (it never changes it).
+
+Failures that are not the agent's (docs/PROTOCOL.md: infra vs agent failure) carry `grader_error` = 1 and `infra_error`
+= 1, so they are no verdict on the agent: a replay process that died or timed out, an image that no longer rebuilds the
+frozen start (`instance_mismatch`), and any exception of the verifier itself. A file the agent got wrong is the agent's
+0 (`bad_trajectory`, infra_error 0). Every reward.json carries success, deterministic, grader_error and infra_error.
"""
import argparse
import json
@@ -22,8 +27,12 @@ TOOL = os.path.join(os.path.dirname(os.path.abspath(__file__)), "rcb_quadruped.p
REPLAY_TIMEOUT_S = 300 # a 1000-step replay takes a few seconds; the video replay renders 4 views every 2nd step
+BASE = {"success": 0, "deterministic": 0, "grader_error": 0, "infra_error": 0}
+
+
def write(ok: bool, out: dict) -> None:
os.makedirs(LOG_DIR, exist_ok=True)
+ out = {**BASE, **out}
open(os.path.join(LOG_DIR, "reward.txt"), "w").write("1\n" if ok else "0\n")
json.dump(out, open(os.path.join(LOG_DIR, "reward.json"), "w"))
print(json.dumps(out), flush=True)
@@ -96,11 +105,11 @@ def main() -> int:
"n_actions": int(acts.shape[0]) if acts.ndim == 2 else 0, "grader_error": 0, **cost})
return 0
runs = [replay(a.instance, a.traj, None), replay(a.instance, a.traj, None)]
- if any("error" in r and not r.get("bad_trajectory") for r in runs):
- write(False, {"success": 0, "deterministic": 0, "grader_error": 1, **cost})
+ if any("error" in r and not r.get("bad_trajectory") for r in runs): # the replay process died / timed out
+ write(False, {"success": 0, "deterministic": 0, "grader_error": 1, "infra_error": 1, "replay_failed": 1, **cost})
return 0
if not all(r.get("instance_ok") for r in runs): # the image no longer builds the frozen start: not the agent's 0
- write(False, {"success": 0, "deterministic": 0, "grader_error": 1, "instance_mismatch": 1, **cost})
+ write(False, {"success": 0, "deterministic": 0, "grader_error": 1, "infra_error": 1, "instance_mismatch": 1, **cost})
return 0
all_success = all(bool(r.get("success")) for r in runs)
deterministic = runs[0].get("state_sha") == runs[1].get("state_sha")
@@ -116,7 +125,7 @@ def main() -> int:
write(ok, out) # the verdict, before any video
except Exception: # noqa: BLE001
traceback.print_exc(file=sys.stderr)
- write(False, {"success": 0, "deterministic": 0, "grader_error": 1, **cost})
+ write(False, {"success": 0, "deterministic": 0, "grader_error": 1, "infra_error": 1, **cost})
return 0
try: # the video, for humans: it never changes the verdict
if replay(a.instance, a.traj, os.path.join(LOG_DIR, "replay.mp4")).get("video"):
diff --git a/tasks/mujoco_playground_go1_getup-standard/README.md b/tasks/mujoco_playground_go1_getup-standard/README.md
index 32d7d9aed3..04a9b006b7 100644
--- a/tasks/mujoco_playground_go1_getup-standard/README.md
+++ b/tasks/mujoco_playground_go1_getup-standard/README.md
@@ -13,14 +13,14 @@
## Task
MuJoCo Playground `Go1Getup` (Unitree Go1) on the same frozen instance as `mujoco_playground_go1_getup-privileged`: The Unitree Go1 has fallen over. Make it get up and stand upright at its standing height.
-It starts lying with its trunk tilted 110 deg (Playground's random drop, then a 0.5 s settle) and has 6 s to stand upright at its standing height. Standard mode (eai-standard/2.2): the agent drives the robot through a client in one episode, with only
-the robot's own sensors and four cameras that follow it.
+It starts lying with its trunk tilted 110 deg (Playground's random drop, then a 0.5 s settle) and has 6 s to stand upright at its standing height. Standard mode (eai-standard/2.2): the agent drives the robot through a client, with 10 resets (only
+the last episode is graded), the robot's own sensors and the cameras that follow it.
## Difficulty
-Same instance and success check as `mujoco_playground_go1_getup-privileged`. In standard mode the agent loses resets, the simulator and
+Same instance and success check as `mujoco_playground_go1_getup-privileged`. In standard mode the agent loses free resets, the simulator and
the world state: it sees camera images and proprioception (joints, IMU, foot contacts), acts in joint position chunks,
-and gets one episode, so it has to close the loop on the robot's own sensors and recover from its own mistakes without
-starting over.
+and gets 10 resets (`RCB_RESET_BUDGET`, the compose file's default; PROTOCOL §4.5), so it has to close the loop
+on the robot's own sensors; only the last episode is graded.
## Reference solution
`solution/oracle_replay.py` streams the privileged reference's rows (`solution/oracle.npy`, 136 rows, float32-exact)
@@ -60,3 +60,4 @@ Generated by `scripts/mujoco-playground/generate.py`, as its privileged twin.
- 2026-10-10 (review 0067): a `final.json` that is missing, unreadable or was never finalized by the collect hook is graded `infra_error = 1` (no verdict on the agent); before, a missing one was a plain 0 and an unfinalized one was graded as it stood. The collect hook writes `final.json` atomically. A request field of the wrong type (`dt` or `seconds` null, `cameras` not a list of names, a `note` that is not a string) is answered 400 before any physics instead of ending the episode with a 500. `meta["time_left_s"]` counts from the agent's first call, not from the service's start.
- 2026-10-10 (review round 2, board 0067): the client only (`environment/rcb_client.py`). `Robot()` waits up to 3600 s for the service (was 600 s) and prints its progress to stderr once a minute, so the first `Robot()` outlasts the simulator's start (no effect on this benchmark, whose service is up within a minute); `cameras="name"` (a string) is one camera, no longer its characters. The service, the collect hook, the grader, the scene and the pass rule are unchanged.
- 2026-10-10 (review round 3, board 0067): the shared service layer and the client only (`environment/sim/rcb_limited.py`, `environment/rcb_client.py`). A camera named more than once in `cameras` is rendered and returned once (N copies of one name made the service build N x N images within the frame budget, which could exhaust its memory). A service that has given up on the episode (it could not start it, or could not resume it after its crashes) answers every request with 409 and the reason, and `Robot()` / `wait_ready()` raise `EpisodeOver` at once; before, nothing answered and they waited out their timeout (up to an hour). The adapter, the collect hook, the grader, the scene and the pass rule are unchanged.
+- 2026-10-14 (owner decision C1, board 0113): 10 resets (PROTOCOL §4.5). The service (`environment/sim/server.py`) gained the optional reset capability: `RCB_RESET_BUDGET` (the compose file's default 10; 0 restores the old single episode), `POST /reset` abandons the running or ended episode and starts a new one from the same frozen start (no physics), refused (409) after a success, once the agent's time is up and once the budget is spent; only the last episode is graded (`trajectory.npz` holds its rows, replayed twice from the frozen start as before); `/status` and every `meta` add `episode`, `resets_left` and `control_steps_total`; `final.json` and `reward.json` add `resets`, `reset_budget` and `episodes`. A crash resumes the current episode (the journal is replayed with the committed resets). The instruction's rules, the client example (`POST /reset` through `Robot._post`) and this README say so. The scene, the instance, the success check and the grader's pass rule are unchanged; results recorded before this change ran with one episode and are reported separately.
diff --git a/tasks/mujoco_playground_go1_getup-standard/environment/docker-compose.yaml b/tasks/mujoco_playground_go1_getup-standard/environment/docker-compose.yaml
index bc0c363241..1226b8c3a6 100644
--- a/tasks/mujoco_playground_go1_getup-standard/environment/docker-compose.yaml
+++ b/tasks/mujoco_playground_go1_getup-standard/environment/docker-compose.yaml
@@ -28,6 +28,7 @@ services:
RCB_STARTUP_TIMEOUT_S: "300" # the scene builds in about a second
PYTHONPATH: /srv
RCB_SESSION_LOG: "" # the sidecar's own Session must not write the agent-side interaction log
+ RCB_RESET_BUDGET: "${RCB_RESET_BUDGET:-10}" # Playground quadrupeds standard: 10 resets (PROTOCOL §4.5); 0: one episode
# >>> rcb_monitor (resource limits monitor, docs/PROTOCOL.md §2.10; generated by scripts/rcb_monitor/inject.py)
rcb_monitor:
image: ghcr.io/mll-lab-nu/eai-monitor:0.1.0@sha256:f2c910f53f79c2bfa4a71298b10c78383cae3afc40720ca562cc845b8d20fb0c
diff --git a/tasks/mujoco_playground_go1_getup-standard/environment/sim/server.py b/tasks/mujoco_playground_go1_getup-standard/environment/sim/server.py
index 108eca3cae..cad8920259 100644
--- a/tasks/mujoco_playground_go1_getup-standard/environment/sim/server.py
+++ b/tasks/mujoco_playground_go1_getup-standard/environment/sim/server.py
@@ -1,16 +1,20 @@
"""eai-standard/2.2 sim service for MuJoCo Playground's quadruped tasks (docs/STANDARD_MODE.md): the adapter of the
shared layer rcb_limited.py; the simulator is rcb_quadruped.Session (plain MuJoCo on the CPU).
-HTTP surface: GET /spec, GET /status, POST /observe, POST /submit, POST /kinematics. Every call is logged to
-/data/episode.jsonl.
+HTTP surface: GET /spec, GET /status, POST /observe, POST /submit, POST /kinematics, and POST /reset when the run grants
+resets (RCB_RESET_BUDGET, the optional reset capability of docs/STANDARD_MODE.md / PROTOCOL §4.5; the quadrupeds'
+standard tasks default to 10 in their compose file, owner 2026-10-14). Every call is logged to /data/episode.jsonl.
The episode is the frozen instance's: one control step = the instance's control period (Go1, Spot: 1/50 s; Barkour:
1/100 s). A joystick episode ends after the whole command schedule
(the horizon), when the robot meets Playground's termination rule (it falls over: a failure), when the agent submits
`done`, or when its time is over; success = the task's check on the steps executed. A getup episode ends the moment the
-check holds (success), after the horizon, on `done` or on timeout.
+check holds (success), after the horizon, on `done` or on timeout. With a reset budget, POST /reset abandons the running
+(or ended) episode and starts a new one from the frozen instance; only the last episode is graded, and a success ends
+the trial.
Crash recovery (rcb_limited.Episode, resume class `exact`: MuJoCo on the CPU repeats bit for bit in a fresh process):
-every executed row is journaled; after a crash the service rebuilds the episode and replays the journal. When the
+every executed row is journaled; after a crash the service rebuilds the episode and replays the journal, resetting where
+the committed resets did (their journal rows are checkpointed, `persist`). When the
episode ends the service writes /data/trajectory.npz + /data/episode.json; the collect hook (finalize.sh) makes it exit
and replays the trajectory in two fresh processes."""
@@ -31,18 +35,21 @@ INSTANCE = os.environ.get("RCB_INSTANCE", os.path.join(os.path.dirname(os.path.a
WALL_CLOCK_S = float(os.environ.get("RCB_WALL_CLOCK_S", "3600"))
MAX_CHUNK_S = float(os.environ.get("RCB_MAX_CHUNK_S", "10"))
MAX_WATCH_S = float(os.environ.get("RCB_MAX_WATCH_S", "30"))
+RESET_BUDGET = max(0, int(os.environ.get("RCB_RESET_BUDGET") or 0)) # the compose file's default: 10 (PROTOCOL §4.5)
CAM_W, CAM_H = 320, 240
LOCK = threading.Lock()
PID_FILE = os.path.join(DATA, "server.pid")
E = {"ready": False, "step": 0, "calls": 0, "done": False, "reason": None, "success": False, "t0": time.time(), "error": None,
- "t_first": None} # t_first: the agent's first call (its clock starts there; kept through a crash by the checkpoints)
+ "t_first": None, # t_first: the agent's first call (its clock starts there; kept through a crash by the checkpoints)
+ "resets": 0, # resets done (the episode is resets + 1)
+ "reset_at": []} # the journal row at which each reset happened: a resume resets there again
os.environ["RCB_SESSION_LOG"] = ""
os.environ["RCB_AGENT_TRACE_DIR"] = ""
import rcb_quadruped as Q # noqa: E402 (first: it picks the rendering backend before mujoco loads)
import mujoco # noqa: E402
-EP = L.Episode(DATA, dim=Q.N, resume="exact", persist=("t_first",))
+EP = L.Episode(DATA, dim=Q.N, resume="exact", persist=("t_first", "resets", "reset_at"))
EP.startup() # start count, back-off after failed starts, the startup watchdog until scene_built()
S = Q.Session(INSTANCE)
ROWS: list = [] # every executed row (a resume replays them first): trajectory.npz
@@ -153,13 +160,15 @@ def schedule_text() -> list:
"wz_radps": round(c["wz"], 4)} for c in S.inst["commands"]]
+ONE = "one episode, no reset" if not RESET_BUDGET else (
+ f"an episode (with up to {RESET_BUDGET} resets: POST /reset starts a new one; only the last is graded)")
TASK_EPISODE = {
- "joystick": f"one episode, no reset: it ends after {S.horizon} control steps (the {S.horizon * DT:g} s command "
+ "joystick": f"{ONE}: it ends after {S.horizon} control steps (the {S.horizon * DT:g} s command "
f"schedule), when the robot meets its termination rule ({S.spec.get('term_text')}: a failure), when you "
"submit `done`, or when your time is over",
- "getup": f"one episode, no reset: it ends the moment the robot has stood upright for {Q.HOLD_STEPS} control steps in a "
+ "getup": f"{ONE}: it ends the moment the robot has stood upright for {Q.HOLD_STEPS} control steps in a "
f"row (success), after {S.horizon} control steps, when you submit `done`, or when your time is over",
- "stand_on_two": f"one episode, no reset: it ends the moment the pose has been held for {Q.POSE_HOLD_STEPS} control steps "
+ "stand_on_two": f"{ONE}: it ends the moment the pose has been held for {Q.POSE_HOLD_STEPS} control steps "
"in a row (success), when the robot meets the termination rule (a failure: its trunk's up axis below "
f"-0.25, or {'a leg' if S.spec.get('hind_leg_contact') else 'a front leg'}'s thigh, calf or hip on the "
f"floor), after {S.horizon} control steps, when you "
@@ -204,7 +213,7 @@ SPEC = {
"world": {"frame": "z up; the robot starts where the episode puts it (no world coordinates are given)", "up": "+z",
"units": {"length": "m", "angle": "rad", "time": "s"}, "quaternion_order": "wxyz",
"camera_axes": "intrinsic_cv / extrinsic_cv: OpenCV camera axes (x right, y down, z forward); cam2world_gl: OpenGL axes"},
- "budget": {"reset": 0, "steps": None, "wall_clock_s": WALL_CLOCK_S}, # + "clock" (rcb_limited.CLOCK_NOTE): what time_left_s counts from
+ "budget": {"reset": RESET_BUDGET, "steps": None, "wall_clock_s": WALL_CLOCK_S}, # + "clock" (rcb_limited.CLOCK_NOTE): what time_left_s counts from
"max_control_steps": S.horizon,
"episode": TASK_EPISODE,
"command_schedule": schedule_text(),
@@ -223,6 +232,15 @@ SPEC = {
"depth": {"modes": list(L.DEPTH_MODES), "default": "none", "kind": "z-depth", "unit": "m", "invalid": 0,
"max_m": L.DEPTH_MAX_M, "frames_encoding": "uint16 mm, zlib", "noise": L.DEPTH_NOISE},
}
+if RESET_BUDGET:
+ SPEC["reset"] = (
+ f"POST /reset (JSON or RCB1 {{cameras, video_fps, depth}}; no physics) abandons the running episode, or follows "
+ f"one that has ended, and starts a new one from the same frozen starting state, {RESET_BUDGET} times in all; it "
+ "returns the new episode's first frames and proprioception. Only the last episode is graded: an episode you "
+ "reset away is never graded, a success ends the trial (409 after it), and when your time is up the episode then "
+ "running is the one graded. The wall clock never restarts. /status and every meta add `episode`, `resets_left` "
+ "and `control_steps_total` (all episodes); `control_steps` counts the current episode's."
+ )
L.check_spec(SPEC) # a spec that misses a 2.2 field or names a forbidden one never starts the episode
EP.scene_built()
@@ -277,6 +295,7 @@ def summary() -> dict:
"reason": E["reason"], "done": E["done"], "env": S.env, "protocol": L.PROTOCOL, "n_actions": len(ROWS),
"fell_at": chk.get("fell_at"), "worst_lin_err": chk.get("worst_lin_err"), "worst_yaw_err": chk.get("worst_yaw_err"),
"tilt_deg": chk.get("tilt_deg"), "imu_height": chk.get("imu_height"), "pose_angle_deg": chk.get("pose_angle_deg"),
+ "episodes": E["resets"] + 1, "resets": E["resets"], "reset_budget": RESET_BUDGET, "control_steps_total": EP.rows,
**EP.final_fields()}
@@ -294,7 +313,8 @@ def finish(reason: str) -> None:
E["success"] = bool(chk["success"])
except Exception: # noqa: BLE001
traceback.print_exc()
- EP.log({"ev": "end", "reason": reason, "step": E["step"], "control_steps": ctrl_steps(), "success": bool(E["success"])})
+ EP.log({"ev": "end", "reason": reason, "step": E["step"], "control_steps": ctrl_steps(), "success": bool(E["success"]),
+ "episode": E["resets"] + 1, "journal_rows": EP.rows})
dump()
@@ -388,7 +408,7 @@ def meta_common(cams) -> dict:
"sim_time_s": round(n * DT, 3), "time_s": round(_agent_time(), 2),
"time_left_s": max(0.0, round(WALL_CLOCK_S - _agent_time(), 1)),
"proprio": {"layout": PROPRIO_LAYOUT, "now": proprio_row().tolist()}, "done": E["done"], "reason": E["reason"],
- "cameras": {c: camera_info(c) for c in cams}}
+ "cameras": {c: camera_info(c) for c in cams}, **reset_counters()}
if S.kind == "joystick":
out["command_now"] = dict(zip(("vx_mps", "vy_mps", "wz_radps"), S.command_now().round(4).tolist()))
return out
@@ -478,10 +498,54 @@ def do_observe(meta: dict):
return out, rec.parts(video if seconds > 0 else ("frames" if video == "mp4" else video), fps)
+def reset_counters() -> dict:
+ if not RESET_BUDGET:
+ return {}
+ left = 0 if (E["success"] or E.get("exiting")) else RESET_BUDGET - E["resets"]
+ return {"episode": E["resets"] + 1, "resets_left": left, "control_steps_total": EP.rows}
+
+
def status() -> dict:
_touch()
return {"step": E["step"], "calls": E["calls"], "control_steps": ctrl_steps(), "done": E["done"], "reason": E["reason"],
- "elapsed_s": round(time.time() - E["t0"], 1)}
+ "elapsed_s": round(time.time() - E["t0"], 1), **reset_counters()}
+
+
+class Refused(Exception):
+ """A /reset the service turns down (409): nothing changes."""
+
+ def __init__(self, body: dict):
+ super().__init__(body.get("error"))
+ self.body = body
+
+
+def do_reset(meta: dict):
+ """POST /reset: a new episode from the frozen instance (the running one, if any, is abandoned, never graded)."""
+ meta = _request(meta)
+ cams, fps = cameras_of(meta), video_fps(meta)
+ L.depth_mode(meta) # a bad depth value is a 400 before anything changes
+ if E["success"]:
+ raise Refused({"error": "trial finished: an episode succeeded", "reason": E["reason"], "resets_left": 0})
+ if E["resets"] >= RESET_BUDGET:
+ raise Refused({"error": "reset budget exhausted", "reason": "reset budget exhausted", "budget": RESET_BUDGET,
+ "resets_left": 0})
+ abandoned = {"episode": E["resets"] + 1, "control_steps": ctrl_steps(), "done": E["done"], "reason": E["reason"]}
+ if not E["done"]: # the abandoned episode's end, for the log (never graded: its rows leave the trajectory)
+ EP.log({"ev": "end", "reason": "reset", "step": E["step"], "control_steps": ctrl_steps(), "success": False,
+ "episode": E["resets"] + 1, "journal_rows": EP.rows})
+ S.reset()
+ ROWS.clear()
+ E["resets"] += 1
+ E["reset_at"] = list(E["reset_at"]) + [EP.rows]
+ E.update(done=False, reason=None, success=False, t_end=None, episode_control_steps=None, dumped=False, check=None)
+ EP._ck = None # the state changed without a journal row: the checkpoint after this call must hash it anew
+ E["calls"] += 1
+ EP.log({"ev": "reset", "call": E["calls"], "episode": E["resets"] + 1, "journal_rows": EP.rows,
+ "abandoned": abandoned, "state_sha": S.state_sha()})
+ rec = recorder(cams, fps, meta)
+ rec.tick(force=True, last=True)
+ out = {**meta_common(cams), "reset_id": E["resets"], "ok": True}
+ return out, rec.parts("frames", fps)
def end_from_signal() -> None:
@@ -498,20 +562,106 @@ def end_from_signal() -> None:
os._exit(0)
+# A resume replays the journal from the frozen instance and resets where the committed resets did: their journal rows
+# are in the last checkpoint (`persist`), the one Episode.begin resumes from.
+_CPS = [e for e in EP.events() if e.get("ev") == "checkpoint"]
+RESUME = {"at": list((_CPS[-1].get("persist") or {}).get("reset_at") or []) if _CPS else [], "i": 0}
+
+
+def _resume_resets() -> None:
+ while RESUME["at"] and int(RESUME["at"][0]) <= RESUME["i"]:
+ RESUME["at"].pop(0)
+ S.reset()
+ ROWS.clear()
+
+
def replay_row(row) -> None:
"""A journaled row during a resume."""
+ _resume_resets()
S.step(row)
ROWS.append(np.asarray(row, dtype=np.float64))
+ RESUME["i"] += 1
-EP.begin(state=E, step=replay_row, state_sha=S.state_sha, proprio=proprio_row, spec=SPEC, on_end=end_from_signal,
- start_event={"env": S.env, "seed": S.inst.get("seed"), "cameras": CAMS})
+def resume_sha() -> str:
+ _resume_resets() # a reset after the last replayed row (or before the first)
+ return S.state_sha()
+
+
+RESUMED = EP.begin(state=E, step=replay_row, state_sha=resume_sha, proprio=proprio_row, spec=SPEC, on_end=end_from_signal,
+ start_event={"env": S.env, "seed": S.inst.get("seed"), "cameras": CAMS})
+if RESUMED and E["resets"]: # Episode.begin looks for an "end" of any episode: only the current one's counts
+ _ends = [e for e in EP.events() if e.get("ev") == "end" and int(e.get("episode") or 1) == E["resets"] + 1
+ and e.get("reason") != "reset" and int(e.get("journal_rows") or 0) <= EP.rows]
+ if _ends:
+ E.update(done=True, reason=_ends[-1].get("reason"), success=bool(_ends[-1].get("success")),
+ t_end=float(_ends[-1]["ts"]), episode_control_steps=int(_ends[-1].get("control_steps") or ctrl_steps()))
+ else:
+ E.update(done=False, reason=None, success=False, t_end=None, episode_control_steps=None)
if E["done"]:
dump() # resumed after the episode had ended: the record is rewritten from the replay
E["ready"] = True
write_final()
open(PID_FILE, "w").write(str(os.getpid()))
print(f"{L.PROTOCOL} MuJoCo Playground quadruped sim ready ({S.env}; startup {EP.n_startups}, resumes {EP.resumes})", flush=True)
-L.serve(L.Service(spec=SPEC, lock=LOCK, state=E, submit=do_submit, observe=do_observe, kinematics=do_kinematics, status=status,
- on_submit_error=lambda exc: finish("error"), log=EP.log, episode=EP, state_sha=S.state_sha,
- proprio=proprio_row), port=int(os.environ.get("RCB_PORT", "8000")))
+
+
+def make_handler(svc: L.Service):
+ """The shared 2.2 handler, plus POST /reset when the run grants resets (the shared layer has no reset)."""
+ base = L.make_handler(svc)
+
+ class H(base):
+ def do_POST(self):
+ if self.path != "/reset" or not RESET_BUDGET:
+ return super().do_POST()
+ try:
+ meta, _ = self._body()
+ except Exception as exc: # noqa: BLE001
+ return self._json(413 if "too large" in str(exc) else 400, {"error": str(exc)[:200]})
+ with LOCK:
+ if E.get("exiting"):
+ return self._json(409, {"error": "trial finished: your time is up", "reason": E.get("reason"),
+ "resets_left": 0})
+ cid = EP.call_begin(self.path, self.connection)
+ out = blobs = err = None
+ try:
+ out, blobs = do_reset(meta)
+ except Refused as exc:
+ err = (409, exc.body)
+ except ValueError as exc:
+ err = (400, {"error": str(exc)})
+ except Exception as exc: # noqa: BLE001
+ traceback.print_exc()
+ E["error"] = repr(exc)
+ EP.log({"ev": "error", "path": self.path, "error": repr(exc)})
+ try:
+ finish("error")
+ except Exception: # noqa: BLE001
+ traceback.print_exc()
+ err = (500, {"error": "simulator error", "detail": repr(exc)[:200]})
+ finally:
+ if out is not None or (err and err[0] == 500):
+ try:
+ EP.checkpoint(E, S.state_sha, proprio_row)
+ except Exception: # noqa: BLE001
+ traceback.print_exc()
+ EP.call_end(cid, self.path)
+ if err is not None:
+ return self._json(*err)
+ return self._rcb1({**out, "stopped": None}, blobs)
+
+ return H
+
+
+def serve(svc: L.Service, port: int) -> None:
+ """L.serve with this service's handler (the same spec defaults and check, and L.bind: a port still held when the
+ service restarts is waited for, not a crash)."""
+ svc.spec.setdefault("on_service_error", L.ON_SERVICE_ERROR)
+ svc.spec.setdefault("max_call_wall_s", EP.max_call_wall_s)
+ L.check_spec(svc.spec)
+ L.bind(port, make_handler(svc)).serve_forever()
+
+
+serve(L.Service(spec=SPEC, lock=LOCK, state=E, submit=do_submit, observe=do_observe, kinematics=do_kinematics, status=status,
+ on_submit_error=lambda exc: finish("error"), log=EP.log, episode=EP, state_sha=S.state_sha,
+ proprio=proprio_row), port=int(os.environ.get("RCB_PORT", "8000")))
diff --git a/tasks/mujoco_playground_go1_getup-standard/instruction.md b/tasks/mujoco_playground_go1_getup-standard/instruction.md
index 5783d84245..1277182f7a 100644
--- a/tasks/mujoco_playground_go1_getup-standard/instruction.md
+++ b/tasks/mujoco_playground_go1_getup-standard/instruction.md
@@ -2,15 +2,23 @@
**Task:** The Unitree Go1 has fallen over. Make it get up and stand upright at its standing height.
-You control the Unitree Go1 quadruped — **as if it were real hardware** — in one episode of MuJoCo Playground's
+You control the Unitree Go1 quadruped — **as if it were real hardware** — in MuJoCo Playground's
`Go1Getup` environment (plain MuJoCo 3.3.7 with Playground's scene, commit 4057c14). The simulator runs in a service you
-cannot open. You get one client and one episode.
+cannot open. You get one client and up to 11 episodes.
## The episode
The robot was dropped from 0.5 m in a random orientation with random joint angles and a random base velocity and settled for 0.5 s with its joints held: at the start it lies with its trunk tilted 110 deg from upright. You have 300 control steps (6 s).
## Rules (real-robot rules)
-- **No reset.** One episode. You see four cameras that follow the robot (RGB, depth if you ask) and the robot's own
+- **Episodes and resets.** You have **10 resets** (`spec["budget"]["reset"]`, which is authoritative: if it
+ says 0, there is no reset and the one episode is the run). A reset (`POST /reset`, below) puts the robot back on the
+ same frozen starting state and starts a new episode, either abandoning the running one or after one has ended; it
+ takes no simulated time and returns the start's images and proprioception. Once an episode has ended, `submit`
+ raises `EpisodeOver` until you reset. **Only the last episode is graded**: an episode you reset away is never graded,
+ a success ends the run at once (no reset after it), and when your time is up the episode then running is the one
+ graded. The wall clock is the run's and never restarts; `meta["episode"]` and `meta["resets_left"]` count the
+ episodes (`meta["control_steps"]` is the current episode's).
+- **No privileged state.** You see the cameras that follow the robot (RGB, depth if you ask) and the robot's own
sensors: joint positions, velocities and torques, the IMU (gyro, accelerometer, gravity's direction, the trunk's
velocity), foot contacts. No world positions, no success check.
- **The episode ends** the moment the robot has been upright for 25 control steps in a row (success), after 300 control steps (`spec["max_control_steps"]`, counted by `meta["control_steps"]`), when you submit `done`, or when your time is up.
@@ -19,8 +27,8 @@ The robot was dropped from 0.5 m in a random orientation with random joint angle
`get_obs(seconds=...)`, which holds the current targets). `meta["time_left_s"]` is the service's count of the limit,
from your first request to the service: it never runs out before your time does, so plan by it.
- Nothing to hand in: the service records the episode and grades it. Leave your script at `/app/output/gen.py`.
-- **If the service crashes**, a call fails with a connection error: it restarts and resumes the SAME episode (never a
- reset) in the state after the last call it completed. `r.wait_ready()` waits for it; then check `r.status()`.
+- **If the service crashes**, a call fails with a connection error: it restarts and resumes the SAME episode (a crash
+ is never a reset and costs none) in the state after the last call it completed. `r.wait_ready()` waits for it; then check `r.status()`.
- Nothing is put into your context automatically: images come back as arrays in Python objects.
## Client (`/opt/tools/rcb_client.py`)
@@ -33,7 +41,8 @@ res = r.submit("joint_position", A, dt=0.02) # A: (T, 12) rows dt apart; res.p
obs = r.get_obs(seconds=1.0, video="mp4") # hold the current targets and watch for 1 s (max 30)
r.fk(q); r.ik({"FR": [x, y, z, 1, 0, 0, 0]}, seed=q) # feet in the trunk frame <-> joints (no physics)
r.submit("done", []) # end the episode now
-r.status() # calls, control steps, done, reason, elapsed time
+r.status() # calls, control steps, done, reason, elapsed time, episode, resets_left
+obs = r._post("/reset", {"cameras": ["track"]}, [], 900) # POST /reset: a new episode (the 2.2 client has no reset())
```
Cameras: `track`, `side`, `back`, `top` (they follow the robot); `spec["cameras"]` gives each one's mounting and intrinsics, and every response carries its pose. Proprioception (`spec["proprio_groups"]`): `joint_pos`, `joint_vel`, `joint_effort` (12 each, action
order), `imu_gyro`, `imu_accelerometer`, `imu_gravity`, `imu_linvel` (3 each, IMU frame: x forward, y left, z up),
diff --git a/tasks/mujoco_playground_go1_getup-standard/task.toml b/tasks/mujoco_playground_go1_getup-standard/task.toml
index c7c83cd599..9fe55a3afa 100644
--- a/tasks/mujoco_playground_go1_getup-standard/task.toml
+++ b/tasks/mujoco_playground_go1_getup-standard/task.toml
@@ -16,7 +16,7 @@ artifacts = [{ source = "/var/rcb-monitor/", service = "rcb_monitor" },
[task]
name = "robot-coding-bench/mujoco_playground_go1_getup-standard"
-description = "STANDARD MODE (robot as a service, eai-standard/2.2): MuJoCo Playground Go1Getup (Unitree Go1) on the same frozen instance as mujoco_playground_go1_getup-privileged: The Unitree Go1 has fallen over. Make it get up and stand upright at its standing height. The agent only gets a client: four cameras that follow the robot (RGB, optional depth), the robot's joints and IMU, joint position chunks at 50 Hz, the feet's kinematics; one episode, no reset, no world state; the sim service records the episode, replays it in fresh processes and grades it."
+description = "STANDARD MODE (robot as a service, eai-standard/2.2): MuJoCo Playground Go1Getup (Unitree Go1) on the same frozen instance as mujoco_playground_go1_getup-privileged: The Unitree Go1 has fallen over. Make it get up and stand upright at its standing height. The agent only gets a client: four cameras that follow the robot (RGB, optional depth), the robot's joints and IMU, joint position chunks at 50 Hz, the feet's kinematics; 10 resets (only the last episode is graded), no world state; the sim service records the episode, replays it in fresh processes and grades it."
authors = [{ name = "Changhe Chen" }]
keywords = ["robotics", "mujoco", "mujoco-playground", "unitree-go1", "quadruped", "locomotion", "getup", "code-as-policy", "trajectory", "standard-mode"]
@@ -28,6 +28,7 @@ simulator = "mujoco"
robot = "unitree-go1"
family = "mujoco_playground/Go1Getup"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "full"
instance_seed = 5
diff --git a/tasks/mujoco_playground_go1_getup-standard/tests/grade.py b/tasks/mujoco_playground_go1_getup-standard/tests/grade.py
index 41ce495421..3d4c23ad7e 100644
--- a/tasks/mujoco_playground_go1_getup-standard/tests/grade.py
+++ b/tasks/mujoco_playground_go1_getup-standard/tests/grade.py
@@ -77,6 +77,9 @@ write(
"sim_error": int(d.get("reason") == "error"),
"sim_startups": int(d.get("sim_startups") or 0),
"resumes": int(d.get("resumes") or 0),
+ "resets": int(d.get("resets") or 0), # PROTOCOL §4.5: resets the agent used; only the last episode is graded
+ "reset_budget": int(d.get("reset_budget") or 0),
+ "episodes": int(d.get("episodes") or 1),
"resume_state_mismatch": int(d.get("resume_state_mismatch") or 0),
"resume_max_joint_dev": num("resume_max_joint_dev"),
},
diff --git a/tasks/mujoco_playground_go1_handstand-privileged/README.md b/tasks/mujoco_playground_go1_handstand-privileged/README.md
index 1ee74d691a..5cbae3aba4 100644
--- a/tasks/mujoco_playground_go1_handstand-privileged/README.md
+++ b/tasks/mujoco_playground_go1_handstand-privileged/README.md
@@ -53,3 +53,11 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
### Provenance
Generated by `scripts/mujoco-playground/generate.py` from `scripts/mujoco-playground/instances/Go1Handstand.json`
(`make_instances.py`, `select_instances.py`) and `scripts/mujoco-playground/oracles/Go1Handstand.npy` (`oracle.py`).
+
+### Change Log
+- 2026-10-14 (review of board 0105, B1-m2): `tests/grade.py` only. A replay process that died or ran out of time, a
+ rebuilt start that does not match the instance's hash (`instance_mismatch`) and an exception of the verifier itself
+ now write `infra_error = 1` beside `grader_error = 1`: no verdict on the agent. Before, they wrote `grader_error = 1`
+ alone, which counted as the agent's 0. Every `reward.json` now carries `infra_error`. The scene, the instance, the
+ pass rule and the reference are unchanged, so the task keeps its name; a past 0 with `grader_error = 1` was an
+ infrastructure failure.
diff --git a/tasks/mujoco_playground_go1_handstand-privileged/task.toml b/tasks/mujoco_playground_go1_handstand-privileged/task.toml
index 111da19841..1ace6bd29f 100644
--- a/tasks/mujoco_playground_go1_handstand-privileged/task.toml
+++ b/tasks/mujoco_playground_go1_handstand-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "mujoco"
robot = "unitree-go1"
family = "mujoco_playground/Go1Handstand"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 0 # the PRNGKey of Playground's reset (and, for the joystick tasks, its command process)
expert_time_estimate_hours = 3.0
diff --git a/tasks/mujoco_playground_go1_handstand-privileged/tests/grade.py b/tasks/mujoco_playground_go1_handstand-privileged/tests/grade.py
index c75c326868..1667d090c8 100644
--- a/tasks/mujoco_playground_go1_handstand-privileged/tests/grade.py
+++ b/tasks/mujoco_playground_go1_handstand-privileged/tests/grade.py
@@ -6,6 +6,11 @@ Two replays, each in its own fresh process (`rcb_quadruped.py replay`: the froze
task's success check on the whole trajectory). success = both replays succeed and end in the same state. Writes
/logs/verifier/reward.txt (1/0) and reward.json (numbers only); then a third replay renders replay.mp4, after the
verdict is on disk (it never changes it).
+
+Failures that are not the agent's (docs/PROTOCOL.md: infra vs agent failure) carry `grader_error` = 1 and `infra_error`
+= 1, so they are no verdict on the agent: a replay process that died or timed out, an image that no longer rebuilds the
+frozen start (`instance_mismatch`), and any exception of the verifier itself. A file the agent got wrong is the agent's
+0 (`bad_trajectory`, infra_error 0). Every reward.json carries success, deterministic, grader_error and infra_error.
"""
import argparse
import json
@@ -22,8 +27,12 @@ TOOL = os.path.join(os.path.dirname(os.path.abspath(__file__)), "rcb_quadruped.p
REPLAY_TIMEOUT_S = 300 # a 1000-step replay takes a few seconds; the video replay renders 4 views every 2nd step
+BASE = {"success": 0, "deterministic": 0, "grader_error": 0, "infra_error": 0}
+
+
def write(ok: bool, out: dict) -> None:
os.makedirs(LOG_DIR, exist_ok=True)
+ out = {**BASE, **out}
open(os.path.join(LOG_DIR, "reward.txt"), "w").write("1\n" if ok else "0\n")
json.dump(out, open(os.path.join(LOG_DIR, "reward.json"), "w"))
print(json.dumps(out), flush=True)
@@ -96,11 +105,11 @@ def main() -> int:
"n_actions": int(acts.shape[0]) if acts.ndim == 2 else 0, "grader_error": 0, **cost})
return 0
runs = [replay(a.instance, a.traj, None), replay(a.instance, a.traj, None)]
- if any("error" in r and not r.get("bad_trajectory") for r in runs):
- write(False, {"success": 0, "deterministic": 0, "grader_error": 1, **cost})
+ if any("error" in r and not r.get("bad_trajectory") for r in runs): # the replay process died / timed out
+ write(False, {"success": 0, "deterministic": 0, "grader_error": 1, "infra_error": 1, "replay_failed": 1, **cost})
return 0
if not all(r.get("instance_ok") for r in runs): # the image no longer builds the frozen start: not the agent's 0
- write(False, {"success": 0, "deterministic": 0, "grader_error": 1, "instance_mismatch": 1, **cost})
+ write(False, {"success": 0, "deterministic": 0, "grader_error": 1, "infra_error": 1, "instance_mismatch": 1, **cost})
return 0
all_success = all(bool(r.get("success")) for r in runs)
deterministic = runs[0].get("state_sha") == runs[1].get("state_sha")
@@ -116,7 +125,7 @@ def main() -> int:
write(ok, out) # the verdict, before any video
except Exception: # noqa: BLE001
traceback.print_exc(file=sys.stderr)
- write(False, {"success": 0, "deterministic": 0, "grader_error": 1, **cost})
+ write(False, {"success": 0, "deterministic": 0, "grader_error": 1, "infra_error": 1, **cost})
return 0
try: # the video, for humans: it never changes the verdict
if replay(a.instance, a.traj, os.path.join(LOG_DIR, "replay.mp4")).get("video"):
diff --git a/tasks/mujoco_playground_go1_handstand-standard/README.md b/tasks/mujoco_playground_go1_handstand-standard/README.md
index fdfa5a04bc..8448dba2db 100644
--- a/tasks/mujoco_playground_go1_handstand-standard/README.md
+++ b/tasks/mujoco_playground_go1_handstand-standard/README.md
@@ -13,14 +13,14 @@
## Task
MuJoCo Playground `Go1Handstand` (Unitree Go1) on the same frozen instance as `mujoco_playground_go1_handstand-privileged`: Make the Unitree Go1 do a handstand: balance on its front legs, nose down, hind legs in the air, for 2 s.
-From standing, it has 10 s to get into a handstand on its front legs and hold it for 2 s (Playground's handstand reward aims at the IMU 0.55 m high). Standard mode (eai-standard/2.2): the agent drives the robot through a client in one episode, with only
-the robot's own sensors and four cameras that follow it.
+From standing, it has 10 s to get into a handstand on its front legs and hold it for 2 s (Playground's handstand reward aims at the IMU 0.55 m high). Standard mode (eai-standard/2.2): the agent drives the robot through a client, with 10 resets (only
+the last episode is graded), the robot's own sensors and the cameras that follow it.
## Difficulty
-Same instance and success check as `mujoco_playground_go1_handstand-privileged`. In standard mode the agent loses resets, the simulator and
+Same instance and success check as `mujoco_playground_go1_handstand-privileged`. In standard mode the agent loses free resets, the simulator and
the world state: it sees camera images and proprioception (joints, IMU, foot contacts), acts in joint position chunks,
-and gets one episode, so it has to close the loop on the robot's own sensors and recover from its own mistakes without
-starting over.
+and gets 10 resets (`RCB_RESET_BUDGET`, the compose file's default; PROTOCOL §4.5), so it has to close the loop
+on the robot's own sensors; only the last episode is graded.
## Reference solution
`solution/oracle_replay.py` streams the privileged reference's rows (`solution/oracle.npy`, 115 rows, float32-exact)
@@ -60,3 +60,4 @@ Generated by `scripts/mujoco-playground/generate.py`, as its privileged twin.
- 2026-10-10 (review 0067): a `final.json` that is missing, unreadable or was never finalized by the collect hook is graded `infra_error = 1` (no verdict on the agent); before, a missing one was a plain 0 and an unfinalized one was graded as it stood. The collect hook writes `final.json` atomically. A request field of the wrong type (`dt` or `seconds` null, `cameras` not a list of names, a `note` that is not a string) is answered 400 before any physics instead of ending the episode with a 500. `meta["time_left_s"]` counts from the agent's first call, not from the service's start.
- 2026-10-10 (review round 2, board 0067): the client only (`environment/rcb_client.py`). `Robot()` waits up to 3600 s for the service (was 600 s) and prints its progress to stderr once a minute, so the first `Robot()` outlasts the simulator's start (no effect on this benchmark, whose service is up within a minute); `cameras="name"` (a string) is one camera, no longer its characters. The service, the collect hook, the grader, the scene and the pass rule are unchanged.
- 2026-10-10 (review round 3, board 0067): the shared service layer and the client only (`environment/sim/rcb_limited.py`, `environment/rcb_client.py`). A camera named more than once in `cameras` is rendered and returned once (N copies of one name made the service build N x N images within the frame budget, which could exhaust its memory). A service that has given up on the episode (it could not start it, or could not resume it after its crashes) answers every request with 409 and the reason, and `Robot()` / `wait_ready()` raise `EpisodeOver` at once; before, nothing answered and they waited out their timeout (up to an hour). The adapter, the collect hook, the grader, the scene and the pass rule are unchanged.
+- 2026-10-14 (owner decision C1, board 0113): 10 resets (PROTOCOL §4.5). The service (`environment/sim/server.py`) gained the optional reset capability: `RCB_RESET_BUDGET` (the compose file's default 10; 0 restores the old single episode), `POST /reset` abandons the running or ended episode and starts a new one from the same frozen start (no physics), refused (409) after a success, once the agent's time is up and once the budget is spent; only the last episode is graded (`trajectory.npz` holds its rows, replayed twice from the frozen start as before); `/status` and every `meta` add `episode`, `resets_left` and `control_steps_total`; `final.json` and `reward.json` add `resets`, `reset_budget` and `episodes`. A crash resumes the current episode (the journal is replayed with the committed resets). The instruction's rules, the client example (`POST /reset` through `Robot._post`) and this README say so. The scene, the instance, the success check and the grader's pass rule are unchanged; results recorded before this change ran with one episode and are reported separately.
diff --git a/tasks/mujoco_playground_go1_handstand-standard/environment/docker-compose.yaml b/tasks/mujoco_playground_go1_handstand-standard/environment/docker-compose.yaml
index bc0c363241..1226b8c3a6 100644
--- a/tasks/mujoco_playground_go1_handstand-standard/environment/docker-compose.yaml
+++ b/tasks/mujoco_playground_go1_handstand-standard/environment/docker-compose.yaml
@@ -28,6 +28,7 @@ services:
RCB_STARTUP_TIMEOUT_S: "300" # the scene builds in about a second
PYTHONPATH: /srv
RCB_SESSION_LOG: "" # the sidecar's own Session must not write the agent-side interaction log
+ RCB_RESET_BUDGET: "${RCB_RESET_BUDGET:-10}" # Playground quadrupeds standard: 10 resets (PROTOCOL §4.5); 0: one episode
# >>> rcb_monitor (resource limits monitor, docs/PROTOCOL.md §2.10; generated by scripts/rcb_monitor/inject.py)
rcb_monitor:
image: ghcr.io/mll-lab-nu/eai-monitor:0.1.0@sha256:f2c910f53f79c2bfa4a71298b10c78383cae3afc40720ca562cc845b8d20fb0c
diff --git a/tasks/mujoco_playground_go1_handstand-standard/environment/sim/server.py b/tasks/mujoco_playground_go1_handstand-standard/environment/sim/server.py
index 108eca3cae..cad8920259 100644
--- a/tasks/mujoco_playground_go1_handstand-standard/environment/sim/server.py
+++ b/tasks/mujoco_playground_go1_handstand-standard/environment/sim/server.py
@@ -1,16 +1,20 @@
"""eai-standard/2.2 sim service for MuJoCo Playground's quadruped tasks (docs/STANDARD_MODE.md): the adapter of the
shared layer rcb_limited.py; the simulator is rcb_quadruped.Session (plain MuJoCo on the CPU).
-HTTP surface: GET /spec, GET /status, POST /observe, POST /submit, POST /kinematics. Every call is logged to
-/data/episode.jsonl.
+HTTP surface: GET /spec, GET /status, POST /observe, POST /submit, POST /kinematics, and POST /reset when the run grants
+resets (RCB_RESET_BUDGET, the optional reset capability of docs/STANDARD_MODE.md / PROTOCOL §4.5; the quadrupeds'
+standard tasks default to 10 in their compose file, owner 2026-10-14). Every call is logged to /data/episode.jsonl.
The episode is the frozen instance's: one control step = the instance's control period (Go1, Spot: 1/50 s; Barkour:
1/100 s). A joystick episode ends after the whole command schedule
(the horizon), when the robot meets Playground's termination rule (it falls over: a failure), when the agent submits
`done`, or when its time is over; success = the task's check on the steps executed. A getup episode ends the moment the
-check holds (success), after the horizon, on `done` or on timeout.
+check holds (success), after the horizon, on `done` or on timeout. With a reset budget, POST /reset abandons the running
+(or ended) episode and starts a new one from the frozen instance; only the last episode is graded, and a success ends
+the trial.
Crash recovery (rcb_limited.Episode, resume class `exact`: MuJoCo on the CPU repeats bit for bit in a fresh process):
-every executed row is journaled; after a crash the service rebuilds the episode and replays the journal. When the
+every executed row is journaled; after a crash the service rebuilds the episode and replays the journal, resetting where
+the committed resets did (their journal rows are checkpointed, `persist`). When the
episode ends the service writes /data/trajectory.npz + /data/episode.json; the collect hook (finalize.sh) makes it exit
and replays the trajectory in two fresh processes."""
@@ -31,18 +35,21 @@ INSTANCE = os.environ.get("RCB_INSTANCE", os.path.join(os.path.dirname(os.path.a
WALL_CLOCK_S = float(os.environ.get("RCB_WALL_CLOCK_S", "3600"))
MAX_CHUNK_S = float(os.environ.get("RCB_MAX_CHUNK_S", "10"))
MAX_WATCH_S = float(os.environ.get("RCB_MAX_WATCH_S", "30"))
+RESET_BUDGET = max(0, int(os.environ.get("RCB_RESET_BUDGET") or 0)) # the compose file's default: 10 (PROTOCOL §4.5)
CAM_W, CAM_H = 320, 240
LOCK = threading.Lock()
PID_FILE = os.path.join(DATA, "server.pid")
E = {"ready": False, "step": 0, "calls": 0, "done": False, "reason": None, "success": False, "t0": time.time(), "error": None,
- "t_first": None} # t_first: the agent's first call (its clock starts there; kept through a crash by the checkpoints)
+ "t_first": None, # t_first: the agent's first call (its clock starts there; kept through a crash by the checkpoints)
+ "resets": 0, # resets done (the episode is resets + 1)
+ "reset_at": []} # the journal row at which each reset happened: a resume resets there again
os.environ["RCB_SESSION_LOG"] = ""
os.environ["RCB_AGENT_TRACE_DIR"] = ""
import rcb_quadruped as Q # noqa: E402 (first: it picks the rendering backend before mujoco loads)
import mujoco # noqa: E402
-EP = L.Episode(DATA, dim=Q.N, resume="exact", persist=("t_first",))
+EP = L.Episode(DATA, dim=Q.N, resume="exact", persist=("t_first", "resets", "reset_at"))
EP.startup() # start count, back-off after failed starts, the startup watchdog until scene_built()
S = Q.Session(INSTANCE)
ROWS: list = [] # every executed row (a resume replays them first): trajectory.npz
@@ -153,13 +160,15 @@ def schedule_text() -> list:
"wz_radps": round(c["wz"], 4)} for c in S.inst["commands"]]
+ONE = "one episode, no reset" if not RESET_BUDGET else (
+ f"an episode (with up to {RESET_BUDGET} resets: POST /reset starts a new one; only the last is graded)")
TASK_EPISODE = {
- "joystick": f"one episode, no reset: it ends after {S.horizon} control steps (the {S.horizon * DT:g} s command "
+ "joystick": f"{ONE}: it ends after {S.horizon} control steps (the {S.horizon * DT:g} s command "
f"schedule), when the robot meets its termination rule ({S.spec.get('term_text')}: a failure), when you "
"submit `done`, or when your time is over",
- "getup": f"one episode, no reset: it ends the moment the robot has stood upright for {Q.HOLD_STEPS} control steps in a "
+ "getup": f"{ONE}: it ends the moment the robot has stood upright for {Q.HOLD_STEPS} control steps in a "
f"row (success), after {S.horizon} control steps, when you submit `done`, or when your time is over",
- "stand_on_two": f"one episode, no reset: it ends the moment the pose has been held for {Q.POSE_HOLD_STEPS} control steps "
+ "stand_on_two": f"{ONE}: it ends the moment the pose has been held for {Q.POSE_HOLD_STEPS} control steps "
"in a row (success), when the robot meets the termination rule (a failure: its trunk's up axis below "
f"-0.25, or {'a leg' if S.spec.get('hind_leg_contact') else 'a front leg'}'s thigh, calf or hip on the "
f"floor), after {S.horizon} control steps, when you "
@@ -204,7 +213,7 @@ SPEC = {
"world": {"frame": "z up; the robot starts where the episode puts it (no world coordinates are given)", "up": "+z",
"units": {"length": "m", "angle": "rad", "time": "s"}, "quaternion_order": "wxyz",
"camera_axes": "intrinsic_cv / extrinsic_cv: OpenCV camera axes (x right, y down, z forward); cam2world_gl: OpenGL axes"},
- "budget": {"reset": 0, "steps": None, "wall_clock_s": WALL_CLOCK_S}, # + "clock" (rcb_limited.CLOCK_NOTE): what time_left_s counts from
+ "budget": {"reset": RESET_BUDGET, "steps": None, "wall_clock_s": WALL_CLOCK_S}, # + "clock" (rcb_limited.CLOCK_NOTE): what time_left_s counts from
"max_control_steps": S.horizon,
"episode": TASK_EPISODE,
"command_schedule": schedule_text(),
@@ -223,6 +232,15 @@ SPEC = {
"depth": {"modes": list(L.DEPTH_MODES), "default": "none", "kind": "z-depth", "unit": "m", "invalid": 0,
"max_m": L.DEPTH_MAX_M, "frames_encoding": "uint16 mm, zlib", "noise": L.DEPTH_NOISE},
}
+if RESET_BUDGET:
+ SPEC["reset"] = (
+ f"POST /reset (JSON or RCB1 {{cameras, video_fps, depth}}; no physics) abandons the running episode, or follows "
+ f"one that has ended, and starts a new one from the same frozen starting state, {RESET_BUDGET} times in all; it "
+ "returns the new episode's first frames and proprioception. Only the last episode is graded: an episode you "
+ "reset away is never graded, a success ends the trial (409 after it), and when your time is up the episode then "
+ "running is the one graded. The wall clock never restarts. /status and every meta add `episode`, `resets_left` "
+ "and `control_steps_total` (all episodes); `control_steps` counts the current episode's."
+ )
L.check_spec(SPEC) # a spec that misses a 2.2 field or names a forbidden one never starts the episode
EP.scene_built()
@@ -277,6 +295,7 @@ def summary() -> dict:
"reason": E["reason"], "done": E["done"], "env": S.env, "protocol": L.PROTOCOL, "n_actions": len(ROWS),
"fell_at": chk.get("fell_at"), "worst_lin_err": chk.get("worst_lin_err"), "worst_yaw_err": chk.get("worst_yaw_err"),
"tilt_deg": chk.get("tilt_deg"), "imu_height": chk.get("imu_height"), "pose_angle_deg": chk.get("pose_angle_deg"),
+ "episodes": E["resets"] + 1, "resets": E["resets"], "reset_budget": RESET_BUDGET, "control_steps_total": EP.rows,
**EP.final_fields()}
@@ -294,7 +313,8 @@ def finish(reason: str) -> None:
E["success"] = bool(chk["success"])
except Exception: # noqa: BLE001
traceback.print_exc()
- EP.log({"ev": "end", "reason": reason, "step": E["step"], "control_steps": ctrl_steps(), "success": bool(E["success"])})
+ EP.log({"ev": "end", "reason": reason, "step": E["step"], "control_steps": ctrl_steps(), "success": bool(E["success"]),
+ "episode": E["resets"] + 1, "journal_rows": EP.rows})
dump()
@@ -388,7 +408,7 @@ def meta_common(cams) -> dict:
"sim_time_s": round(n * DT, 3), "time_s": round(_agent_time(), 2),
"time_left_s": max(0.0, round(WALL_CLOCK_S - _agent_time(), 1)),
"proprio": {"layout": PROPRIO_LAYOUT, "now": proprio_row().tolist()}, "done": E["done"], "reason": E["reason"],
- "cameras": {c: camera_info(c) for c in cams}}
+ "cameras": {c: camera_info(c) for c in cams}, **reset_counters()}
if S.kind == "joystick":
out["command_now"] = dict(zip(("vx_mps", "vy_mps", "wz_radps"), S.command_now().round(4).tolist()))
return out
@@ -478,10 +498,54 @@ def do_observe(meta: dict):
return out, rec.parts(video if seconds > 0 else ("frames" if video == "mp4" else video), fps)
+def reset_counters() -> dict:
+ if not RESET_BUDGET:
+ return {}
+ left = 0 if (E["success"] or E.get("exiting")) else RESET_BUDGET - E["resets"]
+ return {"episode": E["resets"] + 1, "resets_left": left, "control_steps_total": EP.rows}
+
+
def status() -> dict:
_touch()
return {"step": E["step"], "calls": E["calls"], "control_steps": ctrl_steps(), "done": E["done"], "reason": E["reason"],
- "elapsed_s": round(time.time() - E["t0"], 1)}
+ "elapsed_s": round(time.time() - E["t0"], 1), **reset_counters()}
+
+
+class Refused(Exception):
+ """A /reset the service turns down (409): nothing changes."""
+
+ def __init__(self, body: dict):
+ super().__init__(body.get("error"))
+ self.body = body
+
+
+def do_reset(meta: dict):
+ """POST /reset: a new episode from the frozen instance (the running one, if any, is abandoned, never graded)."""
+ meta = _request(meta)
+ cams, fps = cameras_of(meta), video_fps(meta)
+ L.depth_mode(meta) # a bad depth value is a 400 before anything changes
+ if E["success"]:
+ raise Refused({"error": "trial finished: an episode succeeded", "reason": E["reason"], "resets_left": 0})
+ if E["resets"] >= RESET_BUDGET:
+ raise Refused({"error": "reset budget exhausted", "reason": "reset budget exhausted", "budget": RESET_BUDGET,
+ "resets_left": 0})
+ abandoned = {"episode": E["resets"] + 1, "control_steps": ctrl_steps(), "done": E["done"], "reason": E["reason"]}
+ if not E["done"]: # the abandoned episode's end, for the log (never graded: its rows leave the trajectory)
+ EP.log({"ev": "end", "reason": "reset", "step": E["step"], "control_steps": ctrl_steps(), "success": False,
+ "episode": E["resets"] + 1, "journal_rows": EP.rows})
+ S.reset()
+ ROWS.clear()
+ E["resets"] += 1
+ E["reset_at"] = list(E["reset_at"]) + [EP.rows]
+ E.update(done=False, reason=None, success=False, t_end=None, episode_control_steps=None, dumped=False, check=None)
+ EP._ck = None # the state changed without a journal row: the checkpoint after this call must hash it anew
+ E["calls"] += 1
+ EP.log({"ev": "reset", "call": E["calls"], "episode": E["resets"] + 1, "journal_rows": EP.rows,
+ "abandoned": abandoned, "state_sha": S.state_sha()})
+ rec = recorder(cams, fps, meta)
+ rec.tick(force=True, last=True)
+ out = {**meta_common(cams), "reset_id": E["resets"], "ok": True}
+ return out, rec.parts("frames", fps)
def end_from_signal() -> None:
@@ -498,20 +562,106 @@ def end_from_signal() -> None:
os._exit(0)
+# A resume replays the journal from the frozen instance and resets where the committed resets did: their journal rows
+# are in the last checkpoint (`persist`), the one Episode.begin resumes from.
+_CPS = [e for e in EP.events() if e.get("ev") == "checkpoint"]
+RESUME = {"at": list((_CPS[-1].get("persist") or {}).get("reset_at") or []) if _CPS else [], "i": 0}
+
+
+def _resume_resets() -> None:
+ while RESUME["at"] and int(RESUME["at"][0]) <= RESUME["i"]:
+ RESUME["at"].pop(0)
+ S.reset()
+ ROWS.clear()
+
+
def replay_row(row) -> None:
"""A journaled row during a resume."""
+ _resume_resets()
S.step(row)
ROWS.append(np.asarray(row, dtype=np.float64))
+ RESUME["i"] += 1
-EP.begin(state=E, step=replay_row, state_sha=S.state_sha, proprio=proprio_row, spec=SPEC, on_end=end_from_signal,
- start_event={"env": S.env, "seed": S.inst.get("seed"), "cameras": CAMS})
+def resume_sha() -> str:
+ _resume_resets() # a reset after the last replayed row (or before the first)
+ return S.state_sha()
+
+
+RESUMED = EP.begin(state=E, step=replay_row, state_sha=resume_sha, proprio=proprio_row, spec=SPEC, on_end=end_from_signal,
+ start_event={"env": S.env, "seed": S.inst.get("seed"), "cameras": CAMS})
+if RESUMED and E["resets"]: # Episode.begin looks for an "end" of any episode: only the current one's counts
+ _ends = [e for e in EP.events() if e.get("ev") == "end" and int(e.get("episode") or 1) == E["resets"] + 1
+ and e.get("reason") != "reset" and int(e.get("journal_rows") or 0) <= EP.rows]
+ if _ends:
+ E.update(done=True, reason=_ends[-1].get("reason"), success=bool(_ends[-1].get("success")),
+ t_end=float(_ends[-1]["ts"]), episode_control_steps=int(_ends[-1].get("control_steps") or ctrl_steps()))
+ else:
+ E.update(done=False, reason=None, success=False, t_end=None, episode_control_steps=None)
if E["done"]:
dump() # resumed after the episode had ended: the record is rewritten from the replay
E["ready"] = True
write_final()
open(PID_FILE, "w").write(str(os.getpid()))
print(f"{L.PROTOCOL} MuJoCo Playground quadruped sim ready ({S.env}; startup {EP.n_startups}, resumes {EP.resumes})", flush=True)
-L.serve(L.Service(spec=SPEC, lock=LOCK, state=E, submit=do_submit, observe=do_observe, kinematics=do_kinematics, status=status,
- on_submit_error=lambda exc: finish("error"), log=EP.log, episode=EP, state_sha=S.state_sha,
- proprio=proprio_row), port=int(os.environ.get("RCB_PORT", "8000")))
+
+
+def make_handler(svc: L.Service):
+ """The shared 2.2 handler, plus POST /reset when the run grants resets (the shared layer has no reset)."""
+ base = L.make_handler(svc)
+
+ class H(base):
+ def do_POST(self):
+ if self.path != "/reset" or not RESET_BUDGET:
+ return super().do_POST()
+ try:
+ meta, _ = self._body()
+ except Exception as exc: # noqa: BLE001
+ return self._json(413 if "too large" in str(exc) else 400, {"error": str(exc)[:200]})
+ with LOCK:
+ if E.get("exiting"):
+ return self._json(409, {"error": "trial finished: your time is up", "reason": E.get("reason"),
+ "resets_left": 0})
+ cid = EP.call_begin(self.path, self.connection)
+ out = blobs = err = None
+ try:
+ out, blobs = do_reset(meta)
+ except Refused as exc:
+ err = (409, exc.body)
+ except ValueError as exc:
+ err = (400, {"error": str(exc)})
+ except Exception as exc: # noqa: BLE001
+ traceback.print_exc()
+ E["error"] = repr(exc)
+ EP.log({"ev": "error", "path": self.path, "error": repr(exc)})
+ try:
+ finish("error")
+ except Exception: # noqa: BLE001
+ traceback.print_exc()
+ err = (500, {"error": "simulator error", "detail": repr(exc)[:200]})
+ finally:
+ if out is not None or (err and err[0] == 500):
+ try:
+ EP.checkpoint(E, S.state_sha, proprio_row)
+ except Exception: # noqa: BLE001
+ traceback.print_exc()
+ EP.call_end(cid, self.path)
+ if err is not None:
+ return self._json(*err)
+ return self._rcb1({**out, "stopped": None}, blobs)
+
+ return H
+
+
+def serve(svc: L.Service, port: int) -> None:
+ """L.serve with this service's handler (the same spec defaults and check, and L.bind: a port still held when the
+ service restarts is waited for, not a crash)."""
+ svc.spec.setdefault("on_service_error", L.ON_SERVICE_ERROR)
+ svc.spec.setdefault("max_call_wall_s", EP.max_call_wall_s)
+ L.check_spec(svc.spec)
+ L.bind(port, make_handler(svc)).serve_forever()
+
+
+serve(L.Service(spec=SPEC, lock=LOCK, state=E, submit=do_submit, observe=do_observe, kinematics=do_kinematics, status=status,
+ on_submit_error=lambda exc: finish("error"), log=EP.log, episode=EP, state_sha=S.state_sha,
+ proprio=proprio_row), port=int(os.environ.get("RCB_PORT", "8000")))
diff --git a/tasks/mujoco_playground_go1_handstand-standard/instruction.md b/tasks/mujoco_playground_go1_handstand-standard/instruction.md
index 208d22c124..cc7ada6abb 100644
--- a/tasks/mujoco_playground_go1_handstand-standard/instruction.md
+++ b/tasks/mujoco_playground_go1_handstand-standard/instruction.md
@@ -2,15 +2,23 @@
**Task:** Make the Unitree Go1 do a handstand: balance on its front legs, nose down, hind legs in the air, for 2 s.
-You control the Unitree Go1 quadruped — **as if it were real hardware** — in one episode of MuJoCo Playground's
+You control the Unitree Go1 quadruped — **as if it were real hardware** — in MuJoCo Playground's
`Go1Handstand` environment (plain MuJoCo 3.3.7 with Playground's scene, commit 4057c14). The simulator runs in a service you
-cannot open. You get one client and one episode.
+cannot open. You get one client and up to 11 episodes.
## The episode
The robot starts standing at Playground's `home` pose with a random base velocity (Playground's reset). You have 500 control steps (10 s).
## Rules (real-robot rules)
-- **No reset.** One episode. You see four cameras that follow the robot (RGB, depth if you ask) and the robot's own
+- **Episodes and resets.** You have **10 resets** (`spec["budget"]["reset"]`, which is authoritative: if it
+ says 0, there is no reset and the one episode is the run). A reset (`POST /reset`, below) puts the robot back on the
+ same frozen starting state and starts a new episode, either abandoning the running one or after one has ended; it
+ takes no simulated time and returns the start's images and proprioception. Once an episode has ended, `submit`
+ raises `EpisodeOver` until you reset. **Only the last episode is graded**: an episode you reset away is never graded,
+ a success ends the run at once (no reset after it), and when your time is up the episode then running is the one
+ graded. The wall clock is the run's and never restarts; `meta["episode"]` and `meta["resets_left"]` count the
+ episodes (`meta["control_steps"]` is the current episode's).
+- **No privileged state.** You see the cameras that follow the robot (RGB, depth if you ask) and the robot's own
sensors: joint positions, velocities and torques, the IMU (gyro, accelerometer, gravity's direction, the trunk's
velocity), foot contacts. No world positions, no success check.
- **The episode ends** the moment the pose has been held for 100 control steps in a row (success), when the robot meets the termination rule (the trunk's up axis below -0.25 (z component of the IMU frame's z axis), or any part of a front leg other than its foot (thigh, calf, hip) touching the floor: a failure), after 500 control steps (`spec["max_control_steps"]`), when you submit `done`, or when your time is up.
@@ -19,8 +27,8 @@ The robot starts standing at Playground's `home` pose with a random base velocit
`get_obs(seconds=...)`, which holds the current targets). `meta["time_left_s"]` is the service's count of the limit,
from your first request to the service: it never runs out before your time does, so plan by it.
- Nothing to hand in: the service records the episode and grades it. Leave your script at `/app/output/gen.py`.
-- **If the service crashes**, a call fails with a connection error: it restarts and resumes the SAME episode (never a
- reset) in the state after the last call it completed. `r.wait_ready()` waits for it; then check `r.status()`.
+- **If the service crashes**, a call fails with a connection error: it restarts and resumes the SAME episode (a crash
+ is never a reset and costs none) in the state after the last call it completed. `r.wait_ready()` waits for it; then check `r.status()`.
- Nothing is put into your context automatically: images come back as arrays in Python objects.
## Client (`/opt/tools/rcb_client.py`)
@@ -33,7 +41,8 @@ res = r.submit("joint_position", A, dt=0.02) # A: (T, 12) rows dt apart; res.p
obs = r.get_obs(seconds=1.0, video="mp4") # hold the current targets and watch for 1 s (max 30)
r.fk(q); r.ik({"FR": [x, y, z, 1, 0, 0, 0]}, seed=q) # feet in the trunk frame <-> joints (no physics)
r.submit("done", []) # end the episode now
-r.status() # calls, control steps, done, reason, elapsed time
+r.status() # calls, control steps, done, reason, elapsed time, episode, resets_left
+obs = r._post("/reset", {"cameras": ["track"]}, [], 900) # POST /reset: a new episode (the 2.2 client has no reset())
```
Cameras: `track`, `side`, `back`, `top` (they follow the robot); `spec["cameras"]` gives each one's mounting and intrinsics, and every response carries its pose. Proprioception (`spec["proprio_groups"]`): `joint_pos`, `joint_vel`, `joint_effort` (12 each, action
order), `imu_gyro`, `imu_accelerometer`, `imu_gravity`, `imu_linvel` (3 each, IMU frame: x forward, y left, z up),
diff --git a/tasks/mujoco_playground_go1_handstand-standard/task.toml b/tasks/mujoco_playground_go1_handstand-standard/task.toml
index c34df9e3c1..441125d3a8 100644
--- a/tasks/mujoco_playground_go1_handstand-standard/task.toml
+++ b/tasks/mujoco_playground_go1_handstand-standard/task.toml
@@ -16,7 +16,7 @@ artifacts = [{ source = "/var/rcb-monitor/", service = "rcb_monitor" },
[task]
name = "robot-coding-bench/mujoco_playground_go1_handstand-standard"
-description = "STANDARD MODE (robot as a service, eai-standard/2.2): MuJoCo Playground Go1Handstand (Unitree Go1) on the same frozen instance as mujoco_playground_go1_handstand-privileged: Make the Unitree Go1 do a handstand: balance on its front legs, nose down, hind legs in the air, for 2 s. The agent only gets a client: four cameras that follow the robot (RGB, optional depth), the robot's joints and IMU, joint position chunks at 50 Hz, the feet's kinematics; one episode, no reset, no world state; the sim service records the episode, replays it in fresh processes and grades it."
+description = "STANDARD MODE (robot as a service, eai-standard/2.2): MuJoCo Playground Go1Handstand (Unitree Go1) on the same frozen instance as mujoco_playground_go1_handstand-privileged: Make the Unitree Go1 do a handstand: balance on its front legs, nose down, hind legs in the air, for 2 s. The agent only gets a client: four cameras that follow the robot (RGB, optional depth), the robot's joints and IMU, joint position chunks at 50 Hz, the feet's kinematics; 10 resets (only the last episode is graded), no world state; the sim service records the episode, replays it in fresh processes and grades it."
authors = [{ name = "Changhe Chen" }]
keywords = ["robotics", "mujoco", "mujoco-playground", "unitree-go1", "quadruped", "locomotion", "stand_on_two", "code-as-policy", "trajectory", "standard-mode"]
@@ -28,6 +28,7 @@ simulator = "mujoco"
robot = "unitree-go1"
family = "mujoco_playground/Go1Handstand"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "full"
instance_seed = 0
diff --git a/tasks/mujoco_playground_go1_handstand-standard/tests/grade.py b/tasks/mujoco_playground_go1_handstand-standard/tests/grade.py
index 41ce495421..3d4c23ad7e 100644
--- a/tasks/mujoco_playground_go1_handstand-standard/tests/grade.py
+++ b/tasks/mujoco_playground_go1_handstand-standard/tests/grade.py
@@ -77,6 +77,9 @@ write(
"sim_error": int(d.get("reason") == "error"),
"sim_startups": int(d.get("sim_startups") or 0),
"resumes": int(d.get("resumes") or 0),
+ "resets": int(d.get("resets") or 0), # PROTOCOL §4.5: resets the agent used; only the last episode is graded
+ "reset_budget": int(d.get("reset_budget") or 0),
+ "episodes": int(d.get("episodes") or 1),
"resume_state_mismatch": int(d.get("resume_state_mismatch") or 0),
"resume_max_joint_dev": num("resume_max_joint_dev"),
},
diff --git a/tasks/mujoco_playground_go1_joystick_flat_terrain-privileged/README.md b/tasks/mujoco_playground_go1_joystick_flat_terrain-privileged/README.md
index 244997a824..cc4b380882 100644
--- a/tasks/mujoco_playground_go1_joystick_flat_terrain-privileged/README.md
+++ b/tasks/mujoco_playground_go1_joystick_flat_terrain-privileged/README.md
@@ -53,3 +53,11 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
### Provenance
Generated by `scripts/mujoco-playground/generate.py` from `scripts/mujoco-playground/instances/Go1JoystickFlatTerrain.json`
(`make_instances.py`, `select_instances.py`) and `scripts/mujoco-playground/oracles/Go1JoystickFlatTerrain.npy` (`oracle.py`).
+
+### Change Log
+- 2026-10-14 (review of board 0105, B1-m2): `tests/grade.py` only. A replay process that died or ran out of time, a
+ rebuilt start that does not match the instance's hash (`instance_mismatch`) and an exception of the verifier itself
+ now write `infra_error = 1` beside `grader_error = 1`: no verdict on the agent. Before, they wrote `grader_error = 1`
+ alone, which counted as the agent's 0. Every `reward.json` now carries `infra_error`. The scene, the instance, the
+ pass rule and the reference are unchanged, so the task keeps its name; a past 0 with `grader_error = 1` was an
+ infrastructure failure.
diff --git a/tasks/mujoco_playground_go1_joystick_flat_terrain-privileged/task.toml b/tasks/mujoco_playground_go1_joystick_flat_terrain-privileged/task.toml
index 5885a07d30..3652c4c69f 100644
--- a/tasks/mujoco_playground_go1_joystick_flat_terrain-privileged/task.toml
+++ b/tasks/mujoco_playground_go1_joystick_flat_terrain-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "mujoco"
robot = "unitree-go1"
family = "mujoco_playground/Go1JoystickFlatTerrain"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 0 # the PRNGKey of Playground's reset (and, for the joystick tasks, its command process)
expert_time_estimate_hours = 3.0
diff --git a/tasks/mujoco_playground_go1_joystick_flat_terrain-privileged/tests/grade.py b/tasks/mujoco_playground_go1_joystick_flat_terrain-privileged/tests/grade.py
index c75c326868..1667d090c8 100644
--- a/tasks/mujoco_playground_go1_joystick_flat_terrain-privileged/tests/grade.py
+++ b/tasks/mujoco_playground_go1_joystick_flat_terrain-privileged/tests/grade.py
@@ -6,6 +6,11 @@ Two replays, each in its own fresh process (`rcb_quadruped.py replay`: the froze
task's success check on the whole trajectory). success = both replays succeed and end in the same state. Writes
/logs/verifier/reward.txt (1/0) and reward.json (numbers only); then a third replay renders replay.mp4, after the
verdict is on disk (it never changes it).
+
+Failures that are not the agent's (docs/PROTOCOL.md: infra vs agent failure) carry `grader_error` = 1 and `infra_error`
+= 1, so they are no verdict on the agent: a replay process that died or timed out, an image that no longer rebuilds the
+frozen start (`instance_mismatch`), and any exception of the verifier itself. A file the agent got wrong is the agent's
+0 (`bad_trajectory`, infra_error 0). Every reward.json carries success, deterministic, grader_error and infra_error.
"""
import argparse
import json
@@ -22,8 +27,12 @@ TOOL = os.path.join(os.path.dirname(os.path.abspath(__file__)), "rcb_quadruped.p
REPLAY_TIMEOUT_S = 300 # a 1000-step replay takes a few seconds; the video replay renders 4 views every 2nd step
+BASE = {"success": 0, "deterministic": 0, "grader_error": 0, "infra_error": 0}
+
+
def write(ok: bool, out: dict) -> None:
os.makedirs(LOG_DIR, exist_ok=True)
+ out = {**BASE, **out}
open(os.path.join(LOG_DIR, "reward.txt"), "w").write("1\n" if ok else "0\n")
json.dump(out, open(os.path.join(LOG_DIR, "reward.json"), "w"))
print(json.dumps(out), flush=True)
@@ -96,11 +105,11 @@ def main() -> int:
"n_actions": int(acts.shape[0]) if acts.ndim == 2 else 0, "grader_error": 0, **cost})
return 0
runs = [replay(a.instance, a.traj, None), replay(a.instance, a.traj, None)]
- if any("error" in r and not r.get("bad_trajectory") for r in runs):
- write(False, {"success": 0, "deterministic": 0, "grader_error": 1, **cost})
+ if any("error" in r and not r.get("bad_trajectory") for r in runs): # the replay process died / timed out
+ write(False, {"success": 0, "deterministic": 0, "grader_error": 1, "infra_error": 1, "replay_failed": 1, **cost})
return 0
if not all(r.get("instance_ok") for r in runs): # the image no longer builds the frozen start: not the agent's 0
- write(False, {"success": 0, "deterministic": 0, "grader_error": 1, "instance_mismatch": 1, **cost})
+ write(False, {"success": 0, "deterministic": 0, "grader_error": 1, "infra_error": 1, "instance_mismatch": 1, **cost})
return 0
all_success = all(bool(r.get("success")) for r in runs)
deterministic = runs[0].get("state_sha") == runs[1].get("state_sha")
@@ -116,7 +125,7 @@ def main() -> int:
write(ok, out) # the verdict, before any video
except Exception: # noqa: BLE001
traceback.print_exc(file=sys.stderr)
- write(False, {"success": 0, "deterministic": 0, "grader_error": 1, **cost})
+ write(False, {"success": 0, "deterministic": 0, "grader_error": 1, "infra_error": 1, **cost})
return 0
try: # the video, for humans: it never changes the verdict
if replay(a.instance, a.traj, os.path.join(LOG_DIR, "replay.mp4")).get("video"):
diff --git a/tasks/mujoco_playground_go1_joystick_flat_terrain-standard/README.md b/tasks/mujoco_playground_go1_joystick_flat_terrain-standard/README.md
index b73a926e1e..2267e25b1c 100644
--- a/tasks/mujoco_playground_go1_joystick_flat_terrain-standard/README.md
+++ b/tasks/mujoco_playground_go1_joystick_flat_terrain-standard/README.md
@@ -13,14 +13,14 @@
## Task
MuJoCo Playground `Go1JoystickFlatTerrain` (Unitree Go1) on the same frozen instance as `mujoco_playground_go1_joystick_flat_terrain-privileged`: Make the Unitree Go1 follow the joystick command schedule on flat ground for the whole 20 s episode without falling over.
-For 20 s the robot gets a joystick command schedule (4 segments: planar velocity and yaw rate in its own frame) and must track it without falling over. Standard mode (eai-standard/2.2): the agent drives the robot through a client in one episode, with only
-the robot's own sensors and four cameras that follow it.
+For 20 s the robot gets a joystick command schedule (4 segments: planar velocity and yaw rate in its own frame) and must track it without falling over. Standard mode (eai-standard/2.2): the agent drives the robot through a client, with 10 resets (only
+the last episode is graded), the robot's own sensors and the cameras that follow it.
## Difficulty
-Same instance and success check as `mujoco_playground_go1_joystick_flat_terrain-privileged`. In standard mode the agent loses resets, the simulator and
+Same instance and success check as `mujoco_playground_go1_joystick_flat_terrain-privileged`. In standard mode the agent loses free resets, the simulator and
the world state: it sees camera images and proprioception (joints, IMU, foot contacts), acts in joint position chunks,
-and gets one episode, so it has to close the loop on the robot's own sensors and recover from its own mistakes without
-starting over.
+and gets 10 resets (`RCB_RESET_BUDGET`, the compose file's default; PROTOCOL §4.5), so it has to close the loop
+on the robot's own sensors; only the last episode is graded.
## Reference solution
`solution/oracle_replay.py` streams the privileged reference's rows (`solution/oracle.npy`, 1000 rows, float32-exact)
@@ -60,3 +60,4 @@ Generated by `scripts/mujoco-playground/generate.py`, as its privileged twin.
- 2026-10-10 (review 0067): a `final.json` that is missing, unreadable or was never finalized by the collect hook is graded `infra_error = 1` (no verdict on the agent); before, a missing one was a plain 0 and an unfinalized one was graded as it stood. The collect hook writes `final.json` atomically. A request field of the wrong type (`dt` or `seconds` null, `cameras` not a list of names, a `note` that is not a string) is answered 400 before any physics instead of ending the episode with a 500. `meta["time_left_s"]` counts from the agent's first call, not from the service's start.
- 2026-10-10 (review round 2, board 0067): the client only (`environment/rcb_client.py`). `Robot()` waits up to 3600 s for the service (was 600 s) and prints its progress to stderr once a minute, so the first `Robot()` outlasts the simulator's start (no effect on this benchmark, whose service is up within a minute); `cameras="name"` (a string) is one camera, no longer its characters. The service, the collect hook, the grader, the scene and the pass rule are unchanged.
- 2026-10-10 (review round 3, board 0067): the shared service layer and the client only (`environment/sim/rcb_limited.py`, `environment/rcb_client.py`). A camera named more than once in `cameras` is rendered and returned once (N copies of one name made the service build N x N images within the frame budget, which could exhaust its memory). A service that has given up on the episode (it could not start it, or could not resume it after its crashes) answers every request with 409 and the reason, and `Robot()` / `wait_ready()` raise `EpisodeOver` at once; before, nothing answered and they waited out their timeout (up to an hour). The adapter, the collect hook, the grader, the scene and the pass rule are unchanged.
+- 2026-10-14 (owner decision C1, board 0113): 10 resets (PROTOCOL §4.5). The service (`environment/sim/server.py`) gained the optional reset capability: `RCB_RESET_BUDGET` (the compose file's default 10; 0 restores the old single episode), `POST /reset` abandons the running or ended episode and starts a new one from the same frozen start (no physics), refused (409) after a success, once the agent's time is up and once the budget is spent; only the last episode is graded (`trajectory.npz` holds its rows, replayed twice from the frozen start as before); `/status` and every `meta` add `episode`, `resets_left` and `control_steps_total`; `final.json` and `reward.json` add `resets`, `reset_budget` and `episodes`. A crash resumes the current episode (the journal is replayed with the committed resets). The instruction's rules, the client example (`POST /reset` through `Robot._post`) and this README say so. The scene, the instance, the success check and the grader's pass rule are unchanged; results recorded before this change ran with one episode and are reported separately.
diff --git a/tasks/mujoco_playground_go1_joystick_flat_terrain-standard/environment/docker-compose.yaml b/tasks/mujoco_playground_go1_joystick_flat_terrain-standard/environment/docker-compose.yaml
index bc0c363241..1226b8c3a6 100644
--- a/tasks/mujoco_playground_go1_joystick_flat_terrain-standard/environment/docker-compose.yaml
+++ b/tasks/mujoco_playground_go1_joystick_flat_terrain-standard/environment/docker-compose.yaml
@@ -28,6 +28,7 @@ services:
RCB_STARTUP_TIMEOUT_S: "300" # the scene builds in about a second
PYTHONPATH: /srv
RCB_SESSION_LOG: "" # the sidecar's own Session must not write the agent-side interaction log
+ RCB_RESET_BUDGET: "${RCB_RESET_BUDGET:-10}" # Playground quadrupeds standard: 10 resets (PROTOCOL §4.5); 0: one episode
# >>> rcb_monitor (resource limits monitor, docs/PROTOCOL.md §2.10; generated by scripts/rcb_monitor/inject.py)
rcb_monitor:
image: ghcr.io/mll-lab-nu/eai-monitor:0.1.0@sha256:f2c910f53f79c2bfa4a71298b10c78383cae3afc40720ca562cc845b8d20fb0c
diff --git a/tasks/mujoco_playground_go1_joystick_flat_terrain-standard/environment/sim/server.py b/tasks/mujoco_playground_go1_joystick_flat_terrain-standard/environment/sim/server.py
index 108eca3cae..cad8920259 100644
--- a/tasks/mujoco_playground_go1_joystick_flat_terrain-standard/environment/sim/server.py
+++ b/tasks/mujoco_playground_go1_joystick_flat_terrain-standard/environment/sim/server.py
@@ -1,16 +1,20 @@
"""eai-standard/2.2 sim service for MuJoCo Playground's quadruped tasks (docs/STANDARD_MODE.md): the adapter of the
shared layer rcb_limited.py; the simulator is rcb_quadruped.Session (plain MuJoCo on the CPU).
-HTTP surface: GET /spec, GET /status, POST /observe, POST /submit, POST /kinematics. Every call is logged to
-/data/episode.jsonl.
+HTTP surface: GET /spec, GET /status, POST /observe, POST /submit, POST /kinematics, and POST /reset when the run grants
+resets (RCB_RESET_BUDGET, the optional reset capability of docs/STANDARD_MODE.md / PROTOCOL §4.5; the quadrupeds'
+standard tasks default to 10 in their compose file, owner 2026-10-14). Every call is logged to /data/episode.jsonl.
The episode is the frozen instance's: one control step = the instance's control period (Go1, Spot: 1/50 s; Barkour:
1/100 s). A joystick episode ends after the whole command schedule
(the horizon), when the robot meets Playground's termination rule (it falls over: a failure), when the agent submits
`done`, or when its time is over; success = the task's check on the steps executed. A getup episode ends the moment the
-check holds (success), after the horizon, on `done` or on timeout.
+check holds (success), after the horizon, on `done` or on timeout. With a reset budget, POST /reset abandons the running
+(or ended) episode and starts a new one from the frozen instance; only the last episode is graded, and a success ends
+the trial.
Crash recovery (rcb_limited.Episode, resume class `exact`: MuJoCo on the CPU repeats bit for bit in a fresh process):
-every executed row is journaled; after a crash the service rebuilds the episode and replays the journal. When the
+every executed row is journaled; after a crash the service rebuilds the episode and replays the journal, resetting where
+the committed resets did (their journal rows are checkpointed, `persist`). When the
episode ends the service writes /data/trajectory.npz + /data/episode.json; the collect hook (finalize.sh) makes it exit
and replays the trajectory in two fresh processes."""
@@ -31,18 +35,21 @@ INSTANCE = os.environ.get("RCB_INSTANCE", os.path.join(os.path.dirname(os.path.a
WALL_CLOCK_S = float(os.environ.get("RCB_WALL_CLOCK_S", "3600"))
MAX_CHUNK_S = float(os.environ.get("RCB_MAX_CHUNK_S", "10"))
MAX_WATCH_S = float(os.environ.get("RCB_MAX_WATCH_S", "30"))
+RESET_BUDGET = max(0, int(os.environ.get("RCB_RESET_BUDGET") or 0)) # the compose file's default: 10 (PROTOCOL §4.5)
CAM_W, CAM_H = 320, 240
LOCK = threading.Lock()
PID_FILE = os.path.join(DATA, "server.pid")
E = {"ready": False, "step": 0, "calls": 0, "done": False, "reason": None, "success": False, "t0": time.time(), "error": None,
- "t_first": None} # t_first: the agent's first call (its clock starts there; kept through a crash by the checkpoints)
+ "t_first": None, # t_first: the agent's first call (its clock starts there; kept through a crash by the checkpoints)
+ "resets": 0, # resets done (the episode is resets + 1)
+ "reset_at": []} # the journal row at which each reset happened: a resume resets there again
os.environ["RCB_SESSION_LOG"] = ""
os.environ["RCB_AGENT_TRACE_DIR"] = ""
import rcb_quadruped as Q # noqa: E402 (first: it picks the rendering backend before mujoco loads)
import mujoco # noqa: E402
-EP = L.Episode(DATA, dim=Q.N, resume="exact", persist=("t_first",))
+EP = L.Episode(DATA, dim=Q.N, resume="exact", persist=("t_first", "resets", "reset_at"))
EP.startup() # start count, back-off after failed starts, the startup watchdog until scene_built()
S = Q.Session(INSTANCE)
ROWS: list = [] # every executed row (a resume replays them first): trajectory.npz
@@ -153,13 +160,15 @@ def schedule_text() -> list:
"wz_radps": round(c["wz"], 4)} for c in S.inst["commands"]]
+ONE = "one episode, no reset" if not RESET_BUDGET else (
+ f"an episode (with up to {RESET_BUDGET} resets: POST /reset starts a new one; only the last is graded)")
TASK_EPISODE = {
- "joystick": f"one episode, no reset: it ends after {S.horizon} control steps (the {S.horizon * DT:g} s command "
+ "joystick": f"{ONE}: it ends after {S.horizon} control steps (the {S.horizon * DT:g} s command "
f"schedule), when the robot meets its termination rule ({S.spec.get('term_text')}: a failure), when you "
"submit `done`, or when your time is over",
- "getup": f"one episode, no reset: it ends the moment the robot has stood upright for {Q.HOLD_STEPS} control steps in a "
+ "getup": f"{ONE}: it ends the moment the robot has stood upright for {Q.HOLD_STEPS} control steps in a "
f"row (success), after {S.horizon} control steps, when you submit `done`, or when your time is over",
- "stand_on_two": f"one episode, no reset: it ends the moment the pose has been held for {Q.POSE_HOLD_STEPS} control steps "
+ "stand_on_two": f"{ONE}: it ends the moment the pose has been held for {Q.POSE_HOLD_STEPS} control steps "
"in a row (success), when the robot meets the termination rule (a failure: its trunk's up axis below "
f"-0.25, or {'a leg' if S.spec.get('hind_leg_contact') else 'a front leg'}'s thigh, calf or hip on the "
f"floor), after {S.horizon} control steps, when you "
@@ -204,7 +213,7 @@ SPEC = {
"world": {"frame": "z up; the robot starts where the episode puts it (no world coordinates are given)", "up": "+z",
"units": {"length": "m", "angle": "rad", "time": "s"}, "quaternion_order": "wxyz",
"camera_axes": "intrinsic_cv / extrinsic_cv: OpenCV camera axes (x right, y down, z forward); cam2world_gl: OpenGL axes"},
- "budget": {"reset": 0, "steps": None, "wall_clock_s": WALL_CLOCK_S}, # + "clock" (rcb_limited.CLOCK_NOTE): what time_left_s counts from
+ "budget": {"reset": RESET_BUDGET, "steps": None, "wall_clock_s": WALL_CLOCK_S}, # + "clock" (rcb_limited.CLOCK_NOTE): what time_left_s counts from
"max_control_steps": S.horizon,
"episode": TASK_EPISODE,
"command_schedule": schedule_text(),
@@ -223,6 +232,15 @@ SPEC = {
"depth": {"modes": list(L.DEPTH_MODES), "default": "none", "kind": "z-depth", "unit": "m", "invalid": 0,
"max_m": L.DEPTH_MAX_M, "frames_encoding": "uint16 mm, zlib", "noise": L.DEPTH_NOISE},
}
+if RESET_BUDGET:
+ SPEC["reset"] = (
+ f"POST /reset (JSON or RCB1 {{cameras, video_fps, depth}}; no physics) abandons the running episode, or follows "
+ f"one that has ended, and starts a new one from the same frozen starting state, {RESET_BUDGET} times in all; it "
+ "returns the new episode's first frames and proprioception. Only the last episode is graded: an episode you "
+ "reset away is never graded, a success ends the trial (409 after it), and when your time is up the episode then "
+ "running is the one graded. The wall clock never restarts. /status and every meta add `episode`, `resets_left` "
+ "and `control_steps_total` (all episodes); `control_steps` counts the current episode's."
+ )
L.check_spec(SPEC) # a spec that misses a 2.2 field or names a forbidden one never starts the episode
EP.scene_built()
@@ -277,6 +295,7 @@ def summary() -> dict:
"reason": E["reason"], "done": E["done"], "env": S.env, "protocol": L.PROTOCOL, "n_actions": len(ROWS),
"fell_at": chk.get("fell_at"), "worst_lin_err": chk.get("worst_lin_err"), "worst_yaw_err": chk.get("worst_yaw_err"),
"tilt_deg": chk.get("tilt_deg"), "imu_height": chk.get("imu_height"), "pose_angle_deg": chk.get("pose_angle_deg"),
+ "episodes": E["resets"] + 1, "resets": E["resets"], "reset_budget": RESET_BUDGET, "control_steps_total": EP.rows,
**EP.final_fields()}
@@ -294,7 +313,8 @@ def finish(reason: str) -> None:
E["success"] = bool(chk["success"])
except Exception: # noqa: BLE001
traceback.print_exc()
- EP.log({"ev": "end", "reason": reason, "step": E["step"], "control_steps": ctrl_steps(), "success": bool(E["success"])})
+ EP.log({"ev": "end", "reason": reason, "step": E["step"], "control_steps": ctrl_steps(), "success": bool(E["success"]),
+ "episode": E["resets"] + 1, "journal_rows": EP.rows})
dump()
@@ -388,7 +408,7 @@ def meta_common(cams) -> dict:
"sim_time_s": round(n * DT, 3), "time_s": round(_agent_time(), 2),
"time_left_s": max(0.0, round(WALL_CLOCK_S - _agent_time(), 1)),
"proprio": {"layout": PROPRIO_LAYOUT, "now": proprio_row().tolist()}, "done": E["done"], "reason": E["reason"],
- "cameras": {c: camera_info(c) for c in cams}}
+ "cameras": {c: camera_info(c) for c in cams}, **reset_counters()}
if S.kind == "joystick":
out["command_now"] = dict(zip(("vx_mps", "vy_mps", "wz_radps"), S.command_now().round(4).tolist()))
return out
@@ -478,10 +498,54 @@ def do_observe(meta: dict):
return out, rec.parts(video if seconds > 0 else ("frames" if video == "mp4" else video), fps)
+def reset_counters() -> dict:
+ if not RESET_BUDGET:
+ return {}
+ left = 0 if (E["success"] or E.get("exiting")) else RESET_BUDGET - E["resets"]
+ return {"episode": E["resets"] + 1, "resets_left": left, "control_steps_total": EP.rows}
+
+
def status() -> dict:
_touch()
return {"step": E["step"], "calls": E["calls"], "control_steps": ctrl_steps(), "done": E["done"], "reason": E["reason"],
- "elapsed_s": round(time.time() - E["t0"], 1)}
+ "elapsed_s": round(time.time() - E["t0"], 1), **reset_counters()}
+
+
+class Refused(Exception):
+ """A /reset the service turns down (409): nothing changes."""
+
+ def __init__(self, body: dict):
+ super().__init__(body.get("error"))
+ self.body = body
+
+
+def do_reset(meta: dict):
+ """POST /reset: a new episode from the frozen instance (the running one, if any, is abandoned, never graded)."""
+ meta = _request(meta)
+ cams, fps = cameras_of(meta), video_fps(meta)
+ L.depth_mode(meta) # a bad depth value is a 400 before anything changes
+ if E["success"]:
+ raise Refused({"error": "trial finished: an episode succeeded", "reason": E["reason"], "resets_left": 0})
+ if E["resets"] >= RESET_BUDGET:
+ raise Refused({"error": "reset budget exhausted", "reason": "reset budget exhausted", "budget": RESET_BUDGET,
+ "resets_left": 0})
+ abandoned = {"episode": E["resets"] + 1, "control_steps": ctrl_steps(), "done": E["done"], "reason": E["reason"]}
+ if not E["done"]: # the abandoned episode's end, for the log (never graded: its rows leave the trajectory)
+ EP.log({"ev": "end", "reason": "reset", "step": E["step"], "control_steps": ctrl_steps(), "success": False,
+ "episode": E["resets"] + 1, "journal_rows": EP.rows})
+ S.reset()
+ ROWS.clear()
+ E["resets"] += 1
+ E["reset_at"] = list(E["reset_at"]) + [EP.rows]
+ E.update(done=False, reason=None, success=False, t_end=None, episode_control_steps=None, dumped=False, check=None)
+ EP._ck = None # the state changed without a journal row: the checkpoint after this call must hash it anew
+ E["calls"] += 1
+ EP.log({"ev": "reset", "call": E["calls"], "episode": E["resets"] + 1, "journal_rows": EP.rows,
+ "abandoned": abandoned, "state_sha": S.state_sha()})
+ rec = recorder(cams, fps, meta)
+ rec.tick(force=True, last=True)
+ out = {**meta_common(cams), "reset_id": E["resets"], "ok": True}
+ return out, rec.parts("frames", fps)
def end_from_signal() -> None:
@@ -498,20 +562,106 @@ def end_from_signal() -> None:
os._exit(0)
+# A resume replays the journal from the frozen instance and resets where the committed resets did: their journal rows
+# are in the last checkpoint (`persist`), the one Episode.begin resumes from.
+_CPS = [e for e in EP.events() if e.get("ev") == "checkpoint"]
+RESUME = {"at": list((_CPS[-1].get("persist") or {}).get("reset_at") or []) if _CPS else [], "i": 0}
+
+
+def _resume_resets() -> None:
+ while RESUME["at"] and int(RESUME["at"][0]) <= RESUME["i"]:
+ RESUME["at"].pop(0)
+ S.reset()
+ ROWS.clear()
+
+
def replay_row(row) -> None:
"""A journaled row during a resume."""
+ _resume_resets()
S.step(row)
ROWS.append(np.asarray(row, dtype=np.float64))
+ RESUME["i"] += 1
-EP.begin(state=E, step=replay_row, state_sha=S.state_sha, proprio=proprio_row, spec=SPEC, on_end=end_from_signal,
- start_event={"env": S.env, "seed": S.inst.get("seed"), "cameras": CAMS})
+def resume_sha() -> str:
+ _resume_resets() # a reset after the last replayed row (or before the first)
+ return S.state_sha()
+
+
+RESUMED = EP.begin(state=E, step=replay_row, state_sha=resume_sha, proprio=proprio_row, spec=SPEC, on_end=end_from_signal,
+ start_event={"env": S.env, "seed": S.inst.get("seed"), "cameras": CAMS})
+if RESUMED and E["resets"]: # Episode.begin looks for an "end" of any episode: only the current one's counts
+ _ends = [e for e in EP.events() if e.get("ev") == "end" and int(e.get("episode") or 1) == E["resets"] + 1
+ and e.get("reason") != "reset" and int(e.get("journal_rows") or 0) <= EP.rows]
+ if _ends:
+ E.update(done=True, reason=_ends[-1].get("reason"), success=bool(_ends[-1].get("success")),
+ t_end=float(_ends[-1]["ts"]), episode_control_steps=int(_ends[-1].get("control_steps") or ctrl_steps()))
+ else:
+ E.update(done=False, reason=None, success=False, t_end=None, episode_control_steps=None)
if E["done"]:
dump() # resumed after the episode had ended: the record is rewritten from the replay
E["ready"] = True
write_final()
open(PID_FILE, "w").write(str(os.getpid()))
print(f"{L.PROTOCOL} MuJoCo Playground quadruped sim ready ({S.env}; startup {EP.n_startups}, resumes {EP.resumes})", flush=True)
-L.serve(L.Service(spec=SPEC, lock=LOCK, state=E, submit=do_submit, observe=do_observe, kinematics=do_kinematics, status=status,
- on_submit_error=lambda exc: finish("error"), log=EP.log, episode=EP, state_sha=S.state_sha,
- proprio=proprio_row), port=int(os.environ.get("RCB_PORT", "8000")))
+
+
+def make_handler(svc: L.Service):
+ """The shared 2.2 handler, plus POST /reset when the run grants resets (the shared layer has no reset)."""
+ base = L.make_handler(svc)
+
+ class H(base):
+ def do_POST(self):
+ if self.path != "/reset" or not RESET_BUDGET:
+ return super().do_POST()
+ try:
+ meta, _ = self._body()
+ except Exception as exc: # noqa: BLE001
+ return self._json(413 if "too large" in str(exc) else 400, {"error": str(exc)[:200]})
+ with LOCK:
+ if E.get("exiting"):
+ return self._json(409, {"error": "trial finished: your time is up", "reason": E.get("reason"),
+ "resets_left": 0})
+ cid = EP.call_begin(self.path, self.connection)
+ out = blobs = err = None
+ try:
+ out, blobs = do_reset(meta)
+ except Refused as exc:
+ err = (409, exc.body)
+ except ValueError as exc:
+ err = (400, {"error": str(exc)})
+ except Exception as exc: # noqa: BLE001
+ traceback.print_exc()
+ E["error"] = repr(exc)
+ EP.log({"ev": "error", "path": self.path, "error": repr(exc)})
+ try:
+ finish("error")
+ except Exception: # noqa: BLE001
+ traceback.print_exc()
+ err = (500, {"error": "simulator error", "detail": repr(exc)[:200]})
+ finally:
+ if out is not None or (err and err[0] == 500):
+ try:
+ EP.checkpoint(E, S.state_sha, proprio_row)
+ except Exception: # noqa: BLE001
+ traceback.print_exc()
+ EP.call_end(cid, self.path)
+ if err is not None:
+ return self._json(*err)
+ return self._rcb1({**out, "stopped": None}, blobs)
+
+ return H
+
+
+def serve(svc: L.Service, port: int) -> None:
+ """L.serve with this service's handler (the same spec defaults and check, and L.bind: a port still held when the
+ service restarts is waited for, not a crash)."""
+ svc.spec.setdefault("on_service_error", L.ON_SERVICE_ERROR)
+ svc.spec.setdefault("max_call_wall_s", EP.max_call_wall_s)
+ L.check_spec(svc.spec)
+ L.bind(port, make_handler(svc)).serve_forever()
+
+
+serve(L.Service(spec=SPEC, lock=LOCK, state=E, submit=do_submit, observe=do_observe, kinematics=do_kinematics, status=status,
+ on_submit_error=lambda exc: finish("error"), log=EP.log, episode=EP, state_sha=S.state_sha,
+ proprio=proprio_row), port=int(os.environ.get("RCB_PORT", "8000")))
diff --git a/tasks/mujoco_playground_go1_joystick_flat_terrain-standard/instruction.md b/tasks/mujoco_playground_go1_joystick_flat_terrain-standard/instruction.md
index 0d406d88af..b208603642 100644
--- a/tasks/mujoco_playground_go1_joystick_flat_terrain-standard/instruction.md
+++ b/tasks/mujoco_playground_go1_joystick_flat_terrain-standard/instruction.md
@@ -2,9 +2,9 @@
**Task:** Make the Unitree Go1 follow the joystick command schedule on flat ground for the whole 20 s episode without falling over.
-You control the Unitree Go1 quadruped — **as if it were real hardware** — in one episode of MuJoCo Playground's
+You control the Unitree Go1 quadruped — **as if it were real hardware** — in MuJoCo Playground's
`Go1JoystickFlatTerrain` environment (plain MuJoCo 3.3.7 with Playground's scene, commit 4057c14). The simulator runs in a service you
-cannot open. You get one client and one episode.
+cannot open. You get one client and up to 11 episodes.
## The episode
The robot starts at Playground's `home` pose, its base moved to (-0.493, -0.479) m and turned to a yaw of -142.4 deg, with a random base velocity (Playground's reset). The ground is flat.
@@ -20,7 +20,15 @@ For 20 s (1000 control steps) it is given this command schedule, drawn by Playgr
The same schedule is `spec["command_schedule"]`; every response's `meta["command_now"]` is the current command.
## Rules (real-robot rules)
-- **No reset.** One episode. You see four cameras that follow the robot (RGB, depth if you ask) and the robot's own
+- **Episodes and resets.** You have **10 resets** (`spec["budget"]["reset"]`, which is authoritative: if it
+ says 0, there is no reset and the one episode is the run). A reset (`POST /reset`, below) puts the robot back on the
+ same frozen starting state and starts a new episode, either abandoning the running one or after one has ended; it
+ takes no simulated time and returns the start's images and proprioception. Once an episode has ended, `submit`
+ raises `EpisodeOver` until you reset. **Only the last episode is graded**: an episode you reset away is never graded,
+ a success ends the run at once (no reset after it), and when your time is up the episode then running is the one
+ graded. The wall clock is the run's and never restarts; `meta["episode"]` and `meta["resets_left"]` count the
+ episodes (`meta["control_steps"]` is the current episode's).
+- **No privileged state.** You see the cameras that follow the robot (RGB, depth if you ask) and the robot's own
sensors: joint positions, velocities and torques, the IMU (gyro, accelerometer, gravity's direction, the trunk's
velocity), foot contacts. No world positions, no success check.
- **The episode ends** after 1000 control steps, the whole schedule (`spec["max_control_steps"]`, counted by `meta["control_steps"]`), when the robot meets its termination rule (the trunk's up axis pointing below the horizon (the z component of the IMU frame's z axis negative): a failure), when you submit `done`, or when your time is up (before the end of the schedule, both are failures).
@@ -29,8 +37,8 @@ The same schedule is `spec["command_schedule"]`; every response's `meta["command
`get_obs(seconds=...)`, which holds the current targets). `meta["time_left_s"]` is the service's count of the limit,
from your first request to the service: it never runs out before your time does, so plan by it.
- Nothing to hand in: the service records the episode and grades it. Leave your script at `/app/output/gen.py`.
-- **If the service crashes**, a call fails with a connection error: it restarts and resumes the SAME episode (never a
- reset) in the state after the last call it completed. `r.wait_ready()` waits for it; then check `r.status()`.
+- **If the service crashes**, a call fails with a connection error: it restarts and resumes the SAME episode (a crash
+ is never a reset and costs none) in the state after the last call it completed. `r.wait_ready()` waits for it; then check `r.status()`.
- Nothing is put into your context automatically: images come back as arrays in Python objects.
## Client (`/opt/tools/rcb_client.py`)
@@ -43,7 +51,8 @@ res = r.submit("joint_position", A, dt=0.02) # A: (T, 12) rows dt apart; res.p
obs = r.get_obs(seconds=1.0, video="mp4") # hold the current targets and watch for 1 s (max 30)
r.fk(q); r.ik({"FR": [x, y, z, 1, 0, 0, 0]}, seed=q) # feet in the trunk frame <-> joints (no physics)
r.submit("done", []) # end the episode now
-r.status() # calls, control steps, done, reason, elapsed time
+r.status() # calls, control steps, done, reason, elapsed time, episode, resets_left
+obs = r._post("/reset", {"cameras": ["track"]}, [], 900) # POST /reset: a new episode (the 2.2 client has no reset())
```
Cameras: `track`, `side`, `back`, `top` (they follow the robot); `spec["cameras"]` gives each one's mounting and intrinsics, and every response carries its pose. Proprioception (`spec["proprio_groups"]`): `joint_pos`, `joint_vel`, `joint_effort` (12 each, action
order), `imu_gyro`, `imu_accelerometer`, `imu_gravity`, `imu_linvel` (3 each, IMU frame: x forward, y left, z up),
diff --git a/tasks/mujoco_playground_go1_joystick_flat_terrain-standard/task.toml b/tasks/mujoco_playground_go1_joystick_flat_terrain-standard/task.toml
index dbe35965b7..fb5022e344 100644
--- a/tasks/mujoco_playground_go1_joystick_flat_terrain-standard/task.toml
+++ b/tasks/mujoco_playground_go1_joystick_flat_terrain-standard/task.toml
@@ -16,7 +16,7 @@ artifacts = [{ source = "/var/rcb-monitor/", service = "rcb_monitor" },
[task]
name = "robot-coding-bench/mujoco_playground_go1_joystick_flat_terrain-standard"
-description = "STANDARD MODE (robot as a service, eai-standard/2.2): MuJoCo Playground Go1JoystickFlatTerrain (Unitree Go1) on the same frozen instance as mujoco_playground_go1_joystick_flat_terrain-privileged: Make the Unitree Go1 follow the joystick command schedule on flat ground for the whole 20 s episode without falling over. The agent only gets a client: four cameras that follow the robot (RGB, optional depth), the robot's joints and IMU, joint position chunks at 50 Hz, the feet's kinematics; one episode, no reset, no world state; the sim service records the episode, replays it in fresh processes and grades it."
+description = "STANDARD MODE (robot as a service, eai-standard/2.2): MuJoCo Playground Go1JoystickFlatTerrain (Unitree Go1) on the same frozen instance as mujoco_playground_go1_joystick_flat_terrain-privileged: Make the Unitree Go1 follow the joystick command schedule on flat ground for the whole 20 s episode without falling over. The agent only gets a client: four cameras that follow the robot (RGB, optional depth), the robot's joints and IMU, joint position chunks at 50 Hz, the feet's kinematics; 10 resets (only the last episode is graded), no world state; the sim service records the episode, replays it in fresh processes and grades it."
authors = [{ name = "Changhe Chen" }]
keywords = ["robotics", "mujoco", "mujoco-playground", "unitree-go1", "quadruped", "locomotion", "joystick", "code-as-policy", "trajectory", "standard-mode"]
@@ -28,6 +28,7 @@ simulator = "mujoco"
robot = "unitree-go1"
family = "mujoco_playground/Go1JoystickFlatTerrain"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "full"
instance_seed = 0
diff --git a/tasks/mujoco_playground_go1_joystick_flat_terrain-standard/tests/grade.py b/tasks/mujoco_playground_go1_joystick_flat_terrain-standard/tests/grade.py
index 41ce495421..3d4c23ad7e 100644
--- a/tasks/mujoco_playground_go1_joystick_flat_terrain-standard/tests/grade.py
+++ b/tasks/mujoco_playground_go1_joystick_flat_terrain-standard/tests/grade.py
@@ -77,6 +77,9 @@ write(
"sim_error": int(d.get("reason") == "error"),
"sim_startups": int(d.get("sim_startups") or 0),
"resumes": int(d.get("resumes") or 0),
+ "resets": int(d.get("resets") or 0), # PROTOCOL §4.5: resets the agent used; only the last episode is graded
+ "reset_budget": int(d.get("reset_budget") or 0),
+ "episodes": int(d.get("episodes") or 1),
"resume_state_mismatch": int(d.get("resume_state_mismatch") or 0),
"resume_max_joint_dev": num("resume_max_joint_dev"),
},
diff --git a/tasks/mujoco_playground_go1_joystick_rough_terrain-privileged/README.md b/tasks/mujoco_playground_go1_joystick_rough_terrain-privileged/README.md
index a0c0516650..a3526c5182 100644
--- a/tasks/mujoco_playground_go1_joystick_rough_terrain-privileged/README.md
+++ b/tasks/mujoco_playground_go1_joystick_rough_terrain-privileged/README.md
@@ -53,3 +53,11 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
### Provenance
Generated by `scripts/mujoco-playground/generate.py` from `scripts/mujoco-playground/instances/Go1JoystickRoughTerrain.json`
(`make_instances.py`, `select_instances.py`) and `scripts/mujoco-playground/oracles/Go1JoystickRoughTerrain.npy` (`oracle.py`).
+
+### Change Log
+- 2026-10-14 (review of board 0105, B1-m2): `tests/grade.py` only. A replay process that died or ran out of time, a
+ rebuilt start that does not match the instance's hash (`instance_mismatch`) and an exception of the verifier itself
+ now write `infra_error = 1` beside `grader_error = 1`: no verdict on the agent. Before, they wrote `grader_error = 1`
+ alone, which counted as the agent's 0. Every `reward.json` now carries `infra_error`. The scene, the instance, the
+ pass rule and the reference are unchanged, so the task keeps its name; a past 0 with `grader_error = 1` was an
+ infrastructure failure.
diff --git a/tasks/mujoco_playground_go1_joystick_rough_terrain-privileged/task.toml b/tasks/mujoco_playground_go1_joystick_rough_terrain-privileged/task.toml
index 87c528200e..7ded3200cf 100644
--- a/tasks/mujoco_playground_go1_joystick_rough_terrain-privileged/task.toml
+++ b/tasks/mujoco_playground_go1_joystick_rough_terrain-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "mujoco"
robot = "unitree-go1"
family = "mujoco_playground/Go1JoystickRoughTerrain"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 1 # the PRNGKey of Playground's reset (and, for the joystick tasks, its command process)
expert_time_estimate_hours = 3.0
diff --git a/tasks/mujoco_playground_go1_joystick_rough_terrain-privileged/tests/grade.py b/tasks/mujoco_playground_go1_joystick_rough_terrain-privileged/tests/grade.py
index c75c326868..1667d090c8 100644
--- a/tasks/mujoco_playground_go1_joystick_rough_terrain-privileged/tests/grade.py
+++ b/tasks/mujoco_playground_go1_joystick_rough_terrain-privileged/tests/grade.py
@@ -6,6 +6,11 @@ Two replays, each in its own fresh process (`rcb_quadruped.py replay`: the froze
task's success check on the whole trajectory). success = both replays succeed and end in the same state. Writes
/logs/verifier/reward.txt (1/0) and reward.json (numbers only); then a third replay renders replay.mp4, after the
verdict is on disk (it never changes it).
+
+Failures that are not the agent's (docs/PROTOCOL.md: infra vs agent failure) carry `grader_error` = 1 and `infra_error`
+= 1, so they are no verdict on the agent: a replay process that died or timed out, an image that no longer rebuilds the
+frozen start (`instance_mismatch`), and any exception of the verifier itself. A file the agent got wrong is the agent's
+0 (`bad_trajectory`, infra_error 0). Every reward.json carries success, deterministic, grader_error and infra_error.
"""
import argparse
import json
@@ -22,8 +27,12 @@ TOOL = os.path.join(os.path.dirname(os.path.abspath(__file__)), "rcb_quadruped.p
REPLAY_TIMEOUT_S = 300 # a 1000-step replay takes a few seconds; the video replay renders 4 views every 2nd step
+BASE = {"success": 0, "deterministic": 0, "grader_error": 0, "infra_error": 0}
+
+
def write(ok: bool, out: dict) -> None:
os.makedirs(LOG_DIR, exist_ok=True)
+ out = {**BASE, **out}
open(os.path.join(LOG_DIR, "reward.txt"), "w").write("1\n" if ok else "0\n")
json.dump(out, open(os.path.join(LOG_DIR, "reward.json"), "w"))
print(json.dumps(out), flush=True)
@@ -96,11 +105,11 @@ def main() -> int:
"n_actions": int(acts.shape[0]) if acts.ndim == 2 else 0, "grader_error": 0, **cost})
return 0
runs = [replay(a.instance, a.traj, None), replay(a.instance, a.traj, None)]
- if any("error" in r and not r.get("bad_trajectory") for r in runs):
- write(False, {"success": 0, "deterministic": 0, "grader_error": 1, **cost})
+ if any("error" in r and not r.get("bad_trajectory") for r in runs): # the replay process died / timed out
+ write(False, {"success": 0, "deterministic": 0, "grader_error": 1, "infra_error": 1, "replay_failed": 1, **cost})
return 0
if not all(r.get("instance_ok") for r in runs): # the image no longer builds the frozen start: not the agent's 0
- write(False, {"success": 0, "deterministic": 0, "grader_error": 1, "instance_mismatch": 1, **cost})
+ write(False, {"success": 0, "deterministic": 0, "grader_error": 1, "infra_error": 1, "instance_mismatch": 1, **cost})
return 0
all_success = all(bool(r.get("success")) for r in runs)
deterministic = runs[0].get("state_sha") == runs[1].get("state_sha")
@@ -116,7 +125,7 @@ def main() -> int:
write(ok, out) # the verdict, before any video
except Exception: # noqa: BLE001
traceback.print_exc(file=sys.stderr)
- write(False, {"success": 0, "deterministic": 0, "grader_error": 1, **cost})
+ write(False, {"success": 0, "deterministic": 0, "grader_error": 1, "infra_error": 1, **cost})
return 0
try: # the video, for humans: it never changes the verdict
if replay(a.instance, a.traj, os.path.join(LOG_DIR, "replay.mp4")).get("video"):
diff --git a/tasks/mujoco_playground_go1_joystick_rough_terrain-standard/README.md b/tasks/mujoco_playground_go1_joystick_rough_terrain-standard/README.md
index d0129c2aa0..6bbd6c7b94 100644
--- a/tasks/mujoco_playground_go1_joystick_rough_terrain-standard/README.md
+++ b/tasks/mujoco_playground_go1_joystick_rough_terrain-standard/README.md
@@ -13,14 +13,14 @@
## Task
MuJoCo Playground `Go1JoystickRoughTerrain` (Unitree Go1) on the same frozen instance as `mujoco_playground_go1_joystick_rough_terrain-privileged`: Make the Unitree Go1 follow the joystick command schedule on rough terrain (a 20 x 20 m heightfield, up to 5 cm high) for the whole 20 s episode without falling over.
-For 20 s the robot gets a joystick command schedule (7 segments: planar velocity and yaw rate in its own frame) and must track it without falling over. The ground is a 20 x 20 m heightfield with bumps up to 5 cm. Standard mode (eai-standard/2.2): the agent drives the robot through a client in one episode, with only
-the robot's own sensors and four cameras that follow it.
+For 20 s the robot gets a joystick command schedule (7 segments: planar velocity and yaw rate in its own frame) and must track it without falling over. The ground is a 20 x 20 m heightfield with bumps up to 5 cm. Standard mode (eai-standard/2.2): the agent drives the robot through a client, with 10 resets (only
+the last episode is graded), the robot's own sensors and the cameras that follow it.
## Difficulty
-Same instance and success check as `mujoco_playground_go1_joystick_rough_terrain-privileged`. In standard mode the agent loses resets, the simulator and
+Same instance and success check as `mujoco_playground_go1_joystick_rough_terrain-privileged`. In standard mode the agent loses free resets, the simulator and
the world state: it sees camera images and proprioception (joints, IMU, foot contacts), acts in joint position chunks,
-and gets one episode, so it has to close the loop on the robot's own sensors and recover from its own mistakes without
-starting over.
+and gets 10 resets (`RCB_RESET_BUDGET`, the compose file's default; PROTOCOL §4.5), so it has to close the loop
+on the robot's own sensors; only the last episode is graded.
## Reference solution
`solution/oracle_replay.py` streams the privileged reference's rows (`solution/oracle.npy`, 1000 rows, float32-exact)
@@ -60,3 +60,4 @@ Generated by `scripts/mujoco-playground/generate.py`, as its privileged twin.
- 2026-10-10 (review 0067): a `final.json` that is missing, unreadable or was never finalized by the collect hook is graded `infra_error = 1` (no verdict on the agent); before, a missing one was a plain 0 and an unfinalized one was graded as it stood. The collect hook writes `final.json` atomically. A request field of the wrong type (`dt` or `seconds` null, `cameras` not a list of names, a `note` that is not a string) is answered 400 before any physics instead of ending the episode with a 500. `meta["time_left_s"]` counts from the agent's first call, not from the service's start.
- 2026-10-10 (review round 2, board 0067): the client only (`environment/rcb_client.py`). `Robot()` waits up to 3600 s for the service (was 600 s) and prints its progress to stderr once a minute, so the first `Robot()` outlasts the simulator's start (no effect on this benchmark, whose service is up within a minute); `cameras="name"` (a string) is one camera, no longer its characters. The service, the collect hook, the grader, the scene and the pass rule are unchanged.
- 2026-10-10 (review round 3, board 0067): the shared service layer and the client only (`environment/sim/rcb_limited.py`, `environment/rcb_client.py`). A camera named more than once in `cameras` is rendered and returned once (N copies of one name made the service build N x N images within the frame budget, which could exhaust its memory). A service that has given up on the episode (it could not start it, or could not resume it after its crashes) answers every request with 409 and the reason, and `Robot()` / `wait_ready()` raise `EpisodeOver` at once; before, nothing answered and they waited out their timeout (up to an hour). The adapter, the collect hook, the grader, the scene and the pass rule are unchanged.
+- 2026-10-14 (owner decision C1, board 0113): 10 resets (PROTOCOL §4.5). The service (`environment/sim/server.py`) gained the optional reset capability: `RCB_RESET_BUDGET` (the compose file's default 10; 0 restores the old single episode), `POST /reset` abandons the running or ended episode and starts a new one from the same frozen start (no physics), refused (409) after a success, once the agent's time is up and once the budget is spent; only the last episode is graded (`trajectory.npz` holds its rows, replayed twice from the frozen start as before); `/status` and every `meta` add `episode`, `resets_left` and `control_steps_total`; `final.json` and `reward.json` add `resets`, `reset_budget` and `episodes`. A crash resumes the current episode (the journal is replayed with the committed resets). The instruction's rules, the client example (`POST /reset` through `Robot._post`) and this README say so. The scene, the instance, the success check and the grader's pass rule are unchanged; results recorded before this change ran with one episode and are reported separately.
diff --git a/tasks/mujoco_playground_go1_joystick_rough_terrain-standard/environment/docker-compose.yaml b/tasks/mujoco_playground_go1_joystick_rough_terrain-standard/environment/docker-compose.yaml
index bc0c363241..1226b8c3a6 100644
--- a/tasks/mujoco_playground_go1_joystick_rough_terrain-standard/environment/docker-compose.yaml
+++ b/tasks/mujoco_playground_go1_joystick_rough_terrain-standard/environment/docker-compose.yaml
@@ -28,6 +28,7 @@ services:
RCB_STARTUP_TIMEOUT_S: "300" # the scene builds in about a second
PYTHONPATH: /srv
RCB_SESSION_LOG: "" # the sidecar's own Session must not write the agent-side interaction log
+ RCB_RESET_BUDGET: "${RCB_RESET_BUDGET:-10}" # Playground quadrupeds standard: 10 resets (PROTOCOL §4.5); 0: one episode
# >>> rcb_monitor (resource limits monitor, docs/PROTOCOL.md §2.10; generated by scripts/rcb_monitor/inject.py)
rcb_monitor:
image: ghcr.io/mll-lab-nu/eai-monitor:0.1.0@sha256:f2c910f53f79c2bfa4a71298b10c78383cae3afc40720ca562cc845b8d20fb0c
diff --git a/tasks/mujoco_playground_go1_joystick_rough_terrain-standard/environment/sim/server.py b/tasks/mujoco_playground_go1_joystick_rough_terrain-standard/environment/sim/server.py
index 108eca3cae..cad8920259 100644
--- a/tasks/mujoco_playground_go1_joystick_rough_terrain-standard/environment/sim/server.py
+++ b/tasks/mujoco_playground_go1_joystick_rough_terrain-standard/environment/sim/server.py
@@ -1,16 +1,20 @@
"""eai-standard/2.2 sim service for MuJoCo Playground's quadruped tasks (docs/STANDARD_MODE.md): the adapter of the
shared layer rcb_limited.py; the simulator is rcb_quadruped.Session (plain MuJoCo on the CPU).
-HTTP surface: GET /spec, GET /status, POST /observe, POST /submit, POST /kinematics. Every call is logged to
-/data/episode.jsonl.
+HTTP surface: GET /spec, GET /status, POST /observe, POST /submit, POST /kinematics, and POST /reset when the run grants
+resets (RCB_RESET_BUDGET, the optional reset capability of docs/STANDARD_MODE.md / PROTOCOL §4.5; the quadrupeds'
+standard tasks default to 10 in their compose file, owner 2026-10-14). Every call is logged to /data/episode.jsonl.
The episode is the frozen instance's: one control step = the instance's control period (Go1, Spot: 1/50 s; Barkour:
1/100 s). A joystick episode ends after the whole command schedule
(the horizon), when the robot meets Playground's termination rule (it falls over: a failure), when the agent submits
`done`, or when its time is over; success = the task's check on the steps executed. A getup episode ends the moment the
-check holds (success), after the horizon, on `done` or on timeout.
+check holds (success), after the horizon, on `done` or on timeout. With a reset budget, POST /reset abandons the running
+(or ended) episode and starts a new one from the frozen instance; only the last episode is graded, and a success ends
+the trial.
Crash recovery (rcb_limited.Episode, resume class `exact`: MuJoCo on the CPU repeats bit for bit in a fresh process):
-every executed row is journaled; after a crash the service rebuilds the episode and replays the journal. When the
+every executed row is journaled; after a crash the service rebuilds the episode and replays the journal, resetting where
+the committed resets did (their journal rows are checkpointed, `persist`). When the
episode ends the service writes /data/trajectory.npz + /data/episode.json; the collect hook (finalize.sh) makes it exit
and replays the trajectory in two fresh processes."""
@@ -31,18 +35,21 @@ INSTANCE = os.environ.get("RCB_INSTANCE", os.path.join(os.path.dirname(os.path.a
WALL_CLOCK_S = float(os.environ.get("RCB_WALL_CLOCK_S", "3600"))
MAX_CHUNK_S = float(os.environ.get("RCB_MAX_CHUNK_S", "10"))
MAX_WATCH_S = float(os.environ.get("RCB_MAX_WATCH_S", "30"))
+RESET_BUDGET = max(0, int(os.environ.get("RCB_RESET_BUDGET") or 0)) # the compose file's default: 10 (PROTOCOL §4.5)
CAM_W, CAM_H = 320, 240
LOCK = threading.Lock()
PID_FILE = os.path.join(DATA, "server.pid")
E = {"ready": False, "step": 0, "calls": 0, "done": False, "reason": None, "success": False, "t0": time.time(), "error": None,
- "t_first": None} # t_first: the agent's first call (its clock starts there; kept through a crash by the checkpoints)
+ "t_first": None, # t_first: the agent's first call (its clock starts there; kept through a crash by the checkpoints)
+ "resets": 0, # resets done (the episode is resets + 1)
+ "reset_at": []} # the journal row at which each reset happened: a resume resets there again
os.environ["RCB_SESSION_LOG"] = ""
os.environ["RCB_AGENT_TRACE_DIR"] = ""
import rcb_quadruped as Q # noqa: E402 (first: it picks the rendering backend before mujoco loads)
import mujoco # noqa: E402
-EP = L.Episode(DATA, dim=Q.N, resume="exact", persist=("t_first",))
+EP = L.Episode(DATA, dim=Q.N, resume="exact", persist=("t_first", "resets", "reset_at"))
EP.startup() # start count, back-off after failed starts, the startup watchdog until scene_built()
S = Q.Session(INSTANCE)
ROWS: list = [] # every executed row (a resume replays them first): trajectory.npz
@@ -153,13 +160,15 @@ def schedule_text() -> list:
"wz_radps": round(c["wz"], 4)} for c in S.inst["commands"]]
+ONE = "one episode, no reset" if not RESET_BUDGET else (
+ f"an episode (with up to {RESET_BUDGET} resets: POST /reset starts a new one; only the last is graded)")
TASK_EPISODE = {
- "joystick": f"one episode, no reset: it ends after {S.horizon} control steps (the {S.horizon * DT:g} s command "
+ "joystick": f"{ONE}: it ends after {S.horizon} control steps (the {S.horizon * DT:g} s command "
f"schedule), when the robot meets its termination rule ({S.spec.get('term_text')}: a failure), when you "
"submit `done`, or when your time is over",
- "getup": f"one episode, no reset: it ends the moment the robot has stood upright for {Q.HOLD_STEPS} control steps in a "
+ "getup": f"{ONE}: it ends the moment the robot has stood upright for {Q.HOLD_STEPS} control steps in a "
f"row (success), after {S.horizon} control steps, when you submit `done`, or when your time is over",
- "stand_on_two": f"one episode, no reset: it ends the moment the pose has been held for {Q.POSE_HOLD_STEPS} control steps "
+ "stand_on_two": f"{ONE}: it ends the moment the pose has been held for {Q.POSE_HOLD_STEPS} control steps "
"in a row (success), when the robot meets the termination rule (a failure: its trunk's up axis below "
f"-0.25, or {'a leg' if S.spec.get('hind_leg_contact') else 'a front leg'}'s thigh, calf or hip on the "
f"floor), after {S.horizon} control steps, when you "
@@ -204,7 +213,7 @@ SPEC = {
"world": {"frame": "z up; the robot starts where the episode puts it (no world coordinates are given)", "up": "+z",
"units": {"length": "m", "angle": "rad", "time": "s"}, "quaternion_order": "wxyz",
"camera_axes": "intrinsic_cv / extrinsic_cv: OpenCV camera axes (x right, y down, z forward); cam2world_gl: OpenGL axes"},
- "budget": {"reset": 0, "steps": None, "wall_clock_s": WALL_CLOCK_S}, # + "clock" (rcb_limited.CLOCK_NOTE): what time_left_s counts from
+ "budget": {"reset": RESET_BUDGET, "steps": None, "wall_clock_s": WALL_CLOCK_S}, # + "clock" (rcb_limited.CLOCK_NOTE): what time_left_s counts from
"max_control_steps": S.horizon,
"episode": TASK_EPISODE,
"command_schedule": schedule_text(),
@@ -223,6 +232,15 @@ SPEC = {
"depth": {"modes": list(L.DEPTH_MODES), "default": "none", "kind": "z-depth", "unit": "m", "invalid": 0,
"max_m": L.DEPTH_MAX_M, "frames_encoding": "uint16 mm, zlib", "noise": L.DEPTH_NOISE},
}
+if RESET_BUDGET:
+ SPEC["reset"] = (
+ f"POST /reset (JSON or RCB1 {{cameras, video_fps, depth}}; no physics) abandons the running episode, or follows "
+ f"one that has ended, and starts a new one from the same frozen starting state, {RESET_BUDGET} times in all; it "
+ "returns the new episode's first frames and proprioception. Only the last episode is graded: an episode you "
+ "reset away is never graded, a success ends the trial (409 after it), and when your time is up the episode then "
+ "running is the one graded. The wall clock never restarts. /status and every meta add `episode`, `resets_left` "
+ "and `control_steps_total` (all episodes); `control_steps` counts the current episode's."
+ )
L.check_spec(SPEC) # a spec that misses a 2.2 field or names a forbidden one never starts the episode
EP.scene_built()
@@ -277,6 +295,7 @@ def summary() -> dict:
"reason": E["reason"], "done": E["done"], "env": S.env, "protocol": L.PROTOCOL, "n_actions": len(ROWS),
"fell_at": chk.get("fell_at"), "worst_lin_err": chk.get("worst_lin_err"), "worst_yaw_err": chk.get("worst_yaw_err"),
"tilt_deg": chk.get("tilt_deg"), "imu_height": chk.get("imu_height"), "pose_angle_deg": chk.get("pose_angle_deg"),
+ "episodes": E["resets"] + 1, "resets": E["resets"], "reset_budget": RESET_BUDGET, "control_steps_total": EP.rows,
**EP.final_fields()}
@@ -294,7 +313,8 @@ def finish(reason: str) -> None:
E["success"] = bool(chk["success"])
except Exception: # noqa: BLE001
traceback.print_exc()
- EP.log({"ev": "end", "reason": reason, "step": E["step"], "control_steps": ctrl_steps(), "success": bool(E["success"])})
+ EP.log({"ev": "end", "reason": reason, "step": E["step"], "control_steps": ctrl_steps(), "success": bool(E["success"]),
+ "episode": E["resets"] + 1, "journal_rows": EP.rows})
dump()
@@ -388,7 +408,7 @@ def meta_common(cams) -> dict:
"sim_time_s": round(n * DT, 3), "time_s": round(_agent_time(), 2),
"time_left_s": max(0.0, round(WALL_CLOCK_S - _agent_time(), 1)),
"proprio": {"layout": PROPRIO_LAYOUT, "now": proprio_row().tolist()}, "done": E["done"], "reason": E["reason"],
- "cameras": {c: camera_info(c) for c in cams}}
+ "cameras": {c: camera_info(c) for c in cams}, **reset_counters()}
if S.kind == "joystick":
out["command_now"] = dict(zip(("vx_mps", "vy_mps", "wz_radps"), S.command_now().round(4).tolist()))
return out
@@ -478,10 +498,54 @@ def do_observe(meta: dict):
return out, rec.parts(video if seconds > 0 else ("frames" if video == "mp4" else video), fps)
+def reset_counters() -> dict:
+ if not RESET_BUDGET:
+ return {}
+ left = 0 if (E["success"] or E.get("exiting")) else RESET_BUDGET - E["resets"]
+ return {"episode": E["resets"] + 1, "resets_left": left, "control_steps_total": EP.rows}
+
+
def status() -> dict:
_touch()
return {"step": E["step"], "calls": E["calls"], "control_steps": ctrl_steps(), "done": E["done"], "reason": E["reason"],
- "elapsed_s": round(time.time() - E["t0"], 1)}
+ "elapsed_s": round(time.time() - E["t0"], 1), **reset_counters()}
+
+
+class Refused(Exception):
+ """A /reset the service turns down (409): nothing changes."""
+
+ def __init__(self, body: dict):
+ super().__init__(body.get("error"))
+ self.body = body
+
+
+def do_reset(meta: dict):
+ """POST /reset: a new episode from the frozen instance (the running one, if any, is abandoned, never graded)."""
+ meta = _request(meta)
+ cams, fps = cameras_of(meta), video_fps(meta)
+ L.depth_mode(meta) # a bad depth value is a 400 before anything changes
+ if E["success"]:
+ raise Refused({"error": "trial finished: an episode succeeded", "reason": E["reason"], "resets_left": 0})
+ if E["resets"] >= RESET_BUDGET:
+ raise Refused({"error": "reset budget exhausted", "reason": "reset budget exhausted", "budget": RESET_BUDGET,
+ "resets_left": 0})
+ abandoned = {"episode": E["resets"] + 1, "control_steps": ctrl_steps(), "done": E["done"], "reason": E["reason"]}
+ if not E["done"]: # the abandoned episode's end, for the log (never graded: its rows leave the trajectory)
+ EP.log({"ev": "end", "reason": "reset", "step": E["step"], "control_steps": ctrl_steps(), "success": False,
+ "episode": E["resets"] + 1, "journal_rows": EP.rows})
+ S.reset()
+ ROWS.clear()
+ E["resets"] += 1
+ E["reset_at"] = list(E["reset_at"]) + [EP.rows]
+ E.update(done=False, reason=None, success=False, t_end=None, episode_control_steps=None, dumped=False, check=None)
+ EP._ck = None # the state changed without a journal row: the checkpoint after this call must hash it anew
+ E["calls"] += 1
+ EP.log({"ev": "reset", "call": E["calls"], "episode": E["resets"] + 1, "journal_rows": EP.rows,
+ "abandoned": abandoned, "state_sha": S.state_sha()})
+ rec = recorder(cams, fps, meta)
+ rec.tick(force=True, last=True)
+ out = {**meta_common(cams), "reset_id": E["resets"], "ok": True}
+ return out, rec.parts("frames", fps)
def end_from_signal() -> None:
@@ -498,20 +562,106 @@ def end_from_signal() -> None:
os._exit(0)
+# A resume replays the journal from the frozen instance and resets where the committed resets did: their journal rows
+# are in the last checkpoint (`persist`), the one Episode.begin resumes from.
+_CPS = [e for e in EP.events() if e.get("ev") == "checkpoint"]
+RESUME = {"at": list((_CPS[-1].get("persist") or {}).get("reset_at") or []) if _CPS else [], "i": 0}
+
+
+def _resume_resets() -> None:
+ while RESUME["at"] and int(RESUME["at"][0]) <= RESUME["i"]:
+ RESUME["at"].pop(0)
+ S.reset()
+ ROWS.clear()
+
+
def replay_row(row) -> None:
"""A journaled row during a resume."""
+ _resume_resets()
S.step(row)
ROWS.append(np.asarray(row, dtype=np.float64))
+ RESUME["i"] += 1
-EP.begin(state=E, step=replay_row, state_sha=S.state_sha, proprio=proprio_row, spec=SPEC, on_end=end_from_signal,
- start_event={"env": S.env, "seed": S.inst.get("seed"), "cameras": CAMS})
+def resume_sha() -> str:
+ _resume_resets() # a reset after the last replayed row (or before the first)
+ return S.state_sha()
+
+
+RESUMED = EP.begin(state=E, step=replay_row, state_sha=resume_sha, proprio=proprio_row, spec=SPEC, on_end=end_from_signal,
+ start_event={"env": S.env, "seed": S.inst.get("seed"), "cameras": CAMS})
+if RESUMED and E["resets"]: # Episode.begin looks for an "end" of any episode: only the current one's counts
+ _ends = [e for e in EP.events() if e.get("ev") == "end" and int(e.get("episode") or 1) == E["resets"] + 1
+ and e.get("reason") != "reset" and int(e.get("journal_rows") or 0) <= EP.rows]
+ if _ends:
+ E.update(done=True, reason=_ends[-1].get("reason"), success=bool(_ends[-1].get("success")),
+ t_end=float(_ends[-1]["ts"]), episode_control_steps=int(_ends[-1].get("control_steps") or ctrl_steps()))
+ else:
+ E.update(done=False, reason=None, success=False, t_end=None, episode_control_steps=None)
if E["done"]:
dump() # resumed after the episode had ended: the record is rewritten from the replay
E["ready"] = True
write_final()
open(PID_FILE, "w").write(str(os.getpid()))
print(f"{L.PROTOCOL} MuJoCo Playground quadruped sim ready ({S.env}; startup {EP.n_startups}, resumes {EP.resumes})", flush=True)
-L.serve(L.Service(spec=SPEC, lock=LOCK, state=E, submit=do_submit, observe=do_observe, kinematics=do_kinematics, status=status,
- on_submit_error=lambda exc: finish("error"), log=EP.log, episode=EP, state_sha=S.state_sha,
- proprio=proprio_row), port=int(os.environ.get("RCB_PORT", "8000")))
+
+
+def make_handler(svc: L.Service):
+ """The shared 2.2 handler, plus POST /reset when the run grants resets (the shared layer has no reset)."""
+ base = L.make_handler(svc)
+
+ class H(base):
+ def do_POST(self):
+ if self.path != "/reset" or not RESET_BUDGET:
+ return super().do_POST()
+ try:
+ meta, _ = self._body()
+ except Exception as exc: # noqa: BLE001
+ return self._json(413 if "too large" in str(exc) else 400, {"error": str(exc)[:200]})
+ with LOCK:
+ if E.get("exiting"):
+ return self._json(409, {"error": "trial finished: your time is up", "reason": E.get("reason"),
+ "resets_left": 0})
+ cid = EP.call_begin(self.path, self.connection)
+ out = blobs = err = None
+ try:
+ out, blobs = do_reset(meta)
+ except Refused as exc:
+ err = (409, exc.body)
+ except ValueError as exc:
+ err = (400, {"error": str(exc)})
+ except Exception as exc: # noqa: BLE001
+ traceback.print_exc()
+ E["error"] = repr(exc)
+ EP.log({"ev": "error", "path": self.path, "error": repr(exc)})
+ try:
+ finish("error")
+ except Exception: # noqa: BLE001
+ traceback.print_exc()
+ err = (500, {"error": "simulator error", "detail": repr(exc)[:200]})
+ finally:
+ if out is not None or (err and err[0] == 500):
+ try:
+ EP.checkpoint(E, S.state_sha, proprio_row)
+ except Exception: # noqa: BLE001
+ traceback.print_exc()
+ EP.call_end(cid, self.path)
+ if err is not None:
+ return self._json(*err)
+ return self._rcb1({**out, "stopped": None}, blobs)
+
+ return H
+
+
+def serve(svc: L.Service, port: int) -> None:
+ """L.serve with this service's handler (the same spec defaults and check, and L.bind: a port still held when the
+ service restarts is waited for, not a crash)."""
+ svc.spec.setdefault("on_service_error", L.ON_SERVICE_ERROR)
+ svc.spec.setdefault("max_call_wall_s", EP.max_call_wall_s)
+ L.check_spec(svc.spec)
+ L.bind(port, make_handler(svc)).serve_forever()
+
+
+serve(L.Service(spec=SPEC, lock=LOCK, state=E, submit=do_submit, observe=do_observe, kinematics=do_kinematics, status=status,
+ on_submit_error=lambda exc: finish("error"), log=EP.log, episode=EP, state_sha=S.state_sha,
+ proprio=proprio_row), port=int(os.environ.get("RCB_PORT", "8000")))
diff --git a/tasks/mujoco_playground_go1_joystick_rough_terrain-standard/instruction.md b/tasks/mujoco_playground_go1_joystick_rough_terrain-standard/instruction.md
index 35bf3e5f3c..b6917e4e7a 100644
--- a/tasks/mujoco_playground_go1_joystick_rough_terrain-standard/instruction.md
+++ b/tasks/mujoco_playground_go1_joystick_rough_terrain-standard/instruction.md
@@ -2,9 +2,9 @@
**Task:** Make the Unitree Go1 follow the joystick command schedule on rough terrain (a 20 x 20 m heightfield, up to 5 cm high) for the whole 20 s episode without falling over.
-You control the Unitree Go1 quadruped — **as if it were real hardware** — in one episode of MuJoCo Playground's
+You control the Unitree Go1 quadruped — **as if it were real hardware** — in MuJoCo Playground's
`Go1JoystickRoughTerrain` environment (plain MuJoCo 3.3.7 with Playground's scene, commit 4057c14). The simulator runs in a service you
-cannot open. You get one client and one episode.
+cannot open. You get one client and up to 11 episodes.
## The episode
The robot starts at Playground's `home` pose, its base moved to (-0.096, 0.049) m and turned to a yaw of 118.2 deg, with a random base velocity (Playground's reset). The ground is Playground's rough terrain: a 20 x 20 m heightfield (`hfield.png`, up to 5 cm high) centred on the world origin, with nothing beyond its edge.
@@ -23,7 +23,15 @@ For 20 s (1000 control steps) it is given this command schedule, drawn by Playgr
The same schedule is `spec["command_schedule"]`; every response's `meta["command_now"]` is the current command.
## Rules (real-robot rules)
-- **No reset.** One episode. You see four cameras that follow the robot (RGB, depth if you ask) and the robot's own
+- **Episodes and resets.** You have **10 resets** (`spec["budget"]["reset"]`, which is authoritative: if it
+ says 0, there is no reset and the one episode is the run). A reset (`POST /reset`, below) puts the robot back on the
+ same frozen starting state and starts a new episode, either abandoning the running one or after one has ended; it
+ takes no simulated time and returns the start's images and proprioception. Once an episode has ended, `submit`
+ raises `EpisodeOver` until you reset. **Only the last episode is graded**: an episode you reset away is never graded,
+ a success ends the run at once (no reset after it), and when your time is up the episode then running is the one
+ graded. The wall clock is the run's and never restarts; `meta["episode"]` and `meta["resets_left"]` count the
+ episodes (`meta["control_steps"]` is the current episode's).
+- **No privileged state.** You see the cameras that follow the robot (RGB, depth if you ask) and the robot's own
sensors: joint positions, velocities and torques, the IMU (gyro, accelerometer, gravity's direction, the trunk's
velocity), foot contacts. No world positions, no success check.
- **The episode ends** after 1000 control steps, the whole schedule (`spec["max_control_steps"]`, counted by `meta["control_steps"]`), when the robot meets its termination rule (the trunk's up axis pointing below the horizon (the z component of the IMU frame's z axis negative): a failure), when you submit `done`, or when your time is up (before the end of the schedule, both are failures).
@@ -32,8 +40,8 @@ The same schedule is `spec["command_schedule"]`; every response's `meta["command
`get_obs(seconds=...)`, which holds the current targets). `meta["time_left_s"]` is the service's count of the limit,
from your first request to the service: it never runs out before your time does, so plan by it.
- Nothing to hand in: the service records the episode and grades it. Leave your script at `/app/output/gen.py`.
-- **If the service crashes**, a call fails with a connection error: it restarts and resumes the SAME episode (never a
- reset) in the state after the last call it completed. `r.wait_ready()` waits for it; then check `r.status()`.
+- **If the service crashes**, a call fails with a connection error: it restarts and resumes the SAME episode (a crash
+ is never a reset and costs none) in the state after the last call it completed. `r.wait_ready()` waits for it; then check `r.status()`.
- Nothing is put into your context automatically: images come back as arrays in Python objects.
## Client (`/opt/tools/rcb_client.py`)
@@ -46,7 +54,8 @@ res = r.submit("joint_position", A, dt=0.02) # A: (T, 12) rows dt apart; res.p
obs = r.get_obs(seconds=1.0, video="mp4") # hold the current targets and watch for 1 s (max 30)
r.fk(q); r.ik({"FR": [x, y, z, 1, 0, 0, 0]}, seed=q) # feet in the trunk frame <-> joints (no physics)
r.submit("done", []) # end the episode now
-r.status() # calls, control steps, done, reason, elapsed time
+r.status() # calls, control steps, done, reason, elapsed time, episode, resets_left
+obs = r._post("/reset", {"cameras": ["track"]}, [], 900) # POST /reset: a new episode (the 2.2 client has no reset())
```
Cameras: `track`, `side`, `back`, `top` (they follow the robot); `spec["cameras"]` gives each one's mounting and intrinsics, and every response carries its pose. Proprioception (`spec["proprio_groups"]`): `joint_pos`, `joint_vel`, `joint_effort` (12 each, action
order), `imu_gyro`, `imu_accelerometer`, `imu_gravity`, `imu_linvel` (3 each, IMU frame: x forward, y left, z up),
diff --git a/tasks/mujoco_playground_go1_joystick_rough_terrain-standard/task.toml b/tasks/mujoco_playground_go1_joystick_rough_terrain-standard/task.toml
index 664eadadc1..a79348776f 100644
--- a/tasks/mujoco_playground_go1_joystick_rough_terrain-standard/task.toml
+++ b/tasks/mujoco_playground_go1_joystick_rough_terrain-standard/task.toml
@@ -16,7 +16,7 @@ artifacts = [{ source = "/var/rcb-monitor/", service = "rcb_monitor" },
[task]
name = "robot-coding-bench/mujoco_playground_go1_joystick_rough_terrain-standard"
-description = "STANDARD MODE (robot as a service, eai-standard/2.2): MuJoCo Playground Go1JoystickRoughTerrain (Unitree Go1) on the same frozen instance as mujoco_playground_go1_joystick_rough_terrain-privileged: Make the Unitree Go1 follow the joystick command schedule on rough terrain (a 20 x 20 m heightfield, up to 5 cm high) for the whole 20 s episode without falling over. The agent only gets a client: four cameras that follow the robot (RGB, optional depth), the robot's joints and IMU, joint position chunks at 50 Hz, the feet's kinematics; one episode, no reset, no world state; the sim service records the episode, replays it in fresh processes and grades it."
+description = "STANDARD MODE (robot as a service, eai-standard/2.2): MuJoCo Playground Go1JoystickRoughTerrain (Unitree Go1) on the same frozen instance as mujoco_playground_go1_joystick_rough_terrain-privileged: Make the Unitree Go1 follow the joystick command schedule on rough terrain (a 20 x 20 m heightfield, up to 5 cm high) for the whole 20 s episode without falling over. The agent only gets a client: four cameras that follow the robot (RGB, optional depth), the robot's joints and IMU, joint position chunks at 50 Hz, the feet's kinematics; 10 resets (only the last episode is graded), no world state; the sim service records the episode, replays it in fresh processes and grades it."
authors = [{ name = "Changhe Chen" }]
keywords = ["robotics", "mujoco", "mujoco-playground", "unitree-go1", "quadruped", "locomotion", "joystick", "code-as-policy", "trajectory", "standard-mode"]
@@ -28,6 +28,7 @@ simulator = "mujoco"
robot = "unitree-go1"
family = "mujoco_playground/Go1JoystickRoughTerrain"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "full"
instance_seed = 1
diff --git a/tasks/mujoco_playground_go1_joystick_rough_terrain-standard/tests/grade.py b/tasks/mujoco_playground_go1_joystick_rough_terrain-standard/tests/grade.py
index 41ce495421..3d4c23ad7e 100644
--- a/tasks/mujoco_playground_go1_joystick_rough_terrain-standard/tests/grade.py
+++ b/tasks/mujoco_playground_go1_joystick_rough_terrain-standard/tests/grade.py
@@ -77,6 +77,9 @@ write(
"sim_error": int(d.get("reason") == "error"),
"sim_startups": int(d.get("sim_startups") or 0),
"resumes": int(d.get("resumes") or 0),
+ "resets": int(d.get("resets") or 0), # PROTOCOL §4.5: resets the agent used; only the last episode is graded
+ "reset_budget": int(d.get("reset_budget") or 0),
+ "episodes": int(d.get("episodes") or 1),
"resume_state_mismatch": int(d.get("resume_state_mismatch") or 0),
"resume_max_joint_dev": num("resume_max_joint_dev"),
},
diff --git a/tasks/mujoco_playground_panda_open_cabinet-privileged/task.toml b/tasks/mujoco_playground_panda_open_cabinet-privileged/task.toml
index b4aea1d053..a6758016bf 100644
--- a/tasks/mujoco_playground_panda_open_cabinet-privileged/task.toml
+++ b/tasks/mujoco_playground_panda_open_cabinet-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "mujoco"
robot = "franka-panda"
family = "mujoco_playground/PandaOpenCabinet"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 60417 # Playground's reset randomization (target x, arm start perturbation) with this seed
expert_time_estimate_hours = 1.0
diff --git a/tasks/mujoco_playground_panda_open_cabinet-standard/task.toml b/tasks/mujoco_playground_panda_open_cabinet-standard/task.toml
index f3e88a8f14..670d5e0ca5 100644
--- a/tasks/mujoco_playground_panda_open_cabinet-standard/task.toml
+++ b/tasks/mujoco_playground_panda_open_cabinet-standard/task.toml
@@ -24,6 +24,7 @@ family = "mujoco_playground/PandaOpenCabinet"
instance_seed = 60417 # Playground's reset randomization (target x, arm start perturbation) with this seed
expert_time_estimate_hours = 1.0
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
diff --git a/tasks/mujoco_playground_panda_pick_cube_orientation-privileged/task.toml b/tasks/mujoco_playground_panda_pick_cube_orientation-privileged/task.toml
index 4cf9766185..20c93a58bc 100644
--- a/tasks/mujoco_playground_panda_pick_cube_orientation-privileged/task.toml
+++ b/tasks/mujoco_playground_panda_pick_cube_orientation-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "mujoco"
robot = "franka-panda"
family = "mujoco_playground/PandaPickCubeOrientation"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 60417 # Playground's reset randomization (box xy, target position and orientation) with this seed
expert_time_estimate_hours = 1.0
diff --git a/tasks/mujoco_playground_panda_pick_cube_orientation-standard/task.toml b/tasks/mujoco_playground_panda_pick_cube_orientation-standard/task.toml
index f394cc5c0f..325a275274 100644
--- a/tasks/mujoco_playground_panda_pick_cube_orientation-standard/task.toml
+++ b/tasks/mujoco_playground_panda_pick_cube_orientation-standard/task.toml
@@ -24,6 +24,7 @@ family = "mujoco_playground/PandaPickCubeOrientation"
instance_seed = 60417 # Playground's reset randomization (box xy, target position and orientation) with this seed
expert_time_estimate_hours = 1.0
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
diff --git a/tasks/mujoco_playground_spot_flat_terrain_joystick-privileged/README.md b/tasks/mujoco_playground_spot_flat_terrain_joystick-privileged/README.md
index 3944235e14..1c85624f9c 100644
--- a/tasks/mujoco_playground_spot_flat_terrain_joystick-privileged/README.md
+++ b/tasks/mujoco_playground_spot_flat_terrain_joystick-privileged/README.md
@@ -53,3 +53,11 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
### Provenance
Generated by `scripts/mujoco-playground/generate.py` from `scripts/mujoco-playground/instances/SpotFlatTerrainJoystick.json`
(`make_instances.py`, `select_instances.py`) and `scripts/mujoco-playground/oracles/SpotFlatTerrainJoystick.npy` (`oracle.py`).
+
+### Change Log
+- 2026-10-14 (review of board 0105, B1-m2): `tests/grade.py` only. A replay process that died or ran out of time, a
+ rebuilt start that does not match the instance's hash (`instance_mismatch`) and an exception of the verifier itself
+ now write `infra_error = 1` beside `grader_error = 1`: no verdict on the agent. Before, they wrote `grader_error = 1`
+ alone, which counted as the agent's 0. Every `reward.json` now carries `infra_error`. The scene, the instance, the
+ pass rule and the reference are unchanged, so the task keeps its name; a past 0 with `grader_error = 1` was an
+ infrastructure failure.
diff --git a/tasks/mujoco_playground_spot_flat_terrain_joystick-privileged/task.toml b/tasks/mujoco_playground_spot_flat_terrain_joystick-privileged/task.toml
index 215f05cd04..83b2170c28 100644
--- a/tasks/mujoco_playground_spot_flat_terrain_joystick-privileged/task.toml
+++ b/tasks/mujoco_playground_spot_flat_terrain_joystick-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "mujoco"
robot = "boston-dynamics-spot"
family = "mujoco_playground/SpotFlatTerrainJoystick"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 0 # the PRNGKey of Playground's reset (and, for the joystick tasks, its command process)
expert_time_estimate_hours = 3.0
diff --git a/tasks/mujoco_playground_spot_flat_terrain_joystick-privileged/tests/grade.py b/tasks/mujoco_playground_spot_flat_terrain_joystick-privileged/tests/grade.py
index c75c326868..1667d090c8 100644
--- a/tasks/mujoco_playground_spot_flat_terrain_joystick-privileged/tests/grade.py
+++ b/tasks/mujoco_playground_spot_flat_terrain_joystick-privileged/tests/grade.py
@@ -6,6 +6,11 @@ Two replays, each in its own fresh process (`rcb_quadruped.py replay`: the froze
task's success check on the whole trajectory). success = both replays succeed and end in the same state. Writes
/logs/verifier/reward.txt (1/0) and reward.json (numbers only); then a third replay renders replay.mp4, after the
verdict is on disk (it never changes it).
+
+Failures that are not the agent's (docs/PROTOCOL.md: infra vs agent failure) carry `grader_error` = 1 and `infra_error`
+= 1, so they are no verdict on the agent: a replay process that died or timed out, an image that no longer rebuilds the
+frozen start (`instance_mismatch`), and any exception of the verifier itself. A file the agent got wrong is the agent's
+0 (`bad_trajectory`, infra_error 0). Every reward.json carries success, deterministic, grader_error and infra_error.
"""
import argparse
import json
@@ -22,8 +27,12 @@ TOOL = os.path.join(os.path.dirname(os.path.abspath(__file__)), "rcb_quadruped.p
REPLAY_TIMEOUT_S = 300 # a 1000-step replay takes a few seconds; the video replay renders 4 views every 2nd step
+BASE = {"success": 0, "deterministic": 0, "grader_error": 0, "infra_error": 0}
+
+
def write(ok: bool, out: dict) -> None:
os.makedirs(LOG_DIR, exist_ok=True)
+ out = {**BASE, **out}
open(os.path.join(LOG_DIR, "reward.txt"), "w").write("1\n" if ok else "0\n")
json.dump(out, open(os.path.join(LOG_DIR, "reward.json"), "w"))
print(json.dumps(out), flush=True)
@@ -96,11 +105,11 @@ def main() -> int:
"n_actions": int(acts.shape[0]) if acts.ndim == 2 else 0, "grader_error": 0, **cost})
return 0
runs = [replay(a.instance, a.traj, None), replay(a.instance, a.traj, None)]
- if any("error" in r and not r.get("bad_trajectory") for r in runs):
- write(False, {"success": 0, "deterministic": 0, "grader_error": 1, **cost})
+ if any("error" in r and not r.get("bad_trajectory") for r in runs): # the replay process died / timed out
+ write(False, {"success": 0, "deterministic": 0, "grader_error": 1, "infra_error": 1, "replay_failed": 1, **cost})
return 0
if not all(r.get("instance_ok") for r in runs): # the image no longer builds the frozen start: not the agent's 0
- write(False, {"success": 0, "deterministic": 0, "grader_error": 1, "instance_mismatch": 1, **cost})
+ write(False, {"success": 0, "deterministic": 0, "grader_error": 1, "infra_error": 1, "instance_mismatch": 1, **cost})
return 0
all_success = all(bool(r.get("success")) for r in runs)
deterministic = runs[0].get("state_sha") == runs[1].get("state_sha")
@@ -116,7 +125,7 @@ def main() -> int:
write(ok, out) # the verdict, before any video
except Exception: # noqa: BLE001
traceback.print_exc(file=sys.stderr)
- write(False, {"success": 0, "deterministic": 0, "grader_error": 1, **cost})
+ write(False, {"success": 0, "deterministic": 0, "grader_error": 1, "infra_error": 1, **cost})
return 0
try: # the video, for humans: it never changes the verdict
if replay(a.instance, a.traj, os.path.join(LOG_DIR, "replay.mp4")).get("video"):
diff --git a/tasks/mujoco_playground_spot_flat_terrain_joystick-standard/README.md b/tasks/mujoco_playground_spot_flat_terrain_joystick-standard/README.md
index 1744a18086..fe7ff003bb 100644
--- a/tasks/mujoco_playground_spot_flat_terrain_joystick-standard/README.md
+++ b/tasks/mujoco_playground_spot_flat_terrain_joystick-standard/README.md
@@ -13,14 +13,14 @@
## Task
MuJoCo Playground `SpotFlatTerrainJoystick` (Boston Dynamics Spot) on the same frozen instance as `mujoco_playground_spot_flat_terrain_joystick-privileged`: Make Boston Dynamics Spot follow the joystick command schedule on flat ground for the whole 20 s episode without falling over.
-For 20 s the robot gets a joystick command schedule (5 segments: planar velocity and yaw rate in its own frame) and must track it without falling over. Standard mode (eai-standard/2.2): the agent drives the robot through a client in one episode, with only
-the robot's own sensors and four cameras that follow it.
+For 20 s the robot gets a joystick command schedule (5 segments: planar velocity and yaw rate in its own frame) and must track it without falling over. Standard mode (eai-standard/2.2): the agent drives the robot through a client, with 10 resets (only
+the last episode is graded), the robot's own sensors and the cameras that follow it.
## Difficulty
-Same instance and success check as `mujoco_playground_spot_flat_terrain_joystick-privileged`. In standard mode the agent loses resets, the simulator and
+Same instance and success check as `mujoco_playground_spot_flat_terrain_joystick-privileged`. In standard mode the agent loses free resets, the simulator and
the world state: it sees camera images and proprioception (joints, IMU, foot contacts), acts in joint position chunks,
-and gets one episode, so it has to close the loop on the robot's own sensors and recover from its own mistakes without
-starting over.
+and gets 10 resets (`RCB_RESET_BUDGET`, the compose file's default; PROTOCOL §4.5), so it has to close the loop
+on the robot's own sensors; only the last episode is graded.
## Reference solution
`solution/oracle_replay.py` streams the privileged reference's rows (`solution/oracle.npy`, 1000 rows, float32-exact)
@@ -60,3 +60,4 @@ Generated by `scripts/mujoco-playground/generate.py`, as its privileged twin.
- 2026-10-10 (review 0067): a `final.json` that is missing, unreadable or was never finalized by the collect hook is graded `infra_error = 1` (no verdict on the agent); before, a missing one was a plain 0 and an unfinalized one was graded as it stood. The collect hook writes `final.json` atomically. A request field of the wrong type (`dt` or `seconds` null, `cameras` not a list of names, a `note` that is not a string) is answered 400 before any physics instead of ending the episode with a 500. `meta["time_left_s"]` counts from the agent's first call, not from the service's start.
- 2026-10-10 (review round 2, board 0067): the client only (`environment/rcb_client.py`). `Robot()` waits up to 3600 s for the service (was 600 s) and prints its progress to stderr once a minute, so the first `Robot()` outlasts the simulator's start (no effect on this benchmark, whose service is up within a minute); `cameras="name"` (a string) is one camera, no longer its characters. The service, the collect hook, the grader, the scene and the pass rule are unchanged.
- 2026-10-10 (review round 3, board 0067): the shared service layer and the client only (`environment/sim/rcb_limited.py`, `environment/rcb_client.py`). A camera named more than once in `cameras` is rendered and returned once (N copies of one name made the service build N x N images within the frame budget, which could exhaust its memory). A service that has given up on the episode (it could not start it, or could not resume it after its crashes) answers every request with 409 and the reason, and `Robot()` / `wait_ready()` raise `EpisodeOver` at once; before, nothing answered and they waited out their timeout (up to an hour). The adapter, the collect hook, the grader, the scene and the pass rule are unchanged.
+- 2026-10-14 (owner decision C1, board 0113): 10 resets (PROTOCOL §4.5). The service (`environment/sim/server.py`) gained the optional reset capability: `RCB_RESET_BUDGET` (the compose file's default 10; 0 restores the old single episode), `POST /reset` abandons the running or ended episode and starts a new one from the same frozen start (no physics), refused (409) after a success, once the agent's time is up and once the budget is spent; only the last episode is graded (`trajectory.npz` holds its rows, replayed twice from the frozen start as before); `/status` and every `meta` add `episode`, `resets_left` and `control_steps_total`; `final.json` and `reward.json` add `resets`, `reset_budget` and `episodes`. A crash resumes the current episode (the journal is replayed with the committed resets). The instruction's rules, the client example (`POST /reset` through `Robot._post`) and this README say so. The scene, the instance, the success check and the grader's pass rule are unchanged; results recorded before this change ran with one episode and are reported separately.
diff --git a/tasks/mujoco_playground_spot_flat_terrain_joystick-standard/environment/docker-compose.yaml b/tasks/mujoco_playground_spot_flat_terrain_joystick-standard/environment/docker-compose.yaml
index bc0c363241..1226b8c3a6 100644
--- a/tasks/mujoco_playground_spot_flat_terrain_joystick-standard/environment/docker-compose.yaml
+++ b/tasks/mujoco_playground_spot_flat_terrain_joystick-standard/environment/docker-compose.yaml
@@ -28,6 +28,7 @@ services:
RCB_STARTUP_TIMEOUT_S: "300" # the scene builds in about a second
PYTHONPATH: /srv
RCB_SESSION_LOG: "" # the sidecar's own Session must not write the agent-side interaction log
+ RCB_RESET_BUDGET: "${RCB_RESET_BUDGET:-10}" # Playground quadrupeds standard: 10 resets (PROTOCOL §4.5); 0: one episode
# >>> rcb_monitor (resource limits monitor, docs/PROTOCOL.md §2.10; generated by scripts/rcb_monitor/inject.py)
rcb_monitor:
image: ghcr.io/mll-lab-nu/eai-monitor:0.1.0@sha256:f2c910f53f79c2bfa4a71298b10c78383cae3afc40720ca562cc845b8d20fb0c
diff --git a/tasks/mujoco_playground_spot_flat_terrain_joystick-standard/environment/sim/server.py b/tasks/mujoco_playground_spot_flat_terrain_joystick-standard/environment/sim/server.py
index 108eca3cae..cad8920259 100644
--- a/tasks/mujoco_playground_spot_flat_terrain_joystick-standard/environment/sim/server.py
+++ b/tasks/mujoco_playground_spot_flat_terrain_joystick-standard/environment/sim/server.py
@@ -1,16 +1,20 @@
"""eai-standard/2.2 sim service for MuJoCo Playground's quadruped tasks (docs/STANDARD_MODE.md): the adapter of the
shared layer rcb_limited.py; the simulator is rcb_quadruped.Session (plain MuJoCo on the CPU).
-HTTP surface: GET /spec, GET /status, POST /observe, POST /submit, POST /kinematics. Every call is logged to
-/data/episode.jsonl.
+HTTP surface: GET /spec, GET /status, POST /observe, POST /submit, POST /kinematics, and POST /reset when the run grants
+resets (RCB_RESET_BUDGET, the optional reset capability of docs/STANDARD_MODE.md / PROTOCOL §4.5; the quadrupeds'
+standard tasks default to 10 in their compose file, owner 2026-10-14). Every call is logged to /data/episode.jsonl.
The episode is the frozen instance's: one control step = the instance's control period (Go1, Spot: 1/50 s; Barkour:
1/100 s). A joystick episode ends after the whole command schedule
(the horizon), when the robot meets Playground's termination rule (it falls over: a failure), when the agent submits
`done`, or when its time is over; success = the task's check on the steps executed. A getup episode ends the moment the
-check holds (success), after the horizon, on `done` or on timeout.
+check holds (success), after the horizon, on `done` or on timeout. With a reset budget, POST /reset abandons the running
+(or ended) episode and starts a new one from the frozen instance; only the last episode is graded, and a success ends
+the trial.
Crash recovery (rcb_limited.Episode, resume class `exact`: MuJoCo on the CPU repeats bit for bit in a fresh process):
-every executed row is journaled; after a crash the service rebuilds the episode and replays the journal. When the
+every executed row is journaled; after a crash the service rebuilds the episode and replays the journal, resetting where
+the committed resets did (their journal rows are checkpointed, `persist`). When the
episode ends the service writes /data/trajectory.npz + /data/episode.json; the collect hook (finalize.sh) makes it exit
and replays the trajectory in two fresh processes."""
@@ -31,18 +35,21 @@ INSTANCE = os.environ.get("RCB_INSTANCE", os.path.join(os.path.dirname(os.path.a
WALL_CLOCK_S = float(os.environ.get("RCB_WALL_CLOCK_S", "3600"))
MAX_CHUNK_S = float(os.environ.get("RCB_MAX_CHUNK_S", "10"))
MAX_WATCH_S = float(os.environ.get("RCB_MAX_WATCH_S", "30"))
+RESET_BUDGET = max(0, int(os.environ.get("RCB_RESET_BUDGET") or 0)) # the compose file's default: 10 (PROTOCOL §4.5)
CAM_W, CAM_H = 320, 240
LOCK = threading.Lock()
PID_FILE = os.path.join(DATA, "server.pid")
E = {"ready": False, "step": 0, "calls": 0, "done": False, "reason": None, "success": False, "t0": time.time(), "error": None,
- "t_first": None} # t_first: the agent's first call (its clock starts there; kept through a crash by the checkpoints)
+ "t_first": None, # t_first: the agent's first call (its clock starts there; kept through a crash by the checkpoints)
+ "resets": 0, # resets done (the episode is resets + 1)
+ "reset_at": []} # the journal row at which each reset happened: a resume resets there again
os.environ["RCB_SESSION_LOG"] = ""
os.environ["RCB_AGENT_TRACE_DIR"] = ""
import rcb_quadruped as Q # noqa: E402 (first: it picks the rendering backend before mujoco loads)
import mujoco # noqa: E402
-EP = L.Episode(DATA, dim=Q.N, resume="exact", persist=("t_first",))
+EP = L.Episode(DATA, dim=Q.N, resume="exact", persist=("t_first", "resets", "reset_at"))
EP.startup() # start count, back-off after failed starts, the startup watchdog until scene_built()
S = Q.Session(INSTANCE)
ROWS: list = [] # every executed row (a resume replays them first): trajectory.npz
@@ -153,13 +160,15 @@ def schedule_text() -> list:
"wz_radps": round(c["wz"], 4)} for c in S.inst["commands"]]
+ONE = "one episode, no reset" if not RESET_BUDGET else (
+ f"an episode (with up to {RESET_BUDGET} resets: POST /reset starts a new one; only the last is graded)")
TASK_EPISODE = {
- "joystick": f"one episode, no reset: it ends after {S.horizon} control steps (the {S.horizon * DT:g} s command "
+ "joystick": f"{ONE}: it ends after {S.horizon} control steps (the {S.horizon * DT:g} s command "
f"schedule), when the robot meets its termination rule ({S.spec.get('term_text')}: a failure), when you "
"submit `done`, or when your time is over",
- "getup": f"one episode, no reset: it ends the moment the robot has stood upright for {Q.HOLD_STEPS} control steps in a "
+ "getup": f"{ONE}: it ends the moment the robot has stood upright for {Q.HOLD_STEPS} control steps in a "
f"row (success), after {S.horizon} control steps, when you submit `done`, or when your time is over",
- "stand_on_two": f"one episode, no reset: it ends the moment the pose has been held for {Q.POSE_HOLD_STEPS} control steps "
+ "stand_on_two": f"{ONE}: it ends the moment the pose has been held for {Q.POSE_HOLD_STEPS} control steps "
"in a row (success), when the robot meets the termination rule (a failure: its trunk's up axis below "
f"-0.25, or {'a leg' if S.spec.get('hind_leg_contact') else 'a front leg'}'s thigh, calf or hip on the "
f"floor), after {S.horizon} control steps, when you "
@@ -204,7 +213,7 @@ SPEC = {
"world": {"frame": "z up; the robot starts where the episode puts it (no world coordinates are given)", "up": "+z",
"units": {"length": "m", "angle": "rad", "time": "s"}, "quaternion_order": "wxyz",
"camera_axes": "intrinsic_cv / extrinsic_cv: OpenCV camera axes (x right, y down, z forward); cam2world_gl: OpenGL axes"},
- "budget": {"reset": 0, "steps": None, "wall_clock_s": WALL_CLOCK_S}, # + "clock" (rcb_limited.CLOCK_NOTE): what time_left_s counts from
+ "budget": {"reset": RESET_BUDGET, "steps": None, "wall_clock_s": WALL_CLOCK_S}, # + "clock" (rcb_limited.CLOCK_NOTE): what time_left_s counts from
"max_control_steps": S.horizon,
"episode": TASK_EPISODE,
"command_schedule": schedule_text(),
@@ -223,6 +232,15 @@ SPEC = {
"depth": {"modes": list(L.DEPTH_MODES), "default": "none", "kind": "z-depth", "unit": "m", "invalid": 0,
"max_m": L.DEPTH_MAX_M, "frames_encoding": "uint16 mm, zlib", "noise": L.DEPTH_NOISE},
}
+if RESET_BUDGET:
+ SPEC["reset"] = (
+ f"POST /reset (JSON or RCB1 {{cameras, video_fps, depth}}; no physics) abandons the running episode, or follows "
+ f"one that has ended, and starts a new one from the same frozen starting state, {RESET_BUDGET} times in all; it "
+ "returns the new episode's first frames and proprioception. Only the last episode is graded: an episode you "
+ "reset away is never graded, a success ends the trial (409 after it), and when your time is up the episode then "
+ "running is the one graded. The wall clock never restarts. /status and every meta add `episode`, `resets_left` "
+ "and `control_steps_total` (all episodes); `control_steps` counts the current episode's."
+ )
L.check_spec(SPEC) # a spec that misses a 2.2 field or names a forbidden one never starts the episode
EP.scene_built()
@@ -277,6 +295,7 @@ def summary() -> dict:
"reason": E["reason"], "done": E["done"], "env": S.env, "protocol": L.PROTOCOL, "n_actions": len(ROWS),
"fell_at": chk.get("fell_at"), "worst_lin_err": chk.get("worst_lin_err"), "worst_yaw_err": chk.get("worst_yaw_err"),
"tilt_deg": chk.get("tilt_deg"), "imu_height": chk.get("imu_height"), "pose_angle_deg": chk.get("pose_angle_deg"),
+ "episodes": E["resets"] + 1, "resets": E["resets"], "reset_budget": RESET_BUDGET, "control_steps_total": EP.rows,
**EP.final_fields()}
@@ -294,7 +313,8 @@ def finish(reason: str) -> None:
E["success"] = bool(chk["success"])
except Exception: # noqa: BLE001
traceback.print_exc()
- EP.log({"ev": "end", "reason": reason, "step": E["step"], "control_steps": ctrl_steps(), "success": bool(E["success"])})
+ EP.log({"ev": "end", "reason": reason, "step": E["step"], "control_steps": ctrl_steps(), "success": bool(E["success"]),
+ "episode": E["resets"] + 1, "journal_rows": EP.rows})
dump()
@@ -388,7 +408,7 @@ def meta_common(cams) -> dict:
"sim_time_s": round(n * DT, 3), "time_s": round(_agent_time(), 2),
"time_left_s": max(0.0, round(WALL_CLOCK_S - _agent_time(), 1)),
"proprio": {"layout": PROPRIO_LAYOUT, "now": proprio_row().tolist()}, "done": E["done"], "reason": E["reason"],
- "cameras": {c: camera_info(c) for c in cams}}
+ "cameras": {c: camera_info(c) for c in cams}, **reset_counters()}
if S.kind == "joystick":
out["command_now"] = dict(zip(("vx_mps", "vy_mps", "wz_radps"), S.command_now().round(4).tolist()))
return out
@@ -478,10 +498,54 @@ def do_observe(meta: dict):
return out, rec.parts(video if seconds > 0 else ("frames" if video == "mp4" else video), fps)
+def reset_counters() -> dict:
+ if not RESET_BUDGET:
+ return {}
+ left = 0 if (E["success"] or E.get("exiting")) else RESET_BUDGET - E["resets"]
+ return {"episode": E["resets"] + 1, "resets_left": left, "control_steps_total": EP.rows}
+
+
def status() -> dict:
_touch()
return {"step": E["step"], "calls": E["calls"], "control_steps": ctrl_steps(), "done": E["done"], "reason": E["reason"],
- "elapsed_s": round(time.time() - E["t0"], 1)}
+ "elapsed_s": round(time.time() - E["t0"], 1), **reset_counters()}
+
+
+class Refused(Exception):
+ """A /reset the service turns down (409): nothing changes."""
+
+ def __init__(self, body: dict):
+ super().__init__(body.get("error"))
+ self.body = body
+
+
+def do_reset(meta: dict):
+ """POST /reset: a new episode from the frozen instance (the running one, if any, is abandoned, never graded)."""
+ meta = _request(meta)
+ cams, fps = cameras_of(meta), video_fps(meta)
+ L.depth_mode(meta) # a bad depth value is a 400 before anything changes
+ if E["success"]:
+ raise Refused({"error": "trial finished: an episode succeeded", "reason": E["reason"], "resets_left": 0})
+ if E["resets"] >= RESET_BUDGET:
+ raise Refused({"error": "reset budget exhausted", "reason": "reset budget exhausted", "budget": RESET_BUDGET,
+ "resets_left": 0})
+ abandoned = {"episode": E["resets"] + 1, "control_steps": ctrl_steps(), "done": E["done"], "reason": E["reason"]}
+ if not E["done"]: # the abandoned episode's end, for the log (never graded: its rows leave the trajectory)
+ EP.log({"ev": "end", "reason": "reset", "step": E["step"], "control_steps": ctrl_steps(), "success": False,
+ "episode": E["resets"] + 1, "journal_rows": EP.rows})
+ S.reset()
+ ROWS.clear()
+ E["resets"] += 1
+ E["reset_at"] = list(E["reset_at"]) + [EP.rows]
+ E.update(done=False, reason=None, success=False, t_end=None, episode_control_steps=None, dumped=False, check=None)
+ EP._ck = None # the state changed without a journal row: the checkpoint after this call must hash it anew
+ E["calls"] += 1
+ EP.log({"ev": "reset", "call": E["calls"], "episode": E["resets"] + 1, "journal_rows": EP.rows,
+ "abandoned": abandoned, "state_sha": S.state_sha()})
+ rec = recorder(cams, fps, meta)
+ rec.tick(force=True, last=True)
+ out = {**meta_common(cams), "reset_id": E["resets"], "ok": True}
+ return out, rec.parts("frames", fps)
def end_from_signal() -> None:
@@ -498,20 +562,106 @@ def end_from_signal() -> None:
os._exit(0)
+# A resume replays the journal from the frozen instance and resets where the committed resets did: their journal rows
+# are in the last checkpoint (`persist`), the one Episode.begin resumes from.
+_CPS = [e for e in EP.events() if e.get("ev") == "checkpoint"]
+RESUME = {"at": list((_CPS[-1].get("persist") or {}).get("reset_at") or []) if _CPS else [], "i": 0}
+
+
+def _resume_resets() -> None:
+ while RESUME["at"] and int(RESUME["at"][0]) <= RESUME["i"]:
+ RESUME["at"].pop(0)
+ S.reset()
+ ROWS.clear()
+
+
def replay_row(row) -> None:
"""A journaled row during a resume."""
+ _resume_resets()
S.step(row)
ROWS.append(np.asarray(row, dtype=np.float64))
+ RESUME["i"] += 1
-EP.begin(state=E, step=replay_row, state_sha=S.state_sha, proprio=proprio_row, spec=SPEC, on_end=end_from_signal,
- start_event={"env": S.env, "seed": S.inst.get("seed"), "cameras": CAMS})
+def resume_sha() -> str:
+ _resume_resets() # a reset after the last replayed row (or before the first)
+ return S.state_sha()
+
+
+RESUMED = EP.begin(state=E, step=replay_row, state_sha=resume_sha, proprio=proprio_row, spec=SPEC, on_end=end_from_signal,
+ start_event={"env": S.env, "seed": S.inst.get("seed"), "cameras": CAMS})
+if RESUMED and E["resets"]: # Episode.begin looks for an "end" of any episode: only the current one's counts
+ _ends = [e for e in EP.events() if e.get("ev") == "end" and int(e.get("episode") or 1) == E["resets"] + 1
+ and e.get("reason") != "reset" and int(e.get("journal_rows") or 0) <= EP.rows]
+ if _ends:
+ E.update(done=True, reason=_ends[-1].get("reason"), success=bool(_ends[-1].get("success")),
+ t_end=float(_ends[-1]["ts"]), episode_control_steps=int(_ends[-1].get("control_steps") or ctrl_steps()))
+ else:
+ E.update(done=False, reason=None, success=False, t_end=None, episode_control_steps=None)
if E["done"]:
dump() # resumed after the episode had ended: the record is rewritten from the replay
E["ready"] = True
write_final()
open(PID_FILE, "w").write(str(os.getpid()))
print(f"{L.PROTOCOL} MuJoCo Playground quadruped sim ready ({S.env}; startup {EP.n_startups}, resumes {EP.resumes})", flush=True)
-L.serve(L.Service(spec=SPEC, lock=LOCK, state=E, submit=do_submit, observe=do_observe, kinematics=do_kinematics, status=status,
- on_submit_error=lambda exc: finish("error"), log=EP.log, episode=EP, state_sha=S.state_sha,
- proprio=proprio_row), port=int(os.environ.get("RCB_PORT", "8000")))
+
+
+def make_handler(svc: L.Service):
+ """The shared 2.2 handler, plus POST /reset when the run grants resets (the shared layer has no reset)."""
+ base = L.make_handler(svc)
+
+ class H(base):
+ def do_POST(self):
+ if self.path != "/reset" or not RESET_BUDGET:
+ return super().do_POST()
+ try:
+ meta, _ = self._body()
+ except Exception as exc: # noqa: BLE001
+ return self._json(413 if "too large" in str(exc) else 400, {"error": str(exc)[:200]})
+ with LOCK:
+ if E.get("exiting"):
+ return self._json(409, {"error": "trial finished: your time is up", "reason": E.get("reason"),
+ "resets_left": 0})
+ cid = EP.call_begin(self.path, self.connection)
+ out = blobs = err = None
+ try:
+ out, blobs = do_reset(meta)
+ except Refused as exc:
+ err = (409, exc.body)
+ except ValueError as exc:
+ err = (400, {"error": str(exc)})
+ except Exception as exc: # noqa: BLE001
+ traceback.print_exc()
+ E["error"] = repr(exc)
+ EP.log({"ev": "error", "path": self.path, "error": repr(exc)})
+ try:
+ finish("error")
+ except Exception: # noqa: BLE001
+ traceback.print_exc()
+ err = (500, {"error": "simulator error", "detail": repr(exc)[:200]})
+ finally:
+ if out is not None or (err and err[0] == 500):
+ try:
+ EP.checkpoint(E, S.state_sha, proprio_row)
+ except Exception: # noqa: BLE001
+ traceback.print_exc()
+ EP.call_end(cid, self.path)
+ if err is not None:
+ return self._json(*err)
+ return self._rcb1({**out, "stopped": None}, blobs)
+
+ return H
+
+
+def serve(svc: L.Service, port: int) -> None:
+ """L.serve with this service's handler (the same spec defaults and check, and L.bind: a port still held when the
+ service restarts is waited for, not a crash)."""
+ svc.spec.setdefault("on_service_error", L.ON_SERVICE_ERROR)
+ svc.spec.setdefault("max_call_wall_s", EP.max_call_wall_s)
+ L.check_spec(svc.spec)
+ L.bind(port, make_handler(svc)).serve_forever()
+
+
+serve(L.Service(spec=SPEC, lock=LOCK, state=E, submit=do_submit, observe=do_observe, kinematics=do_kinematics, status=status,
+ on_submit_error=lambda exc: finish("error"), log=EP.log, episode=EP, state_sha=S.state_sha,
+ proprio=proprio_row), port=int(os.environ.get("RCB_PORT", "8000")))
diff --git a/tasks/mujoco_playground_spot_flat_terrain_joystick-standard/instruction.md b/tasks/mujoco_playground_spot_flat_terrain_joystick-standard/instruction.md
index ecf4365c6f..918741349c 100644
--- a/tasks/mujoco_playground_spot_flat_terrain_joystick-standard/instruction.md
+++ b/tasks/mujoco_playground_spot_flat_terrain_joystick-standard/instruction.md
@@ -2,9 +2,9 @@
**Task:** Make Boston Dynamics Spot follow the joystick command schedule on flat ground for the whole 20 s episode without falling over.
-You control the Boston Dynamics Spot quadruped — **as if it were real hardware** — in one episode of MuJoCo Playground's
+You control the Boston Dynamics Spot quadruped — **as if it were real hardware** — in MuJoCo Playground's
`SpotFlatTerrainJoystick` environment (plain MuJoCo 3.3.7 with Playground's scene, commit 4057c14). The simulator runs in a service you
-cannot open. You get one client and one episode.
+cannot open. You get one client and up to 11 episodes.
## The episode
The robot starts standing still at Playground's `home` pose (this environment's reset does not move it). The ground is flat.
@@ -21,7 +21,15 @@ For 20 s (1000 control steps) it is given this command schedule, drawn by Playgr
The same schedule is `spec["command_schedule"]`; every response's `meta["command_now"]` is the current command.
## Rules (real-robot rules)
-- **No reset.** One episode. You see four cameras that follow the robot (RGB, depth if you ask) and the robot's own
+- **Episodes and resets.** You have **10 resets** (`spec["budget"]["reset"]`, which is authoritative: if it
+ says 0, there is no reset and the one episode is the run). A reset (`POST /reset`, below) puts the robot back on the
+ same frozen starting state and starts a new episode, either abandoning the running one or after one has ended; it
+ takes no simulated time and returns the start's images and proprioception. Once an episode has ended, `submit`
+ raises `EpisodeOver` until you reset. **Only the last episode is graded**: an episode you reset away is never graded,
+ a success ends the run at once (no reset after it), and when your time is up the episode then running is the one
+ graded. The wall clock is the run's and never restarts; `meta["episode"]` and `meta["resets_left"]` count the
+ episodes (`meta["control_steps"]` is the current episode's).
+- **No privileged state.** You see the cameras that follow the robot (RGB, depth if you ask) and the robot's own
sensors: joint positions, velocities and torques, the IMU (gyro, accelerometer, gravity's direction, the trunk's
velocity), foot contacts. No world positions, no success check.
- **The episode ends** after 1000 control steps, the whole schedule (`spec["max_control_steps"]`, counted by `meta["control_steps"]`), when the robot meets its termination rule (the trunk tilted by more than about 32 deg (the z component of the IMU frame's z axis below 0.85): a failure), when you submit `done`, or when your time is up (before the end of the schedule, both are failures).
@@ -30,8 +38,8 @@ The same schedule is `spec["command_schedule"]`; every response's `meta["command
`get_obs(seconds=...)`, which holds the current targets). `meta["time_left_s"]` is the service's count of the limit,
from your first request to the service: it never runs out before your time does, so plan by it.
- Nothing to hand in: the service records the episode and grades it. Leave your script at `/app/output/gen.py`.
-- **If the service crashes**, a call fails with a connection error: it restarts and resumes the SAME episode (never a
- reset) in the state after the last call it completed. `r.wait_ready()` waits for it; then check `r.status()`.
+- **If the service crashes**, a call fails with a connection error: it restarts and resumes the SAME episode (a crash
+ is never a reset and costs none) in the state after the last call it completed. `r.wait_ready()` waits for it; then check `r.status()`.
- Nothing is put into your context automatically: images come back as arrays in Python objects.
## Client (`/opt/tools/rcb_client.py`)
@@ -44,7 +52,8 @@ res = r.submit("joint_position", A, dt=0.02) # A: (T, 12) rows dt apart; res.p
obs = r.get_obs(seconds=1.0, video="mp4") # hold the current targets and watch for 1 s (max 30)
r.fk(q); r.ik({"FL": [x, y, z, 1, 0, 0, 0]}, seed=q) # feet in the trunk frame <-> joints (no physics)
r.submit("done", []) # end the episode now
-r.status() # calls, control steps, done, reason, elapsed time
+r.status() # calls, control steps, done, reason, elapsed time, episode, resets_left
+obs = r._post("/reset", {"cameras": ["track"]}, [], 900) # POST /reset: a new episode (the 2.2 client has no reset())
```
Cameras: `track`, `side`, `back`, `top` (they follow the robot); `spec["cameras"]` gives each one's mounting and intrinsics, and every response carries its pose. Proprioception (`spec["proprio_groups"]`): `joint_pos`, `joint_vel`, `joint_effort` (12 each, action
order), `imu_gyro`, `imu_accelerometer`, `imu_gravity`, `imu_linvel` (3 each, IMU frame: x forward, y left, z up),
diff --git a/tasks/mujoco_playground_spot_flat_terrain_joystick-standard/task.toml b/tasks/mujoco_playground_spot_flat_terrain_joystick-standard/task.toml
index 5ffcc74cae..30c0889c0e 100644
--- a/tasks/mujoco_playground_spot_flat_terrain_joystick-standard/task.toml
+++ b/tasks/mujoco_playground_spot_flat_terrain_joystick-standard/task.toml
@@ -16,7 +16,7 @@ artifacts = [{ source = "/var/rcb-monitor/", service = "rcb_monitor" },
[task]
name = "robot-coding-bench/mujoco_playground_spot_flat_terrain_joystick-standard"
-description = "STANDARD MODE (robot as a service, eai-standard/2.2): MuJoCo Playground SpotFlatTerrainJoystick (Boston Dynamics Spot) on the same frozen instance as mujoco_playground_spot_flat_terrain_joystick-privileged: Make Boston Dynamics Spot follow the joystick command schedule on flat ground for the whole 20 s episode without falling over. The agent only gets a client: four cameras that follow the robot (RGB, optional depth), the robot's joints and IMU, joint position chunks at 50 Hz, the feet's kinematics; one episode, no reset, no world state; the sim service records the episode, replays it in fresh processes and grades it."
+description = "STANDARD MODE (robot as a service, eai-standard/2.2): MuJoCo Playground SpotFlatTerrainJoystick (Boston Dynamics Spot) on the same frozen instance as mujoco_playground_spot_flat_terrain_joystick-privileged: Make Boston Dynamics Spot follow the joystick command schedule on flat ground for the whole 20 s episode without falling over. The agent only gets a client: four cameras that follow the robot (RGB, optional depth), the robot's joints and IMU, joint position chunks at 50 Hz, the feet's kinematics; 10 resets (only the last episode is graded), no world state; the sim service records the episode, replays it in fresh processes and grades it."
authors = [{ name = "Changhe Chen" }]
keywords = ["robotics", "mujoco", "mujoco-playground", "boston-dynamics-spot", "quadruped", "locomotion", "joystick", "code-as-policy", "trajectory", "standard-mode"]
@@ -28,6 +28,7 @@ simulator = "mujoco"
robot = "boston-dynamics-spot"
family = "mujoco_playground/SpotFlatTerrainJoystick"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "full"
instance_seed = 0
diff --git a/tasks/mujoco_playground_spot_flat_terrain_joystick-standard/tests/grade.py b/tasks/mujoco_playground_spot_flat_terrain_joystick-standard/tests/grade.py
index 41ce495421..3d4c23ad7e 100644
--- a/tasks/mujoco_playground_spot_flat_terrain_joystick-standard/tests/grade.py
+++ b/tasks/mujoco_playground_spot_flat_terrain_joystick-standard/tests/grade.py
@@ -77,6 +77,9 @@ write(
"sim_error": int(d.get("reason") == "error"),
"sim_startups": int(d.get("sim_startups") or 0),
"resumes": int(d.get("resumes") or 0),
+ "resets": int(d.get("resets") or 0), # PROTOCOL §4.5: resets the agent used; only the last episode is graded
+ "reset_budget": int(d.get("reset_budget") or 0),
+ "episodes": int(d.get("episodes") or 1),
"resume_state_mismatch": int(d.get("resume_state_mismatch") or 0),
"resume_max_joint_dev": num("resume_max_joint_dev"),
},
diff --git a/tasks/mujoco_playground_spot_getup-privileged/README.md b/tasks/mujoco_playground_spot_getup-privileged/README.md
index c1c3f00ce9..161a30dbe3 100644
--- a/tasks/mujoco_playground_spot_getup-privileged/README.md
+++ b/tasks/mujoco_playground_spot_getup-privileged/README.md
@@ -53,3 +53,11 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
### Provenance
Generated by `scripts/mujoco-playground/generate.py` from `scripts/mujoco-playground/instances/SpotGetup.json`
(`make_instances.py`, `select_instances.py`) and `scripts/mujoco-playground/oracles/SpotGetup.npy` (`oracle.py`).
+
+### Change Log
+- 2026-10-14 (review of board 0105, B1-m2): `tests/grade.py` only. A replay process that died or ran out of time, a
+ rebuilt start that does not match the instance's hash (`instance_mismatch`) and an exception of the verifier itself
+ now write `infra_error = 1` beside `grader_error = 1`: no verdict on the agent. Before, they wrote `grader_error = 1`
+ alone, which counted as the agent's 0. Every `reward.json` now carries `infra_error`. The scene, the instance, the
+ pass rule and the reference are unchanged, so the task keeps its name; a past 0 with `grader_error = 1` was an
+ infrastructure failure.
diff --git a/tasks/mujoco_playground_spot_getup-privileged/task.toml b/tasks/mujoco_playground_spot_getup-privileged/task.toml
index 434b829d74..97775aeccb 100644
--- a/tasks/mujoco_playground_spot_getup-privileged/task.toml
+++ b/tasks/mujoco_playground_spot_getup-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "mujoco"
robot = "boston-dynamics-spot"
family = "mujoco_playground/SpotGetup"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 1 # the PRNGKey of Playground's reset (and, for the joystick tasks, its command process)
expert_time_estimate_hours = 2.0
diff --git a/tasks/mujoco_playground_spot_getup-privileged/tests/grade.py b/tasks/mujoco_playground_spot_getup-privileged/tests/grade.py
index c75c326868..1667d090c8 100644
--- a/tasks/mujoco_playground_spot_getup-privileged/tests/grade.py
+++ b/tasks/mujoco_playground_spot_getup-privileged/tests/grade.py
@@ -6,6 +6,11 @@ Two replays, each in its own fresh process (`rcb_quadruped.py replay`: the froze
task's success check on the whole trajectory). success = both replays succeed and end in the same state. Writes
/logs/verifier/reward.txt (1/0) and reward.json (numbers only); then a third replay renders replay.mp4, after the
verdict is on disk (it never changes it).
+
+Failures that are not the agent's (docs/PROTOCOL.md: infra vs agent failure) carry `grader_error` = 1 and `infra_error`
+= 1, so they are no verdict on the agent: a replay process that died or timed out, an image that no longer rebuilds the
+frozen start (`instance_mismatch`), and any exception of the verifier itself. A file the agent got wrong is the agent's
+0 (`bad_trajectory`, infra_error 0). Every reward.json carries success, deterministic, grader_error and infra_error.
"""
import argparse
import json
@@ -22,8 +27,12 @@ TOOL = os.path.join(os.path.dirname(os.path.abspath(__file__)), "rcb_quadruped.p
REPLAY_TIMEOUT_S = 300 # a 1000-step replay takes a few seconds; the video replay renders 4 views every 2nd step
+BASE = {"success": 0, "deterministic": 0, "grader_error": 0, "infra_error": 0}
+
+
def write(ok: bool, out: dict) -> None:
os.makedirs(LOG_DIR, exist_ok=True)
+ out = {**BASE, **out}
open(os.path.join(LOG_DIR, "reward.txt"), "w").write("1\n" if ok else "0\n")
json.dump(out, open(os.path.join(LOG_DIR, "reward.json"), "w"))
print(json.dumps(out), flush=True)
@@ -96,11 +105,11 @@ def main() -> int:
"n_actions": int(acts.shape[0]) if acts.ndim == 2 else 0, "grader_error": 0, **cost})
return 0
runs = [replay(a.instance, a.traj, None), replay(a.instance, a.traj, None)]
- if any("error" in r and not r.get("bad_trajectory") for r in runs):
- write(False, {"success": 0, "deterministic": 0, "grader_error": 1, **cost})
+ if any("error" in r and not r.get("bad_trajectory") for r in runs): # the replay process died / timed out
+ write(False, {"success": 0, "deterministic": 0, "grader_error": 1, "infra_error": 1, "replay_failed": 1, **cost})
return 0
if not all(r.get("instance_ok") for r in runs): # the image no longer builds the frozen start: not the agent's 0
- write(False, {"success": 0, "deterministic": 0, "grader_error": 1, "instance_mismatch": 1, **cost})
+ write(False, {"success": 0, "deterministic": 0, "grader_error": 1, "infra_error": 1, "instance_mismatch": 1, **cost})
return 0
all_success = all(bool(r.get("success")) for r in runs)
deterministic = runs[0].get("state_sha") == runs[1].get("state_sha")
@@ -116,7 +125,7 @@ def main() -> int:
write(ok, out) # the verdict, before any video
except Exception: # noqa: BLE001
traceback.print_exc(file=sys.stderr)
- write(False, {"success": 0, "deterministic": 0, "grader_error": 1, **cost})
+ write(False, {"success": 0, "deterministic": 0, "grader_error": 1, "infra_error": 1, **cost})
return 0
try: # the video, for humans: it never changes the verdict
if replay(a.instance, a.traj, os.path.join(LOG_DIR, "replay.mp4")).get("video"):
diff --git a/tasks/mujoco_playground_spot_getup-standard/README.md b/tasks/mujoco_playground_spot_getup-standard/README.md
index f383a9de59..c34c9654be 100644
--- a/tasks/mujoco_playground_spot_getup-standard/README.md
+++ b/tasks/mujoco_playground_spot_getup-standard/README.md
@@ -13,14 +13,14 @@
## Task
MuJoCo Playground `SpotGetup` (Boston Dynamics Spot) on the same frozen instance as `mujoco_playground_spot_getup-privileged`: Boston Dynamics Spot has fallen over. Make it get up and stand upright at its standing height.
-It starts lying with its trunk tilted 128.4 deg (Playground's random drop, then a 0.5 s settle) and has 6 s to stand upright at its standing height. Standard mode (eai-standard/2.2): the agent drives the robot through a client in one episode, with only
-the robot's own sensors and four cameras that follow it.
+It starts lying with its trunk tilted 128.4 deg (Playground's random drop, then a 0.5 s settle) and has 6 s to stand upright at its standing height. Standard mode (eai-standard/2.2): the agent drives the robot through a client, with 10 resets (only
+the last episode is graded), the robot's own sensors and the cameras that follow it.
## Difficulty
-Same instance and success check as `mujoco_playground_spot_getup-privileged`. In standard mode the agent loses resets, the simulator and
+Same instance and success check as `mujoco_playground_spot_getup-privileged`. In standard mode the agent loses free resets, the simulator and
the world state: it sees camera images and proprioception (joints, IMU, foot contacts), acts in joint position chunks,
-and gets one episode, so it has to close the loop on the robot's own sensors and recover from its own mistakes without
-starting over.
+and gets 10 resets (`RCB_RESET_BUDGET`, the compose file's default; PROTOCOL §4.5), so it has to close the loop
+on the robot's own sensors; only the last episode is graded.
## Reference solution
`solution/oracle_replay.py` streams the privileged reference's rows (`solution/oracle.npy`, 63 rows, float32-exact)
@@ -60,3 +60,4 @@ Generated by `scripts/mujoco-playground/generate.py`, as its privileged twin.
- 2026-10-10 (review 0067): a `final.json` that is missing, unreadable or was never finalized by the collect hook is graded `infra_error = 1` (no verdict on the agent); before, a missing one was a plain 0 and an unfinalized one was graded as it stood. The collect hook writes `final.json` atomically. A request field of the wrong type (`dt` or `seconds` null, `cameras` not a list of names, a `note` that is not a string) is answered 400 before any physics instead of ending the episode with a 500. `meta["time_left_s"]` counts from the agent's first call, not from the service's start.
- 2026-10-10 (review round 2, board 0067): the client only (`environment/rcb_client.py`). `Robot()` waits up to 3600 s for the service (was 600 s) and prints its progress to stderr once a minute, so the first `Robot()` outlasts the simulator's start (no effect on this benchmark, whose service is up within a minute); `cameras="name"` (a string) is one camera, no longer its characters. The service, the collect hook, the grader, the scene and the pass rule are unchanged.
- 2026-10-10 (review round 3, board 0067): the shared service layer and the client only (`environment/sim/rcb_limited.py`, `environment/rcb_client.py`). A camera named more than once in `cameras` is rendered and returned once (N copies of one name made the service build N x N images within the frame budget, which could exhaust its memory). A service that has given up on the episode (it could not start it, or could not resume it after its crashes) answers every request with 409 and the reason, and `Robot()` / `wait_ready()` raise `EpisodeOver` at once; before, nothing answered and they waited out their timeout (up to an hour). The adapter, the collect hook, the grader, the scene and the pass rule are unchanged.
+- 2026-10-14 (owner decision C1, board 0113): 10 resets (PROTOCOL §4.5). The service (`environment/sim/server.py`) gained the optional reset capability: `RCB_RESET_BUDGET` (the compose file's default 10; 0 restores the old single episode), `POST /reset` abandons the running or ended episode and starts a new one from the same frozen start (no physics), refused (409) after a success, once the agent's time is up and once the budget is spent; only the last episode is graded (`trajectory.npz` holds its rows, replayed twice from the frozen start as before); `/status` and every `meta` add `episode`, `resets_left` and `control_steps_total`; `final.json` and `reward.json` add `resets`, `reset_budget` and `episodes`. A crash resumes the current episode (the journal is replayed with the committed resets). The instruction's rules, the client example (`POST /reset` through `Robot._post`) and this README say so. The scene, the instance, the success check and the grader's pass rule are unchanged; results recorded before this change ran with one episode and are reported separately.
diff --git a/tasks/mujoco_playground_spot_getup-standard/environment/docker-compose.yaml b/tasks/mujoco_playground_spot_getup-standard/environment/docker-compose.yaml
index bc0c363241..1226b8c3a6 100644
--- a/tasks/mujoco_playground_spot_getup-standard/environment/docker-compose.yaml
+++ b/tasks/mujoco_playground_spot_getup-standard/environment/docker-compose.yaml
@@ -28,6 +28,7 @@ services:
RCB_STARTUP_TIMEOUT_S: "300" # the scene builds in about a second
PYTHONPATH: /srv
RCB_SESSION_LOG: "" # the sidecar's own Session must not write the agent-side interaction log
+ RCB_RESET_BUDGET: "${RCB_RESET_BUDGET:-10}" # Playground quadrupeds standard: 10 resets (PROTOCOL §4.5); 0: one episode
# >>> rcb_monitor (resource limits monitor, docs/PROTOCOL.md §2.10; generated by scripts/rcb_monitor/inject.py)
rcb_monitor:
image: ghcr.io/mll-lab-nu/eai-monitor:0.1.0@sha256:f2c910f53f79c2bfa4a71298b10c78383cae3afc40720ca562cc845b8d20fb0c
diff --git a/tasks/mujoco_playground_spot_getup-standard/environment/sim/server.py b/tasks/mujoco_playground_spot_getup-standard/environment/sim/server.py
index 108eca3cae..cad8920259 100644
--- a/tasks/mujoco_playground_spot_getup-standard/environment/sim/server.py
+++ b/tasks/mujoco_playground_spot_getup-standard/environment/sim/server.py
@@ -1,16 +1,20 @@
"""eai-standard/2.2 sim service for MuJoCo Playground's quadruped tasks (docs/STANDARD_MODE.md): the adapter of the
shared layer rcb_limited.py; the simulator is rcb_quadruped.Session (plain MuJoCo on the CPU).
-HTTP surface: GET /spec, GET /status, POST /observe, POST /submit, POST /kinematics. Every call is logged to
-/data/episode.jsonl.
+HTTP surface: GET /spec, GET /status, POST /observe, POST /submit, POST /kinematics, and POST /reset when the run grants
+resets (RCB_RESET_BUDGET, the optional reset capability of docs/STANDARD_MODE.md / PROTOCOL §4.5; the quadrupeds'
+standard tasks default to 10 in their compose file, owner 2026-10-14). Every call is logged to /data/episode.jsonl.
The episode is the frozen instance's: one control step = the instance's control period (Go1, Spot: 1/50 s; Barkour:
1/100 s). A joystick episode ends after the whole command schedule
(the horizon), when the robot meets Playground's termination rule (it falls over: a failure), when the agent submits
`done`, or when its time is over; success = the task's check on the steps executed. A getup episode ends the moment the
-check holds (success), after the horizon, on `done` or on timeout.
+check holds (success), after the horizon, on `done` or on timeout. With a reset budget, POST /reset abandons the running
+(or ended) episode and starts a new one from the frozen instance; only the last episode is graded, and a success ends
+the trial.
Crash recovery (rcb_limited.Episode, resume class `exact`: MuJoCo on the CPU repeats bit for bit in a fresh process):
-every executed row is journaled; after a crash the service rebuilds the episode and replays the journal. When the
+every executed row is journaled; after a crash the service rebuilds the episode and replays the journal, resetting where
+the committed resets did (their journal rows are checkpointed, `persist`). When the
episode ends the service writes /data/trajectory.npz + /data/episode.json; the collect hook (finalize.sh) makes it exit
and replays the trajectory in two fresh processes."""
@@ -31,18 +35,21 @@ INSTANCE = os.environ.get("RCB_INSTANCE", os.path.join(os.path.dirname(os.path.a
WALL_CLOCK_S = float(os.environ.get("RCB_WALL_CLOCK_S", "3600"))
MAX_CHUNK_S = float(os.environ.get("RCB_MAX_CHUNK_S", "10"))
MAX_WATCH_S = float(os.environ.get("RCB_MAX_WATCH_S", "30"))
+RESET_BUDGET = max(0, int(os.environ.get("RCB_RESET_BUDGET") or 0)) # the compose file's default: 10 (PROTOCOL §4.5)
CAM_W, CAM_H = 320, 240
LOCK = threading.Lock()
PID_FILE = os.path.join(DATA, "server.pid")
E = {"ready": False, "step": 0, "calls": 0, "done": False, "reason": None, "success": False, "t0": time.time(), "error": None,
- "t_first": None} # t_first: the agent's first call (its clock starts there; kept through a crash by the checkpoints)
+ "t_first": None, # t_first: the agent's first call (its clock starts there; kept through a crash by the checkpoints)
+ "resets": 0, # resets done (the episode is resets + 1)
+ "reset_at": []} # the journal row at which each reset happened: a resume resets there again
os.environ["RCB_SESSION_LOG"] = ""
os.environ["RCB_AGENT_TRACE_DIR"] = ""
import rcb_quadruped as Q # noqa: E402 (first: it picks the rendering backend before mujoco loads)
import mujoco # noqa: E402
-EP = L.Episode(DATA, dim=Q.N, resume="exact", persist=("t_first",))
+EP = L.Episode(DATA, dim=Q.N, resume="exact", persist=("t_first", "resets", "reset_at"))
EP.startup() # start count, back-off after failed starts, the startup watchdog until scene_built()
S = Q.Session(INSTANCE)
ROWS: list = [] # every executed row (a resume replays them first): trajectory.npz
@@ -153,13 +160,15 @@ def schedule_text() -> list:
"wz_radps": round(c["wz"], 4)} for c in S.inst["commands"]]
+ONE = "one episode, no reset" if not RESET_BUDGET else (
+ f"an episode (with up to {RESET_BUDGET} resets: POST /reset starts a new one; only the last is graded)")
TASK_EPISODE = {
- "joystick": f"one episode, no reset: it ends after {S.horizon} control steps (the {S.horizon * DT:g} s command "
+ "joystick": f"{ONE}: it ends after {S.horizon} control steps (the {S.horizon * DT:g} s command "
f"schedule), when the robot meets its termination rule ({S.spec.get('term_text')}: a failure), when you "
"submit `done`, or when your time is over",
- "getup": f"one episode, no reset: it ends the moment the robot has stood upright for {Q.HOLD_STEPS} control steps in a "
+ "getup": f"{ONE}: it ends the moment the robot has stood upright for {Q.HOLD_STEPS} control steps in a "
f"row (success), after {S.horizon} control steps, when you submit `done`, or when your time is over",
- "stand_on_two": f"one episode, no reset: it ends the moment the pose has been held for {Q.POSE_HOLD_STEPS} control steps "
+ "stand_on_two": f"{ONE}: it ends the moment the pose has been held for {Q.POSE_HOLD_STEPS} control steps "
"in a row (success), when the robot meets the termination rule (a failure: its trunk's up axis below "
f"-0.25, or {'a leg' if S.spec.get('hind_leg_contact') else 'a front leg'}'s thigh, calf or hip on the "
f"floor), after {S.horizon} control steps, when you "
@@ -204,7 +213,7 @@ SPEC = {
"world": {"frame": "z up; the robot starts where the episode puts it (no world coordinates are given)", "up": "+z",
"units": {"length": "m", "angle": "rad", "time": "s"}, "quaternion_order": "wxyz",
"camera_axes": "intrinsic_cv / extrinsic_cv: OpenCV camera axes (x right, y down, z forward); cam2world_gl: OpenGL axes"},
- "budget": {"reset": 0, "steps": None, "wall_clock_s": WALL_CLOCK_S}, # + "clock" (rcb_limited.CLOCK_NOTE): what time_left_s counts from
+ "budget": {"reset": RESET_BUDGET, "steps": None, "wall_clock_s": WALL_CLOCK_S}, # + "clock" (rcb_limited.CLOCK_NOTE): what time_left_s counts from
"max_control_steps": S.horizon,
"episode": TASK_EPISODE,
"command_schedule": schedule_text(),
@@ -223,6 +232,15 @@ SPEC = {
"depth": {"modes": list(L.DEPTH_MODES), "default": "none", "kind": "z-depth", "unit": "m", "invalid": 0,
"max_m": L.DEPTH_MAX_M, "frames_encoding": "uint16 mm, zlib", "noise": L.DEPTH_NOISE},
}
+if RESET_BUDGET:
+ SPEC["reset"] = (
+ f"POST /reset (JSON or RCB1 {{cameras, video_fps, depth}}; no physics) abandons the running episode, or follows "
+ f"one that has ended, and starts a new one from the same frozen starting state, {RESET_BUDGET} times in all; it "
+ "returns the new episode's first frames and proprioception. Only the last episode is graded: an episode you "
+ "reset away is never graded, a success ends the trial (409 after it), and when your time is up the episode then "
+ "running is the one graded. The wall clock never restarts. /status and every meta add `episode`, `resets_left` "
+ "and `control_steps_total` (all episodes); `control_steps` counts the current episode's."
+ )
L.check_spec(SPEC) # a spec that misses a 2.2 field or names a forbidden one never starts the episode
EP.scene_built()
@@ -277,6 +295,7 @@ def summary() -> dict:
"reason": E["reason"], "done": E["done"], "env": S.env, "protocol": L.PROTOCOL, "n_actions": len(ROWS),
"fell_at": chk.get("fell_at"), "worst_lin_err": chk.get("worst_lin_err"), "worst_yaw_err": chk.get("worst_yaw_err"),
"tilt_deg": chk.get("tilt_deg"), "imu_height": chk.get("imu_height"), "pose_angle_deg": chk.get("pose_angle_deg"),
+ "episodes": E["resets"] + 1, "resets": E["resets"], "reset_budget": RESET_BUDGET, "control_steps_total": EP.rows,
**EP.final_fields()}
@@ -294,7 +313,8 @@ def finish(reason: str) -> None:
E["success"] = bool(chk["success"])
except Exception: # noqa: BLE001
traceback.print_exc()
- EP.log({"ev": "end", "reason": reason, "step": E["step"], "control_steps": ctrl_steps(), "success": bool(E["success"])})
+ EP.log({"ev": "end", "reason": reason, "step": E["step"], "control_steps": ctrl_steps(), "success": bool(E["success"]),
+ "episode": E["resets"] + 1, "journal_rows": EP.rows})
dump()
@@ -388,7 +408,7 @@ def meta_common(cams) -> dict:
"sim_time_s": round(n * DT, 3), "time_s": round(_agent_time(), 2),
"time_left_s": max(0.0, round(WALL_CLOCK_S - _agent_time(), 1)),
"proprio": {"layout": PROPRIO_LAYOUT, "now": proprio_row().tolist()}, "done": E["done"], "reason": E["reason"],
- "cameras": {c: camera_info(c) for c in cams}}
+ "cameras": {c: camera_info(c) for c in cams}, **reset_counters()}
if S.kind == "joystick":
out["command_now"] = dict(zip(("vx_mps", "vy_mps", "wz_radps"), S.command_now().round(4).tolist()))
return out
@@ -478,10 +498,54 @@ def do_observe(meta: dict):
return out, rec.parts(video if seconds > 0 else ("frames" if video == "mp4" else video), fps)
+def reset_counters() -> dict:
+ if not RESET_BUDGET:
+ return {}
+ left = 0 if (E["success"] or E.get("exiting")) else RESET_BUDGET - E["resets"]
+ return {"episode": E["resets"] + 1, "resets_left": left, "control_steps_total": EP.rows}
+
+
def status() -> dict:
_touch()
return {"step": E["step"], "calls": E["calls"], "control_steps": ctrl_steps(), "done": E["done"], "reason": E["reason"],
- "elapsed_s": round(time.time() - E["t0"], 1)}
+ "elapsed_s": round(time.time() - E["t0"], 1), **reset_counters()}
+
+
+class Refused(Exception):
+ """A /reset the service turns down (409): nothing changes."""
+
+ def __init__(self, body: dict):
+ super().__init__(body.get("error"))
+ self.body = body
+
+
+def do_reset(meta: dict):
+ """POST /reset: a new episode from the frozen instance (the running one, if any, is abandoned, never graded)."""
+ meta = _request(meta)
+ cams, fps = cameras_of(meta), video_fps(meta)
+ L.depth_mode(meta) # a bad depth value is a 400 before anything changes
+ if E["success"]:
+ raise Refused({"error": "trial finished: an episode succeeded", "reason": E["reason"], "resets_left": 0})
+ if E["resets"] >= RESET_BUDGET:
+ raise Refused({"error": "reset budget exhausted", "reason": "reset budget exhausted", "budget": RESET_BUDGET,
+ "resets_left": 0})
+ abandoned = {"episode": E["resets"] + 1, "control_steps": ctrl_steps(), "done": E["done"], "reason": E["reason"]}
+ if not E["done"]: # the abandoned episode's end, for the log (never graded: its rows leave the trajectory)
+ EP.log({"ev": "end", "reason": "reset", "step": E["step"], "control_steps": ctrl_steps(), "success": False,
+ "episode": E["resets"] + 1, "journal_rows": EP.rows})
+ S.reset()
+ ROWS.clear()
+ E["resets"] += 1
+ E["reset_at"] = list(E["reset_at"]) + [EP.rows]
+ E.update(done=False, reason=None, success=False, t_end=None, episode_control_steps=None, dumped=False, check=None)
+ EP._ck = None # the state changed without a journal row: the checkpoint after this call must hash it anew
+ E["calls"] += 1
+ EP.log({"ev": "reset", "call": E["calls"], "episode": E["resets"] + 1, "journal_rows": EP.rows,
+ "abandoned": abandoned, "state_sha": S.state_sha()})
+ rec = recorder(cams, fps, meta)
+ rec.tick(force=True, last=True)
+ out = {**meta_common(cams), "reset_id": E["resets"], "ok": True}
+ return out, rec.parts("frames", fps)
def end_from_signal() -> None:
@@ -498,20 +562,106 @@ def end_from_signal() -> None:
os._exit(0)
+# A resume replays the journal from the frozen instance and resets where the committed resets did: their journal rows
+# are in the last checkpoint (`persist`), the one Episode.begin resumes from.
+_CPS = [e for e in EP.events() if e.get("ev") == "checkpoint"]
+RESUME = {"at": list((_CPS[-1].get("persist") or {}).get("reset_at") or []) if _CPS else [], "i": 0}
+
+
+def _resume_resets() -> None:
+ while RESUME["at"] and int(RESUME["at"][0]) <= RESUME["i"]:
+ RESUME["at"].pop(0)
+ S.reset()
+ ROWS.clear()
+
+
def replay_row(row) -> None:
"""A journaled row during a resume."""
+ _resume_resets()
S.step(row)
ROWS.append(np.asarray(row, dtype=np.float64))
+ RESUME["i"] += 1
-EP.begin(state=E, step=replay_row, state_sha=S.state_sha, proprio=proprio_row, spec=SPEC, on_end=end_from_signal,
- start_event={"env": S.env, "seed": S.inst.get("seed"), "cameras": CAMS})
+def resume_sha() -> str:
+ _resume_resets() # a reset after the last replayed row (or before the first)
+ return S.state_sha()
+
+
+RESUMED = EP.begin(state=E, step=replay_row, state_sha=resume_sha, proprio=proprio_row, spec=SPEC, on_end=end_from_signal,
+ start_event={"env": S.env, "seed": S.inst.get("seed"), "cameras": CAMS})
+if RESUMED and E["resets"]: # Episode.begin looks for an "end" of any episode: only the current one's counts
+ _ends = [e for e in EP.events() if e.get("ev") == "end" and int(e.get("episode") or 1) == E["resets"] + 1
+ and e.get("reason") != "reset" and int(e.get("journal_rows") or 0) <= EP.rows]
+ if _ends:
+ E.update(done=True, reason=_ends[-1].get("reason"), success=bool(_ends[-1].get("success")),
+ t_end=float(_ends[-1]["ts"]), episode_control_steps=int(_ends[-1].get("control_steps") or ctrl_steps()))
+ else:
+ E.update(done=False, reason=None, success=False, t_end=None, episode_control_steps=None)
if E["done"]:
dump() # resumed after the episode had ended: the record is rewritten from the replay
E["ready"] = True
write_final()
open(PID_FILE, "w").write(str(os.getpid()))
print(f"{L.PROTOCOL} MuJoCo Playground quadruped sim ready ({S.env}; startup {EP.n_startups}, resumes {EP.resumes})", flush=True)
-L.serve(L.Service(spec=SPEC, lock=LOCK, state=E, submit=do_submit, observe=do_observe, kinematics=do_kinematics, status=status,
- on_submit_error=lambda exc: finish("error"), log=EP.log, episode=EP, state_sha=S.state_sha,
- proprio=proprio_row), port=int(os.environ.get("RCB_PORT", "8000")))
+
+
+def make_handler(svc: L.Service):
+ """The shared 2.2 handler, plus POST /reset when the run grants resets (the shared layer has no reset)."""
+ base = L.make_handler(svc)
+
+ class H(base):
+ def do_POST(self):
+ if self.path != "/reset" or not RESET_BUDGET:
+ return super().do_POST()
+ try:
+ meta, _ = self._body()
+ except Exception as exc: # noqa: BLE001
+ return self._json(413 if "too large" in str(exc) else 400, {"error": str(exc)[:200]})
+ with LOCK:
+ if E.get("exiting"):
+ return self._json(409, {"error": "trial finished: your time is up", "reason": E.get("reason"),
+ "resets_left": 0})
+ cid = EP.call_begin(self.path, self.connection)
+ out = blobs = err = None
+ try:
+ out, blobs = do_reset(meta)
+ except Refused as exc:
+ err = (409, exc.body)
+ except ValueError as exc:
+ err = (400, {"error": str(exc)})
+ except Exception as exc: # noqa: BLE001
+ traceback.print_exc()
+ E["error"] = repr(exc)
+ EP.log({"ev": "error", "path": self.path, "error": repr(exc)})
+ try:
+ finish("error")
+ except Exception: # noqa: BLE001
+ traceback.print_exc()
+ err = (500, {"error": "simulator error", "detail": repr(exc)[:200]})
+ finally:
+ if out is not None or (err and err[0] == 500):
+ try:
+ EP.checkpoint(E, S.state_sha, proprio_row)
+ except Exception: # noqa: BLE001
+ traceback.print_exc()
+ EP.call_end(cid, self.path)
+ if err is not None:
+ return self._json(*err)
+ return self._rcb1({**out, "stopped": None}, blobs)
+
+ return H
+
+
+def serve(svc: L.Service, port: int) -> None:
+ """L.serve with this service's handler (the same spec defaults and check, and L.bind: a port still held when the
+ service restarts is waited for, not a crash)."""
+ svc.spec.setdefault("on_service_error", L.ON_SERVICE_ERROR)
+ svc.spec.setdefault("max_call_wall_s", EP.max_call_wall_s)
+ L.check_spec(svc.spec)
+ L.bind(port, make_handler(svc)).serve_forever()
+
+
+serve(L.Service(spec=SPEC, lock=LOCK, state=E, submit=do_submit, observe=do_observe, kinematics=do_kinematics, status=status,
+ on_submit_error=lambda exc: finish("error"), log=EP.log, episode=EP, state_sha=S.state_sha,
+ proprio=proprio_row), port=int(os.environ.get("RCB_PORT", "8000")))
diff --git a/tasks/mujoco_playground_spot_getup-standard/instruction.md b/tasks/mujoco_playground_spot_getup-standard/instruction.md
index 3d0c99f0a9..598578b307 100644
--- a/tasks/mujoco_playground_spot_getup-standard/instruction.md
+++ b/tasks/mujoco_playground_spot_getup-standard/instruction.md
@@ -2,15 +2,23 @@
**Task:** Boston Dynamics Spot has fallen over. Make it get up and stand upright at its standing height.
-You control the Boston Dynamics Spot quadruped — **as if it were real hardware** — in one episode of MuJoCo Playground's
+You control the Boston Dynamics Spot quadruped — **as if it were real hardware** — in MuJoCo Playground's
`SpotGetup` environment (plain MuJoCo 3.3.7 with Playground's scene, commit 4057c14). The simulator runs in a service you
-cannot open. You get one client and one episode.
+cannot open. You get one client and up to 11 episodes.
## The episode
The robot was dropped from 1 m in a random orientation with random joint angles (no base velocity) and settled for 0.5 s with its joints held: at the start it lies with its trunk tilted 128.4 deg from upright. You have 300 control steps (6 s).
## Rules (real-robot rules)
-- **No reset.** One episode. You see four cameras that follow the robot (RGB, depth if you ask) and the robot's own
+- **Episodes and resets.** You have **10 resets** (`spec["budget"]["reset"]`, which is authoritative: if it
+ says 0, there is no reset and the one episode is the run). A reset (`POST /reset`, below) puts the robot back on the
+ same frozen starting state and starts a new episode, either abandoning the running one or after one has ended; it
+ takes no simulated time and returns the start's images and proprioception. Once an episode has ended, `submit`
+ raises `EpisodeOver` until you reset. **Only the last episode is graded**: an episode you reset away is never graded,
+ a success ends the run at once (no reset after it), and when your time is up the episode then running is the one
+ graded. The wall clock is the run's and never restarts; `meta["episode"]` and `meta["resets_left"]` count the
+ episodes (`meta["control_steps"]` is the current episode's).
+- **No privileged state.** You see the cameras that follow the robot (RGB, depth if you ask) and the robot's own
sensors: joint positions, velocities and torques, the IMU (gyro, accelerometer, gravity's direction, the trunk's
velocity), foot contacts. No world positions, no success check.
- **The episode ends** the moment the robot has been upright for 25 control steps in a row (success), after 300 control steps (`spec["max_control_steps"]`, counted by `meta["control_steps"]`), when you submit `done`, or when your time is up.
@@ -19,8 +27,8 @@ The robot was dropped from 1 m in a random orientation with random joint angles
`get_obs(seconds=...)`, which holds the current targets). `meta["time_left_s"]` is the service's count of the limit,
from your first request to the service: it never runs out before your time does, so plan by it.
- Nothing to hand in: the service records the episode and grades it. Leave your script at `/app/output/gen.py`.
-- **If the service crashes**, a call fails with a connection error: it restarts and resumes the SAME episode (never a
- reset) in the state after the last call it completed. `r.wait_ready()` waits for it; then check `r.status()`.
+- **If the service crashes**, a call fails with a connection error: it restarts and resumes the SAME episode (a crash
+ is never a reset and costs none) in the state after the last call it completed. `r.wait_ready()` waits for it; then check `r.status()`.
- Nothing is put into your context automatically: images come back as arrays in Python objects.
## Client (`/opt/tools/rcb_client.py`)
@@ -33,7 +41,8 @@ res = r.submit("joint_position", A, dt=0.02) # A: (T, 12) rows dt apart; res.p
obs = r.get_obs(seconds=1.0, video="mp4") # hold the current targets and watch for 1 s (max 30)
r.fk(q); r.ik({"FL": [x, y, z, 1, 0, 0, 0]}, seed=q) # feet in the trunk frame <-> joints (no physics)
r.submit("done", []) # end the episode now
-r.status() # calls, control steps, done, reason, elapsed time
+r.status() # calls, control steps, done, reason, elapsed time, episode, resets_left
+obs = r._post("/reset", {"cameras": ["track"]}, [], 900) # POST /reset: a new episode (the 2.2 client has no reset())
```
Cameras: `track`, `side`, `back`, `top` (they follow the robot); `spec["cameras"]` gives each one's mounting and intrinsics, and every response carries its pose. Proprioception (`spec["proprio_groups"]`): `joint_pos`, `joint_vel`, `joint_effort` (12 each, action
order), `imu_gyro`, `imu_accelerometer`, `imu_gravity`, `imu_linvel` (3 each, IMU frame: x forward, y left, z up),
diff --git a/tasks/mujoco_playground_spot_getup-standard/task.toml b/tasks/mujoco_playground_spot_getup-standard/task.toml
index 36c86f6d10..df8d4e34d4 100644
--- a/tasks/mujoco_playground_spot_getup-standard/task.toml
+++ b/tasks/mujoco_playground_spot_getup-standard/task.toml
@@ -16,7 +16,7 @@ artifacts = [{ source = "/var/rcb-monitor/", service = "rcb_monitor" },
[task]
name = "robot-coding-bench/mujoco_playground_spot_getup-standard"
-description = "STANDARD MODE (robot as a service, eai-standard/2.2): MuJoCo Playground SpotGetup (Boston Dynamics Spot) on the same frozen instance as mujoco_playground_spot_getup-privileged: Boston Dynamics Spot has fallen over. Make it get up and stand upright at its standing height. The agent only gets a client: four cameras that follow the robot (RGB, optional depth), the robot's joints and IMU, joint position chunks at 50 Hz, the feet's kinematics; one episode, no reset, no world state; the sim service records the episode, replays it in fresh processes and grades it."
+description = "STANDARD MODE (robot as a service, eai-standard/2.2): MuJoCo Playground SpotGetup (Boston Dynamics Spot) on the same frozen instance as mujoco_playground_spot_getup-privileged: Boston Dynamics Spot has fallen over. Make it get up and stand upright at its standing height. The agent only gets a client: four cameras that follow the robot (RGB, optional depth), the robot's joints and IMU, joint position chunks at 50 Hz, the feet's kinematics; 10 resets (only the last episode is graded), no world state; the sim service records the episode, replays it in fresh processes and grades it."
authors = [{ name = "Changhe Chen" }]
keywords = ["robotics", "mujoco", "mujoco-playground", "boston-dynamics-spot", "quadruped", "locomotion", "getup", "code-as-policy", "trajectory", "standard-mode"]
@@ -28,6 +28,7 @@ simulator = "mujoco"
robot = "boston-dynamics-spot"
family = "mujoco_playground/SpotGetup"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "full"
instance_seed = 1
diff --git a/tasks/mujoco_playground_spot_getup-standard/tests/grade.py b/tasks/mujoco_playground_spot_getup-standard/tests/grade.py
index 41ce495421..3d4c23ad7e 100644
--- a/tasks/mujoco_playground_spot_getup-standard/tests/grade.py
+++ b/tasks/mujoco_playground_spot_getup-standard/tests/grade.py
@@ -77,6 +77,9 @@ write(
"sim_error": int(d.get("reason") == "error"),
"sim_startups": int(d.get("sim_startups") or 0),
"resumes": int(d.get("resumes") or 0),
+ "resets": int(d.get("resets") or 0), # PROTOCOL §4.5: resets the agent used; only the last episode is graded
+ "reset_budget": int(d.get("reset_budget") or 0),
+ "episodes": int(d.get("episodes") or 1),
"resume_state_mismatch": int(d.get("resume_state_mismatch") or 0),
"resume_max_joint_dev": num("resume_max_joint_dev"),
},
diff --git a/tasks/mujoco_playground_spot_joystick_gait_tracking-privileged/README.md b/tasks/mujoco_playground_spot_joystick_gait_tracking-privileged/README.md
index 0a994d0bae..926884f613 100644
--- a/tasks/mujoco_playground_spot_joystick_gait_tracking-privileged/README.md
+++ b/tasks/mujoco_playground_spot_joystick_gait_tracking-privileged/README.md
@@ -53,3 +53,11 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
### Provenance
Generated by `scripts/mujoco-playground/generate.py` from `scripts/mujoco-playground/instances/SpotJoystickGaitTracking.json`
(`make_instances.py`, `select_instances.py`) and `scripts/mujoco-playground/oracles/SpotJoystickGaitTracking.npy` (`oracle.py`).
+
+### Change Log
+- 2026-10-14 (review of board 0105, B1-m2): `tests/grade.py` only. A replay process that died or ran out of time, a
+ rebuilt start that does not match the instance's hash (`instance_mismatch`) and an exception of the verifier itself
+ now write `infra_error = 1` beside `grader_error = 1`: no verdict on the agent. Before, they wrote `grader_error = 1`
+ alone, which counted as the agent's 0. Every `reward.json` now carries `infra_error`. The scene, the instance, the
+ pass rule and the reference are unchanged, so the task keeps its name; a past 0 with `grader_error = 1` was an
+ infrastructure failure.
diff --git a/tasks/mujoco_playground_spot_joystick_gait_tracking-privileged/task.toml b/tasks/mujoco_playground_spot_joystick_gait_tracking-privileged/task.toml
index a6bd4a27eb..8c8858c691 100644
--- a/tasks/mujoco_playground_spot_joystick_gait_tracking-privileged/task.toml
+++ b/tasks/mujoco_playground_spot_joystick_gait_tracking-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "mujoco"
robot = "boston-dynamics-spot"
family = "mujoco_playground/SpotJoystickGaitTracking"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 0 # the PRNGKey of Playground's reset (and, for the joystick tasks, its command process)
expert_time_estimate_hours = 3.0
diff --git a/tasks/mujoco_playground_spot_joystick_gait_tracking-privileged/tests/grade.py b/tasks/mujoco_playground_spot_joystick_gait_tracking-privileged/tests/grade.py
index c75c326868..1667d090c8 100644
--- a/tasks/mujoco_playground_spot_joystick_gait_tracking-privileged/tests/grade.py
+++ b/tasks/mujoco_playground_spot_joystick_gait_tracking-privileged/tests/grade.py
@@ -6,6 +6,11 @@ Two replays, each in its own fresh process (`rcb_quadruped.py replay`: the froze
task's success check on the whole trajectory). success = both replays succeed and end in the same state. Writes
/logs/verifier/reward.txt (1/0) and reward.json (numbers only); then a third replay renders replay.mp4, after the
verdict is on disk (it never changes it).
+
+Failures that are not the agent's (docs/PROTOCOL.md: infra vs agent failure) carry `grader_error` = 1 and `infra_error`
+= 1, so they are no verdict on the agent: a replay process that died or timed out, an image that no longer rebuilds the
+frozen start (`instance_mismatch`), and any exception of the verifier itself. A file the agent got wrong is the agent's
+0 (`bad_trajectory`, infra_error 0). Every reward.json carries success, deterministic, grader_error and infra_error.
"""
import argparse
import json
@@ -22,8 +27,12 @@ TOOL = os.path.join(os.path.dirname(os.path.abspath(__file__)), "rcb_quadruped.p
REPLAY_TIMEOUT_S = 300 # a 1000-step replay takes a few seconds; the video replay renders 4 views every 2nd step
+BASE = {"success": 0, "deterministic": 0, "grader_error": 0, "infra_error": 0}
+
+
def write(ok: bool, out: dict) -> None:
os.makedirs(LOG_DIR, exist_ok=True)
+ out = {**BASE, **out}
open(os.path.join(LOG_DIR, "reward.txt"), "w").write("1\n" if ok else "0\n")
json.dump(out, open(os.path.join(LOG_DIR, "reward.json"), "w"))
print(json.dumps(out), flush=True)
@@ -96,11 +105,11 @@ def main() -> int:
"n_actions": int(acts.shape[0]) if acts.ndim == 2 else 0, "grader_error": 0, **cost})
return 0
runs = [replay(a.instance, a.traj, None), replay(a.instance, a.traj, None)]
- if any("error" in r and not r.get("bad_trajectory") for r in runs):
- write(False, {"success": 0, "deterministic": 0, "grader_error": 1, **cost})
+ if any("error" in r and not r.get("bad_trajectory") for r in runs): # the replay process died / timed out
+ write(False, {"success": 0, "deterministic": 0, "grader_error": 1, "infra_error": 1, "replay_failed": 1, **cost})
return 0
if not all(r.get("instance_ok") for r in runs): # the image no longer builds the frozen start: not the agent's 0
- write(False, {"success": 0, "deterministic": 0, "grader_error": 1, "instance_mismatch": 1, **cost})
+ write(False, {"success": 0, "deterministic": 0, "grader_error": 1, "infra_error": 1, "instance_mismatch": 1, **cost})
return 0
all_success = all(bool(r.get("success")) for r in runs)
deterministic = runs[0].get("state_sha") == runs[1].get("state_sha")
@@ -116,7 +125,7 @@ def main() -> int:
write(ok, out) # the verdict, before any video
except Exception: # noqa: BLE001
traceback.print_exc(file=sys.stderr)
- write(False, {"success": 0, "deterministic": 0, "grader_error": 1, **cost})
+ write(False, {"success": 0, "deterministic": 0, "grader_error": 1, "infra_error": 1, **cost})
return 0
try: # the video, for humans: it never changes the verdict
if replay(a.instance, a.traj, os.path.join(LOG_DIR, "replay.mp4")).get("video"):
diff --git a/tasks/mujoco_playground_spot_joystick_gait_tracking-standard/README.md b/tasks/mujoco_playground_spot_joystick_gait_tracking-standard/README.md
index a4585784ba..997a748d81 100644
--- a/tasks/mujoco_playground_spot_joystick_gait_tracking-standard/README.md
+++ b/tasks/mujoco_playground_spot_joystick_gait_tracking-standard/README.md
@@ -13,14 +13,14 @@
## Task
MuJoCo Playground `SpotJoystickGaitTracking` (Boston Dynamics Spot) on the same frozen instance as `mujoco_playground_spot_joystick_gait_tracking-privileged`: Make Boston Dynamics Spot follow the joystick command schedule for the whole 20 s episode with the commanded gait (the feet's swing pattern, frequency and height) without falling over.
-For 20 s the robot gets a joystick command schedule (5 segments: planar velocity and yaw rate in its own frame) and must track it without falling over. Its feet must follow a pace gait. Standard mode (eai-standard/2.2): the agent drives the robot through a client in one episode, with only
-the robot's own sensors and four cameras that follow it.
+For 20 s the robot gets a joystick command schedule (5 segments: planar velocity and yaw rate in its own frame) and must track it without falling over. Its feet must follow a pace gait. Standard mode (eai-standard/2.2): the agent drives the robot through a client, with 10 resets (only
+the last episode is graded), the robot's own sensors and the cameras that follow it.
## Difficulty
-Same instance and success check as `mujoco_playground_spot_joystick_gait_tracking-privileged`. In standard mode the agent loses resets, the simulator and
+Same instance and success check as `mujoco_playground_spot_joystick_gait_tracking-privileged`. In standard mode the agent loses free resets, the simulator and
the world state: it sees camera images and proprioception (joints, IMU, foot contacts), acts in joint position chunks,
-and gets one episode, so it has to close the loop on the robot's own sensors and recover from its own mistakes without
-starting over.
+and gets 10 resets (`RCB_RESET_BUDGET`, the compose file's default; PROTOCOL §4.5), so it has to close the loop
+on the robot's own sensors; only the last episode is graded.
## Reference solution
`solution/oracle_replay.py` streams the privileged reference's rows (`solution/oracle.npy`, 1000 rows, float32-exact)
@@ -60,3 +60,4 @@ Generated by `scripts/mujoco-playground/generate.py`, as its privileged twin.
- 2026-10-10 (review 0067): a `final.json` that is missing, unreadable or was never finalized by the collect hook is graded `infra_error = 1` (no verdict on the agent); before, a missing one was a plain 0 and an unfinalized one was graded as it stood. The collect hook writes `final.json` atomically. A request field of the wrong type (`dt` or `seconds` null, `cameras` not a list of names, a `note` that is not a string) is answered 400 before any physics instead of ending the episode with a 500. `meta["time_left_s"]` counts from the agent's first call, not from the service's start.
- 2026-10-10 (review round 2, board 0067): the client only (`environment/rcb_client.py`). `Robot()` waits up to 3600 s for the service (was 600 s) and prints its progress to stderr once a minute, so the first `Robot()` outlasts the simulator's start (no effect on this benchmark, whose service is up within a minute); `cameras="name"` (a string) is one camera, no longer its characters. The service, the collect hook, the grader, the scene and the pass rule are unchanged.
- 2026-10-10 (review round 3, board 0067): the shared service layer and the client only (`environment/sim/rcb_limited.py`, `environment/rcb_client.py`). A camera named more than once in `cameras` is rendered and returned once (N copies of one name made the service build N x N images within the frame budget, which could exhaust its memory). A service that has given up on the episode (it could not start it, or could not resume it after its crashes) answers every request with 409 and the reason, and `Robot()` / `wait_ready()` raise `EpisodeOver` at once; before, nothing answered and they waited out their timeout (up to an hour). The adapter, the collect hook, the grader, the scene and the pass rule are unchanged.
+- 2026-10-14 (owner decision C1, board 0113): 10 resets (PROTOCOL §4.5). The service (`environment/sim/server.py`) gained the optional reset capability: `RCB_RESET_BUDGET` (the compose file's default 10; 0 restores the old single episode), `POST /reset` abandons the running or ended episode and starts a new one from the same frozen start (no physics), refused (409) after a success, once the agent's time is up and once the budget is spent; only the last episode is graded (`trajectory.npz` holds its rows, replayed twice from the frozen start as before); `/status` and every `meta` add `episode`, `resets_left` and `control_steps_total`; `final.json` and `reward.json` add `resets`, `reset_budget` and `episodes`. A crash resumes the current episode (the journal is replayed with the committed resets). The instruction's rules, the client example (`POST /reset` through `Robot._post`) and this README say so. The scene, the instance, the success check and the grader's pass rule are unchanged; results recorded before this change ran with one episode and are reported separately.
diff --git a/tasks/mujoco_playground_spot_joystick_gait_tracking-standard/environment/docker-compose.yaml b/tasks/mujoco_playground_spot_joystick_gait_tracking-standard/environment/docker-compose.yaml
index bc0c363241..1226b8c3a6 100644
--- a/tasks/mujoco_playground_spot_joystick_gait_tracking-standard/environment/docker-compose.yaml
+++ b/tasks/mujoco_playground_spot_joystick_gait_tracking-standard/environment/docker-compose.yaml
@@ -28,6 +28,7 @@ services:
RCB_STARTUP_TIMEOUT_S: "300" # the scene builds in about a second
PYTHONPATH: /srv
RCB_SESSION_LOG: "" # the sidecar's own Session must not write the agent-side interaction log
+ RCB_RESET_BUDGET: "${RCB_RESET_BUDGET:-10}" # Playground quadrupeds standard: 10 resets (PROTOCOL §4.5); 0: one episode
# >>> rcb_monitor (resource limits monitor, docs/PROTOCOL.md §2.10; generated by scripts/rcb_monitor/inject.py)
rcb_monitor:
image: ghcr.io/mll-lab-nu/eai-monitor:0.1.0@sha256:f2c910f53f79c2bfa4a71298b10c78383cae3afc40720ca562cc845b8d20fb0c
diff --git a/tasks/mujoco_playground_spot_joystick_gait_tracking-standard/environment/sim/server.py b/tasks/mujoco_playground_spot_joystick_gait_tracking-standard/environment/sim/server.py
index 108eca3cae..cad8920259 100644
--- a/tasks/mujoco_playground_spot_joystick_gait_tracking-standard/environment/sim/server.py
+++ b/tasks/mujoco_playground_spot_joystick_gait_tracking-standard/environment/sim/server.py
@@ -1,16 +1,20 @@
"""eai-standard/2.2 sim service for MuJoCo Playground's quadruped tasks (docs/STANDARD_MODE.md): the adapter of the
shared layer rcb_limited.py; the simulator is rcb_quadruped.Session (plain MuJoCo on the CPU).
-HTTP surface: GET /spec, GET /status, POST /observe, POST /submit, POST /kinematics. Every call is logged to
-/data/episode.jsonl.
+HTTP surface: GET /spec, GET /status, POST /observe, POST /submit, POST /kinematics, and POST /reset when the run grants
+resets (RCB_RESET_BUDGET, the optional reset capability of docs/STANDARD_MODE.md / PROTOCOL §4.5; the quadrupeds'
+standard tasks default to 10 in their compose file, owner 2026-10-14). Every call is logged to /data/episode.jsonl.
The episode is the frozen instance's: one control step = the instance's control period (Go1, Spot: 1/50 s; Barkour:
1/100 s). A joystick episode ends after the whole command schedule
(the horizon), when the robot meets Playground's termination rule (it falls over: a failure), when the agent submits
`done`, or when its time is over; success = the task's check on the steps executed. A getup episode ends the moment the
-check holds (success), after the horizon, on `done` or on timeout.
+check holds (success), after the horizon, on `done` or on timeout. With a reset budget, POST /reset abandons the running
+(or ended) episode and starts a new one from the frozen instance; only the last episode is graded, and a success ends
+the trial.
Crash recovery (rcb_limited.Episode, resume class `exact`: MuJoCo on the CPU repeats bit for bit in a fresh process):
-every executed row is journaled; after a crash the service rebuilds the episode and replays the journal. When the
+every executed row is journaled; after a crash the service rebuilds the episode and replays the journal, resetting where
+the committed resets did (their journal rows are checkpointed, `persist`). When the
episode ends the service writes /data/trajectory.npz + /data/episode.json; the collect hook (finalize.sh) makes it exit
and replays the trajectory in two fresh processes."""
@@ -31,18 +35,21 @@ INSTANCE = os.environ.get("RCB_INSTANCE", os.path.join(os.path.dirname(os.path.a
WALL_CLOCK_S = float(os.environ.get("RCB_WALL_CLOCK_S", "3600"))
MAX_CHUNK_S = float(os.environ.get("RCB_MAX_CHUNK_S", "10"))
MAX_WATCH_S = float(os.environ.get("RCB_MAX_WATCH_S", "30"))
+RESET_BUDGET = max(0, int(os.environ.get("RCB_RESET_BUDGET") or 0)) # the compose file's default: 10 (PROTOCOL §4.5)
CAM_W, CAM_H = 320, 240
LOCK = threading.Lock()
PID_FILE = os.path.join(DATA, "server.pid")
E = {"ready": False, "step": 0, "calls": 0, "done": False, "reason": None, "success": False, "t0": time.time(), "error": None,
- "t_first": None} # t_first: the agent's first call (its clock starts there; kept through a crash by the checkpoints)
+ "t_first": None, # t_first: the agent's first call (its clock starts there; kept through a crash by the checkpoints)
+ "resets": 0, # resets done (the episode is resets + 1)
+ "reset_at": []} # the journal row at which each reset happened: a resume resets there again
os.environ["RCB_SESSION_LOG"] = ""
os.environ["RCB_AGENT_TRACE_DIR"] = ""
import rcb_quadruped as Q # noqa: E402 (first: it picks the rendering backend before mujoco loads)
import mujoco # noqa: E402
-EP = L.Episode(DATA, dim=Q.N, resume="exact", persist=("t_first",))
+EP = L.Episode(DATA, dim=Q.N, resume="exact", persist=("t_first", "resets", "reset_at"))
EP.startup() # start count, back-off after failed starts, the startup watchdog until scene_built()
S = Q.Session(INSTANCE)
ROWS: list = [] # every executed row (a resume replays them first): trajectory.npz
@@ -153,13 +160,15 @@ def schedule_text() -> list:
"wz_radps": round(c["wz"], 4)} for c in S.inst["commands"]]
+ONE = "one episode, no reset" if not RESET_BUDGET else (
+ f"an episode (with up to {RESET_BUDGET} resets: POST /reset starts a new one; only the last is graded)")
TASK_EPISODE = {
- "joystick": f"one episode, no reset: it ends after {S.horizon} control steps (the {S.horizon * DT:g} s command "
+ "joystick": f"{ONE}: it ends after {S.horizon} control steps (the {S.horizon * DT:g} s command "
f"schedule), when the robot meets its termination rule ({S.spec.get('term_text')}: a failure), when you "
"submit `done`, or when your time is over",
- "getup": f"one episode, no reset: it ends the moment the robot has stood upright for {Q.HOLD_STEPS} control steps in a "
+ "getup": f"{ONE}: it ends the moment the robot has stood upright for {Q.HOLD_STEPS} control steps in a "
f"row (success), after {S.horizon} control steps, when you submit `done`, or when your time is over",
- "stand_on_two": f"one episode, no reset: it ends the moment the pose has been held for {Q.POSE_HOLD_STEPS} control steps "
+ "stand_on_two": f"{ONE}: it ends the moment the pose has been held for {Q.POSE_HOLD_STEPS} control steps "
"in a row (success), when the robot meets the termination rule (a failure: its trunk's up axis below "
f"-0.25, or {'a leg' if S.spec.get('hind_leg_contact') else 'a front leg'}'s thigh, calf or hip on the "
f"floor), after {S.horizon} control steps, when you "
@@ -204,7 +213,7 @@ SPEC = {
"world": {"frame": "z up; the robot starts where the episode puts it (no world coordinates are given)", "up": "+z",
"units": {"length": "m", "angle": "rad", "time": "s"}, "quaternion_order": "wxyz",
"camera_axes": "intrinsic_cv / extrinsic_cv: OpenCV camera axes (x right, y down, z forward); cam2world_gl: OpenGL axes"},
- "budget": {"reset": 0, "steps": None, "wall_clock_s": WALL_CLOCK_S}, # + "clock" (rcb_limited.CLOCK_NOTE): what time_left_s counts from
+ "budget": {"reset": RESET_BUDGET, "steps": None, "wall_clock_s": WALL_CLOCK_S}, # + "clock" (rcb_limited.CLOCK_NOTE): what time_left_s counts from
"max_control_steps": S.horizon,
"episode": TASK_EPISODE,
"command_schedule": schedule_text(),
@@ -223,6 +232,15 @@ SPEC = {
"depth": {"modes": list(L.DEPTH_MODES), "default": "none", "kind": "z-depth", "unit": "m", "invalid": 0,
"max_m": L.DEPTH_MAX_M, "frames_encoding": "uint16 mm, zlib", "noise": L.DEPTH_NOISE},
}
+if RESET_BUDGET:
+ SPEC["reset"] = (
+ f"POST /reset (JSON or RCB1 {{cameras, video_fps, depth}}; no physics) abandons the running episode, or follows "
+ f"one that has ended, and starts a new one from the same frozen starting state, {RESET_BUDGET} times in all; it "
+ "returns the new episode's first frames and proprioception. Only the last episode is graded: an episode you "
+ "reset away is never graded, a success ends the trial (409 after it), and when your time is up the episode then "
+ "running is the one graded. The wall clock never restarts. /status and every meta add `episode`, `resets_left` "
+ "and `control_steps_total` (all episodes); `control_steps` counts the current episode's."
+ )
L.check_spec(SPEC) # a spec that misses a 2.2 field or names a forbidden one never starts the episode
EP.scene_built()
@@ -277,6 +295,7 @@ def summary() -> dict:
"reason": E["reason"], "done": E["done"], "env": S.env, "protocol": L.PROTOCOL, "n_actions": len(ROWS),
"fell_at": chk.get("fell_at"), "worst_lin_err": chk.get("worst_lin_err"), "worst_yaw_err": chk.get("worst_yaw_err"),
"tilt_deg": chk.get("tilt_deg"), "imu_height": chk.get("imu_height"), "pose_angle_deg": chk.get("pose_angle_deg"),
+ "episodes": E["resets"] + 1, "resets": E["resets"], "reset_budget": RESET_BUDGET, "control_steps_total": EP.rows,
**EP.final_fields()}
@@ -294,7 +313,8 @@ def finish(reason: str) -> None:
E["success"] = bool(chk["success"])
except Exception: # noqa: BLE001
traceback.print_exc()
- EP.log({"ev": "end", "reason": reason, "step": E["step"], "control_steps": ctrl_steps(), "success": bool(E["success"])})
+ EP.log({"ev": "end", "reason": reason, "step": E["step"], "control_steps": ctrl_steps(), "success": bool(E["success"]),
+ "episode": E["resets"] + 1, "journal_rows": EP.rows})
dump()
@@ -388,7 +408,7 @@ def meta_common(cams) -> dict:
"sim_time_s": round(n * DT, 3), "time_s": round(_agent_time(), 2),
"time_left_s": max(0.0, round(WALL_CLOCK_S - _agent_time(), 1)),
"proprio": {"layout": PROPRIO_LAYOUT, "now": proprio_row().tolist()}, "done": E["done"], "reason": E["reason"],
- "cameras": {c: camera_info(c) for c in cams}}
+ "cameras": {c: camera_info(c) for c in cams}, **reset_counters()}
if S.kind == "joystick":
out["command_now"] = dict(zip(("vx_mps", "vy_mps", "wz_radps"), S.command_now().round(4).tolist()))
return out
@@ -478,10 +498,54 @@ def do_observe(meta: dict):
return out, rec.parts(video if seconds > 0 else ("frames" if video == "mp4" else video), fps)
+def reset_counters() -> dict:
+ if not RESET_BUDGET:
+ return {}
+ left = 0 if (E["success"] or E.get("exiting")) else RESET_BUDGET - E["resets"]
+ return {"episode": E["resets"] + 1, "resets_left": left, "control_steps_total": EP.rows}
+
+
def status() -> dict:
_touch()
return {"step": E["step"], "calls": E["calls"], "control_steps": ctrl_steps(), "done": E["done"], "reason": E["reason"],
- "elapsed_s": round(time.time() - E["t0"], 1)}
+ "elapsed_s": round(time.time() - E["t0"], 1), **reset_counters()}
+
+
+class Refused(Exception):
+ """A /reset the service turns down (409): nothing changes."""
+
+ def __init__(self, body: dict):
+ super().__init__(body.get("error"))
+ self.body = body
+
+
+def do_reset(meta: dict):
+ """POST /reset: a new episode from the frozen instance (the running one, if any, is abandoned, never graded)."""
+ meta = _request(meta)
+ cams, fps = cameras_of(meta), video_fps(meta)
+ L.depth_mode(meta) # a bad depth value is a 400 before anything changes
+ if E["success"]:
+ raise Refused({"error": "trial finished: an episode succeeded", "reason": E["reason"], "resets_left": 0})
+ if E["resets"] >= RESET_BUDGET:
+ raise Refused({"error": "reset budget exhausted", "reason": "reset budget exhausted", "budget": RESET_BUDGET,
+ "resets_left": 0})
+ abandoned = {"episode": E["resets"] + 1, "control_steps": ctrl_steps(), "done": E["done"], "reason": E["reason"]}
+ if not E["done"]: # the abandoned episode's end, for the log (never graded: its rows leave the trajectory)
+ EP.log({"ev": "end", "reason": "reset", "step": E["step"], "control_steps": ctrl_steps(), "success": False,
+ "episode": E["resets"] + 1, "journal_rows": EP.rows})
+ S.reset()
+ ROWS.clear()
+ E["resets"] += 1
+ E["reset_at"] = list(E["reset_at"]) + [EP.rows]
+ E.update(done=False, reason=None, success=False, t_end=None, episode_control_steps=None, dumped=False, check=None)
+ EP._ck = None # the state changed without a journal row: the checkpoint after this call must hash it anew
+ E["calls"] += 1
+ EP.log({"ev": "reset", "call": E["calls"], "episode": E["resets"] + 1, "journal_rows": EP.rows,
+ "abandoned": abandoned, "state_sha": S.state_sha()})
+ rec = recorder(cams, fps, meta)
+ rec.tick(force=True, last=True)
+ out = {**meta_common(cams), "reset_id": E["resets"], "ok": True}
+ return out, rec.parts("frames", fps)
def end_from_signal() -> None:
@@ -498,20 +562,106 @@ def end_from_signal() -> None:
os._exit(0)
+# A resume replays the journal from the frozen instance and resets where the committed resets did: their journal rows
+# are in the last checkpoint (`persist`), the one Episode.begin resumes from.
+_CPS = [e for e in EP.events() if e.get("ev") == "checkpoint"]
+RESUME = {"at": list((_CPS[-1].get("persist") or {}).get("reset_at") or []) if _CPS else [], "i": 0}
+
+
+def _resume_resets() -> None:
+ while RESUME["at"] and int(RESUME["at"][0]) <= RESUME["i"]:
+ RESUME["at"].pop(0)
+ S.reset()
+ ROWS.clear()
+
+
def replay_row(row) -> None:
"""A journaled row during a resume."""
+ _resume_resets()
S.step(row)
ROWS.append(np.asarray(row, dtype=np.float64))
+ RESUME["i"] += 1
-EP.begin(state=E, step=replay_row, state_sha=S.state_sha, proprio=proprio_row, spec=SPEC, on_end=end_from_signal,
- start_event={"env": S.env, "seed": S.inst.get("seed"), "cameras": CAMS})
+def resume_sha() -> str:
+ _resume_resets() # a reset after the last replayed row (or before the first)
+ return S.state_sha()
+
+
+RESUMED = EP.begin(state=E, step=replay_row, state_sha=resume_sha, proprio=proprio_row, spec=SPEC, on_end=end_from_signal,
+ start_event={"env": S.env, "seed": S.inst.get("seed"), "cameras": CAMS})
+if RESUMED and E["resets"]: # Episode.begin looks for an "end" of any episode: only the current one's counts
+ _ends = [e for e in EP.events() if e.get("ev") == "end" and int(e.get("episode") or 1) == E["resets"] + 1
+ and e.get("reason") != "reset" and int(e.get("journal_rows") or 0) <= EP.rows]
+ if _ends:
+ E.update(done=True, reason=_ends[-1].get("reason"), success=bool(_ends[-1].get("success")),
+ t_end=float(_ends[-1]["ts"]), episode_control_steps=int(_ends[-1].get("control_steps") or ctrl_steps()))
+ else:
+ E.update(done=False, reason=None, success=False, t_end=None, episode_control_steps=None)
if E["done"]:
dump() # resumed after the episode had ended: the record is rewritten from the replay
E["ready"] = True
write_final()
open(PID_FILE, "w").write(str(os.getpid()))
print(f"{L.PROTOCOL} MuJoCo Playground quadruped sim ready ({S.env}; startup {EP.n_startups}, resumes {EP.resumes})", flush=True)
-L.serve(L.Service(spec=SPEC, lock=LOCK, state=E, submit=do_submit, observe=do_observe, kinematics=do_kinematics, status=status,
- on_submit_error=lambda exc: finish("error"), log=EP.log, episode=EP, state_sha=S.state_sha,
- proprio=proprio_row), port=int(os.environ.get("RCB_PORT", "8000")))
+
+
+def make_handler(svc: L.Service):
+ """The shared 2.2 handler, plus POST /reset when the run grants resets (the shared layer has no reset)."""
+ base = L.make_handler(svc)
+
+ class H(base):
+ def do_POST(self):
+ if self.path != "/reset" or not RESET_BUDGET:
+ return super().do_POST()
+ try:
+ meta, _ = self._body()
+ except Exception as exc: # noqa: BLE001
+ return self._json(413 if "too large" in str(exc) else 400, {"error": str(exc)[:200]})
+ with LOCK:
+ if E.get("exiting"):
+ return self._json(409, {"error": "trial finished: your time is up", "reason": E.get("reason"),
+ "resets_left": 0})
+ cid = EP.call_begin(self.path, self.connection)
+ out = blobs = err = None
+ try:
+ out, blobs = do_reset(meta)
+ except Refused as exc:
+ err = (409, exc.body)
+ except ValueError as exc:
+ err = (400, {"error": str(exc)})
+ except Exception as exc: # noqa: BLE001
+ traceback.print_exc()
+ E["error"] = repr(exc)
+ EP.log({"ev": "error", "path": self.path, "error": repr(exc)})
+ try:
+ finish("error")
+ except Exception: # noqa: BLE001
+ traceback.print_exc()
+ err = (500, {"error": "simulator error", "detail": repr(exc)[:200]})
+ finally:
+ if out is not None or (err and err[0] == 500):
+ try:
+ EP.checkpoint(E, S.state_sha, proprio_row)
+ except Exception: # noqa: BLE001
+ traceback.print_exc()
+ EP.call_end(cid, self.path)
+ if err is not None:
+ return self._json(*err)
+ return self._rcb1({**out, "stopped": None}, blobs)
+
+ return H
+
+
+def serve(svc: L.Service, port: int) -> None:
+ """L.serve with this service's handler (the same spec defaults and check, and L.bind: a port still held when the
+ service restarts is waited for, not a crash)."""
+ svc.spec.setdefault("on_service_error", L.ON_SERVICE_ERROR)
+ svc.spec.setdefault("max_call_wall_s", EP.max_call_wall_s)
+ L.check_spec(svc.spec)
+ L.bind(port, make_handler(svc)).serve_forever()
+
+
+serve(L.Service(spec=SPEC, lock=LOCK, state=E, submit=do_submit, observe=do_observe, kinematics=do_kinematics, status=status,
+ on_submit_error=lambda exc: finish("error"), log=EP.log, episode=EP, state_sha=S.state_sha,
+ proprio=proprio_row), port=int(os.environ.get("RCB_PORT", "8000")))
diff --git a/tasks/mujoco_playground_spot_joystick_gait_tracking-standard/instruction.md b/tasks/mujoco_playground_spot_joystick_gait_tracking-standard/instruction.md
index 2e2fa85a20..df686da5dc 100644
--- a/tasks/mujoco_playground_spot_joystick_gait_tracking-standard/instruction.md
+++ b/tasks/mujoco_playground_spot_joystick_gait_tracking-standard/instruction.md
@@ -2,9 +2,9 @@
**Task:** Make Boston Dynamics Spot follow the joystick command schedule for the whole 20 s episode with the commanded gait (the feet's swing pattern, frequency and height) without falling over.
-You control the Boston Dynamics Spot quadruped — **as if it were real hardware** — in one episode of MuJoCo Playground's
+You control the Boston Dynamics Spot quadruped — **as if it were real hardware** — in MuJoCo Playground's
`SpotJoystickGaitTracking` environment (plain MuJoCo 3.3.7 with Playground's scene, commit 4057c14). The simulator runs in a service you
-cannot open. You get one client and one episode.
+cannot open. You get one client and up to 11 episodes.
## The episode
The robot starts standing still at Playground's `home` pose (this environment's reset does not move it). The ground is flat.
@@ -23,7 +23,15 @@ The feet must also follow a **pace** gait at 3.66 Hz with a swing height of 0.12
The same schedule is `spec["command_schedule"]`; every response's `meta["command_now"]` is the current command. The gait is `spec["gait"]`.
## Rules (real-robot rules)
-- **No reset.** One episode. You see four cameras that follow the robot (RGB, depth if you ask) and the robot's own
+- **Episodes and resets.** You have **10 resets** (`spec["budget"]["reset"]`, which is authoritative: if it
+ says 0, there is no reset and the one episode is the run). A reset (`POST /reset`, below) puts the robot back on the
+ same frozen starting state and starts a new episode, either abandoning the running one or after one has ended; it
+ takes no simulated time and returns the start's images and proprioception. Once an episode has ended, `submit`
+ raises `EpisodeOver` until you reset. **Only the last episode is graded**: an episode you reset away is never graded,
+ a success ends the run at once (no reset after it), and when your time is up the episode then running is the one
+ graded. The wall clock is the run's and never restarts; `meta["episode"]` and `meta["resets_left"]` count the
+ episodes (`meta["control_steps"]` is the current episode's).
+- **No privileged state.** You see the cameras that follow the robot (RGB, depth if you ask) and the robot's own
sensors: joint positions, velocities and torques, the IMU (gyro, accelerometer, gravity's direction, the trunk's
velocity), foot contacts. No world positions, no success check.
- **The episode ends** after 1000 control steps, the whole schedule (`spec["max_control_steps"]`, counted by `meta["control_steps"]`), when the robot meets its termination rule (the trunk tilted by more than about 32 deg (the z component of the IMU frame's z axis below 0.85): a failure), when you submit `done`, or when your time is up (before the end of the schedule, both are failures).
@@ -32,8 +40,8 @@ The same schedule is `spec["command_schedule"]`; every response's `meta["command
`get_obs(seconds=...)`, which holds the current targets). `meta["time_left_s"]` is the service's count of the limit,
from your first request to the service: it never runs out before your time does, so plan by it.
- Nothing to hand in: the service records the episode and grades it. Leave your script at `/app/output/gen.py`.
-- **If the service crashes**, a call fails with a connection error: it restarts and resumes the SAME episode (never a
- reset) in the state after the last call it completed. `r.wait_ready()` waits for it; then check `r.status()`.
+- **If the service crashes**, a call fails with a connection error: it restarts and resumes the SAME episode (a crash
+ is never a reset and costs none) in the state after the last call it completed. `r.wait_ready()` waits for it; then check `r.status()`.
- Nothing is put into your context automatically: images come back as arrays in Python objects.
## Client (`/opt/tools/rcb_client.py`)
@@ -46,7 +54,8 @@ res = r.submit("joint_position", A, dt=0.02) # A: (T, 12) rows dt apart; res.p
obs = r.get_obs(seconds=1.0, video="mp4") # hold the current targets and watch for 1 s (max 30)
r.fk(q); r.ik({"FL": [x, y, z, 1, 0, 0, 0]}, seed=q) # feet in the trunk frame <-> joints (no physics)
r.submit("done", []) # end the episode now
-r.status() # calls, control steps, done, reason, elapsed time
+r.status() # calls, control steps, done, reason, elapsed time, episode, resets_left
+obs = r._post("/reset", {"cameras": ["track"]}, [], 900) # POST /reset: a new episode (the 2.2 client has no reset())
```
Cameras: `track`, `side`, `back`, `top` (they follow the robot); `spec["cameras"]` gives each one's mounting and intrinsics, and every response carries its pose. Proprioception (`spec["proprio_groups"]`): `joint_pos`, `joint_vel`, `joint_effort` (12 each, action
order), `imu_gyro`, `imu_accelerometer`, `imu_gravity`, `imu_linvel` (3 each, IMU frame: x forward, y left, z up),
diff --git a/tasks/mujoco_playground_spot_joystick_gait_tracking-standard/task.toml b/tasks/mujoco_playground_spot_joystick_gait_tracking-standard/task.toml
index a2d197dcc0..3b04e42be3 100644
--- a/tasks/mujoco_playground_spot_joystick_gait_tracking-standard/task.toml
+++ b/tasks/mujoco_playground_spot_joystick_gait_tracking-standard/task.toml
@@ -16,7 +16,7 @@ artifacts = [{ source = "/var/rcb-monitor/", service = "rcb_monitor" },
[task]
name = "robot-coding-bench/mujoco_playground_spot_joystick_gait_tracking-standard"
-description = "STANDARD MODE (robot as a service, eai-standard/2.2): MuJoCo Playground SpotJoystickGaitTracking (Boston Dynamics Spot) on the same frozen instance as mujoco_playground_spot_joystick_gait_tracking-privileged: Make Boston Dynamics Spot follow the joystick command schedule for the whole 20 s episode with the commanded gait (the feet's swing pattern, frequency and height) without falling over. The agent only gets a client: four cameras that follow the robot (RGB, optional depth), the robot's joints and IMU, joint position chunks at 50 Hz, the feet's kinematics; one episode, no reset, no world state; the sim service records the episode, replays it in fresh processes and grades it."
+description = "STANDARD MODE (robot as a service, eai-standard/2.2): MuJoCo Playground SpotJoystickGaitTracking (Boston Dynamics Spot) on the same frozen instance as mujoco_playground_spot_joystick_gait_tracking-privileged: Make Boston Dynamics Spot follow the joystick command schedule for the whole 20 s episode with the commanded gait (the feet's swing pattern, frequency and height) without falling over. The agent only gets a client: four cameras that follow the robot (RGB, optional depth), the robot's joints and IMU, joint position chunks at 50 Hz, the feet's kinematics; 10 resets (only the last episode is graded), no world state; the sim service records the episode, replays it in fresh processes and grades it."
authors = [{ name = "Changhe Chen" }]
keywords = ["robotics", "mujoco", "mujoco-playground", "boston-dynamics-spot", "quadruped", "locomotion", "joystick", "code-as-policy", "trajectory", "standard-mode"]
@@ -28,6 +28,7 @@ simulator = "mujoco"
robot = "boston-dynamics-spot"
family = "mujoco_playground/SpotJoystickGaitTracking"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "full"
instance_seed = 0
diff --git a/tasks/mujoco_playground_spot_joystick_gait_tracking-standard/tests/grade.py b/tasks/mujoco_playground_spot_joystick_gait_tracking-standard/tests/grade.py
index 41ce495421..3d4c23ad7e 100644
--- a/tasks/mujoco_playground_spot_joystick_gait_tracking-standard/tests/grade.py
+++ b/tasks/mujoco_playground_spot_joystick_gait_tracking-standard/tests/grade.py
@@ -77,6 +77,9 @@ write(
"sim_error": int(d.get("reason") == "error"),
"sim_startups": int(d.get("sim_startups") or 0),
"resumes": int(d.get("resumes") or 0),
+ "resets": int(d.get("resets") or 0), # PROTOCOL §4.5: resets the agent used; only the last episode is graded
+ "reset_budget": int(d.get("reset_budget") or 0),
+ "episodes": int(d.get("episodes") or 1),
"resume_state_mismatch": int(d.get("resume_state_mismatch") or 0),
"resume_max_joint_dev": num("resume_max_joint_dev"),
},
diff --git a/tasks/myosuite-arm-reach-i00-privileged/environment/rcb_myosuite.py b/tasks/myosuite-arm-reach-i00-privileged/environment/rcb_myosuite.py
index 0f2fca5acd..930c0c73ba 100644
--- a/tasks/myosuite-arm-reach-i00-privileged/environment/rcb_myosuite.py
+++ b/tasks/myosuite-arm-reach-i00-privileged/environment/rcb_myosuite.py
@@ -465,6 +465,12 @@ class Session:
self.done, self.success, self.reason = True, True, "success"
elif failed:
self.done, self.reason = True, "failure: " + _FAILURE.get(self.task, "task terminated")
+ # Baoding's targets come from a precomputed trajectory of 1000 points (baoding_v1.create_goal_trajectory,
+ # "assumes that it's greater than env horizon"): with MyoSuite's step budget lifted, step 1001 would index past
+ # it (IndexError, board 0086). The episode ends there instead, as a failure unless it succeeded on that step.
+ n_goal = len(getattr(self.base, "goal", ())) if self.task == "baoding" else 0
+ if not self.done and n_goal and self.steps >= n_goal:
+ self.done, self.reason = True, f"failure: the targets' trajectory ends after {n_goal} control steps"
return not self.done
def _bimanual_contacts_ok(self) -> bool:
diff --git a/tasks/myosuite-arm-reach-i00-privileged/task.toml b/tasks/myosuite-arm-reach-i00-privileged/task.toml
index 6aa630582b..e6ed49c444 100644
--- a/tasks/myosuite-arm-reach-i00-privileged/task.toml
+++ b/tasks/myosuite-arm-reach-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "myosuite" # MyoSuite 2.12.2 on MuJoCo 3.6.0 (CPU)
robot = "myoarm"
family = "arm_reach"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 0 # myoArmReachFixed-v0: seed(0), then reset(seed=0)
expert_time_estimate_hours = 1.5
diff --git a/tasks/myosuite-arm-reach-i00-privileged/tests/rcb_myosuite.py b/tasks/myosuite-arm-reach-i00-privileged/tests/rcb_myosuite.py
index 0f2fca5acd..930c0c73ba 100644
--- a/tasks/myosuite-arm-reach-i00-privileged/tests/rcb_myosuite.py
+++ b/tasks/myosuite-arm-reach-i00-privileged/tests/rcb_myosuite.py
@@ -465,6 +465,12 @@ class Session:
self.done, self.success, self.reason = True, True, "success"
elif failed:
self.done, self.reason = True, "failure: " + _FAILURE.get(self.task, "task terminated")
+ # Baoding's targets come from a precomputed trajectory of 1000 points (baoding_v1.create_goal_trajectory,
+ # "assumes that it's greater than env horizon"): with MyoSuite's step budget lifted, step 1001 would index past
+ # it (IndexError, board 0086). The episode ends there instead, as a failure unless it succeeded on that step.
+ n_goal = len(getattr(self.base, "goal", ())) if self.task == "baoding" else 0
+ if not self.done and n_goal and self.steps >= n_goal:
+ self.done, self.reason = True, f"failure: the targets' trajectory ends after {n_goal} control steps"
return not self.done
def _bimanual_contacts_ok(self) -> bool:
diff --git a/tasks/myosuite-arm-reach-i00-standard/environment/sim/rcb_myosuite.py b/tasks/myosuite-arm-reach-i00-standard/environment/sim/rcb_myosuite.py
index 0f2fca5acd..930c0c73ba 100644
--- a/tasks/myosuite-arm-reach-i00-standard/environment/sim/rcb_myosuite.py
+++ b/tasks/myosuite-arm-reach-i00-standard/environment/sim/rcb_myosuite.py
@@ -465,6 +465,12 @@ class Session:
self.done, self.success, self.reason = True, True, "success"
elif failed:
self.done, self.reason = True, "failure: " + _FAILURE.get(self.task, "task terminated")
+ # Baoding's targets come from a precomputed trajectory of 1000 points (baoding_v1.create_goal_trajectory,
+ # "assumes that it's greater than env horizon"): with MyoSuite's step budget lifted, step 1001 would index past
+ # it (IndexError, board 0086). The episode ends there instead, as a failure unless it succeeded on that step.
+ n_goal = len(getattr(self.base, "goal", ())) if self.task == "baoding" else 0
+ if not self.done and n_goal and self.steps >= n_goal:
+ self.done, self.reason = True, f"failure: the targets' trajectory ends after {n_goal} control steps"
return not self.done
def _bimanual_contacts_ok(self) -> bool:
diff --git a/tasks/myosuite-arm-reach-i00-standard/task.toml b/tasks/myosuite-arm-reach-i00-standard/task.toml
index 8c56f54109..07e9f6ab98 100644
--- a/tasks/myosuite-arm-reach-i00-standard/task.toml
+++ b/tasks/myosuite-arm-reach-i00-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "myosuite" # MyoSuite 2.12.2 on MuJoCo 3.6.0 (CPU)
robot = "myoarm"
family = "arm_reach"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "full"
instance_seed = 0 # myoArmReachFixed-v0: seed(0), then reset(seed=0)
diff --git a/tasks/myosuite-baoding-i00-privileged/README.md b/tasks/myosuite-baoding-i00-privileged/README.md
index 058611dac8..02f3a74454 100644
--- a/tasks/myosuite-baoding-i00-privileged/README.md
+++ b/tasks/myosuite-baoding-i00-privileged/README.md
@@ -12,7 +12,7 @@
| Base image | `ghcr.io/mll-lab-nu/eai-myosuite:0.1.1` |
## Task
-MyoSuite `myoChallengeBaodingP1-v1` (MyoHand, a musculoskeletal model of a human forearm and hand, 39 muscles): Baoding balls: rotate the two balls around each other in the palm so that each follows its moving target (the targets circle the palm once every 5 s). Success: both balls within 0.015 m of their targets for 200 consecutive control steps (5.0 s: one full orbit of the targets). Failure (the episode ends): a ball's height (world z) drops below 1.25 m, i.e. it falls out of the palm.
+MyoSuite `myoChallengeBaodingP1-v1` (MyoHand, a musculoskeletal model of a human forearm and hand, 39 muscles): Baoding balls: rotate the two balls around each other in the palm so that each follows its moving target (the targets circle the palm once every 5 s). Success: both balls within 0.015 m of their targets for 200 consecutive control steps (5.0 s: one full orbit of the targets). Failure (the episode ends): a ball's height (world z) drops below 1.25 m, i.e. it falls out of the palm. MyoSuite's target trajectory has 1000 points: an episode that has not succeeded after 1000 control steps (25 s) ends there, as a failure.
Two balls rest in the palm of the upturned hand; two small markers show where each ball should be now. The deliverable is one replayable trajectory of actuator commands (`/app/output/trajectory.npz`, `(T, 39)`, one row per 25 ms control step).
## Difficulty
@@ -61,3 +61,9 @@ Generated by `scripts/myosuite/generate.py` from `scripts/myosuite/instances/bao
`bad_trajectory`. Every `reward.json` now carries `infra_error` (1 next to every `grader_error`, so also when the
verifier's host does not rebuild the frozen instance). The pass rule is unchanged; a trial graded `bad_trajectory`
before this date whose rows had the right shape was an infrastructure failure.
+- 2026-10-14 (board 0113, from board 0086): an episode that has not succeeded after 1000 control steps (25 s) ends
+ there as a failure, and the task text says so. MyoSuite's baoding targets are a precomputed trajectory of 1000
+ points; with MyoSuite's step budget lifted, the step after it raised an IndexError (the service answered 500 and
+ the episode ended as a simulator error; a privileged replay that long stopped with an error). Nothing before that
+ step changes: the instance, the tooling's physics and the pass rule are the same, and every verdict reached before
+ is the same now.
diff --git a/tasks/myosuite-baoding-i00-privileged/environment/rcb_myosuite.py b/tasks/myosuite-baoding-i00-privileged/environment/rcb_myosuite.py
index 0f2fca5acd..930c0c73ba 100644
--- a/tasks/myosuite-baoding-i00-privileged/environment/rcb_myosuite.py
+++ b/tasks/myosuite-baoding-i00-privileged/environment/rcb_myosuite.py
@@ -465,6 +465,12 @@ class Session:
self.done, self.success, self.reason = True, True, "success"
elif failed:
self.done, self.reason = True, "failure: " + _FAILURE.get(self.task, "task terminated")
+ # Baoding's targets come from a precomputed trajectory of 1000 points (baoding_v1.create_goal_trajectory,
+ # "assumes that it's greater than env horizon"): with MyoSuite's step budget lifted, step 1001 would index past
+ # it (IndexError, board 0086). The episode ends there instead, as a failure unless it succeeded on that step.
+ n_goal = len(getattr(self.base, "goal", ())) if self.task == "baoding" else 0
+ if not self.done and n_goal and self.steps >= n_goal:
+ self.done, self.reason = True, f"failure: the targets' trajectory ends after {n_goal} control steps"
return not self.done
def _bimanual_contacts_ok(self) -> bool:
diff --git a/tasks/myosuite-baoding-i00-privileged/instruction.md b/tasks/myosuite-baoding-i00-privileged/instruction.md
index 8947a1830f..66768f6a11 100644
--- a/tasks/myosuite-baoding-i00-privileged/instruction.md
+++ b/tasks/myosuite-baoding-i00-privileged/instruction.md
@@ -2,7 +2,7 @@
You control the **MyoHand**, a musculoskeletal model of a human forearm and hand (MyoSuite 2.12.2 on MuJoCo), via Python, through its actuators — 39
muscles: there is no joint-level controller.
-**Task:** Baoding balls: rotate the two balls around each other in the palm so that each follows its moving target (the targets circle the palm once every 5 s). Success: both balls within 0.015 m of their targets for 200 consecutive control steps (5.0 s: one full orbit of the targets). Failure (the episode ends): a ball's height (world z) drops below 1.25 m, i.e. it falls out of the palm.
+**Task:** Baoding balls: rotate the two balls around each other in the palm so that each follows its moving target (the targets circle the palm once every 5 s). Success: both balls within 0.015 m of their targets for 200 consecutive control steps (5.0 s: one full orbit of the targets). Failure (the episode ends): a ball's height (world z) drops below 1.25 m, i.e. it falls out of the palm. MyoSuite's target trajectory has 1000 points: an episode that has not succeeded after 1000 control steps (25 s) ends there, as a failure.
Two balls rest in the palm of the upturned hand; two small markers show where each ball should be now. `s.state()` gives the task's observation dict (privileged): `hand_pos`, `object1_pos`, `object2_pos`, `object1_velp`, `object2_velp`, `target1_pos`, `target2_pos`, `target1_err`, `target2_err`, `act`.
MyoSuite's reward code (from `/usr/local/lib/python3.11/site-packages/myosuite/envs/myo/myochallenge/baoding_v1.py`; `solved` is the goal test,
diff --git a/tasks/myosuite-baoding-i00-privileged/task.toml b/tasks/myosuite-baoding-i00-privileged/task.toml
index 1074552556..1ae2341e3d 100644
--- a/tasks/myosuite-baoding-i00-privileged/task.toml
+++ b/tasks/myosuite-baoding-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "myosuite" # MyoSuite 2.12.2 on MuJoCo 3.6.0 (CPU)
robot = "myohand"
family = "baoding"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "a successful model trial on this frozen instance: Claude Code + Claude Opus 5.5 (run `claude_code-claude_opus_5_5-default`, 2026-09-28), trial `myosuite-baoding-i00__J6BJVtG` (215 rows); today's verifier scores its trajectory 1 in two fresh trials (docs/benchmarks/myosuite.md); human review in PR #65"
instance_seed = 0 # myoChallengeBaodingP1-v1: seed(0), then reset(seed=0)
diff --git a/tasks/myosuite-baoding-i00-privileged/tests/rcb_myosuite.py b/tasks/myosuite-baoding-i00-privileged/tests/rcb_myosuite.py
index 0f2fca5acd..930c0c73ba 100644
--- a/tasks/myosuite-baoding-i00-privileged/tests/rcb_myosuite.py
+++ b/tasks/myosuite-baoding-i00-privileged/tests/rcb_myosuite.py
@@ -465,6 +465,12 @@ class Session:
self.done, self.success, self.reason = True, True, "success"
elif failed:
self.done, self.reason = True, "failure: " + _FAILURE.get(self.task, "task terminated")
+ # Baoding's targets come from a precomputed trajectory of 1000 points (baoding_v1.create_goal_trajectory,
+ # "assumes that it's greater than env horizon"): with MyoSuite's step budget lifted, step 1001 would index past
+ # it (IndexError, board 0086). The episode ends there instead, as a failure unless it succeeded on that step.
+ n_goal = len(getattr(self.base, "goal", ())) if self.task == "baoding" else 0
+ if not self.done and n_goal and self.steps >= n_goal:
+ self.done, self.reason = True, f"failure: the targets' trajectory ends after {n_goal} control steps"
return not self.done
def _bimanual_contacts_ok(self) -> bool:
diff --git a/tasks/myosuite-baoding-i00-standard/README.md b/tasks/myosuite-baoding-i00-standard/README.md
index 35a6df40da..e3fbb7f8f7 100644
--- a/tasks/myosuite-baoding-i00-standard/README.md
+++ b/tasks/myosuite-baoding-i00-standard/README.md
@@ -12,7 +12,7 @@
| Base image | `ghcr.io/mll-lab-nu/eai-myosuite:0.1.1` |
## Task
-MyoSuite `myoChallengeBaodingP1-v1` (MyoHand, 39 muscles) on the same frozen instance as `myosuite-baoding-i00-privileged`: Baoding balls: rotate the two balls around each other in the palm so that each follows its moving target (the targets circle the palm once every 5 s). Success: both balls within 0.015 m of their targets for 200 consecutive control steps (5.0 s: one full orbit of the targets). Failure (the episode ends): a ball's height (world z) drops below 1.25 m, i.e. it falls out of the palm.
+MyoSuite `myoChallengeBaodingP1-v1` (MyoHand, 39 muscles) on the same frozen instance as `myosuite-baoding-i00-privileged`: Baoding balls: rotate the two balls around each other in the palm so that each follows its moving target (the targets circle the palm once every 5 s). Success: both balls within 0.015 m of their targets for 200 consecutive control steps (5.0 s: one full orbit of the targets). Failure (the episode ends): a ball's height (world z) drops below 1.25 m, i.e. it falls out of the palm. MyoSuite's target trajectory has 1000 points: an episode that has not succeeded after 1000 control steps (25 s) ends there, as a failure.
Two balls rest in the palm of the upturned hand; two small markers show where each ball should be now. Standard mode (eai-standard/2.2): the agent drives the body through a client, with only its own proprioception and the scene's cameras, and 10 resets (only the last episode is graded).
## Difficulty
@@ -90,3 +90,9 @@ Generated by `scripts/myosuite/generate.py` from `scripts/myosuite/instances/bao
still had that time. `/spec` says what the clock counts from (`budget.clock`), and the instruction's `## Time limit`
tells the agent to plan by `time_left_s`. The service never ended an episode on this clock: the scene, the frozen
instance, the pass rule and the grader are unchanged, so the task keeps its name and earlier results stand.
+- 2026-10-14 (board 0113, from board 0086): an episode that has not succeeded after 1000 control steps (25 s) ends
+ there as a failure, and the task text says so. MyoSuite's baoding targets are a precomputed trajectory of 1000
+ points; with MyoSuite's step budget lifted, the step after it raised an IndexError (the service answered 500 and
+ the episode ended as a simulator error; a privileged replay that long stopped with an error). Nothing before that
+ step changes: the instance, the tooling's physics and the pass rule are the same, and every verdict reached before
+ is the same now.
diff --git a/tasks/myosuite-baoding-i00-standard/environment/sim/rcb_myosuite.py b/tasks/myosuite-baoding-i00-standard/environment/sim/rcb_myosuite.py
index 0f2fca5acd..930c0c73ba 100644
--- a/tasks/myosuite-baoding-i00-standard/environment/sim/rcb_myosuite.py
+++ b/tasks/myosuite-baoding-i00-standard/environment/sim/rcb_myosuite.py
@@ -465,6 +465,12 @@ class Session:
self.done, self.success, self.reason = True, True, "success"
elif failed:
self.done, self.reason = True, "failure: " + _FAILURE.get(self.task, "task terminated")
+ # Baoding's targets come from a precomputed trajectory of 1000 points (baoding_v1.create_goal_trajectory,
+ # "assumes that it's greater than env horizon"): with MyoSuite's step budget lifted, step 1001 would index past
+ # it (IndexError, board 0086). The episode ends there instead, as a failure unless it succeeded on that step.
+ n_goal = len(getattr(self.base, "goal", ())) if self.task == "baoding" else 0
+ if not self.done and n_goal and self.steps >= n_goal:
+ self.done, self.reason = True, f"failure: the targets' trajectory ends after {n_goal} control steps"
return not self.done
def _bimanual_contacts_ok(self) -> bool:
diff --git a/tasks/myosuite-baoding-i00-standard/environment/sim/task.json b/tasks/myosuite-baoding-i00-standard/environment/sim/task.json
index d371997122..8bf0b21c1c 100644
--- a/tasks/myosuite-baoding-i00-standard/environment/sim/task.json
+++ b/tasks/myosuite-baoding-i00-standard/environment/sim/task.json
@@ -1,5 +1,5 @@
{
- "task_text": "Baoding balls: rotate the two balls around each other in the palm so that each follows its moving target (the targets circle the palm once every 5 s). Success: both balls within 0.015 m of their targets for 200 consecutive control steps (5.0 s: one full orbit of the targets). Failure (the episode ends): a ball's height (world z) drops below 1.25 m, i.e. it falls out of the palm.",
+ "task_text": "Baoding balls: rotate the two balls around each other in the palm so that each follows its moving target (the targets circle the palm once every 5 s). Success: both balls within 0.015 m of their targets for 200 consecutive control steps (5.0 s: one full orbit of the targets). Failure (the episode ends): a ball's height (world z) drops below 1.25 m, i.e. it falls out of the palm. MyoSuite's target trajectory has 1000 points: an episode that has not succeeded after 1000 control steps (25 s) ends there, as a failure.",
"default_cameras": [
"hand_top"
],
diff --git a/tasks/myosuite-baoding-i00-standard/instruction.md b/tasks/myosuite-baoding-i00-standard/instruction.md
index c99aa89d60..cf3d3971f2 100644
--- a/tasks/myosuite-baoding-i00-standard/instruction.md
+++ b/tasks/myosuite-baoding-i00-standard/instruction.md
@@ -2,7 +2,7 @@
You control the **MyoHand**, a musculoskeletal model of a human forearm and hand, **as if it were real**: the simulator (MyoSuite 2.12.2 on MuJoCo) runs in a service
you cannot open. You get one client and up to 11 episodes, and you act only through its actuators (39 muscles).
-**Task:** Baoding balls: rotate the two balls around each other in the palm so that each follows its moving target (the targets circle the palm once every 5 s). Success: both balls within 0.015 m of their targets for 200 consecutive control steps (5.0 s: one full orbit of the targets). Failure (the episode ends): a ball's height (world z) drops below 1.25 m, i.e. it falls out of the palm.
+**Task:** Baoding balls: rotate the two balls around each other in the palm so that each follows its moving target (the targets circle the palm once every 5 s). Success: both balls within 0.015 m of their targets for 200 consecutive control steps (5.0 s: one full orbit of the targets). Failure (the episode ends): a ball's height (world z) drops below 1.25 m, i.e. it falls out of the palm. MyoSuite's target trajectory has 1000 points: an episode that has not succeeded after 1000 control steps (25 s) ends there, as a failure.
Two balls rest in the palm of the upturned hand; two small markers show where each ball should be now.
diff --git a/tasks/myosuite-baoding-i00-standard/task.toml b/tasks/myosuite-baoding-i00-standard/task.toml
index 38f2462c0f..faca5d612e 100644
--- a/tasks/myosuite-baoding-i00-standard/task.toml
+++ b/tasks/myosuite-baoding-i00-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "myosuite" # MyoSuite 2.12.2 on MuJoCo 3.6.0 (CPU)
robot = "myohand"
family = "baoding"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "no positive example in standard mode yet: the privileged twin's (Claude Code + Claude Opus 5.5 (run `claude_code-claude_opus_5_5-default`, 2026-09-28), trial `myosuite-baoding-i00__J6BJVtG`) fails once its rows are rounded to the client's float32 (a ball drops after 206 control steps; 400 random float32 roundings of its rows fail too); the family's solvability rests on the privileged twin; human review in PR #65"
diff --git a/tasks/myosuite-bimanual-i00-privileged/environment/rcb_myosuite.py b/tasks/myosuite-bimanual-i00-privileged/environment/rcb_myosuite.py
index 0f2fca5acd..930c0c73ba 100644
--- a/tasks/myosuite-bimanual-i00-privileged/environment/rcb_myosuite.py
+++ b/tasks/myosuite-bimanual-i00-privileged/environment/rcb_myosuite.py
@@ -465,6 +465,12 @@ class Session:
self.done, self.success, self.reason = True, True, "success"
elif failed:
self.done, self.reason = True, "failure: " + _FAILURE.get(self.task, "task terminated")
+ # Baoding's targets come from a precomputed trajectory of 1000 points (baoding_v1.create_goal_trajectory,
+ # "assumes that it's greater than env horizon"): with MyoSuite's step budget lifted, step 1001 would index past
+ # it (IndexError, board 0086). The episode ends there instead, as a failure unless it succeeded on that step.
+ n_goal = len(getattr(self.base, "goal", ())) if self.task == "baoding" else 0
+ if not self.done and n_goal and self.steps >= n_goal:
+ self.done, self.reason = True, f"failure: the targets' trajectory ends after {n_goal} control steps"
return not self.done
def _bimanual_contacts_ok(self) -> bool:
diff --git a/tasks/myosuite-bimanual-i00-privileged/task.toml b/tasks/myosuite-bimanual-i00-privileged/task.toml
index d89cfa7791..c79ca854fc 100644
--- a/tasks/myosuite-bimanual-i00-privileged/task.toml
+++ b/tasks/myosuite-bimanual-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "myosuite" # MyoSuite 2.12.2 on MuJoCo 3.6.0 (CPU)
robot = "myoarm-mpl"
family = "bimanual"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "no positive example yet (two model attempts failed: Codex + GPT-6 Luna, effort high, run `codex-gpt6_luna-high` and Claude Code + Claude Opus 5.5 (run `claude_code-claude_opus_5_5-default`, 2026-10-06)); the task is reported separately until one exists (PROTOCOL §7, Open 4); human review in PR #65"
instance_seed = 0 # myoChallengeBimanual-v0: seed(0), then reset(seed=0)
diff --git a/tasks/myosuite-bimanual-i00-privileged/tests/rcb_myosuite.py b/tasks/myosuite-bimanual-i00-privileged/tests/rcb_myosuite.py
index 0f2fca5acd..930c0c73ba 100644
--- a/tasks/myosuite-bimanual-i00-privileged/tests/rcb_myosuite.py
+++ b/tasks/myosuite-bimanual-i00-privileged/tests/rcb_myosuite.py
@@ -465,6 +465,12 @@ class Session:
self.done, self.success, self.reason = True, True, "success"
elif failed:
self.done, self.reason = True, "failure: " + _FAILURE.get(self.task, "task terminated")
+ # Baoding's targets come from a precomputed trajectory of 1000 points (baoding_v1.create_goal_trajectory,
+ # "assumes that it's greater than env horizon"): with MyoSuite's step budget lifted, step 1001 would index past
+ # it (IndexError, board 0086). The episode ends there instead, as a failure unless it succeeded on that step.
+ n_goal = len(getattr(self.base, "goal", ())) if self.task == "baoding" else 0
+ if not self.done and n_goal and self.steps >= n_goal:
+ self.done, self.reason = True, f"failure: the targets' trajectory ends after {n_goal} control steps"
return not self.done
def _bimanual_contacts_ok(self) -> bool:
diff --git a/tasks/myosuite-bimanual-i00-standard/environment/sim/rcb_myosuite.py b/tasks/myosuite-bimanual-i00-standard/environment/sim/rcb_myosuite.py
index 0f2fca5acd..930c0c73ba 100644
--- a/tasks/myosuite-bimanual-i00-standard/environment/sim/rcb_myosuite.py
+++ b/tasks/myosuite-bimanual-i00-standard/environment/sim/rcb_myosuite.py
@@ -465,6 +465,12 @@ class Session:
self.done, self.success, self.reason = True, True, "success"
elif failed:
self.done, self.reason = True, "failure: " + _FAILURE.get(self.task, "task terminated")
+ # Baoding's targets come from a precomputed trajectory of 1000 points (baoding_v1.create_goal_trajectory,
+ # "assumes that it's greater than env horizon"): with MyoSuite's step budget lifted, step 1001 would index past
+ # it (IndexError, board 0086). The episode ends there instead, as a failure unless it succeeded on that step.
+ n_goal = len(getattr(self.base, "goal", ())) if self.task == "baoding" else 0
+ if not self.done and n_goal and self.steps >= n_goal:
+ self.done, self.reason = True, f"failure: the targets' trajectory ends after {n_goal} control steps"
return not self.done
def _bimanual_contacts_ok(self) -> bool:
diff --git a/tasks/myosuite-bimanual-i00-standard/task.toml b/tasks/myosuite-bimanual-i00-standard/task.toml
index 4fef9d8158..c60951f2be 100644
--- a/tasks/myosuite-bimanual-i00-standard/task.toml
+++ b/tasks/myosuite-bimanual-i00-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "myosuite" # MyoSuite 2.12.2 on MuJoCo 3.6.0 (CPU)
robot = "myoarm-mpl"
family = "bimanual"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "no positive example yet (two model attempts failed: Codex + GPT-6 Luna, effort high, run `codex-gpt6_luna-high` and Claude Code + Claude Opus 5.5 (run `claude_code-claude_opus_5_5-default`, 2026-10-06)); the task is reported separately until one exists (PROTOCOL §7, Open 4); human review in PR #65"
diff --git a/tasks/myosuite-chase-tag-i00-privileged/environment/rcb_myosuite.py b/tasks/myosuite-chase-tag-i00-privileged/environment/rcb_myosuite.py
index 0f2fca5acd..930c0c73ba 100644
--- a/tasks/myosuite-chase-tag-i00-privileged/environment/rcb_myosuite.py
+++ b/tasks/myosuite-chase-tag-i00-privileged/environment/rcb_myosuite.py
@@ -465,6 +465,12 @@ class Session:
self.done, self.success, self.reason = True, True, "success"
elif failed:
self.done, self.reason = True, "failure: " + _FAILURE.get(self.task, "task terminated")
+ # Baoding's targets come from a precomputed trajectory of 1000 points (baoding_v1.create_goal_trajectory,
+ # "assumes that it's greater than env horizon"): with MyoSuite's step budget lifted, step 1001 would index past
+ # it (IndexError, board 0086). The episode ends there instead, as a failure unless it succeeded on that step.
+ n_goal = len(getattr(self.base, "goal", ())) if self.task == "baoding" else 0
+ if not self.done and n_goal and self.steps >= n_goal:
+ self.done, self.reason = True, f"failure: the targets' trajectory ends after {n_goal} control steps"
return not self.done
def _bimanual_contacts_ok(self) -> bool:
diff --git a/tasks/myosuite-chase-tag-i00-privileged/task.toml b/tasks/myosuite-chase-tag-i00-privileged/task.toml
index aa24be47ef..0e596c76b4 100644
--- a/tasks/myosuite-chase-tag-i00-privileged/task.toml
+++ b/tasks/myosuite-chase-tag-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "myosuite" # MyoSuite 2.12.2 on MuJoCo 3.6.0 (CPU)
robot = "myolegs"
family = "chase_tag"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "a successful model trial on this frozen instance: Claude Code + Claude Opus 5.5 (run `claude_code-claude_opus_5_5-default`, 2026-10-06), trial `myosuite-chase-tag-i00-privilege__6GUXvYA` (329 rows); today's verifier scores its trajectory 1 in two fresh trials (docs/benchmarks/myosuite.md); human review in PR #65"
instance_seed = 0 # myoChallengeChaseTagP1-v0: seed(0), then reset(seed=0)
diff --git a/tasks/myosuite-chase-tag-i00-privileged/tests/rcb_myosuite.py b/tasks/myosuite-chase-tag-i00-privileged/tests/rcb_myosuite.py
index 0f2fca5acd..930c0c73ba 100644
--- a/tasks/myosuite-chase-tag-i00-privileged/tests/rcb_myosuite.py
+++ b/tasks/myosuite-chase-tag-i00-privileged/tests/rcb_myosuite.py
@@ -465,6 +465,12 @@ class Session:
self.done, self.success, self.reason = True, True, "success"
elif failed:
self.done, self.reason = True, "failure: " + _FAILURE.get(self.task, "task terminated")
+ # Baoding's targets come from a precomputed trajectory of 1000 points (baoding_v1.create_goal_trajectory,
+ # "assumes that it's greater than env horizon"): with MyoSuite's step budget lifted, step 1001 would index past
+ # it (IndexError, board 0086). The episode ends there instead, as a failure unless it succeeded on that step.
+ n_goal = len(getattr(self.base, "goal", ())) if self.task == "baoding" else 0
+ if not self.done and n_goal and self.steps >= n_goal:
+ self.done, self.reason = True, f"failure: the targets' trajectory ends after {n_goal} control steps"
return not self.done
def _bimanual_contacts_ok(self) -> bool:
diff --git a/tasks/myosuite-chase-tag-i00-standard/environment/sim/rcb_myosuite.py b/tasks/myosuite-chase-tag-i00-standard/environment/sim/rcb_myosuite.py
index 0f2fca5acd..930c0c73ba 100644
--- a/tasks/myosuite-chase-tag-i00-standard/environment/sim/rcb_myosuite.py
+++ b/tasks/myosuite-chase-tag-i00-standard/environment/sim/rcb_myosuite.py
@@ -465,6 +465,12 @@ class Session:
self.done, self.success, self.reason = True, True, "success"
elif failed:
self.done, self.reason = True, "failure: " + _FAILURE.get(self.task, "task terminated")
+ # Baoding's targets come from a precomputed trajectory of 1000 points (baoding_v1.create_goal_trajectory,
+ # "assumes that it's greater than env horizon"): with MyoSuite's step budget lifted, step 1001 would index past
+ # it (IndexError, board 0086). The episode ends there instead, as a failure unless it succeeded on that step.
+ n_goal = len(getattr(self.base, "goal", ())) if self.task == "baoding" else 0
+ if not self.done and n_goal and self.steps >= n_goal:
+ self.done, self.reason = True, f"failure: the targets' trajectory ends after {n_goal} control steps"
return not self.done
def _bimanual_contacts_ok(self) -> bool:
diff --git a/tasks/myosuite-chase-tag-i00-standard/task.toml b/tasks/myosuite-chase-tag-i00-standard/task.toml
index 56416f7fc3..be52222a17 100644
--- a/tasks/myosuite-chase-tag-i00-standard/task.toml
+++ b/tasks/myosuite-chase-tag-i00-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "myosuite" # MyoSuite 2.12.2 on MuJoCo 3.6.0 (CPU)
robot = "myolegs"
family = "chase_tag"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "no positive example in standard mode yet: the privileged twin's (Claude Code + Claude Opus 5.5 (run `claude_code-claude_opus_5_5-default`, 2026-10-06), trial `myosuite-chase-tag-i00-privilege__6GUXvYA`) fails once its rows are rounded to the client's float32 (the walk no longer reaches the opponent within its 329 rows); the family's solvability rests on the privileged twin; human review in PR #65"
diff --git a/tasks/myosuite-dex-airplane-fly-i00-privileged/environment/rcb_myosuite.py b/tasks/myosuite-dex-airplane-fly-i00-privileged/environment/rcb_myosuite.py
index 0f2fca5acd..930c0c73ba 100644
--- a/tasks/myosuite-dex-airplane-fly-i00-privileged/environment/rcb_myosuite.py
+++ b/tasks/myosuite-dex-airplane-fly-i00-privileged/environment/rcb_myosuite.py
@@ -465,6 +465,12 @@ class Session:
self.done, self.success, self.reason = True, True, "success"
elif failed:
self.done, self.reason = True, "failure: " + _FAILURE.get(self.task, "task terminated")
+ # Baoding's targets come from a precomputed trajectory of 1000 points (baoding_v1.create_goal_trajectory,
+ # "assumes that it's greater than env horizon"): with MyoSuite's step budget lifted, step 1001 would index past
+ # it (IndexError, board 0086). The episode ends there instead, as a failure unless it succeeded on that step.
+ n_goal = len(getattr(self.base, "goal", ())) if self.task == "baoding" else 0
+ if not self.done and n_goal and self.steps >= n_goal:
+ self.done, self.reason = True, f"failure: the targets' trajectory ends after {n_goal} control steps"
return not self.done
def _bimanual_contacts_ok(self) -> bool:
diff --git a/tasks/myosuite-dex-airplane-fly-i00-privileged/task.toml b/tasks/myosuite-dex-airplane-fly-i00-privileged/task.toml
index 7dffe96c50..a9f37b35d7 100644
--- a/tasks/myosuite-dex-airplane-fly-i00-privileged/task.toml
+++ b/tasks/myosuite-dex-airplane-fly-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "myosuite" # MyoSuite 2.12.2 on MuJoCo 3.6.0 (CPU)
robot = "myohand"
family = "dex_airplane_fly"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "no positive example yet (two model attempts failed: Claude Code + Claude Opus 5.5 (run `claude_code-claude_opus_5_5-default`, 2026-10-06) and Codex + GPT-6 Luna, effort high, run `codex-gpt6_luna-high`); the task is reported separately until one exists (PROTOCOL §7, Open 4); human review in PR #65"
instance_seed = 0 # MyoHandAirplaneFly-v0: seed(0), then reset(seed=0)
diff --git a/tasks/myosuite-dex-airplane-fly-i00-privileged/tests/rcb_myosuite.py b/tasks/myosuite-dex-airplane-fly-i00-privileged/tests/rcb_myosuite.py
index 0f2fca5acd..930c0c73ba 100644
--- a/tasks/myosuite-dex-airplane-fly-i00-privileged/tests/rcb_myosuite.py
+++ b/tasks/myosuite-dex-airplane-fly-i00-privileged/tests/rcb_myosuite.py
@@ -465,6 +465,12 @@ class Session:
self.done, self.success, self.reason = True, True, "success"
elif failed:
self.done, self.reason = True, "failure: " + _FAILURE.get(self.task, "task terminated")
+ # Baoding's targets come from a precomputed trajectory of 1000 points (baoding_v1.create_goal_trajectory,
+ # "assumes that it's greater than env horizon"): with MyoSuite's step budget lifted, step 1001 would index past
+ # it (IndexError, board 0086). The episode ends there instead, as a failure unless it succeeded on that step.
+ n_goal = len(getattr(self.base, "goal", ())) if self.task == "baoding" else 0
+ if not self.done and n_goal and self.steps >= n_goal:
+ self.done, self.reason = True, f"failure: the targets' trajectory ends after {n_goal} control steps"
return not self.done
def _bimanual_contacts_ok(self) -> bool:
diff --git a/tasks/myosuite-dex-airplane-fly-i00-standard/environment/sim/rcb_myosuite.py b/tasks/myosuite-dex-airplane-fly-i00-standard/environment/sim/rcb_myosuite.py
index 0f2fca5acd..930c0c73ba 100644
--- a/tasks/myosuite-dex-airplane-fly-i00-standard/environment/sim/rcb_myosuite.py
+++ b/tasks/myosuite-dex-airplane-fly-i00-standard/environment/sim/rcb_myosuite.py
@@ -465,6 +465,12 @@ class Session:
self.done, self.success, self.reason = True, True, "success"
elif failed:
self.done, self.reason = True, "failure: " + _FAILURE.get(self.task, "task terminated")
+ # Baoding's targets come from a precomputed trajectory of 1000 points (baoding_v1.create_goal_trajectory,
+ # "assumes that it's greater than env horizon"): with MyoSuite's step budget lifted, step 1001 would index past
+ # it (IndexError, board 0086). The episode ends there instead, as a failure unless it succeeded on that step.
+ n_goal = len(getattr(self.base, "goal", ())) if self.task == "baoding" else 0
+ if not self.done and n_goal and self.steps >= n_goal:
+ self.done, self.reason = True, f"failure: the targets' trajectory ends after {n_goal} control steps"
return not self.done
def _bimanual_contacts_ok(self) -> bool:
diff --git a/tasks/myosuite-dex-airplane-fly-i00-standard/task.toml b/tasks/myosuite-dex-airplane-fly-i00-standard/task.toml
index 6bd3751a3a..4715e74e45 100644
--- a/tasks/myosuite-dex-airplane-fly-i00-standard/task.toml
+++ b/tasks/myosuite-dex-airplane-fly-i00-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "myosuite" # MyoSuite 2.12.2 on MuJoCo 3.6.0 (CPU)
robot = "myohand"
family = "dex_airplane_fly"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "no positive example yet (two model attempts failed: Claude Code + Claude Opus 5.5 (run `claude_code-claude_opus_5_5-default`, 2026-10-06) and Codex + GPT-6 Luna, effort high, run `codex-gpt6_luna-high`); the task is reported separately until one exists (PROTOCOL §7, Open 4); human review in PR #65"
diff --git a/tasks/myosuite-dex-apple-lift-i00-privileged/environment/rcb_myosuite.py b/tasks/myosuite-dex-apple-lift-i00-privileged/environment/rcb_myosuite.py
index 0f2fca5acd..930c0c73ba 100644
--- a/tasks/myosuite-dex-apple-lift-i00-privileged/environment/rcb_myosuite.py
+++ b/tasks/myosuite-dex-apple-lift-i00-privileged/environment/rcb_myosuite.py
@@ -465,6 +465,12 @@ class Session:
self.done, self.success, self.reason = True, True, "success"
elif failed:
self.done, self.reason = True, "failure: " + _FAILURE.get(self.task, "task terminated")
+ # Baoding's targets come from a precomputed trajectory of 1000 points (baoding_v1.create_goal_trajectory,
+ # "assumes that it's greater than env horizon"): with MyoSuite's step budget lifted, step 1001 would index past
+ # it (IndexError, board 0086). The episode ends there instead, as a failure unless it succeeded on that step.
+ n_goal = len(getattr(self.base, "goal", ())) if self.task == "baoding" else 0
+ if not self.done and n_goal and self.steps >= n_goal:
+ self.done, self.reason = True, f"failure: the targets' trajectory ends after {n_goal} control steps"
return not self.done
def _bimanual_contacts_ok(self) -> bool:
diff --git a/tasks/myosuite-dex-apple-lift-i00-privileged/task.toml b/tasks/myosuite-dex-apple-lift-i00-privileged/task.toml
index 407e0b77c2..7e8e9f801a 100644
--- a/tasks/myosuite-dex-apple-lift-i00-privileged/task.toml
+++ b/tasks/myosuite-dex-apple-lift-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "myosuite" # MyoSuite 2.12.2 on MuJoCo 3.6.0 (CPU)
robot = "myohand"
family = "dex_apple_lift"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "a successful model trial on this frozen instance: Claude Code + Claude Opus 5.5 (run `claude_code-claude_opus_5_5-default`, 2026-10-06), trial `myosuite-dex-apple-lift-i00-priv__j5qaRCZ` (57 rows); today's verifier scores its trajectory 1 in two fresh trials (docs/benchmarks/myosuite.md); human review in PR #65"
instance_seed = 0 # MyoHandAppleLift-v0: seed(0), then reset(seed=0)
diff --git a/tasks/myosuite-dex-apple-lift-i00-privileged/tests/rcb_myosuite.py b/tasks/myosuite-dex-apple-lift-i00-privileged/tests/rcb_myosuite.py
index 0f2fca5acd..930c0c73ba 100644
--- a/tasks/myosuite-dex-apple-lift-i00-privileged/tests/rcb_myosuite.py
+++ b/tasks/myosuite-dex-apple-lift-i00-privileged/tests/rcb_myosuite.py
@@ -465,6 +465,12 @@ class Session:
self.done, self.success, self.reason = True, True, "success"
elif failed:
self.done, self.reason = True, "failure: " + _FAILURE.get(self.task, "task terminated")
+ # Baoding's targets come from a precomputed trajectory of 1000 points (baoding_v1.create_goal_trajectory,
+ # "assumes that it's greater than env horizon"): with MyoSuite's step budget lifted, step 1001 would index past
+ # it (IndexError, board 0086). The episode ends there instead, as a failure unless it succeeded on that step.
+ n_goal = len(getattr(self.base, "goal", ())) if self.task == "baoding" else 0
+ if not self.done and n_goal and self.steps >= n_goal:
+ self.done, self.reason = True, f"failure: the targets' trajectory ends after {n_goal} control steps"
return not self.done
def _bimanual_contacts_ok(self) -> bool:
diff --git a/tasks/myosuite-dex-apple-lift-i00-standard/environment/sim/rcb_myosuite.py b/tasks/myosuite-dex-apple-lift-i00-standard/environment/sim/rcb_myosuite.py
index 0f2fca5acd..930c0c73ba 100644
--- a/tasks/myosuite-dex-apple-lift-i00-standard/environment/sim/rcb_myosuite.py
+++ b/tasks/myosuite-dex-apple-lift-i00-standard/environment/sim/rcb_myosuite.py
@@ -465,6 +465,12 @@ class Session:
self.done, self.success, self.reason = True, True, "success"
elif failed:
self.done, self.reason = True, "failure: " + _FAILURE.get(self.task, "task terminated")
+ # Baoding's targets come from a precomputed trajectory of 1000 points (baoding_v1.create_goal_trajectory,
+ # "assumes that it's greater than env horizon"): with MyoSuite's step budget lifted, step 1001 would index past
+ # it (IndexError, board 0086). The episode ends there instead, as a failure unless it succeeded on that step.
+ n_goal = len(getattr(self.base, "goal", ())) if self.task == "baoding" else 0
+ if not self.done and n_goal and self.steps >= n_goal:
+ self.done, self.reason = True, f"failure: the targets' trajectory ends after {n_goal} control steps"
return not self.done
def _bimanual_contacts_ok(self) -> bool:
diff --git a/tasks/myosuite-dex-apple-lift-i00-standard/task.toml b/tasks/myosuite-dex-apple-lift-i00-standard/task.toml
index 93c186d861..b64d8f9e9f 100644
--- a/tasks/myosuite-dex-apple-lift-i00-standard/task.toml
+++ b/tasks/myosuite-dex-apple-lift-i00-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "myosuite" # MyoSuite 2.12.2 on MuJoCo 3.6.0 (CPU)
robot = "myohand"
family = "dex_apple_lift"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "the privileged twin's positive example (Claude Code + Claude Opus 5.5 (run `claude_code-claude_opus_5_5-default`, 2026-10-06), trial `myosuite-dex-apple-lift-i00-priv__j5qaRCZ`, 57 rows) rounded to float32, as the client sends it, and streamed through the client: the service scores it 1 live and in both replays, in two fresh trials (docs/benchmarks/myosuite.md); human review in PR #65"
diff --git a/tasks/myosuite-dex-bunny-inspect-i00-privileged/environment/rcb_myosuite.py b/tasks/myosuite-dex-bunny-inspect-i00-privileged/environment/rcb_myosuite.py
index 0f2fca5acd..930c0c73ba 100644
--- a/tasks/myosuite-dex-bunny-inspect-i00-privileged/environment/rcb_myosuite.py
+++ b/tasks/myosuite-dex-bunny-inspect-i00-privileged/environment/rcb_myosuite.py
@@ -465,6 +465,12 @@ class Session:
self.done, self.success, self.reason = True, True, "success"
elif failed:
self.done, self.reason = True, "failure: " + _FAILURE.get(self.task, "task terminated")
+ # Baoding's targets come from a precomputed trajectory of 1000 points (baoding_v1.create_goal_trajectory,
+ # "assumes that it's greater than env horizon"): with MyoSuite's step budget lifted, step 1001 would index past
+ # it (IndexError, board 0086). The episode ends there instead, as a failure unless it succeeded on that step.
+ n_goal = len(getattr(self.base, "goal", ())) if self.task == "baoding" else 0
+ if not self.done and n_goal and self.steps >= n_goal:
+ self.done, self.reason = True, f"failure: the targets' trajectory ends after {n_goal} control steps"
return not self.done
def _bimanual_contacts_ok(self) -> bool:
diff --git a/tasks/myosuite-dex-bunny-inspect-i00-privileged/task.toml b/tasks/myosuite-dex-bunny-inspect-i00-privileged/task.toml
index 90f0746d39..7bff04a75c 100644
--- a/tasks/myosuite-dex-bunny-inspect-i00-privileged/task.toml
+++ b/tasks/myosuite-dex-bunny-inspect-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "myosuite" # MyoSuite 2.12.2 on MuJoCo 3.6.0 (CPU)
robot = "myohand"
family = "dex_bunny_inspect"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "no positive example yet (two model attempts failed: Claude Code + Claude Opus 5.5 (run `claude_code-claude_opus_5_5-default`, 2026-10-06) and Codex + GPT-6 Luna, effort high, run `codex-gpt6_luna-high`); the task is reported separately until one exists (PROTOCOL §7, Open 4); human review in PR #65"
instance_seed = 0 # MyoHandStanfordbunnyInspect-v0: seed(0), then reset(seed=0)
diff --git a/tasks/myosuite-dex-bunny-inspect-i00-privileged/tests/rcb_myosuite.py b/tasks/myosuite-dex-bunny-inspect-i00-privileged/tests/rcb_myosuite.py
index 0f2fca5acd..930c0c73ba 100644
--- a/tasks/myosuite-dex-bunny-inspect-i00-privileged/tests/rcb_myosuite.py
+++ b/tasks/myosuite-dex-bunny-inspect-i00-privileged/tests/rcb_myosuite.py
@@ -465,6 +465,12 @@ class Session:
self.done, self.success, self.reason = True, True, "success"
elif failed:
self.done, self.reason = True, "failure: " + _FAILURE.get(self.task, "task terminated")
+ # Baoding's targets come from a precomputed trajectory of 1000 points (baoding_v1.create_goal_trajectory,
+ # "assumes that it's greater than env horizon"): with MyoSuite's step budget lifted, step 1001 would index past
+ # it (IndexError, board 0086). The episode ends there instead, as a failure unless it succeeded on that step.
+ n_goal = len(getattr(self.base, "goal", ())) if self.task == "baoding" else 0
+ if not self.done and n_goal and self.steps >= n_goal:
+ self.done, self.reason = True, f"failure: the targets' trajectory ends after {n_goal} control steps"
return not self.done
def _bimanual_contacts_ok(self) -> bool:
diff --git a/tasks/myosuite-dex-bunny-inspect-i00-standard/environment/sim/rcb_myosuite.py b/tasks/myosuite-dex-bunny-inspect-i00-standard/environment/sim/rcb_myosuite.py
index 0f2fca5acd..930c0c73ba 100644
--- a/tasks/myosuite-dex-bunny-inspect-i00-standard/environment/sim/rcb_myosuite.py
+++ b/tasks/myosuite-dex-bunny-inspect-i00-standard/environment/sim/rcb_myosuite.py
@@ -465,6 +465,12 @@ class Session:
self.done, self.success, self.reason = True, True, "success"
elif failed:
self.done, self.reason = True, "failure: " + _FAILURE.get(self.task, "task terminated")
+ # Baoding's targets come from a precomputed trajectory of 1000 points (baoding_v1.create_goal_trajectory,
+ # "assumes that it's greater than env horizon"): with MyoSuite's step budget lifted, step 1001 would index past
+ # it (IndexError, board 0086). The episode ends there instead, as a failure unless it succeeded on that step.
+ n_goal = len(getattr(self.base, "goal", ())) if self.task == "baoding" else 0
+ if not self.done and n_goal and self.steps >= n_goal:
+ self.done, self.reason = True, f"failure: the targets' trajectory ends after {n_goal} control steps"
return not self.done
def _bimanual_contacts_ok(self) -> bool:
diff --git a/tasks/myosuite-dex-bunny-inspect-i00-standard/task.toml b/tasks/myosuite-dex-bunny-inspect-i00-standard/task.toml
index 6b5b7dab46..dca551a1a6 100644
--- a/tasks/myosuite-dex-bunny-inspect-i00-standard/task.toml
+++ b/tasks/myosuite-dex-bunny-inspect-i00-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "myosuite" # MyoSuite 2.12.2 on MuJoCo 3.6.0 (CPU)
robot = "myohand"
family = "dex_bunny_inspect"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "no positive example yet (two model attempts failed: Claude Code + Claude Opus 5.5 (run `claude_code-claude_opus_5_5-default`, 2026-10-06) and Codex + GPT-6 Luna, effort high, run `codex-gpt6_luna-high`); the task is reported separately until one exists (PROTOCOL §7, Open 4); human review in PR #65"
diff --git a/tasks/myosuite-dex-cup-pour-i00-privileged/environment/rcb_myosuite.py b/tasks/myosuite-dex-cup-pour-i00-privileged/environment/rcb_myosuite.py
index 0f2fca5acd..930c0c73ba 100644
--- a/tasks/myosuite-dex-cup-pour-i00-privileged/environment/rcb_myosuite.py
+++ b/tasks/myosuite-dex-cup-pour-i00-privileged/environment/rcb_myosuite.py
@@ -465,6 +465,12 @@ class Session:
self.done, self.success, self.reason = True, True, "success"
elif failed:
self.done, self.reason = True, "failure: " + _FAILURE.get(self.task, "task terminated")
+ # Baoding's targets come from a precomputed trajectory of 1000 points (baoding_v1.create_goal_trajectory,
+ # "assumes that it's greater than env horizon"): with MyoSuite's step budget lifted, step 1001 would index past
+ # it (IndexError, board 0086). The episode ends there instead, as a failure unless it succeeded on that step.
+ n_goal = len(getattr(self.base, "goal", ())) if self.task == "baoding" else 0
+ if not self.done and n_goal and self.steps >= n_goal:
+ self.done, self.reason = True, f"failure: the targets' trajectory ends after {n_goal} control steps"
return not self.done
def _bimanual_contacts_ok(self) -> bool:
diff --git a/tasks/myosuite-dex-cup-pour-i00-privileged/task.toml b/tasks/myosuite-dex-cup-pour-i00-privileged/task.toml
index f2f7cdde72..cf0d0d5e60 100644
--- a/tasks/myosuite-dex-cup-pour-i00-privileged/task.toml
+++ b/tasks/myosuite-dex-cup-pour-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "myosuite" # MyoSuite 2.12.2 on MuJoCo 3.6.0 (CPU)
robot = "myohand"
family = "dex_cup_pour"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "no positive example yet (two model attempts failed: Claude Code + Claude Opus 5.5 (run `claude_code-claude_opus_5_5-default`, 2026-10-06) and Codex + GPT-6 Luna, effort high, run `codex-gpt6_luna-high`); the task is reported separately until one exists (PROTOCOL §7, Open 4); human review in PR #65"
instance_seed = 0 # MyoHandCupPour-v0: seed(0), then reset(seed=0)
diff --git a/tasks/myosuite-dex-cup-pour-i00-privileged/tests/rcb_myosuite.py b/tasks/myosuite-dex-cup-pour-i00-privileged/tests/rcb_myosuite.py
index 0f2fca5acd..930c0c73ba 100644
--- a/tasks/myosuite-dex-cup-pour-i00-privileged/tests/rcb_myosuite.py
+++ b/tasks/myosuite-dex-cup-pour-i00-privileged/tests/rcb_myosuite.py
@@ -465,6 +465,12 @@ class Session:
self.done, self.success, self.reason = True, True, "success"
elif failed:
self.done, self.reason = True, "failure: " + _FAILURE.get(self.task, "task terminated")
+ # Baoding's targets come from a precomputed trajectory of 1000 points (baoding_v1.create_goal_trajectory,
+ # "assumes that it's greater than env horizon"): with MyoSuite's step budget lifted, step 1001 would index past
+ # it (IndexError, board 0086). The episode ends there instead, as a failure unless it succeeded on that step.
+ n_goal = len(getattr(self.base, "goal", ())) if self.task == "baoding" else 0
+ if not self.done and n_goal and self.steps >= n_goal:
+ self.done, self.reason = True, f"failure: the targets' trajectory ends after {n_goal} control steps"
return not self.done
def _bimanual_contacts_ok(self) -> bool:
diff --git a/tasks/myosuite-dex-cup-pour-i00-standard/environment/sim/rcb_myosuite.py b/tasks/myosuite-dex-cup-pour-i00-standard/environment/sim/rcb_myosuite.py
index 0f2fca5acd..930c0c73ba 100644
--- a/tasks/myosuite-dex-cup-pour-i00-standard/environment/sim/rcb_myosuite.py
+++ b/tasks/myosuite-dex-cup-pour-i00-standard/environment/sim/rcb_myosuite.py
@@ -465,6 +465,12 @@ class Session:
self.done, self.success, self.reason = True, True, "success"
elif failed:
self.done, self.reason = True, "failure: " + _FAILURE.get(self.task, "task terminated")
+ # Baoding's targets come from a precomputed trajectory of 1000 points (baoding_v1.create_goal_trajectory,
+ # "assumes that it's greater than env horizon"): with MyoSuite's step budget lifted, step 1001 would index past
+ # it (IndexError, board 0086). The episode ends there instead, as a failure unless it succeeded on that step.
+ n_goal = len(getattr(self.base, "goal", ())) if self.task == "baoding" else 0
+ if not self.done and n_goal and self.steps >= n_goal:
+ self.done, self.reason = True, f"failure: the targets' trajectory ends after {n_goal} control steps"
return not self.done
def _bimanual_contacts_ok(self) -> bool:
diff --git a/tasks/myosuite-dex-cup-pour-i00-standard/task.toml b/tasks/myosuite-dex-cup-pour-i00-standard/task.toml
index dc5ac5f999..aede19e7fc 100644
--- a/tasks/myosuite-dex-cup-pour-i00-standard/task.toml
+++ b/tasks/myosuite-dex-cup-pour-i00-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "myosuite" # MyoSuite 2.12.2 on MuJoCo 3.6.0 (CPU)
robot = "myohand"
family = "dex_cup_pour"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "no positive example yet (two model attempts failed: Claude Code + Claude Opus 5.5 (run `claude_code-claude_opus_5_5-default`, 2026-10-06) and Codex + GPT-6 Luna, effort high, run `codex-gpt6_luna-high`); the task is reported separately until one exists (PROTOCOL §7, Open 4); human review in PR #65"
diff --git a/tasks/myosuite-dex-cylinder-pass-i00-privileged/environment/rcb_myosuite.py b/tasks/myosuite-dex-cylinder-pass-i00-privileged/environment/rcb_myosuite.py
index 0f2fca5acd..930c0c73ba 100644
--- a/tasks/myosuite-dex-cylinder-pass-i00-privileged/environment/rcb_myosuite.py
+++ b/tasks/myosuite-dex-cylinder-pass-i00-privileged/environment/rcb_myosuite.py
@@ -465,6 +465,12 @@ class Session:
self.done, self.success, self.reason = True, True, "success"
elif failed:
self.done, self.reason = True, "failure: " + _FAILURE.get(self.task, "task terminated")
+ # Baoding's targets come from a precomputed trajectory of 1000 points (baoding_v1.create_goal_trajectory,
+ # "assumes that it's greater than env horizon"): with MyoSuite's step budget lifted, step 1001 would index past
+ # it (IndexError, board 0086). The episode ends there instead, as a failure unless it succeeded on that step.
+ n_goal = len(getattr(self.base, "goal", ())) if self.task == "baoding" else 0
+ if not self.done and n_goal and self.steps >= n_goal:
+ self.done, self.reason = True, f"failure: the targets' trajectory ends after {n_goal} control steps"
return not self.done
def _bimanual_contacts_ok(self) -> bool:
diff --git a/tasks/myosuite-dex-cylinder-pass-i00-privileged/task.toml b/tasks/myosuite-dex-cylinder-pass-i00-privileged/task.toml
index e64593c875..adc95e91df 100644
--- a/tasks/myosuite-dex-cylinder-pass-i00-privileged/task.toml
+++ b/tasks/myosuite-dex-cylinder-pass-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "myosuite" # MyoSuite 2.12.2 on MuJoCo 3.6.0 (CPU)
robot = "myohand"
family = "dex_cylinder_pass"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "no positive example yet (two model attempts failed: Codex + GPT-6 Luna, effort high, run `codex-gpt6_luna-high` and Claude Code + Claude Opus 5.5 (run `claude_code-claude_opus_5_5-default`, 2026-10-06)); the task is reported separately until one exists (PROTOCOL §7, Open 4); human review in PR #65"
instance_seed = 0 # MyoHandCylindersmallPass-v0: seed(0), then reset(seed=0)
diff --git a/tasks/myosuite-dex-cylinder-pass-i00-privileged/tests/rcb_myosuite.py b/tasks/myosuite-dex-cylinder-pass-i00-privileged/tests/rcb_myosuite.py
index 0f2fca5acd..930c0c73ba 100644
--- a/tasks/myosuite-dex-cylinder-pass-i00-privileged/tests/rcb_myosuite.py
+++ b/tasks/myosuite-dex-cylinder-pass-i00-privileged/tests/rcb_myosuite.py
@@ -465,6 +465,12 @@ class Session:
self.done, self.success, self.reason = True, True, "success"
elif failed:
self.done, self.reason = True, "failure: " + _FAILURE.get(self.task, "task terminated")
+ # Baoding's targets come from a precomputed trajectory of 1000 points (baoding_v1.create_goal_trajectory,
+ # "assumes that it's greater than env horizon"): with MyoSuite's step budget lifted, step 1001 would index past
+ # it (IndexError, board 0086). The episode ends there instead, as a failure unless it succeeded on that step.
+ n_goal = len(getattr(self.base, "goal", ())) if self.task == "baoding" else 0
+ if not self.done and n_goal and self.steps >= n_goal:
+ self.done, self.reason = True, f"failure: the targets' trajectory ends after {n_goal} control steps"
return not self.done
def _bimanual_contacts_ok(self) -> bool:
diff --git a/tasks/myosuite-dex-cylinder-pass-i00-standard/environment/sim/rcb_myosuite.py b/tasks/myosuite-dex-cylinder-pass-i00-standard/environment/sim/rcb_myosuite.py
index 0f2fca5acd..930c0c73ba 100644
--- a/tasks/myosuite-dex-cylinder-pass-i00-standard/environment/sim/rcb_myosuite.py
+++ b/tasks/myosuite-dex-cylinder-pass-i00-standard/environment/sim/rcb_myosuite.py
@@ -465,6 +465,12 @@ class Session:
self.done, self.success, self.reason = True, True, "success"
elif failed:
self.done, self.reason = True, "failure: " + _FAILURE.get(self.task, "task terminated")
+ # Baoding's targets come from a precomputed trajectory of 1000 points (baoding_v1.create_goal_trajectory,
+ # "assumes that it's greater than env horizon"): with MyoSuite's step budget lifted, step 1001 would index past
+ # it (IndexError, board 0086). The episode ends there instead, as a failure unless it succeeded on that step.
+ n_goal = len(getattr(self.base, "goal", ())) if self.task == "baoding" else 0
+ if not self.done and n_goal and self.steps >= n_goal:
+ self.done, self.reason = True, f"failure: the targets' trajectory ends after {n_goal} control steps"
return not self.done
def _bimanual_contacts_ok(self) -> bool:
diff --git a/tasks/myosuite-dex-cylinder-pass-i00-standard/task.toml b/tasks/myosuite-dex-cylinder-pass-i00-standard/task.toml
index 824c1fa231..f0df4aea34 100644
--- a/tasks/myosuite-dex-cylinder-pass-i00-standard/task.toml
+++ b/tasks/myosuite-dex-cylinder-pass-i00-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "myosuite" # MyoSuite 2.12.2 on MuJoCo 3.6.0 (CPU)
robot = "myohand"
family = "dex_cylinder_pass"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "no positive example yet (two model attempts failed: Codex + GPT-6 Luna, effort high, run `codex-gpt6_luna-high` and Claude Code + Claude Opus 5.5 (run `claude_code-claude_opus_5_5-default`, 2026-10-06)); the task is reported separately until one exists (PROTOCOL §7, Open 4); human review in PR #65"
diff --git a/tasks/myosuite-dex-hammer-use-i00-privileged/environment/rcb_myosuite.py b/tasks/myosuite-dex-hammer-use-i00-privileged/environment/rcb_myosuite.py
index 0f2fca5acd..930c0c73ba 100644
--- a/tasks/myosuite-dex-hammer-use-i00-privileged/environment/rcb_myosuite.py
+++ b/tasks/myosuite-dex-hammer-use-i00-privileged/environment/rcb_myosuite.py
@@ -465,6 +465,12 @@ class Session:
self.done, self.success, self.reason = True, True, "success"
elif failed:
self.done, self.reason = True, "failure: " + _FAILURE.get(self.task, "task terminated")
+ # Baoding's targets come from a precomputed trajectory of 1000 points (baoding_v1.create_goal_trajectory,
+ # "assumes that it's greater than env horizon"): with MyoSuite's step budget lifted, step 1001 would index past
+ # it (IndexError, board 0086). The episode ends there instead, as a failure unless it succeeded on that step.
+ n_goal = len(getattr(self.base, "goal", ())) if self.task == "baoding" else 0
+ if not self.done and n_goal and self.steps >= n_goal:
+ self.done, self.reason = True, f"failure: the targets' trajectory ends after {n_goal} control steps"
return not self.done
def _bimanual_contacts_ok(self) -> bool:
diff --git a/tasks/myosuite-dex-hammer-use-i00-privileged/task.toml b/tasks/myosuite-dex-hammer-use-i00-privileged/task.toml
index 11fbda7431..c1da69cd50 100644
--- a/tasks/myosuite-dex-hammer-use-i00-privileged/task.toml
+++ b/tasks/myosuite-dex-hammer-use-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "myosuite" # MyoSuite 2.12.2 on MuJoCo 3.6.0 (CPU)
robot = "myohand"
family = "dex_hammer_use"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "no positive example yet (two model attempts failed: Codex + GPT-6 Luna, effort high, run `codex-gpt6_luna-high` and Claude Code + Claude Opus 5.5 (run `claude_code-claude_opus_5_5-default`, 2026-10-06)); the task is reported separately until one exists (PROTOCOL §7, Open 4); human review in PR #65"
instance_seed = 0 # MyoHandHammerUse-v0: seed(0), then reset(seed=0)
diff --git a/tasks/myosuite-dex-hammer-use-i00-privileged/tests/rcb_myosuite.py b/tasks/myosuite-dex-hammer-use-i00-privileged/tests/rcb_myosuite.py
index 0f2fca5acd..930c0c73ba 100644
--- a/tasks/myosuite-dex-hammer-use-i00-privileged/tests/rcb_myosuite.py
+++ b/tasks/myosuite-dex-hammer-use-i00-privileged/tests/rcb_myosuite.py
@@ -465,6 +465,12 @@ class Session:
self.done, self.success, self.reason = True, True, "success"
elif failed:
self.done, self.reason = True, "failure: " + _FAILURE.get(self.task, "task terminated")
+ # Baoding's targets come from a precomputed trajectory of 1000 points (baoding_v1.create_goal_trajectory,
+ # "assumes that it's greater than env horizon"): with MyoSuite's step budget lifted, step 1001 would index past
+ # it (IndexError, board 0086). The episode ends there instead, as a failure unless it succeeded on that step.
+ n_goal = len(getattr(self.base, "goal", ())) if self.task == "baoding" else 0
+ if not self.done and n_goal and self.steps >= n_goal:
+ self.done, self.reason = True, f"failure: the targets' trajectory ends after {n_goal} control steps"
return not self.done
def _bimanual_contacts_ok(self) -> bool:
diff --git a/tasks/myosuite-dex-hammer-use-i00-standard/environment/sim/rcb_myosuite.py b/tasks/myosuite-dex-hammer-use-i00-standard/environment/sim/rcb_myosuite.py
index 0f2fca5acd..930c0c73ba 100644
--- a/tasks/myosuite-dex-hammer-use-i00-standard/environment/sim/rcb_myosuite.py
+++ b/tasks/myosuite-dex-hammer-use-i00-standard/environment/sim/rcb_myosuite.py
@@ -465,6 +465,12 @@ class Session:
self.done, self.success, self.reason = True, True, "success"
elif failed:
self.done, self.reason = True, "failure: " + _FAILURE.get(self.task, "task terminated")
+ # Baoding's targets come from a precomputed trajectory of 1000 points (baoding_v1.create_goal_trajectory,
+ # "assumes that it's greater than env horizon"): with MyoSuite's step budget lifted, step 1001 would index past
+ # it (IndexError, board 0086). The episode ends there instead, as a failure unless it succeeded on that step.
+ n_goal = len(getattr(self.base, "goal", ())) if self.task == "baoding" else 0
+ if not self.done and n_goal and self.steps >= n_goal:
+ self.done, self.reason = True, f"failure: the targets' trajectory ends after {n_goal} control steps"
return not self.done
def _bimanual_contacts_ok(self) -> bool:
diff --git a/tasks/myosuite-dex-hammer-use-i00-standard/task.toml b/tasks/myosuite-dex-hammer-use-i00-standard/task.toml
index 8a3391faa1..55d74c56af 100644
--- a/tasks/myosuite-dex-hammer-use-i00-standard/task.toml
+++ b/tasks/myosuite-dex-hammer-use-i00-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "myosuite" # MyoSuite 2.12.2 on MuJoCo 3.6.0 (CPU)
robot = "myohand"
family = "dex_hammer_use"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "no positive example yet (two model attempts failed: Codex + GPT-6 Luna, effort high, run `codex-gpt6_luna-high` and Claude Code + Claude Opus 5.5 (run `claude_code-claude_opus_5_5-default`, 2026-10-06)); the task is reported separately until one exists (PROTOCOL §7, Open 4); human review in PR #65"
diff --git a/tasks/myosuite-die-reorient-i00-privileged/environment/rcb_myosuite.py b/tasks/myosuite-die-reorient-i00-privileged/environment/rcb_myosuite.py
index 0f2fca5acd..930c0c73ba 100644
--- a/tasks/myosuite-die-reorient-i00-privileged/environment/rcb_myosuite.py
+++ b/tasks/myosuite-die-reorient-i00-privileged/environment/rcb_myosuite.py
@@ -465,6 +465,12 @@ class Session:
self.done, self.success, self.reason = True, True, "success"
elif failed:
self.done, self.reason = True, "failure: " + _FAILURE.get(self.task, "task terminated")
+ # Baoding's targets come from a precomputed trajectory of 1000 points (baoding_v1.create_goal_trajectory,
+ # "assumes that it's greater than env horizon"): with MyoSuite's step budget lifted, step 1001 would index past
+ # it (IndexError, board 0086). The episode ends there instead, as a failure unless it succeeded on that step.
+ n_goal = len(getattr(self.base, "goal", ())) if self.task == "baoding" else 0
+ if not self.done and n_goal and self.steps >= n_goal:
+ self.done, self.reason = True, f"failure: the targets' trajectory ends after {n_goal} control steps"
return not self.done
def _bimanual_contacts_ok(self) -> bool:
diff --git a/tasks/myosuite-die-reorient-i00-privileged/task.toml b/tasks/myosuite-die-reorient-i00-privileged/task.toml
index bb8d528ce6..ee70fb9535 100644
--- a/tasks/myosuite-die-reorient-i00-privileged/task.toml
+++ b/tasks/myosuite-die-reorient-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "myosuite" # MyoSuite 2.12.2 on MuJoCo 3.6.0 (CPU)
robot = "myohand"
family = "die_reorient"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "a successful model trial on this frozen instance: Claude Code + Claude Opus 5.5 (run `claude_code-claude_opus_5_5-default`, 2026-09-28), trial `myosuite-die-reorient-i00__EtyJ3X7` (225 rows); today's verifier scores its trajectory 1 in two fresh trials (docs/benchmarks/myosuite.md); human review in PR #65"
instance_seed = 0 # myoChallengeDieReorientP1-v0: seed(0), then reset(seed=0)
diff --git a/tasks/myosuite-die-reorient-i00-privileged/tests/rcb_myosuite.py b/tasks/myosuite-die-reorient-i00-privileged/tests/rcb_myosuite.py
index 0f2fca5acd..930c0c73ba 100644
--- a/tasks/myosuite-die-reorient-i00-privileged/tests/rcb_myosuite.py
+++ b/tasks/myosuite-die-reorient-i00-privileged/tests/rcb_myosuite.py
@@ -465,6 +465,12 @@ class Session:
self.done, self.success, self.reason = True, True, "success"
elif failed:
self.done, self.reason = True, "failure: " + _FAILURE.get(self.task, "task terminated")
+ # Baoding's targets come from a precomputed trajectory of 1000 points (baoding_v1.create_goal_trajectory,
+ # "assumes that it's greater than env horizon"): with MyoSuite's step budget lifted, step 1001 would index past
+ # it (IndexError, board 0086). The episode ends there instead, as a failure unless it succeeded on that step.
+ n_goal = len(getattr(self.base, "goal", ())) if self.task == "baoding" else 0
+ if not self.done and n_goal and self.steps >= n_goal:
+ self.done, self.reason = True, f"failure: the targets' trajectory ends after {n_goal} control steps"
return not self.done
def _bimanual_contacts_ok(self) -> bool:
diff --git a/tasks/myosuite-die-reorient-i00-standard/environment/sim/rcb_myosuite.py b/tasks/myosuite-die-reorient-i00-standard/environment/sim/rcb_myosuite.py
index 0f2fca5acd..930c0c73ba 100644
--- a/tasks/myosuite-die-reorient-i00-standard/environment/sim/rcb_myosuite.py
+++ b/tasks/myosuite-die-reorient-i00-standard/environment/sim/rcb_myosuite.py
@@ -465,6 +465,12 @@ class Session:
self.done, self.success, self.reason = True, True, "success"
elif failed:
self.done, self.reason = True, "failure: " + _FAILURE.get(self.task, "task terminated")
+ # Baoding's targets come from a precomputed trajectory of 1000 points (baoding_v1.create_goal_trajectory,
+ # "assumes that it's greater than env horizon"): with MyoSuite's step budget lifted, step 1001 would index past
+ # it (IndexError, board 0086). The episode ends there instead, as a failure unless it succeeded on that step.
+ n_goal = len(getattr(self.base, "goal", ())) if self.task == "baoding" else 0
+ if not self.done and n_goal and self.steps >= n_goal:
+ self.done, self.reason = True, f"failure: the targets' trajectory ends after {n_goal} control steps"
return not self.done
def _bimanual_contacts_ok(self) -> bool:
diff --git a/tasks/myosuite-die-reorient-i00-standard/task.toml b/tasks/myosuite-die-reorient-i00-standard/task.toml
index 61e246301c..29e32ff993 100644
--- a/tasks/myosuite-die-reorient-i00-standard/task.toml
+++ b/tasks/myosuite-die-reorient-i00-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "myosuite" # MyoSuite 2.12.2 on MuJoCo 3.6.0 (CPU)
robot = "myohand"
family = "die_reorient"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "the privileged twin's positive example (Claude Code + Claude Opus 5.5 (run `claude_code-claude_opus_5_5-default`, 2026-09-28), trial `myosuite-die-reorient-i00__EtyJ3X7`, 225 rows) rounded to float32, as the client sends it, and streamed through the client: the service scores it 1 live and in both replays, in two fresh trials (docs/benchmarks/myosuite.md); human review in PR #65"
diff --git a/tasks/myosuite-hand-pose-i00-privileged/environment/rcb_myosuite.py b/tasks/myosuite-hand-pose-i00-privileged/environment/rcb_myosuite.py
index 0f2fca5acd..930c0c73ba 100644
--- a/tasks/myosuite-hand-pose-i00-privileged/environment/rcb_myosuite.py
+++ b/tasks/myosuite-hand-pose-i00-privileged/environment/rcb_myosuite.py
@@ -465,6 +465,12 @@ class Session:
self.done, self.success, self.reason = True, True, "success"
elif failed:
self.done, self.reason = True, "failure: " + _FAILURE.get(self.task, "task terminated")
+ # Baoding's targets come from a precomputed trajectory of 1000 points (baoding_v1.create_goal_trajectory,
+ # "assumes that it's greater than env horizon"): with MyoSuite's step budget lifted, step 1001 would index past
+ # it (IndexError, board 0086). The episode ends there instead, as a failure unless it succeeded on that step.
+ n_goal = len(getattr(self.base, "goal", ())) if self.task == "baoding" else 0
+ if not self.done and n_goal and self.steps >= n_goal:
+ self.done, self.reason = True, f"failure: the targets' trajectory ends after {n_goal} control steps"
return not self.done
def _bimanual_contacts_ok(self) -> bool:
diff --git a/tasks/myosuite-hand-pose-i00-privileged/task.toml b/tasks/myosuite-hand-pose-i00-privileged/task.toml
index ab7595f7f5..781d397d2c 100644
--- a/tasks/myosuite-hand-pose-i00-privileged/task.toml
+++ b/tasks/myosuite-hand-pose-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "myosuite" # MyoSuite 2.12.2 on MuJoCo 3.6.0 (CPU)
robot = "myohand"
family = "hand_pose"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "a successful model trial on this frozen instance: Codex + GPT-6 Luna (run `codex-gpt6_luna-high`, via an OpenRouter adapter before protocol v1.0, 2026-09-29), trial `myosuite-hand-pose-i00__fC7US9G` (360 rows); today's verifier scores its trajectory 1 in two fresh trials (docs/benchmarks/myosuite.md); human review in PR #65"
instance_seed = 0 # myoHandPoseFixed-v0: seed(0), then reset(seed=0)
diff --git a/tasks/myosuite-hand-pose-i00-privileged/tests/rcb_myosuite.py b/tasks/myosuite-hand-pose-i00-privileged/tests/rcb_myosuite.py
index 0f2fca5acd..930c0c73ba 100644
--- a/tasks/myosuite-hand-pose-i00-privileged/tests/rcb_myosuite.py
+++ b/tasks/myosuite-hand-pose-i00-privileged/tests/rcb_myosuite.py
@@ -465,6 +465,12 @@ class Session:
self.done, self.success, self.reason = True, True, "success"
elif failed:
self.done, self.reason = True, "failure: " + _FAILURE.get(self.task, "task terminated")
+ # Baoding's targets come from a precomputed trajectory of 1000 points (baoding_v1.create_goal_trajectory,
+ # "assumes that it's greater than env horizon"): with MyoSuite's step budget lifted, step 1001 would index past
+ # it (IndexError, board 0086). The episode ends there instead, as a failure unless it succeeded on that step.
+ n_goal = len(getattr(self.base, "goal", ())) if self.task == "baoding" else 0
+ if not self.done and n_goal and self.steps >= n_goal:
+ self.done, self.reason = True, f"failure: the targets' trajectory ends after {n_goal} control steps"
return not self.done
def _bimanual_contacts_ok(self) -> bool:
diff --git a/tasks/myosuite-hand-pose-i00-standard/environment/sim/rcb_myosuite.py b/tasks/myosuite-hand-pose-i00-standard/environment/sim/rcb_myosuite.py
index 0f2fca5acd..930c0c73ba 100644
--- a/tasks/myosuite-hand-pose-i00-standard/environment/sim/rcb_myosuite.py
+++ b/tasks/myosuite-hand-pose-i00-standard/environment/sim/rcb_myosuite.py
@@ -465,6 +465,12 @@ class Session:
self.done, self.success, self.reason = True, True, "success"
elif failed:
self.done, self.reason = True, "failure: " + _FAILURE.get(self.task, "task terminated")
+ # Baoding's targets come from a precomputed trajectory of 1000 points (baoding_v1.create_goal_trajectory,
+ # "assumes that it's greater than env horizon"): with MyoSuite's step budget lifted, step 1001 would index past
+ # it (IndexError, board 0086). The episode ends there instead, as a failure unless it succeeded on that step.
+ n_goal = len(getattr(self.base, "goal", ())) if self.task == "baoding" else 0
+ if not self.done and n_goal and self.steps >= n_goal:
+ self.done, self.reason = True, f"failure: the targets' trajectory ends after {n_goal} control steps"
return not self.done
def _bimanual_contacts_ok(self) -> bool:
diff --git a/tasks/myosuite-hand-pose-i00-standard/task.toml b/tasks/myosuite-hand-pose-i00-standard/task.toml
index ddbc932d72..57e34e9b2d 100644
--- a/tasks/myosuite-hand-pose-i00-standard/task.toml
+++ b/tasks/myosuite-hand-pose-i00-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "myosuite" # MyoSuite 2.12.2 on MuJoCo 3.6.0 (CPU)
robot = "myohand"
family = "hand_pose"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "the privileged twin's positive example (Codex + GPT-6 Luna (run `codex-gpt6_luna-high`, via an OpenRouter adapter before protocol v1.0, 2026-09-29), trial `myosuite-hand-pose-i00__fC7US9G`, 360 rows) rounded to float32, as the client sends it, and streamed through the client: the service scores it 1 live and in both replays, in two fresh trials (docs/benchmarks/myosuite.md); human review in PR #65"
diff --git a/tasks/myosuite-hand-reach-i00-privileged/environment/rcb_myosuite.py b/tasks/myosuite-hand-reach-i00-privileged/environment/rcb_myosuite.py
index 0f2fca5acd..930c0c73ba 100644
--- a/tasks/myosuite-hand-reach-i00-privileged/environment/rcb_myosuite.py
+++ b/tasks/myosuite-hand-reach-i00-privileged/environment/rcb_myosuite.py
@@ -465,6 +465,12 @@ class Session:
self.done, self.success, self.reason = True, True, "success"
elif failed:
self.done, self.reason = True, "failure: " + _FAILURE.get(self.task, "task terminated")
+ # Baoding's targets come from a precomputed trajectory of 1000 points (baoding_v1.create_goal_trajectory,
+ # "assumes that it's greater than env horizon"): with MyoSuite's step budget lifted, step 1001 would index past
+ # it (IndexError, board 0086). The episode ends there instead, as a failure unless it succeeded on that step.
+ n_goal = len(getattr(self.base, "goal", ())) if self.task == "baoding" else 0
+ if not self.done and n_goal and self.steps >= n_goal:
+ self.done, self.reason = True, f"failure: the targets' trajectory ends after {n_goal} control steps"
return not self.done
def _bimanual_contacts_ok(self) -> bool:
diff --git a/tasks/myosuite-hand-reach-i00-privileged/task.toml b/tasks/myosuite-hand-reach-i00-privileged/task.toml
index 904674e390..a460527ca1 100644
--- a/tasks/myosuite-hand-reach-i00-privileged/task.toml
+++ b/tasks/myosuite-hand-reach-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "myosuite" # MyoSuite 2.12.2 on MuJoCo 3.6.0 (CPU)
robot = "myohand"
family = "hand_reach"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "a successful model trial on this frozen instance: Codex + GPT-6 Luna (run `codex-gpt6_luna-high`, via an OpenRouter adapter before protocol v1.0, 2026-09-29), trial `myosuite-hand-reach-i00__w2ryUU7` (70 rows); today's verifier scores its trajectory 1 in two fresh trials (docs/benchmarks/myosuite.md); human review in PR #65"
instance_seed = 0 # myoHandReachFixed-v0: seed(0), then reset(seed=0)
diff --git a/tasks/myosuite-hand-reach-i00-privileged/tests/rcb_myosuite.py b/tasks/myosuite-hand-reach-i00-privileged/tests/rcb_myosuite.py
index 0f2fca5acd..930c0c73ba 100644
--- a/tasks/myosuite-hand-reach-i00-privileged/tests/rcb_myosuite.py
+++ b/tasks/myosuite-hand-reach-i00-privileged/tests/rcb_myosuite.py
@@ -465,6 +465,12 @@ class Session:
self.done, self.success, self.reason = True, True, "success"
elif failed:
self.done, self.reason = True, "failure: " + _FAILURE.get(self.task, "task terminated")
+ # Baoding's targets come from a precomputed trajectory of 1000 points (baoding_v1.create_goal_trajectory,
+ # "assumes that it's greater than env horizon"): with MyoSuite's step budget lifted, step 1001 would index past
+ # it (IndexError, board 0086). The episode ends there instead, as a failure unless it succeeded on that step.
+ n_goal = len(getattr(self.base, "goal", ())) if self.task == "baoding" else 0
+ if not self.done and n_goal and self.steps >= n_goal:
+ self.done, self.reason = True, f"failure: the targets' trajectory ends after {n_goal} control steps"
return not self.done
def _bimanual_contacts_ok(self) -> bool:
diff --git a/tasks/myosuite-hand-reach-i00-standard/environment/sim/rcb_myosuite.py b/tasks/myosuite-hand-reach-i00-standard/environment/sim/rcb_myosuite.py
index 0f2fca5acd..930c0c73ba 100644
--- a/tasks/myosuite-hand-reach-i00-standard/environment/sim/rcb_myosuite.py
+++ b/tasks/myosuite-hand-reach-i00-standard/environment/sim/rcb_myosuite.py
@@ -465,6 +465,12 @@ class Session:
self.done, self.success, self.reason = True, True, "success"
elif failed:
self.done, self.reason = True, "failure: " + _FAILURE.get(self.task, "task terminated")
+ # Baoding's targets come from a precomputed trajectory of 1000 points (baoding_v1.create_goal_trajectory,
+ # "assumes that it's greater than env horizon"): with MyoSuite's step budget lifted, step 1001 would index past
+ # it (IndexError, board 0086). The episode ends there instead, as a failure unless it succeeded on that step.
+ n_goal = len(getattr(self.base, "goal", ())) if self.task == "baoding" else 0
+ if not self.done and n_goal and self.steps >= n_goal:
+ self.done, self.reason = True, f"failure: the targets' trajectory ends after {n_goal} control steps"
return not self.done
def _bimanual_contacts_ok(self) -> bool:
diff --git a/tasks/myosuite-hand-reach-i00-standard/task.toml b/tasks/myosuite-hand-reach-i00-standard/task.toml
index 5dea73bfab..8766d6c429 100644
--- a/tasks/myosuite-hand-reach-i00-standard/task.toml
+++ b/tasks/myosuite-hand-reach-i00-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "myosuite" # MyoSuite 2.12.2 on MuJoCo 3.6.0 (CPU)
robot = "myohand"
family = "hand_reach"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "the privileged twin's positive example (Codex + GPT-6 Luna (run `codex-gpt6_luna-high`, via an OpenRouter adapter before protocol v1.0, 2026-09-29), trial `myosuite-hand-reach-i00__w2ryUU7`, 70 rows) rounded to float32, as the client sends it, and streamed through the client: the service scores it 1 live and in both replays, in two fresh trials (docs/benchmarks/myosuite.md); human review in PR #65"
diff --git a/tasks/myosuite-key-turn-i00-privileged/environment/rcb_myosuite.py b/tasks/myosuite-key-turn-i00-privileged/environment/rcb_myosuite.py
index 0f2fca5acd..930c0c73ba 100644
--- a/tasks/myosuite-key-turn-i00-privileged/environment/rcb_myosuite.py
+++ b/tasks/myosuite-key-turn-i00-privileged/environment/rcb_myosuite.py
@@ -465,6 +465,12 @@ class Session:
self.done, self.success, self.reason = True, True, "success"
elif failed:
self.done, self.reason = True, "failure: " + _FAILURE.get(self.task, "task terminated")
+ # Baoding's targets come from a precomputed trajectory of 1000 points (baoding_v1.create_goal_trajectory,
+ # "assumes that it's greater than env horizon"): with MyoSuite's step budget lifted, step 1001 would index past
+ # it (IndexError, board 0086). The episode ends there instead, as a failure unless it succeeded on that step.
+ n_goal = len(getattr(self.base, "goal", ())) if self.task == "baoding" else 0
+ if not self.done and n_goal and self.steps >= n_goal:
+ self.done, self.reason = True, f"failure: the targets' trajectory ends after {n_goal} control steps"
return not self.done
def _bimanual_contacts_ok(self) -> bool:
diff --git a/tasks/myosuite-key-turn-i00-privileged/task.toml b/tasks/myosuite-key-turn-i00-privileged/task.toml
index 702fe0ffba..dcc680fbfc 100644
--- a/tasks/myosuite-key-turn-i00-privileged/task.toml
+++ b/tasks/myosuite-key-turn-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "myosuite" # MyoSuite 2.12.2 on MuJoCo 3.6.0 (CPU)
robot = "myohand"
family = "key_turn"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 0 # myoHandKeyTurnFixed-v0: seed(0), then reset(seed=0)
expert_time_estimate_hours = 1.5
diff --git a/tasks/myosuite-key-turn-i00-privileged/tests/rcb_myosuite.py b/tasks/myosuite-key-turn-i00-privileged/tests/rcb_myosuite.py
index 0f2fca5acd..930c0c73ba 100644
--- a/tasks/myosuite-key-turn-i00-privileged/tests/rcb_myosuite.py
+++ b/tasks/myosuite-key-turn-i00-privileged/tests/rcb_myosuite.py
@@ -465,6 +465,12 @@ class Session:
self.done, self.success, self.reason = True, True, "success"
elif failed:
self.done, self.reason = True, "failure: " + _FAILURE.get(self.task, "task terminated")
+ # Baoding's targets come from a precomputed trajectory of 1000 points (baoding_v1.create_goal_trajectory,
+ # "assumes that it's greater than env horizon"): with MyoSuite's step budget lifted, step 1001 would index past
+ # it (IndexError, board 0086). The episode ends there instead, as a failure unless it succeeded on that step.
+ n_goal = len(getattr(self.base, "goal", ())) if self.task == "baoding" else 0
+ if not self.done and n_goal and self.steps >= n_goal:
+ self.done, self.reason = True, f"failure: the targets' trajectory ends after {n_goal} control steps"
return not self.done
def _bimanual_contacts_ok(self) -> bool:
diff --git a/tasks/myosuite-key-turn-i00-standard/environment/sim/rcb_myosuite.py b/tasks/myosuite-key-turn-i00-standard/environment/sim/rcb_myosuite.py
index 0f2fca5acd..930c0c73ba 100644
--- a/tasks/myosuite-key-turn-i00-standard/environment/sim/rcb_myosuite.py
+++ b/tasks/myosuite-key-turn-i00-standard/environment/sim/rcb_myosuite.py
@@ -465,6 +465,12 @@ class Session:
self.done, self.success, self.reason = True, True, "success"
elif failed:
self.done, self.reason = True, "failure: " + _FAILURE.get(self.task, "task terminated")
+ # Baoding's targets come from a precomputed trajectory of 1000 points (baoding_v1.create_goal_trajectory,
+ # "assumes that it's greater than env horizon"): with MyoSuite's step budget lifted, step 1001 would index past
+ # it (IndexError, board 0086). The episode ends there instead, as a failure unless it succeeded on that step.
+ n_goal = len(getattr(self.base, "goal", ())) if self.task == "baoding" else 0
+ if not self.done and n_goal and self.steps >= n_goal:
+ self.done, self.reason = True, f"failure: the targets' trajectory ends after {n_goal} control steps"
return not self.done
def _bimanual_contacts_ok(self) -> bool:
diff --git a/tasks/myosuite-key-turn-i00-standard/task.toml b/tasks/myosuite-key-turn-i00-standard/task.toml
index 5ae6a8b707..41cdaed89b 100644
--- a/tasks/myosuite-key-turn-i00-standard/task.toml
+++ b/tasks/myosuite-key-turn-i00-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "myosuite" # MyoSuite 2.12.2 on MuJoCo 3.6.0 (CPU)
robot = "myohand"
family = "key_turn"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "full"
instance_seed = 0 # myoHandKeyTurnFixed-v0: seed(0), then reset(seed=0)
diff --git a/tasks/myosuite-obj-hold-i00-privileged/environment/rcb_myosuite.py b/tasks/myosuite-obj-hold-i00-privileged/environment/rcb_myosuite.py
index 0f2fca5acd..930c0c73ba 100644
--- a/tasks/myosuite-obj-hold-i00-privileged/environment/rcb_myosuite.py
+++ b/tasks/myosuite-obj-hold-i00-privileged/environment/rcb_myosuite.py
@@ -465,6 +465,12 @@ class Session:
self.done, self.success, self.reason = True, True, "success"
elif failed:
self.done, self.reason = True, "failure: " + _FAILURE.get(self.task, "task terminated")
+ # Baoding's targets come from a precomputed trajectory of 1000 points (baoding_v1.create_goal_trajectory,
+ # "assumes that it's greater than env horizon"): with MyoSuite's step budget lifted, step 1001 would index past
+ # it (IndexError, board 0086). The episode ends there instead, as a failure unless it succeeded on that step.
+ n_goal = len(getattr(self.base, "goal", ())) if self.task == "baoding" else 0
+ if not self.done and n_goal and self.steps >= n_goal:
+ self.done, self.reason = True, f"failure: the targets' trajectory ends after {n_goal} control steps"
return not self.done
def _bimanual_contacts_ok(self) -> bool:
diff --git a/tasks/myosuite-obj-hold-i00-privileged/task.toml b/tasks/myosuite-obj-hold-i00-privileged/task.toml
index f9a515ee0c..c626deb68b 100644
--- a/tasks/myosuite-obj-hold-i00-privileged/task.toml
+++ b/tasks/myosuite-obj-hold-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "myosuite" # MyoSuite 2.12.2 on MuJoCo 3.6.0 (CPU)
robot = "myohand"
family = "obj_hold"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 0 # myoHandObjHoldFixed-v0: seed(0), then reset(seed=0)
expert_time_estimate_hours = 1.5
diff --git a/tasks/myosuite-obj-hold-i00-privileged/tests/rcb_myosuite.py b/tasks/myosuite-obj-hold-i00-privileged/tests/rcb_myosuite.py
index 0f2fca5acd..930c0c73ba 100644
--- a/tasks/myosuite-obj-hold-i00-privileged/tests/rcb_myosuite.py
+++ b/tasks/myosuite-obj-hold-i00-privileged/tests/rcb_myosuite.py
@@ -465,6 +465,12 @@ class Session:
self.done, self.success, self.reason = True, True, "success"
elif failed:
self.done, self.reason = True, "failure: " + _FAILURE.get(self.task, "task terminated")
+ # Baoding's targets come from a precomputed trajectory of 1000 points (baoding_v1.create_goal_trajectory,
+ # "assumes that it's greater than env horizon"): with MyoSuite's step budget lifted, step 1001 would index past
+ # it (IndexError, board 0086). The episode ends there instead, as a failure unless it succeeded on that step.
+ n_goal = len(getattr(self.base, "goal", ())) if self.task == "baoding" else 0
+ if not self.done and n_goal and self.steps >= n_goal:
+ self.done, self.reason = True, f"failure: the targets' trajectory ends after {n_goal} control steps"
return not self.done
def _bimanual_contacts_ok(self) -> bool:
diff --git a/tasks/myosuite-obj-hold-i00-standard/environment/sim/rcb_myosuite.py b/tasks/myosuite-obj-hold-i00-standard/environment/sim/rcb_myosuite.py
index 0f2fca5acd..930c0c73ba 100644
--- a/tasks/myosuite-obj-hold-i00-standard/environment/sim/rcb_myosuite.py
+++ b/tasks/myosuite-obj-hold-i00-standard/environment/sim/rcb_myosuite.py
@@ -465,6 +465,12 @@ class Session:
self.done, self.success, self.reason = True, True, "success"
elif failed:
self.done, self.reason = True, "failure: " + _FAILURE.get(self.task, "task terminated")
+ # Baoding's targets come from a precomputed trajectory of 1000 points (baoding_v1.create_goal_trajectory,
+ # "assumes that it's greater than env horizon"): with MyoSuite's step budget lifted, step 1001 would index past
+ # it (IndexError, board 0086). The episode ends there instead, as a failure unless it succeeded on that step.
+ n_goal = len(getattr(self.base, "goal", ())) if self.task == "baoding" else 0
+ if not self.done and n_goal and self.steps >= n_goal:
+ self.done, self.reason = True, f"failure: the targets' trajectory ends after {n_goal} control steps"
return not self.done
def _bimanual_contacts_ok(self) -> bool:
diff --git a/tasks/myosuite-obj-hold-i00-standard/task.toml b/tasks/myosuite-obj-hold-i00-standard/task.toml
index 41bb046632..6c71c530fd 100644
--- a/tasks/myosuite-obj-hold-i00-standard/task.toml
+++ b/tasks/myosuite-obj-hold-i00-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "myosuite" # MyoSuite 2.12.2 on MuJoCo 3.6.0 (CPU)
robot = "myohand"
family = "obj_hold"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "full"
instance_seed = 0 # myoHandObjHoldFixed-v0: seed(0), then reset(seed=0)
diff --git a/tasks/myosuite-obj-reorient-i00-privileged/environment/rcb_myosuite.py b/tasks/myosuite-obj-reorient-i00-privileged/environment/rcb_myosuite.py
index 0f2fca5acd..930c0c73ba 100644
--- a/tasks/myosuite-obj-reorient-i00-privileged/environment/rcb_myosuite.py
+++ b/tasks/myosuite-obj-reorient-i00-privileged/environment/rcb_myosuite.py
@@ -465,6 +465,12 @@ class Session:
self.done, self.success, self.reason = True, True, "success"
elif failed:
self.done, self.reason = True, "failure: " + _FAILURE.get(self.task, "task terminated")
+ # Baoding's targets come from a precomputed trajectory of 1000 points (baoding_v1.create_goal_trajectory,
+ # "assumes that it's greater than env horizon"): with MyoSuite's step budget lifted, step 1001 would index past
+ # it (IndexError, board 0086). The episode ends there instead, as a failure unless it succeeded on that step.
+ n_goal = len(getattr(self.base, "goal", ())) if self.task == "baoding" else 0
+ if not self.done and n_goal and self.steps >= n_goal:
+ self.done, self.reason = True, f"failure: the targets' trajectory ends after {n_goal} control steps"
return not self.done
def _bimanual_contacts_ok(self) -> bool:
diff --git a/tasks/myosuite-obj-reorient-i00-privileged/task.toml b/tasks/myosuite-obj-reorient-i00-privileged/task.toml
index 284496966b..8c4d6bb543 100644
--- a/tasks/myosuite-obj-reorient-i00-privileged/task.toml
+++ b/tasks/myosuite-obj-reorient-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "myosuite" # MyoSuite 2.12.2 on MuJoCo 3.6.0 (CPU)
robot = "myohand"
family = "obj_reorient"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "a successful model trial on this frozen instance: Codex + GPT-6 Luna (run `codex-gpt6_luna-high`, via an OpenRouter adapter before protocol v1.0, 2026-09-29), trial `myosuite-obj-reorient-i00__5dLP3ah` (305 rows); today's verifier scores its trajectory 1 in two fresh trials (docs/benchmarks/myosuite.md); human review in PR #65"
instance_seed = 0 # myoHandReorient8-v0: seed(0), then reset(seed=0)
diff --git a/tasks/myosuite-obj-reorient-i00-privileged/tests/rcb_myosuite.py b/tasks/myosuite-obj-reorient-i00-privileged/tests/rcb_myosuite.py
index 0f2fca5acd..930c0c73ba 100644
--- a/tasks/myosuite-obj-reorient-i00-privileged/tests/rcb_myosuite.py
+++ b/tasks/myosuite-obj-reorient-i00-privileged/tests/rcb_myosuite.py
@@ -465,6 +465,12 @@ class Session:
self.done, self.success, self.reason = True, True, "success"
elif failed:
self.done, self.reason = True, "failure: " + _FAILURE.get(self.task, "task terminated")
+ # Baoding's targets come from a precomputed trajectory of 1000 points (baoding_v1.create_goal_trajectory,
+ # "assumes that it's greater than env horizon"): with MyoSuite's step budget lifted, step 1001 would index past
+ # it (IndexError, board 0086). The episode ends there instead, as a failure unless it succeeded on that step.
+ n_goal = len(getattr(self.base, "goal", ())) if self.task == "baoding" else 0
+ if not self.done and n_goal and self.steps >= n_goal:
+ self.done, self.reason = True, f"failure: the targets' trajectory ends after {n_goal} control steps"
return not self.done
def _bimanual_contacts_ok(self) -> bool:
diff --git a/tasks/myosuite-obj-reorient-i00-standard/environment/sim/rcb_myosuite.py b/tasks/myosuite-obj-reorient-i00-standard/environment/sim/rcb_myosuite.py
index 0f2fca5acd..930c0c73ba 100644
--- a/tasks/myosuite-obj-reorient-i00-standard/environment/sim/rcb_myosuite.py
+++ b/tasks/myosuite-obj-reorient-i00-standard/environment/sim/rcb_myosuite.py
@@ -465,6 +465,12 @@ class Session:
self.done, self.success, self.reason = True, True, "success"
elif failed:
self.done, self.reason = True, "failure: " + _FAILURE.get(self.task, "task terminated")
+ # Baoding's targets come from a precomputed trajectory of 1000 points (baoding_v1.create_goal_trajectory,
+ # "assumes that it's greater than env horizon"): with MyoSuite's step budget lifted, step 1001 would index past
+ # it (IndexError, board 0086). The episode ends there instead, as a failure unless it succeeded on that step.
+ n_goal = len(getattr(self.base, "goal", ())) if self.task == "baoding" else 0
+ if not self.done and n_goal and self.steps >= n_goal:
+ self.done, self.reason = True, f"failure: the targets' trajectory ends after {n_goal} control steps"
return not self.done
def _bimanual_contacts_ok(self) -> bool:
diff --git a/tasks/myosuite-obj-reorient-i00-standard/task.toml b/tasks/myosuite-obj-reorient-i00-standard/task.toml
index d4a391de27..126466a151 100644
--- a/tasks/myosuite-obj-reorient-i00-standard/task.toml
+++ b/tasks/myosuite-obj-reorient-i00-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "myosuite" # MyoSuite 2.12.2 on MuJoCo 3.6.0 (CPU)
robot = "myohand"
family = "obj_reorient"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "the privileged twin's positive example (Codex + GPT-6 Luna (run `codex-gpt6_luna-high`, via an OpenRouter adapter before protocol v1.0, 2026-09-29), trial `myosuite-obj-reorient-i00__5dLP3ah`, 305 rows) rounded to float32, as the client sends it, and streamed through the client: the service scores it 1 live and in both replays, in two fresh trials (docs/benchmarks/myosuite.md); human review in PR #65"
diff --git a/tasks/myosuite-osl-run-i00-privileged/README.md b/tasks/myosuite-osl-run-i00-privileged/README.md
index 1745648224..8072f63567 100644
--- a/tasks/myosuite-osl-run-i00-privileged/README.md
+++ b/tasks/myosuite-osl-run-i00-privileged/README.md
@@ -12,8 +12,8 @@
| Base image | `ghcr.io/mll-lab-nu/eai-myosuite:0.1.1` |
## Task
-MyoSuite `myoChallengeOslRunFixed-v0` (MyoLeg + OSL, a free-standing musculoskeletal model of a human whose left leg is an Open-Source Leg knee-ankle prosthesis, 54 muscles and 2 other actuators): Run the track: move forward 29 m along it to the finish (MyoChallenge 2024 OSL Run, its own win condition), staying within 1 m of the track's centre line and on your feet; the left leg is the prosthesis, driven by its own controller, so you command only the 54 muscles. Failure (the episode ends): falling, leaving the track sideways, or moving 2 m backwards. There is no time limit.
-A musculoskeletal body whose left leg is an Open-Source Leg prosthesis (knee and ankle motors driven by its own controller) stands at the start of a straight, flat running track. The deliverable is one replayable trajectory of actuator commands (`/app/output/trajectory.npz`, `(T, 56)`, one row per 5 ms control step).
+MyoSuite `myoChallengeOslRunFixed-v0` (MyoLeg + OSL, a free-standing musculoskeletal model of a human whose right leg is an Open-Source Leg knee-ankle prosthesis, 54 muscles and 2 other actuators): Run the track: move forward 29 m along it to the finish (MyoChallenge 2024 OSL Run, its own win condition), staying within 1 m of the track's centre line and on your feet; the right leg is the prosthesis (`osl_knee_angle_r`, `osl_ankle_angle_r`), driven by its own controller, so you command only the 54 muscles. Failure (the episode ends): falling, leaving the track sideways, or moving 2 m backwards. There is no time limit.
+A musculoskeletal body whose right leg is an Open-Source Leg prosthesis (knee and ankle motors driven by its own controller) stands at the start of a straight, flat running track. The deliverable is one replayable trajectory of actuator commands (`/app/output/trajectory.npz`, `(T, 56)`, one row per 5 ms control step).
## Difficulty
First-pass rating: hard. The action is muscle-level (54 muscles, each only pulling, several crossing more than one joint, with activation delays): the agent has to build its own joint control from muscle excitations. Success is the challenge's own event; nothing short of it counts, and the challenge's failure conditions and time limit apply.
@@ -61,3 +61,6 @@ Generated by `scripts/myosuite/generate.py` from `scripts/myosuite/instances/osl
`bad_trajectory`. Every `reward.json` now carries `infra_error` (1 next to every `grader_error`, so also when the
verifier's host does not rebuild the frozen instance). The pass rule is unchanged; a trial graded `bad_trajectory`
before this date whose rows had the right shape was an infrastructure failure.
+- 2026-10-14 (board 0113, from board 0086): the texts said that the prosthesis is the left leg; MyoSuite's model
+ puts the Open-Source Leg on the right (its joints are `osl_knee_angle_r` and `osl_ankle_angle_r`, under `femur_r`),
+ and the texts now say so. Text only: the instance, the tooling and the pass rule are the same.
diff --git a/tasks/myosuite-osl-run-i00-privileged/environment/rcb_myosuite.py b/tasks/myosuite-osl-run-i00-privileged/environment/rcb_myosuite.py
index 0f2fca5acd..930c0c73ba 100644
--- a/tasks/myosuite-osl-run-i00-privileged/environment/rcb_myosuite.py
+++ b/tasks/myosuite-osl-run-i00-privileged/environment/rcb_myosuite.py
@@ -465,6 +465,12 @@ class Session:
self.done, self.success, self.reason = True, True, "success"
elif failed:
self.done, self.reason = True, "failure: " + _FAILURE.get(self.task, "task terminated")
+ # Baoding's targets come from a precomputed trajectory of 1000 points (baoding_v1.create_goal_trajectory,
+ # "assumes that it's greater than env horizon"): with MyoSuite's step budget lifted, step 1001 would index past
+ # it (IndexError, board 0086). The episode ends there instead, as a failure unless it succeeded on that step.
+ n_goal = len(getattr(self.base, "goal", ())) if self.task == "baoding" else 0
+ if not self.done and n_goal and self.steps >= n_goal:
+ self.done, self.reason = True, f"failure: the targets' trajectory ends after {n_goal} control steps"
return not self.done
def _bimanual_contacts_ok(self) -> bool:
diff --git a/tasks/myosuite-osl-run-i00-privileged/instruction.md b/tasks/myosuite-osl-run-i00-privileged/instruction.md
index 029bbc941f..6440650090 100644
--- a/tasks/myosuite-osl-run-i00-privileged/instruction.md
+++ b/tasks/myosuite-osl-run-i00-privileged/instruction.md
@@ -1,10 +1,10 @@
# MyoSuite · osl_run — frozen instance
-You control the **MyoLeg + OSL**, a free-standing musculoskeletal model of a human whose left leg is an Open-Source Leg knee-ankle prosthesis (MyoSuite 2.12.2 on MuJoCo), via Python, through its actuators — 54
+You control the **MyoLeg + OSL**, a free-standing musculoskeletal model of a human whose right leg is an Open-Source Leg knee-ankle prosthesis (MyoSuite 2.12.2 on MuJoCo), via Python, through its actuators — 54
muscles and 2 non-muscle actuators: there is no joint-level controller.
-**Task:** Run the track: move forward 29 m along it to the finish (MyoChallenge 2024 OSL Run, its own win condition), staying within 1 m of the track's centre line and on your feet; the left leg is the prosthesis, driven by its own controller, so you command only the 54 muscles. Failure (the episode ends): falling, leaving the track sideways, or moving 2 m backwards. There is no time limit.
+**Task:** Run the track: move forward 29 m along it to the finish (MyoChallenge 2024 OSL Run, its own win condition), staying within 1 m of the track's centre line and on your feet; the right leg is the prosthesis (`osl_knee_angle_r`, `osl_ankle_angle_r`), driven by its own controller, so you command only the 54 muscles. Failure (the episode ends): falling, leaving the track sideways, or moving 2 m backwards. There is no time limit.
-A musculoskeletal body whose left leg is an Open-Source Leg prosthesis (knee and ankle motors driven by its own controller) stands at the start of a straight, flat running track. `s.state()` gives the task's observation dict (privileged): `terrain`, `internal_qpos`, `internal_qvel`, `grf`, `socket_force`, `torso_angle`, `muscle_length`, `muscle_velocity`, `muscle_force`, `act`, `model_root_pos`, `model_root_vel`, `hfield`.
+A musculoskeletal body whose right leg is an Open-Source Leg prosthesis (knee and ankle motors driven by its own controller) stands at the start of a straight, flat running track. `s.state()` gives the task's observation dict (privileged): `terrain`, `internal_qpos`, `internal_qvel`, `grf`, `socket_force`, `torso_angle`, `muscle_length`, `muscle_velocity`, `muscle_force`, `act`, `model_root_pos`, `model_root_vel`, `hfield`.
MyoSuite's reward code (from `/usr/local/lib/python3.11/site-packages/myosuite/envs/myo/myochallenge/run_track_v0.py`; `solved` is the goal test,
`done` the failure test):
```python
diff --git a/tasks/myosuite-osl-run-i00-privileged/task.toml b/tasks/myosuite-osl-run-i00-privileged/task.toml
index 62f9a21433..ac6badbe17 100644
--- a/tasks/myosuite-osl-run-i00-privileged/task.toml
+++ b/tasks/myosuite-osl-run-i00-privileged/task.toml
@@ -4,7 +4,7 @@ artifacts = [{ source = "/var/rcb-monitor/", service = "rcb_monitor" }, "/app/ou
[task]
name = "robot-coding-bench/myosuite-osl-run-i00-privileged"
-description = "MyoSuite osl_run (MyoLeg + OSL, 54 muscles, MuJoCo): produce one replayable trajectory of actuator commands. Run the track: move forward 29 m along it to the finish (MyoChallenge 2024 OSL Run, its own win condition), staying within 1 m of the track's centre line and on your feet; the left leg is the prosthesis, driven by its own controller, so you command only the 54 muscles. Frozen scene instance (myoChallengeOslRunFixed-v0, reset seed 0); graded by two fresh-process replays."
+description = "MyoSuite osl_run (MyoLeg + OSL, 54 muscles, MuJoCo): produce one replayable trajectory of actuator commands. Run the track: move forward 29 m along it to the finish (MyoChallenge 2024 OSL Run, its own win condition), staying within 1 m of the track's centre line and on your feet; the right leg is the prosthesis (`osl_knee_angle_r`, `osl_ankle_angle_r`), driven by its own controller, so you command only the 54 muscles. Frozen scene instance (myoChallengeOslRunFixed-v0, reset seed 0); graded by two fresh-process replays."
authors = [{ name = "Chi Gui" }]
keywords = ["robotics", "myosuite", "mujoco", "musculoskeletal", "myoleg-osl", "code-as-policy", "trajectory", "osl_run", "privileged"]
@@ -15,6 +15,7 @@ simulator = "myosuite" # MyoSuite 2.12.2 on MuJoCo 3.6.0 (CPU)
robot = "myoleg-osl"
family = "osl_run"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "a successful model trial on this frozen instance: Claude Code + Claude Opus 5.5 (run `claude_code-claude_opus_5_5-default`, 2026-10-06), trial `myosuite-osl-run-i00-privileged__cWgLuKv` (6714 rows); today's verifier scores its trajectory 1 in two fresh trials (docs/benchmarks/myosuite.md); human review in PR #65"
instance_seed = 0 # myoChallengeOslRunFixed-v0: seed(0), then reset(seed=0)
diff --git a/tasks/myosuite-osl-run-i00-privileged/tests/rcb_myosuite.py b/tasks/myosuite-osl-run-i00-privileged/tests/rcb_myosuite.py
index 0f2fca5acd..930c0c73ba 100644
--- a/tasks/myosuite-osl-run-i00-privileged/tests/rcb_myosuite.py
+++ b/tasks/myosuite-osl-run-i00-privileged/tests/rcb_myosuite.py
@@ -465,6 +465,12 @@ class Session:
self.done, self.success, self.reason = True, True, "success"
elif failed:
self.done, self.reason = True, "failure: " + _FAILURE.get(self.task, "task terminated")
+ # Baoding's targets come from a precomputed trajectory of 1000 points (baoding_v1.create_goal_trajectory,
+ # "assumes that it's greater than env horizon"): with MyoSuite's step budget lifted, step 1001 would index past
+ # it (IndexError, board 0086). The episode ends there instead, as a failure unless it succeeded on that step.
+ n_goal = len(getattr(self.base, "goal", ())) if self.task == "baoding" else 0
+ if not self.done and n_goal and self.steps >= n_goal:
+ self.done, self.reason = True, f"failure: the targets' trajectory ends after {n_goal} control steps"
return not self.done
def _bimanual_contacts_ok(self) -> bool:
diff --git a/tasks/myosuite-osl-run-i00-standard/README.md b/tasks/myosuite-osl-run-i00-standard/README.md
index 8fc0407787..80d5c1581d 100644
--- a/tasks/myosuite-osl-run-i00-standard/README.md
+++ b/tasks/myosuite-osl-run-i00-standard/README.md
@@ -12,8 +12,8 @@
| Base image | `ghcr.io/mll-lab-nu/eai-myosuite:0.1.1` |
## Task
-MyoSuite `myoChallengeOslRunFixed-v0` (MyoLeg + OSL, 54 muscles) on the same frozen instance as `myosuite-osl-run-i00-privileged`: Run the track: move forward 29 m along it to the finish (MyoChallenge 2024 OSL Run, its own win condition), staying within 1 m of the track's centre line and on your feet; the left leg is the prosthesis, driven by its own controller, so you command only the 54 muscles. Failure (the episode ends): falling, leaving the track sideways, or moving 2 m backwards. There is no time limit.
-A musculoskeletal body whose left leg is an Open-Source Leg prosthesis (knee and ankle motors driven by its own controller) stands at the start of a straight, flat running track. Standard mode (eai-standard/2.2): the agent drives the body through a client, with only its own proprioception and the scene's cameras, and 10 resets (only the last episode is graded).
+MyoSuite `myoChallengeOslRunFixed-v0` (MyoLeg + OSL, 54 muscles) on the same frozen instance as `myosuite-osl-run-i00-privileged`: Run the track: move forward 29 m along it to the finish (MyoChallenge 2024 OSL Run, its own win condition), staying within 1 m of the track's centre line and on your feet; the right leg is the prosthesis (`osl_knee_angle_r`, `osl_ankle_angle_r`), driven by its own controller, so you command only the 54 muscles. Failure (the episode ends): falling, leaving the track sideways, or moving 2 m backwards. There is no time limit.
+A musculoskeletal body whose right leg is an Open-Source Leg prosthesis (knee and ankle motors driven by its own controller) stands at the start of a straight, flat running track. Standard mode (eai-standard/2.2): the agent drives the body through a client, with only its own proprioception and the scene's cameras, and 10 resets (only the last episode is graded).
## Difficulty
Same instance and success rule as `myosuite-osl-run-i00-privileged` (first-pass rating: hard). In standard mode the agent loses free resets, the simulator and the object state: it sees camera images and the limb's proprioception, acts in actuator-command chunks, and gets 10 resets, so it has to close the loop on what it sees and learn the body from a few real attempts.
@@ -90,3 +90,6 @@ Generated by `scripts/myosuite/generate.py` from `scripts/myosuite/instances/osl
still had that time. `/spec` says what the clock counts from (`budget.clock`), and the instruction's `## Time limit`
tells the agent to plan by `time_left_s`. The service never ended an episode on this clock: the scene, the frozen
instance, the pass rule and the grader are unchanged, so the task keeps its name and earlier results stand.
+- 2026-10-14 (board 0113, from board 0086): the texts said that the prosthesis is the left leg; MyoSuite's model
+ puts the Open-Source Leg on the right (its joints are `osl_knee_angle_r` and `osl_ankle_angle_r`, under `femur_r`),
+ and the texts now say so. Text only: the instance, the tooling and the pass rule are the same.
diff --git a/tasks/myosuite-osl-run-i00-standard/environment/sim/rcb_myosuite.py b/tasks/myosuite-osl-run-i00-standard/environment/sim/rcb_myosuite.py
index 0f2fca5acd..930c0c73ba 100644
--- a/tasks/myosuite-osl-run-i00-standard/environment/sim/rcb_myosuite.py
+++ b/tasks/myosuite-osl-run-i00-standard/environment/sim/rcb_myosuite.py
@@ -465,6 +465,12 @@ class Session:
self.done, self.success, self.reason = True, True, "success"
elif failed:
self.done, self.reason = True, "failure: " + _FAILURE.get(self.task, "task terminated")
+ # Baoding's targets come from a precomputed trajectory of 1000 points (baoding_v1.create_goal_trajectory,
+ # "assumes that it's greater than env horizon"): with MyoSuite's step budget lifted, step 1001 would index past
+ # it (IndexError, board 0086). The episode ends there instead, as a failure unless it succeeded on that step.
+ n_goal = len(getattr(self.base, "goal", ())) if self.task == "baoding" else 0
+ if not self.done and n_goal and self.steps >= n_goal:
+ self.done, self.reason = True, f"failure: the targets' trajectory ends after {n_goal} control steps"
return not self.done
def _bimanual_contacts_ok(self) -> bool:
diff --git a/tasks/myosuite-osl-run-i00-standard/environment/sim/task.json b/tasks/myosuite-osl-run-i00-standard/environment/sim/task.json
index e3f1da911f..9139bdb057 100644
--- a/tasks/myosuite-osl-run-i00-standard/environment/sim/task.json
+++ b/tasks/myosuite-osl-run-i00-standard/environment/sim/task.json
@@ -1,5 +1,5 @@
{
- "task_text": "Run the track: move forward 29 m along it to the finish (MyoChallenge 2024 OSL Run, its own win condition), staying within 1 m of the track's centre line and on your feet; the left leg is the prosthesis, driven by its own controller, so you command only the 54 muscles. Failure (the episode ends): falling, leaving the track sideways, or moving 2 m backwards. There is no time limit.",
+ "task_text": "Run the track: move forward 29 m along it to the finish (MyoChallenge 2024 OSL Run, its own win condition), staying within 1 m of the track's centre line and on your feet; the right leg is the prosthesis (`osl_knee_angle_r`, `osl_ankle_angle_r`), driven by its own controller, so you command only the 54 muscles. Failure (the episode ends): falling, leaving the track sideways, or moving 2 m backwards. There is no time limit.",
"default_cameras": [
"side_view"
],
diff --git a/tasks/myosuite-osl-run-i00-standard/instruction.md b/tasks/myosuite-osl-run-i00-standard/instruction.md
index 2b99ab74c2..de1deb52e4 100644
--- a/tasks/myosuite-osl-run-i00-standard/instruction.md
+++ b/tasks/myosuite-osl-run-i00-standard/instruction.md
@@ -1,10 +1,10 @@
# MyoSuite · osl_run — STANDARD MODE (robot as a service, eai-standard/2.2)
-You control the **MyoLeg + OSL**, a free-standing musculoskeletal model of a human whose left leg is an Open-Source Leg knee-ankle prosthesis, **as if it were real**: the simulator (MyoSuite 2.12.2 on MuJoCo) runs in a service
+You control the **MyoLeg + OSL**, a free-standing musculoskeletal model of a human whose right leg is an Open-Source Leg knee-ankle prosthesis, **as if it were real**: the simulator (MyoSuite 2.12.2 on MuJoCo) runs in a service
you cannot open. You get one client and up to 11 episodes, and you act only through its actuators (54 muscles and 2 other actuators).
-**Task:** Run the track: move forward 29 m along it to the finish (MyoChallenge 2024 OSL Run, its own win condition), staying within 1 m of the track's centre line and on your feet; the left leg is the prosthesis, driven by its own controller, so you command only the 54 muscles. Failure (the episode ends): falling, leaving the track sideways, or moving 2 m backwards. There is no time limit.
+**Task:** Run the track: move forward 29 m along it to the finish (MyoChallenge 2024 OSL Run, its own win condition), staying within 1 m of the track's centre line and on your feet; the right leg is the prosthesis (`osl_knee_angle_r`, `osl_ankle_angle_r`), driven by its own controller, so you command only the 54 muscles. Failure (the episode ends): falling, leaving the track sideways, or moving 2 m backwards. There is no time limit.
-A musculoskeletal body whose left leg is an Open-Source Leg prosthesis (knee and ankle motors driven by its own controller) stands at the start of a straight, flat running track.
+A musculoskeletal body whose right leg is an Open-Source Leg prosthesis (knee and ankle motors driven by its own controller) stands at the start of a straight, flat running track.
## Rules
- **Episodes and resets.** The episode starts when the service comes up and ends on success (the challenge's own success event), on its failure conditions or its own time limit, when your time is up, or on
diff --git a/tasks/myosuite-osl-run-i00-standard/task.toml b/tasks/myosuite-osl-run-i00-standard/task.toml
index 07992cb775..f802710e6d 100644
--- a/tasks/myosuite-osl-run-i00-standard/task.toml
+++ b/tasks/myosuite-osl-run-i00-standard/task.toml
@@ -16,7 +16,7 @@ artifacts = [{ source = "/var/rcb-monitor/", service = "rcb_monitor" },
[task]
name = "robot-coding-bench/myosuite-osl-run-i00-standard"
-description = "STANDARD MODE (robot as a service, eai-standard/2.2): MyoSuite osl_run (MyoLeg + OSL, 54 muscles) on the same frozen instance as myosuite-osl-run-i00-privileged: Run the track: move forward 29 m along it to the finish (MyoChallenge 2024 OSL Run, its own win condition), staying within 1 m of the track's centre line and on your feet; the left leg is the prosthesis, driven by its own controller, so you command only the 54 muscles. The agent only gets a client: cameras (RGB, optional depth, calibration), the body's joints and muscles, actuator command chunks, fingertip kinematics; 10 resets, only the last episode graded; no object state; the sim service records the episode, replays it in fresh processes and grades it."
+description = "STANDARD MODE (robot as a service, eai-standard/2.2): MyoSuite osl_run (MyoLeg + OSL, 54 muscles) on the same frozen instance as myosuite-osl-run-i00-privileged: Run the track: move forward 29 m along it to the finish (MyoChallenge 2024 OSL Run, its own win condition), staying within 1 m of the track's centre line and on your feet; the right leg is the prosthesis (`osl_knee_angle_r`, `osl_ankle_angle_r`), driven by its own controller, so you command only the 54 muscles. The agent only gets a client: cameras (RGB, optional depth, calibration), the body's joints and muscles, actuator command chunks, fingertip kinematics; 10 resets, only the last episode graded; no object state; the sim service records the episode, replays it in fresh processes and grades it."
authors = [{ name = "Chi Gui" }]
keywords = ["robotics", "myosuite", "mujoco", "musculoskeletal", "myoleg-osl", "code-as-policy", "trajectory", "osl_run", "standard-mode"]
@@ -27,6 +27,7 @@ simulator = "myosuite" # MyoSuite 2.12.2 on MuJoCo 3.6.0 (CPU)
robot = "myoleg-osl"
family = "osl_run"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "no positive example in standard mode yet: the privileged twin's (Claude Code + Claude Opus 5.5 (run `claude_code-claude_opus_5_5-default`, 2026-10-06), trial `myosuite-osl-run-i00-privileged__cWgLuKv`) fails once its rows are rounded to the client's float32 (the runner falls after 629 control steps); the family's solvability rests on the privileged twin; human review in PR #65"
diff --git a/tasks/myosuite-pen-twirl-i00-privileged/environment/rcb_myosuite.py b/tasks/myosuite-pen-twirl-i00-privileged/environment/rcb_myosuite.py
index 0f2fca5acd..930c0c73ba 100644
--- a/tasks/myosuite-pen-twirl-i00-privileged/environment/rcb_myosuite.py
+++ b/tasks/myosuite-pen-twirl-i00-privileged/environment/rcb_myosuite.py
@@ -465,6 +465,12 @@ class Session:
self.done, self.success, self.reason = True, True, "success"
elif failed:
self.done, self.reason = True, "failure: " + _FAILURE.get(self.task, "task terminated")
+ # Baoding's targets come from a precomputed trajectory of 1000 points (baoding_v1.create_goal_trajectory,
+ # "assumes that it's greater than env horizon"): with MyoSuite's step budget lifted, step 1001 would index past
+ # it (IndexError, board 0086). The episode ends there instead, as a failure unless it succeeded on that step.
+ n_goal = len(getattr(self.base, "goal", ())) if self.task == "baoding" else 0
+ if not self.done and n_goal and self.steps >= n_goal:
+ self.done, self.reason = True, f"failure: the targets' trajectory ends after {n_goal} control steps"
return not self.done
def _bimanual_contacts_ok(self) -> bool:
diff --git a/tasks/myosuite-pen-twirl-i00-privileged/task.toml b/tasks/myosuite-pen-twirl-i00-privileged/task.toml
index 238dc3b276..4e5fe0ae68 100644
--- a/tasks/myosuite-pen-twirl-i00-privileged/task.toml
+++ b/tasks/myosuite-pen-twirl-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "myosuite" # MyoSuite 2.12.2 on MuJoCo 3.6.0 (CPU)
robot = "myohand"
family = "pen_twirl"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "a successful model trial on this frozen instance: Claude Code + Claude Opus 5.5 (run `claude_code-claude_opus_5_5-default`, 2026-09-28), trial `myosuite-pen-twirl-i00__UtTYzQE` (513 rows); today's verifier scores its trajectory 1 in two fresh trials (docs/benchmarks/myosuite.md); human review in PR #65"
instance_seed = 0 # myoHandPenTwirlFixed-v0: seed(0), then reset(seed=0)
diff --git a/tasks/myosuite-pen-twirl-i00-privileged/tests/rcb_myosuite.py b/tasks/myosuite-pen-twirl-i00-privileged/tests/rcb_myosuite.py
index 0f2fca5acd..930c0c73ba 100644
--- a/tasks/myosuite-pen-twirl-i00-privileged/tests/rcb_myosuite.py
+++ b/tasks/myosuite-pen-twirl-i00-privileged/tests/rcb_myosuite.py
@@ -465,6 +465,12 @@ class Session:
self.done, self.success, self.reason = True, True, "success"
elif failed:
self.done, self.reason = True, "failure: " + _FAILURE.get(self.task, "task terminated")
+ # Baoding's targets come from a precomputed trajectory of 1000 points (baoding_v1.create_goal_trajectory,
+ # "assumes that it's greater than env horizon"): with MyoSuite's step budget lifted, step 1001 would index past
+ # it (IndexError, board 0086). The episode ends there instead, as a failure unless it succeeded on that step.
+ n_goal = len(getattr(self.base, "goal", ())) if self.task == "baoding" else 0
+ if not self.done and n_goal and self.steps >= n_goal:
+ self.done, self.reason = True, f"failure: the targets' trajectory ends after {n_goal} control steps"
return not self.done
def _bimanual_contacts_ok(self) -> bool:
diff --git a/tasks/myosuite-pen-twirl-i00-standard/environment/sim/rcb_myosuite.py b/tasks/myosuite-pen-twirl-i00-standard/environment/sim/rcb_myosuite.py
index 0f2fca5acd..930c0c73ba 100644
--- a/tasks/myosuite-pen-twirl-i00-standard/environment/sim/rcb_myosuite.py
+++ b/tasks/myosuite-pen-twirl-i00-standard/environment/sim/rcb_myosuite.py
@@ -465,6 +465,12 @@ class Session:
self.done, self.success, self.reason = True, True, "success"
elif failed:
self.done, self.reason = True, "failure: " + _FAILURE.get(self.task, "task terminated")
+ # Baoding's targets come from a precomputed trajectory of 1000 points (baoding_v1.create_goal_trajectory,
+ # "assumes that it's greater than env horizon"): with MyoSuite's step budget lifted, step 1001 would index past
+ # it (IndexError, board 0086). The episode ends there instead, as a failure unless it succeeded on that step.
+ n_goal = len(getattr(self.base, "goal", ())) if self.task == "baoding" else 0
+ if not self.done and n_goal and self.steps >= n_goal:
+ self.done, self.reason = True, f"failure: the targets' trajectory ends after {n_goal} control steps"
return not self.done
def _bimanual_contacts_ok(self) -> bool:
diff --git a/tasks/myosuite-pen-twirl-i00-standard/task.toml b/tasks/myosuite-pen-twirl-i00-standard/task.toml
index d0e62dc07f..ed79b1c7d7 100644
--- a/tasks/myosuite-pen-twirl-i00-standard/task.toml
+++ b/tasks/myosuite-pen-twirl-i00-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "myosuite" # MyoSuite 2.12.2 on MuJoCo 3.6.0 (CPU)
robot = "myohand"
family = "pen_twirl"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "the privileged twin's positive example (Claude Code + Claude Opus 5.5 (run `claude_code-claude_opus_5_5-default`, 2026-09-28), trial `myosuite-pen-twirl-i00__UtTYzQE`, 513 rows) rounded to float32, as the client sends it, and streamed through the client: the service scores it 1 live and in both replays, in two fresh trials (docs/benchmarks/myosuite.md); human review in PR #65"
diff --git a/tasks/myosuite-relocate-i00-privileged/environment/rcb_myosuite.py b/tasks/myosuite-relocate-i00-privileged/environment/rcb_myosuite.py
index 0f2fca5acd..930c0c73ba 100644
--- a/tasks/myosuite-relocate-i00-privileged/environment/rcb_myosuite.py
+++ b/tasks/myosuite-relocate-i00-privileged/environment/rcb_myosuite.py
@@ -465,6 +465,12 @@ class Session:
self.done, self.success, self.reason = True, True, "success"
elif failed:
self.done, self.reason = True, "failure: " + _FAILURE.get(self.task, "task terminated")
+ # Baoding's targets come from a precomputed trajectory of 1000 points (baoding_v1.create_goal_trajectory,
+ # "assumes that it's greater than env horizon"): with MyoSuite's step budget lifted, step 1001 would index past
+ # it (IndexError, board 0086). The episode ends there instead, as a failure unless it succeeded on that step.
+ n_goal = len(getattr(self.base, "goal", ())) if self.task == "baoding" else 0
+ if not self.done and n_goal and self.steps >= n_goal:
+ self.done, self.reason = True, f"failure: the targets' trajectory ends after {n_goal} control steps"
return not self.done
def _bimanual_contacts_ok(self) -> bool:
diff --git a/tasks/myosuite-relocate-i00-privileged/task.toml b/tasks/myosuite-relocate-i00-privileged/task.toml
index 5d7b0bc53c..1a5ce5c5f1 100644
--- a/tasks/myosuite-relocate-i00-privileged/task.toml
+++ b/tasks/myosuite-relocate-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "myosuite" # MyoSuite 2.12.2 on MuJoCo 3.6.0 (CPU)
robot = "myoarm"
family = "relocate"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "a successful model trial on this frozen instance: Codex + GPT-6 Luna (run `codex-gpt6_luna-high`, via an OpenRouter adapter before protocol v1.0, 2026-09-29), trial `myosuite-relocate-i00__PjqTvZ2` (231 rows); today's verifier scores its trajectory 1 in two fresh trials (docs/benchmarks/myosuite.md); human review in PR #65"
instance_seed = 0 # myoChallengeRelocateP1-v0: seed(0), then reset(seed=0)
diff --git a/tasks/myosuite-relocate-i00-privileged/tests/rcb_myosuite.py b/tasks/myosuite-relocate-i00-privileged/tests/rcb_myosuite.py
index 0f2fca5acd..930c0c73ba 100644
--- a/tasks/myosuite-relocate-i00-privileged/tests/rcb_myosuite.py
+++ b/tasks/myosuite-relocate-i00-privileged/tests/rcb_myosuite.py
@@ -465,6 +465,12 @@ class Session:
self.done, self.success, self.reason = True, True, "success"
elif failed:
self.done, self.reason = True, "failure: " + _FAILURE.get(self.task, "task terminated")
+ # Baoding's targets come from a precomputed trajectory of 1000 points (baoding_v1.create_goal_trajectory,
+ # "assumes that it's greater than env horizon"): with MyoSuite's step budget lifted, step 1001 would index past
+ # it (IndexError, board 0086). The episode ends there instead, as a failure unless it succeeded on that step.
+ n_goal = len(getattr(self.base, "goal", ())) if self.task == "baoding" else 0
+ if not self.done and n_goal and self.steps >= n_goal:
+ self.done, self.reason = True, f"failure: the targets' trajectory ends after {n_goal} control steps"
return not self.done
def _bimanual_contacts_ok(self) -> bool:
diff --git a/tasks/myosuite-relocate-i00-standard/environment/sim/rcb_myosuite.py b/tasks/myosuite-relocate-i00-standard/environment/sim/rcb_myosuite.py
index 0f2fca5acd..930c0c73ba 100644
--- a/tasks/myosuite-relocate-i00-standard/environment/sim/rcb_myosuite.py
+++ b/tasks/myosuite-relocate-i00-standard/environment/sim/rcb_myosuite.py
@@ -465,6 +465,12 @@ class Session:
self.done, self.success, self.reason = True, True, "success"
elif failed:
self.done, self.reason = True, "failure: " + _FAILURE.get(self.task, "task terminated")
+ # Baoding's targets come from a precomputed trajectory of 1000 points (baoding_v1.create_goal_trajectory,
+ # "assumes that it's greater than env horizon"): with MyoSuite's step budget lifted, step 1001 would index past
+ # it (IndexError, board 0086). The episode ends there instead, as a failure unless it succeeded on that step.
+ n_goal = len(getattr(self.base, "goal", ())) if self.task == "baoding" else 0
+ if not self.done and n_goal and self.steps >= n_goal:
+ self.done, self.reason = True, f"failure: the targets' trajectory ends after {n_goal} control steps"
return not self.done
def _bimanual_contacts_ok(self) -> bool:
diff --git a/tasks/myosuite-relocate-i00-standard/task.toml b/tasks/myosuite-relocate-i00-standard/task.toml
index 3d33c35a9a..05938a6885 100644
--- a/tasks/myosuite-relocate-i00-standard/task.toml
+++ b/tasks/myosuite-relocate-i00-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "myosuite" # MyoSuite 2.12.2 on MuJoCo 3.6.0 (CPU)
robot = "myoarm"
family = "relocate"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "the privileged twin's positive example (Codex + GPT-6 Luna (run `codex-gpt6_luna-high`, via an OpenRouter adapter before protocol v1.0, 2026-09-29), trial `myosuite-relocate-i00__PjqTvZ2`, 231 rows) rounded to float32, as the client sends it, and streamed through the client: the service scores it 1 live and in both replays, in two fresh trials (docs/benchmarks/myosuite.md); human review in PR #65"
diff --git a/tasks/myosuite-sarc-key-turn-i00-privileged/environment/rcb_myosuite.py b/tasks/myosuite-sarc-key-turn-i00-privileged/environment/rcb_myosuite.py
index 0f2fca5acd..930c0c73ba 100644
--- a/tasks/myosuite-sarc-key-turn-i00-privileged/environment/rcb_myosuite.py
+++ b/tasks/myosuite-sarc-key-turn-i00-privileged/environment/rcb_myosuite.py
@@ -465,6 +465,12 @@ class Session:
self.done, self.success, self.reason = True, True, "success"
elif failed:
self.done, self.reason = True, "failure: " + _FAILURE.get(self.task, "task terminated")
+ # Baoding's targets come from a precomputed trajectory of 1000 points (baoding_v1.create_goal_trajectory,
+ # "assumes that it's greater than env horizon"): with MyoSuite's step budget lifted, step 1001 would index past
+ # it (IndexError, board 0086). The episode ends there instead, as a failure unless it succeeded on that step.
+ n_goal = len(getattr(self.base, "goal", ())) if self.task == "baoding" else 0
+ if not self.done and n_goal and self.steps >= n_goal:
+ self.done, self.reason = True, f"failure: the targets' trajectory ends after {n_goal} control steps"
return not self.done
def _bimanual_contacts_ok(self) -> bool:
diff --git a/tasks/myosuite-sarc-key-turn-i00-privileged/task.toml b/tasks/myosuite-sarc-key-turn-i00-privileged/task.toml
index b0bc4c2cb4..6aa70eb96c 100644
--- a/tasks/myosuite-sarc-key-turn-i00-privileged/task.toml
+++ b/tasks/myosuite-sarc-key-turn-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "myosuite" # MyoSuite 2.12.2 on MuJoCo 3.6.0 (CPU)
robot = "myohand"
family = "sarc_key_turn"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "a successful model trial on this frozen instance: Claude Code + Claude Opus 5.5 (run `claude_code-claude_opus_5_5-default`, 2026-10-06), trial `myosuite-sarc-key-turn-i00-privi__iVBaqTS` (230 rows); today's verifier scores its trajectory 1 in two fresh trials (docs/benchmarks/myosuite.md); human review in PR #65"
instance_seed = 0 # myoSarcHandKeyTurnFixed-v0: seed(0), then reset(seed=0)
diff --git a/tasks/myosuite-sarc-key-turn-i00-privileged/tests/rcb_myosuite.py b/tasks/myosuite-sarc-key-turn-i00-privileged/tests/rcb_myosuite.py
index 0f2fca5acd..930c0c73ba 100644
--- a/tasks/myosuite-sarc-key-turn-i00-privileged/tests/rcb_myosuite.py
+++ b/tasks/myosuite-sarc-key-turn-i00-privileged/tests/rcb_myosuite.py
@@ -465,6 +465,12 @@ class Session:
self.done, self.success, self.reason = True, True, "success"
elif failed:
self.done, self.reason = True, "failure: " + _FAILURE.get(self.task, "task terminated")
+ # Baoding's targets come from a precomputed trajectory of 1000 points (baoding_v1.create_goal_trajectory,
+ # "assumes that it's greater than env horizon"): with MyoSuite's step budget lifted, step 1001 would index past
+ # it (IndexError, board 0086). The episode ends there instead, as a failure unless it succeeded on that step.
+ n_goal = len(getattr(self.base, "goal", ())) if self.task == "baoding" else 0
+ if not self.done and n_goal and self.steps >= n_goal:
+ self.done, self.reason = True, f"failure: the targets' trajectory ends after {n_goal} control steps"
return not self.done
def _bimanual_contacts_ok(self) -> bool:
diff --git a/tasks/myosuite-sarc-key-turn-i00-standard/environment/sim/rcb_myosuite.py b/tasks/myosuite-sarc-key-turn-i00-standard/environment/sim/rcb_myosuite.py
index 0f2fca5acd..930c0c73ba 100644
--- a/tasks/myosuite-sarc-key-turn-i00-standard/environment/sim/rcb_myosuite.py
+++ b/tasks/myosuite-sarc-key-turn-i00-standard/environment/sim/rcb_myosuite.py
@@ -465,6 +465,12 @@ class Session:
self.done, self.success, self.reason = True, True, "success"
elif failed:
self.done, self.reason = True, "failure: " + _FAILURE.get(self.task, "task terminated")
+ # Baoding's targets come from a precomputed trajectory of 1000 points (baoding_v1.create_goal_trajectory,
+ # "assumes that it's greater than env horizon"): with MyoSuite's step budget lifted, step 1001 would index past
+ # it (IndexError, board 0086). The episode ends there instead, as a failure unless it succeeded on that step.
+ n_goal = len(getattr(self.base, "goal", ())) if self.task == "baoding" else 0
+ if not self.done and n_goal and self.steps >= n_goal:
+ self.done, self.reason = True, f"failure: the targets' trajectory ends after {n_goal} control steps"
return not self.done
def _bimanual_contacts_ok(self) -> bool:
diff --git a/tasks/myosuite-sarc-key-turn-i00-standard/task.toml b/tasks/myosuite-sarc-key-turn-i00-standard/task.toml
index 4b628e41d8..6bbab92ecb 100644
--- a/tasks/myosuite-sarc-key-turn-i00-standard/task.toml
+++ b/tasks/myosuite-sarc-key-turn-i00-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "myosuite" # MyoSuite 2.12.2 on MuJoCo 3.6.0 (CPU)
robot = "myohand"
family = "sarc_key_turn"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "the privileged twin's positive example (Claude Code + Claude Opus 5.5 (run `claude_code-claude_opus_5_5-default`, 2026-10-06), trial `myosuite-sarc-key-turn-i00-privi__iVBaqTS`, 230 rows) rounded to float32, as the client sends it, and streamed through the client: the service scores it 1 live and in both replays, in two fresh trials (docs/benchmarks/myosuite.md); human review in PR #65"
diff --git a/tasks/myosuite-soccer-i00-privileged/environment/rcb_myosuite.py b/tasks/myosuite-soccer-i00-privileged/environment/rcb_myosuite.py
index 0f2fca5acd..930c0c73ba 100644
--- a/tasks/myosuite-soccer-i00-privileged/environment/rcb_myosuite.py
+++ b/tasks/myosuite-soccer-i00-privileged/environment/rcb_myosuite.py
@@ -465,6 +465,12 @@ class Session:
self.done, self.success, self.reason = True, True, "success"
elif failed:
self.done, self.reason = True, "failure: " + _FAILURE.get(self.task, "task terminated")
+ # Baoding's targets come from a precomputed trajectory of 1000 points (baoding_v1.create_goal_trajectory,
+ # "assumes that it's greater than env horizon"): with MyoSuite's step budget lifted, step 1001 would index past
+ # it (IndexError, board 0086). The episode ends there instead, as a failure unless it succeeded on that step.
+ n_goal = len(getattr(self.base, "goal", ())) if self.task == "baoding" else 0
+ if not self.done and n_goal and self.steps >= n_goal:
+ self.done, self.reason = True, f"failure: the targets' trajectory ends after {n_goal} control steps"
return not self.done
def _bimanual_contacts_ok(self) -> bool:
diff --git a/tasks/myosuite-soccer-i00-privileged/task.toml b/tasks/myosuite-soccer-i00-privileged/task.toml
index fd07d2d281..853fb57432 100644
--- a/tasks/myosuite-soccer-i00-privileged/task.toml
+++ b/tasks/myosuite-soccer-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "myosuite" # MyoSuite 2.12.2 on MuJoCo 3.6.0 (CPU)
robot = "myolegs"
family = "soccer"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "a successful model trial on this frozen instance: Claude Code + Claude Opus 5.5 (run `claude_code-claude_opus_5_5-default`, 2026-10-06), trial `myosuite-soccer-i00-privileged__DuXNvn4` (308 rows); today's verifier scores its trajectory 1 in two fresh trials (docs/benchmarks/myosuite.md); human review in PR #65"
instance_seed = 0 # myoChallengeSoccerP1-v0: seed(0), then reset(seed=0)
diff --git a/tasks/myosuite-soccer-i00-privileged/tests/rcb_myosuite.py b/tasks/myosuite-soccer-i00-privileged/tests/rcb_myosuite.py
index 0f2fca5acd..930c0c73ba 100644
--- a/tasks/myosuite-soccer-i00-privileged/tests/rcb_myosuite.py
+++ b/tasks/myosuite-soccer-i00-privileged/tests/rcb_myosuite.py
@@ -465,6 +465,12 @@ class Session:
self.done, self.success, self.reason = True, True, "success"
elif failed:
self.done, self.reason = True, "failure: " + _FAILURE.get(self.task, "task terminated")
+ # Baoding's targets come from a precomputed trajectory of 1000 points (baoding_v1.create_goal_trajectory,
+ # "assumes that it's greater than env horizon"): with MyoSuite's step budget lifted, step 1001 would index past
+ # it (IndexError, board 0086). The episode ends there instead, as a failure unless it succeeded on that step.
+ n_goal = len(getattr(self.base, "goal", ())) if self.task == "baoding" else 0
+ if not self.done and n_goal and self.steps >= n_goal:
+ self.done, self.reason = True, f"failure: the targets' trajectory ends after {n_goal} control steps"
return not self.done
def _bimanual_contacts_ok(self) -> bool:
diff --git a/tasks/myosuite-soccer-i00-standard/environment/sim/rcb_myosuite.py b/tasks/myosuite-soccer-i00-standard/environment/sim/rcb_myosuite.py
index 0f2fca5acd..930c0c73ba 100644
--- a/tasks/myosuite-soccer-i00-standard/environment/sim/rcb_myosuite.py
+++ b/tasks/myosuite-soccer-i00-standard/environment/sim/rcb_myosuite.py
@@ -465,6 +465,12 @@ class Session:
self.done, self.success, self.reason = True, True, "success"
elif failed:
self.done, self.reason = True, "failure: " + _FAILURE.get(self.task, "task terminated")
+ # Baoding's targets come from a precomputed trajectory of 1000 points (baoding_v1.create_goal_trajectory,
+ # "assumes that it's greater than env horizon"): with MyoSuite's step budget lifted, step 1001 would index past
+ # it (IndexError, board 0086). The episode ends there instead, as a failure unless it succeeded on that step.
+ n_goal = len(getattr(self.base, "goal", ())) if self.task == "baoding" else 0
+ if not self.done and n_goal and self.steps >= n_goal:
+ self.done, self.reason = True, f"failure: the targets' trajectory ends after {n_goal} control steps"
return not self.done
def _bimanual_contacts_ok(self) -> bool:
diff --git a/tasks/myosuite-soccer-i00-standard/task.toml b/tasks/myosuite-soccer-i00-standard/task.toml
index 553946a8fd..103c0d844c 100644
--- a/tasks/myosuite-soccer-i00-standard/task.toml
+++ b/tasks/myosuite-soccer-i00-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "myosuite" # MyoSuite 2.12.2 on MuJoCo 3.6.0 (CPU)
robot = "myolegs"
family = "soccer"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "the privileged twin's positive example (Claude Code + Claude Opus 5.5 (run `claude_code-claude_opus_5_5-default`, 2026-10-06), trial `myosuite-soccer-i00-privileged__DuXNvn4`, 308 rows) rounded to float32, as the client sends it, and streamed through the client: the service scores it 1 live and in both replays, in two fresh trials (docs/benchmarks/myosuite.md); human review in PR #65"
diff --git a/tasks/myosuite-table-tennis-i00-privileged/environment/rcb_myosuite.py b/tasks/myosuite-table-tennis-i00-privileged/environment/rcb_myosuite.py
index 0f2fca5acd..930c0c73ba 100644
--- a/tasks/myosuite-table-tennis-i00-privileged/environment/rcb_myosuite.py
+++ b/tasks/myosuite-table-tennis-i00-privileged/environment/rcb_myosuite.py
@@ -465,6 +465,12 @@ class Session:
self.done, self.success, self.reason = True, True, "success"
elif failed:
self.done, self.reason = True, "failure: " + _FAILURE.get(self.task, "task terminated")
+ # Baoding's targets come from a precomputed trajectory of 1000 points (baoding_v1.create_goal_trajectory,
+ # "assumes that it's greater than env horizon"): with MyoSuite's step budget lifted, step 1001 would index past
+ # it (IndexError, board 0086). The episode ends there instead, as a failure unless it succeeded on that step.
+ n_goal = len(getattr(self.base, "goal", ())) if self.task == "baoding" else 0
+ if not self.done and n_goal and self.steps >= n_goal:
+ self.done, self.reason = True, f"failure: the targets' trajectory ends after {n_goal} control steps"
return not self.done
def _bimanual_contacts_ok(self) -> bool:
diff --git a/tasks/myosuite-table-tennis-i00-privileged/task.toml b/tasks/myosuite-table-tennis-i00-privileged/task.toml
index 9280e4a385..8fedb18523 100644
--- a/tasks/myosuite-table-tennis-i00-privileged/task.toml
+++ b/tasks/myosuite-table-tennis-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "myosuite" # MyoSuite 2.12.2 on MuJoCo 3.6.0 (CPU)
robot = "myoarm-torso"
family = "table_tennis"
mode = "privileged"
+dynamics = "dynamic" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "a successful model trial on this frozen instance: Claude Code + Claude Opus 5.5 (run `claude_code-claude_opus_5_5-default`, 2026-10-06), trial `myosuite-table-tennis-i00-privil__kdQdy89` (111 rows); today's verifier scores its trajectory 1 in two fresh trials (docs/benchmarks/myosuite.md); human review in PR #65"
instance_seed = 0 # myoChallengeTableTennisP1-v0: seed(0), then reset(seed=0)
diff --git a/tasks/myosuite-table-tennis-i00-privileged/tests/rcb_myosuite.py b/tasks/myosuite-table-tennis-i00-privileged/tests/rcb_myosuite.py
index 0f2fca5acd..930c0c73ba 100644
--- a/tasks/myosuite-table-tennis-i00-privileged/tests/rcb_myosuite.py
+++ b/tasks/myosuite-table-tennis-i00-privileged/tests/rcb_myosuite.py
@@ -465,6 +465,12 @@ class Session:
self.done, self.success, self.reason = True, True, "success"
elif failed:
self.done, self.reason = True, "failure: " + _FAILURE.get(self.task, "task terminated")
+ # Baoding's targets come from a precomputed trajectory of 1000 points (baoding_v1.create_goal_trajectory,
+ # "assumes that it's greater than env horizon"): with MyoSuite's step budget lifted, step 1001 would index past
+ # it (IndexError, board 0086). The episode ends there instead, as a failure unless it succeeded on that step.
+ n_goal = len(getattr(self.base, "goal", ())) if self.task == "baoding" else 0
+ if not self.done and n_goal and self.steps >= n_goal:
+ self.done, self.reason = True, f"failure: the targets' trajectory ends after {n_goal} control steps"
return not self.done
def _bimanual_contacts_ok(self) -> bool:
diff --git a/tasks/myosuite-table-tennis-i00-standard-realtime/README.md b/tasks/myosuite-table-tennis-i00-standard-realtime/README.md
new file mode 100644
index 0000000000..4cb58b0233
--- /dev/null
+++ b/tasks/myosuite-table-tennis-i00-standard-realtime/README.md
@@ -0,0 +1,56 @@
+# myosuite-table-tennis-i00-standard-realtime
+
+| | |
+|---|---|
+| Benchmark | MyoSuite |
+| Mode | standard-realtime (`eai-standard/2.3`) |
+| Twin | `myosuite-table-tennis-i00-privileged` |
+| Simulator / robot | `myosuite` / `myoarm-torso` |
+| Family / instance | `table_tennis` / seed 0 |
+| Oracle | `none` — a successful episode of the privileged twin (GPT-6 Luna xhigh, trial `myosuite-table-tennis-i00-privil__SMNgbHk`, 82 rows) planned from step 0 in one `submit(..., at=0)` at S = 1 succeeds live and in the fresh replay, with no stall (board 0112, 2026-10-14); the standard twin's evidence holds for the same instance and rule; human review in PR #201 |
+| Agent time budget | 3600 s; tokens: protocol default (PROTOCOL §5) |
+| Real time | slowdown S = 1 (the world runs at 1/S of real speed: one control step every S x control_dt s of wall time; `RCB_SLOWDOWN`'s default); reset budget 10 (`RCB_RESET_BUDGET`'s default) (PROTOCOL §4b) |
+| Base image | `ghcr.io/mll-lab-nu/eai-myosuite:0.1.1` |
+
+## Task
+MyoSuite `myoChallengeTableTennisP1-v0` (MyoArm with torso, 273 muscles) on the same frozen instance as `myosuite-table-tennis-i00-privileged`: Return the served ball: hit it with the paddle so that it lands on the opponent's half of the table (MyoChallenge 2025 TableTennis P1's own success test on the ball's contact history: after the serve's bounce on your half, the paddle touches it once, then it touches the opponent's half, with no second paddle touch and no further bounce on your half). Failure (the episode ends): any other outcome of the rally, the ball falling below 0.3 m, or 3 s of simulated time.
+You stand at a table-tennis table holding a paddle in the right hand; a ball is served toward you from the far side. Standard-realtime mode (eai-standard/2.3, slowdown S = 1): once an episode's clock starts the body and the ball move whatever the agent does; it drives the body through the realtime client from a program of its own, with only its own proprioception and the scene's cameras, and 10 resets (only the last episode is graded).
+
+## Difficulty
+Same instance and success rule as the standard twin (first-pass rating: hard, harder here). The scene is dynamic (scripts/task_dynamics.toml: the ball is served with an initial velocity): the serve reaches you and the rally is decided within about 1.2 s of robot time after the clock starts (an episode that does nothing ends after 121 control steps), so nothing can be steered call by call; the agent has to plan or stream the muscle commands from a program, and learn the body and the serve from 11 real attempts.
+
+## Reference solution
+None in the repository. Solvability evidence: a successful episode of the privileged twin (82 rows, GPT-6 Luna xhigh) planned from step 0 through this service at S = 1 succeeded live and in the fresh replay, with no stall (board 0112, 2026-10-14).
+
+## Verification
+The sim service (`environment/sim/rcb_realtime.py` + `rcb_service.py`, the shared eai-standard/2.3 layer, with `environment/sim/server.py`, MyoSuite's adapter) judges the episode live and records every control step it ran (`trajectory_live.f64`); after the agent's time the collect hook stops the agent's programs and signals the service, which replays the graded episode in a fresh MyoSuite session and writes `final.json`; `tests/grade.py` (separate sandbox, no network) passes iff the live episode succeeded, the replay succeeds and ends in the same state. The clock's record (`stalls`, `max_lag_s`, ...) is copied into `reward.json` and decides nothing. `nop`: the episode never starts (its clock starts at the first submit), reward 0.
+
+## Compute resources
+Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]` in `task.toml`). The agent phase counts every container of the trial at once (the agent's, and in standard modes the simulator sidecar); the verifier phase runs after it.
+
+| | Agent phase | Verifier |
+|---|---|---|
+| CPU cores | 2 | 2 |
+| Memory | 4 GB | 3 GB |
+| GPU memory | 1 GB | 1 GB |
+
+- GPU: required, any NVIDIA GPU; recommended: any NVIDIA GPU with EGL (RTX 4090 / L4).
+- Measured: as the standard twin (sim 1.2-1.4 GB, docker stats, board 0112); the GPU only renders the cameras over EGL (5 ms a frame; board 0112, 2026-10-14, ws-1); the sim keeps the GPU through the collect hook (fresh replay and video)
+- Limits in the agent phase (docs/PROTOCOL.md §2.10; `[metadata.resources.limits]`, shown to the agent in `instruction.md`; always measured and reported, and exceeding one fails the trial only when the run sets `RCB_LIMITS_ENFORCE=1`), the agent's container: no GPU, 8 GiB memory, 512 processes, 20 GiB of new files under `/app`, `/tmp`, `/root`; 60 s within 120 s. Expected peak of the whole trial: 4 GiB GPU memory, 14 GiB memory. Limits measured: GPU: EGL rendering only, under 1 GiB (board 0112) -> peak + 4; memory: no samples of the agent's container; floor 8 GiB (at most the 8 GiB cap); whole trial: + the sim service's 6 GiB (docker peak 1.4 GiB, 5 samples); processes 512 and disk 20 GiB: design defaults, not yet measured.
+
+## Details
+| | |
+|---|---|
+| Upstream | MyoSuite 2.12.2 (Apache-2.0) on MuJoCo 3.6.0, the models that ship in its wheel |
+| Slowdown | S = 1 (`RCB_SLOWDOWN`'s default): measured RTF 2.1-2.7 (2026-10-14, board 0112, ws-1 (32 cores, load 30-60), the service capped at 4 CPUs: 2.08 with the realtime service's clock flat out (RCB_SLOWDOWN = 0.01, 11 episodes, 13.3 s of robot time), 2.6-2.7 for the bare physics; S = ceil(1.5 / 2.08) = 1. At S = 1 a controller at the control rate with a camera image every 0.25 s ran 11 episodes with no stall (largest lag 0.23 s), and an EGL frame takes 5 ms (an OSMesa one 1.3 s)) |
+| Service | the realtime layer of HumanoidBench, BiGym 2.0 and RoboOlympics (`rcb_realtime.py`, `rcb_service.py`, unchanged) with `server.py`, which implements `rcb_service.Sim` on `rcb_myosuite.Session` (the standard twin's scene, action, success rule and replay); cameras `topview`, `default`, `close_up`, `hand_top`, `hand_bottom`, `hand_side_inter`, `hand_side_exter`, `plam_lookat` (640 x 480, EGL on the GPU) |
+| Resources | 4 CPUs and a GPU for rendering (the physics is on the CPU) |
+| Reset budget | 10 (`RCB_RESET_BUDGET`, compose default; PROTOCOL §4b): `POST /reset` starts a new episode from the frozen instance, paused; only the last episode is graded, a success ends the trial; results under another budget or slowdown are reported separately |
+
+### Provenance
+Generated by `scripts/myosuite/generate.py` from `scripts/myosuite/instances/table_tennis.npz` and `table_tennis.json`
+(`make_facts.py`) and `scripts/myosuite/shared/realtime/`; see `docs/benchmarks/myosuite.md`, "Standard-realtime".
+
+### Change Log
+- 2026-10-14 (owner, board 0112): added. The family's scene is dynamic (PROTOCOL §2.14), so it gets a standard-realtime
+ twin with S = 1 and 10 resets.
diff --git a/tasks/myosuite-table-tennis-i00-standard-realtime/environment/Dockerfile b/tasks/myosuite-table-tennis-i00-standard-realtime/environment/Dockerfile
new file mode 100644
index 0000000000..a16dd87255
--- /dev/null
+++ b/tasks/myosuite-table-tennis-i00-standard-realtime/environment/Dockerfile
@@ -0,0 +1,9 @@
+# Standard-realtime mode: the agent container has NO simulator and no MyoSuite code, only Python and the realtime client
+# (eai-standard/2.3), on the shared pinned client base with the shared pins (scripts/eai_standard/client.env).
+FROM docker.io/library/python:3.12.14-slim@sha256:f77ac9e44ae96ef2c90b8053ea08c31f8be030f824196b0ae4db6d462c84e51f
+COPY client-requirements.txt /tmp/client-requirements.txt
+RUN pip install --no-cache-dir -r /tmp/client-requirements.txt
+COPY rcb_client.py rcb1.py /opt/tools/
+ENV PYTHONPATH=/opt/tools
+RUN mkdir -p /app/output
+WORKDIR /app
diff --git a/tasks/myosuite-table-tennis-i00-standard-realtime/environment/client-requirements.txt b/tasks/myosuite-table-tennis-i00-standard-realtime/environment/client-requirements.txt
new file mode 100644
index 0000000000..8aa73a968f
--- /dev/null
+++ b/tasks/myosuite-table-tennis-i00-standard-realtime/environment/client-requirements.txt
@@ -0,0 +1,11 @@
+# Exact pins for what the standard agent container installs (and a verifier, if it installs anything): the closure of
+# `pip install requests numpy pillow scipy`, the unpinned line of every standard task at 9eb18edf, resolved on
+# 2026-10-01 in CLIENT_IMAGE (client.env: Python 3.12.14, pip 25.0.1). A change here is a shared-files PR.
+certifi==2026.7.22
+charset-normalizer==3.5.2
+idna==3.20
+numpy==2.5.3
+pillow==12.3.0
+requests==2.34.2
+scipy==1.18.1
+urllib3==2.8.0
diff --git a/tasks/myosuite-table-tennis-i00-standard-realtime/environment/docker-compose.yaml b/tasks/myosuite-table-tennis-i00-standard-realtime/environment/docker-compose.yaml
new file mode 100644
index 0000000000..afb3ddabb7
--- /dev/null
+++ b/tasks/myosuite-table-tennis-i00-standard-realtime/environment/docker-compose.yaml
@@ -0,0 +1,76 @@
+# Standard-realtime mode (PROTOCOL §4b): the simulator runs in the `sim` sidecar (the agent cannot open it), on its own clock
+# once an episode starts (sim/rcb_realtime.py, the shared eai-standard/2.3 layer, with sim/server.py, MyoSuite's adapter);
+# the agent only reaches port 8000. No `expose`/`ports` on sim: under Harbor's egress control it shares main's network
+# namespace (localhost:8000).
+# GPU on sim: the physics is MuJoCo on the CPU, but every image the agent asks for is rendered here, and on a clock that
+# does not stop a render is time the simulator cannot step: EGL on the GPU takes ~5 ms a frame, OSMesa ~1.3 s
+# (table_tennis, 640x480). `graphics` is the capability that mounts the EGL driver (server.py writes its ICD manifest).
+# Crash recovery: compose restarts the service, which resumes the same episode, paused, never a reset.
+# Knobs: ${NAME:-default} takes NAME from the shell that runs harbor. The defaults are the task's configuration;
+# results under other settings are reported separately.
+services:
+ main:
+ pids_limit: 8192 # rcb-monitor
+ volumes: # rcb-monitor
+ - rcb_limits:/run/rcb-limits:ro # rcb-monitor
+ depends_on: [sim]
+ sim:
+ cpus: 4 # rcb-monitor
+ mem_limit: 8192m # rcb-monitor
+ pids_limit: 8192 # rcb-monitor
+ image: ghcr.io/mll-lab-nu/eai-myosuite:0.1.1
+ volumes:
+ - rcb_observe:/run/rcb-observe:ro # rcb-monitor
+ - ./sim:/srv:ro
+ working_dir: /tmp
+ command: python /srv/rcb_realtime.py
+ restart: on-failure:6 # N >= 3 (PROTOCOL §4.1); a resume continues the episode, paused
+ environment:
+ RCB_INSTANCE: /srv/instance.npz
+ RCB_TASK_INFO: /srv/task.json
+ RCB_DATA: /data
+ RCB_WALL_CLOCK_S: "3600"
+ RCB_STARTUP_TIMEOUT_S: "300" # the scene builds in a few seconds
+ PYTHONPATH: /srv
+ RCB_SESSION_LOG: "" # the sidecar's own Session must not write the agent-side interaction log
+ RCB_SLOWDOWN: "${RCB_SLOWDOWN:-1}" # S = 1: the world runs at 1/S of real speed (this family's measured RTF; PROTOCOL §4b)
+ RCB_RESET_BUDGET: "${RCB_RESET_BUDGET:-10}" # 10 resets (PROTOCOL §4b default); 0: one episode
+ RCB_GL: "${RCB_GL:-auto}" # auto: EGL on the GPU when it works, else OSMesa
+ NVIDIA_DRIVER_CAPABILITIES: compute,graphics,utility
+ deploy:
+ resources:
+ reservations:
+ devices:
+ - driver: nvidia
+ count: 1
+ capabilities: [gpu]
+ # >>> rcb_monitor (resource limits monitor, docs/PROTOCOL.md §2.10; generated by scripts/rcb_monitor/inject.py)
+ rcb_monitor:
+ image: ghcr.io/mll-lab-nu/eai-monitor:0.1.0@sha256:f2c910f53f79c2bfa4a71298b10c78383cae3afc40720ca562cc845b8d20fb0c
+ pid: host # sees every container's processes (and NVML's host PIDs); the agent never sees it
+ cgroup: host # reads each container's own cgroup (memory, OOM kills, caps)
+ network_mode: none
+ cap_add: [SYS_PTRACE] # reads /proc of processes running as other users
+ cpus: 0.5
+ mem_limit: 256m
+ pids_limit: 64
+ restart: on-failure:3
+ working_dir: /var/rcb-monitor
+ volumes:
+ - ./rcb_monitor:/opt/rcb_monitor:ro
+ - rcb_limits:/run/rcb-limits
+ - rcb_observe:/run/rcb-observe
+ environment:
+ RCB_LIMITS_ENFORCE: ${RCB_LIMITS_ENFORCE:-0} # 1: stop the agent at a confirmed violation, and fail the trial
+ command: ["python3", "/opt/rcb_monitor/monitor.py", "run"]
+ deploy:
+ resources:
+ reservations:
+ devices:
+ - driver: nvidia
+ count: 1
+ capabilities: [gpu]
+ # <<< rcb_monitor
+volumes: # rcb-monitor
+ rcb_limits: # rcb-monitor
+ rcb_observe: # rcb-monitor
diff --git a/tasks/myosuite-table-tennis-i00-standard-realtime/environment/rcb1.py b/tasks/myosuite-table-tennis-i00-standard-realtime/environment/rcb1.py
new file mode 100644
index 0000000000..6938c57e38
--- /dev/null
+++ b/tasks/myosuite-table-tennis-i00-standard-realtime/environment/rcb1.py
@@ -0,0 +1,83 @@
+"""RCB1 — a tiny binary frame for one JSON header plus N raw blobs (videos, image arrays, depth, proprio).
+
+ frame = b"RCB1" | header_len uint32 LE | header JSON (UTF-8) | blob_0 | blob_1 | ...
+ header = {"v": 1, "meta": {...}, "parts": [{"name", "kind", "len", ...}, ...]} # blobs in the same order
+
+kinds: "mp4" (fps, n), "jpeg", "array" (dtype, shape; C-order raw bytes), "json" (utf-8), "bytes".
+Zero dependencies for bytes/json parts; numpy only when you ask for arrays. Used on both sides (server and client),
+over HTTP bodies (Content-Type: application/x-rcb1) and equally usable over a raw socket. `python -m rcb1 unpack f dir`.
+"""
+from __future__ import annotations
+import json, struct, sys, pathlib
+
+MAGIC = b"RCB1"
+
+
+def encode(meta: dict, parts: list[tuple[str, str, bytes, dict]]) -> bytes:
+ """parts: [(name, kind, blob, extra), ...]; extra carries dtype/shape/fps/n."""
+ hdr_parts = []
+ for name, kind, blob, extra in parts:
+ d = {"name": name, "kind": kind, "len": len(blob)}
+ d.update({k: v for k, v in (extra or {}).items() if v is not None})
+ hdr_parts.append(d)
+ header = json.dumps({"v": 1, "meta": meta, "parts": hdr_parts}, separators=(",", ":")).encode()
+ return b"".join([MAGIC, struct.pack("<I", len(header)), header] + [blob for _, _, blob, _ in parts])
+
+
+def array_part(name: str, arr, extra: dict | None = None) -> tuple[str, str, bytes, dict]:
+ import numpy as np
+ a = np.ascontiguousarray(arr)
+ return (name, "array", a.tobytes(), {"dtype": str(a.dtype), "shape": list(a.shape), **(extra or {})})
+
+
+def decode(buf: bytes, arrays: bool = True) -> tuple[dict, dict]:
+ """-> (meta, parts) where parts[name] = {"kind", ..., "data"}; arrays become numpy arrays when arrays=True."""
+ if buf[:4] != MAGIC:
+ raise ValueError("not an RCB1 frame")
+ n = struct.unpack("<I", buf[4:8])[0]
+ header = json.loads(buf[8:8 + n].decode()); off = 8 + n
+ out = {}
+ for p in header["parts"]:
+ blob = buf[off:off + p["len"]]; off += p["len"]
+ d = dict(p)
+ if p["kind"] == "array" and arrays:
+ import numpy as np
+ d["data"] = np.frombuffer(blob, dtype=np.dtype(p["dtype"])).reshape(p["shape"])
+ elif p["kind"] == "json":
+ d["data"] = json.loads(blob.decode())
+ else:
+ d["data"] = blob
+ out[p["name"]] = d
+ return header.get("meta", {}), out
+
+
+def unpack_to_dir(buf: bytes, out_dir: str) -> list[str]:
+ """Write a frame's parts as files: mp4/jpeg/bytes verbatim, arrays as .npy, meta as meta.json."""
+ import numpy as np
+ out = pathlib.Path(out_dir); out.mkdir(parents=True, exist_ok=True)
+ meta, parts = decode(buf, arrays=True); written = []
+ (out / "meta.json").write_text(json.dumps(meta, indent=1)); written.append("meta.json")
+ for name, p in parts.items():
+ safe = name.replace("/", "__")
+ if p["kind"] == "array":
+ f = out / f"{safe}.npy"; np.save(f, p["data"])
+ elif p["kind"] == "mp4":
+ f = out / f"{safe}.mp4"; f.write_bytes(p["data"])
+ elif p["kind"] == "jpeg":
+ f = out / f"{safe}.jpg"; f.write_bytes(p["data"])
+ elif p["kind"] == "json":
+ f = out / f"{safe}.json"; f.write_text(json.dumps(p["data"], indent=1))
+ else:
+ f = out / f"{safe}.bin"; f.write_bytes(p["data"])
+ written.append(f.name)
+ return written
+
+
+if __name__ == "__main__":
+ if len(sys.argv) == 4 and sys.argv[1] == "unpack":
+ print("\n".join(unpack_to_dir(open(sys.argv[2], "rb").read(), sys.argv[3])))
+ elif len(sys.argv) == 3 and sys.argv[1] == "info":
+ meta, parts = decode(open(sys.argv[2], "rb").read(), arrays=False)
+ print(json.dumps(meta, indent=1)); [print(n, {k: v for k, v in p.items() if k != "data"}) for n, p in parts.items()]
+ else:
+ print(__doc__)
diff --git a/tasks/myosuite-table-tennis-i00-standard-realtime/environment/rcb_client.py b/tasks/myosuite-table-tennis-i00-standard-realtime/environment/rcb_client.py
new file mode 100644
index 0000000000..f2b0559bf9
--- /dev/null
+++ b/tasks/myosuite-table-tennis-i00-standard-realtime/environment/rcb_client.py
@@ -0,0 +1,209 @@
+"""eai-standard/2.3 client (standard-realtime mode): your whole interface to the robot (installed at /opt/tools/rcb_client.py, on
+PYTHONPATH).
+
+ from rcb_client import Robot, EpisodeOver
+ r = Robot() # finds the service (SIM_URL, http://sim:8000, http://localhost:8000), waits until it is ready
+
+TIME DOES NOT STOP (spec()["timing"]). An episode starts paused; its clock starts at its first submit() or
+get_obs(seconds > 0) (again after every reset()) and from then on runs one control step every spec()["control_dt"] x
+spec()["timing"]["slowdown"] seconds of wall time, whether or not you call, until the episode ends. The service keeps a
+plan, one servo-target row per coming control step: submit() writes into it and answers at once; when the plan runs out
+the last target is held. Whatever you do in between -- think, print, run a model -- the robot keeps executing.
+
+Calls (each is one HTTP request; every answer that carries data is an RCB1 frame, decoded into a Result):
+ r.spec() GET /spec, fetched when connecting: the robot's facts. "action_types"
+ ("joint_position": dim, layout (actuator order), lower/upper, range_rad,
+ units, semantics), "control_dt" (0.02 s), "max_control_steps" (the
+ episode), "timing" (realtime, slowdown, max_ahead_steps), "cameras" {name:
+ rgb, depth, hw, fovy_deg, pose, mounted}, "proprio_layout" and "proprio"
+ (joints, axes, positions), "imu", "links" (masses, centres of mass, the
+ feet's boxes), "depth", "budget" {"reset", "wall_clock_s"}, "episode"
+ r.status() GET /status: calls, rows, control_steps, done, reason, elapsed_s, running,
+ planned_steps, lag_s (and episode, resets_left with a reset budget)
+ r.get_obs(cameras=None, seconds=0.0, video_fps=10, include_depth=False, video="frames")
+ POST /observe. seconds=0: the robot now, at once. seconds>0 (at most
+ spec()["max_watch_s"]): waits while that much robot time runs (starting the
+ clock if it is not running) and returns what the cameras saw; it never
+ changes the plan
+ r.submit("joint_position", actions, dt=None, interpolate="linear", at=None, cameras=None, video_fps=10,
+ video="none", include_depth=False, note="")
+ POST /submit: answers at once with the robot as it is now (the rows
+ have not run yet). actions (T, dim): rows dt apart (a multiple of
+ control_dt, the default; T x dt at most spec()["max_chunk_s"]); each step's
+ target is the row (interpolate="hold") or a straight line from the previous
+ row's ("linear", starting from the target in force just before `at`). They
+ become the plan from control step `at` (an episode step number; None = the
+ next step), replacing whatever was planned from `at` on and keeping what is
+ planned before it; at most spec()["timing"]["max_ahead_steps"] ahead. Rows
+ for steps already gone when the request arrives are dropped
+ (meta["late_steps"]). The first submit of an episode starts its clock
+ r.reset(cameras=None, include_depth=False)
+ POST /reset, only when spec()["budget"]["reset"] > 0: abandons the
+ running episode and starts a new one, paused, from the same starting state;
+ only the last episode is graded, and a success ends the trial
+
+cameras=None means every camera in spec()["cameras"]; cameras=[] means no images at all, only proprioception (the
+fastest call: what a controller loop wants). video: "mp4" (one H.264 file per camera), "frames" (the images as arrays),
+"both" or "none" (only the last frame); a submit returns no video (no time passes inside it). include_depth adds depth
+for the cameras that have it.
+
+Result (what get_obs / submit / reset return):
+ res.meta dict: calls, rows, control_steps (steps run in this episode: the next one is this number),
+ sim_time_s, time_s, time_left_s, proprio {"layout", "now"}, done, reason ("success" |
+ "failed" | "timeout" | "error"; None while the episode runs), running, planned_steps,
+ lag_s; per submit: start_step, late_steps, plan_end_step, clipped [{row, index, reason}]
+ (values outside [-1, 1] are clipped, not refused); with a reset budget also episode,
+ resets_left
+ res.last[cam] uint8 H x W x 3: the last frame (unless cameras=[])
+ res.frames[cam] uint8 N x H x W x 3 (get_obs, video="frames"/"both") res.video[cam] mp4 bytes
+ res.depth[cam] float16 N x H x W, metres, 0 = nothing in range (include_depth)
+ res.proprio float32 rows of spec()["proprio_layout"]: one at the call's start, one every 1/video_fps s
+ of a watch, one at its end
+ res.done the episode is over
+ res.save(dir) meta.json, <cam>.mp4, last/<cam>.jpg, frames/<cam>/NNN.jpg, depth_<cam>.npy, proprio.npy
+
+Errors: ValueError -- refused, nothing changed (400: wrong type or shape, NaN, a bad dt or at, too many frames, an
+unknown camera; or a request over 64 MB, refused before it is sent); EpisodeOver -- the episode, or the trial, is over
+(409); RuntimeError -- the service failed (5xx: see spec()["on_service_error"]) or the endpoint does not exist in this
+run (404).
+"""
+from __future__ import annotations
+
+import json
+import os
+import pathlib
+import time
+
+import numpy as np
+import requests
+
+import rcb1
+
+PROTOCOLS = ("eai-standard/2.", "rcb-limited/2.") # any 2.x service, under either name (rcb-limited is the old one)
+MAX_BODY = 64 * 1024 * 1024 # the protocol's request limit: a larger request is refused
+
+
+class EpisodeOver(RuntimeError):
+ pass
+
+
+class Result:
+ """A decoded RCB1 answer: meta, video, frames, last, depth, proprio; save(dir) writes them all out."""
+
+ def __init__(self, meta: dict, parts: dict):
+ self.meta, self.raw = meta, parts
+ self.video, self.frames, self.last, self.depth, self.proprio = {}, {}, {}, {}, None
+ for name, p in parts.items():
+ kind, data = p["kind"], p["data"]
+ if name.startswith("video/") and kind == "mp4":
+ self.video[name[6:]] = data
+ elif name.startswith("frames/"):
+ self.frames[name[7:]] = data
+ elif name.startswith("last/"):
+ self.last[name[5:]] = data
+ elif name.startswith("depth/"):
+ self.depth[name[6:]] = data
+ elif name == "proprio":
+ self.proprio = data
+
+ @property
+ def done(self) -> bool:
+ return bool(self.meta.get("done"))
+
+ def save(self, out_dir: str) -> str:
+ from PIL import Image
+ out = pathlib.Path(out_dir)
+ out.mkdir(parents=True, exist_ok=True)
+ (out / "meta.json").write_text(json.dumps(self.meta, indent=1))
+ for cam, b in self.video.items():
+ (out / f"{cam}.mp4").write_bytes(b)
+ for cam, img in self.last.items():
+ (out / "last").mkdir(exist_ok=True)
+ Image.fromarray(np.ascontiguousarray(img)).save(out / "last" / f"{cam}.jpg", quality=90)
+ for cam, arr in self.frames.items():
+ d = out / "frames" / cam
+ d.mkdir(parents=True, exist_ok=True)
+ for i, f in enumerate(arr):
+ Image.fromarray(np.ascontiguousarray(f)).save(d / f"{i:03d}.jpg", quality=85)
+ for cam, arr in self.depth.items():
+ np.save(out / f"depth_{cam}.npy", arr)
+ if self.proprio is not None:
+ np.save(out / "proprio.npy", self.proprio)
+ return str(out)
+
+
+class Robot:
+ def __init__(self, base: str | None = None, timeout: float = 600.0):
+ candidates = [c for c in (base or os.environ.get("SIM_URL"), "http://sim:8000", "http://localhost:8000") if c]
+ t0, last = time.time(), None
+ while time.time() - t0 < timeout:
+ for b in candidates:
+ try:
+ r = requests.get(b + "/spec", timeout=(5, 60))
+ if r.status_code == 503:
+ last = "warming up"
+ continue
+ r.raise_for_status()
+ self.base, self._spec = b, r.json()
+ if not str(self._spec.get("protocol", "")).startswith(PROTOCOLS):
+ raise RuntimeError(f"not an eai-standard/2.x service: {self._spec.get('protocol')}")
+ return
+ except (requests.exceptions.ConnectionError, requests.exceptions.Timeout) as e:
+ last = e # not up yet (the scene takes about a minute), or busy
+ time.sleep(2)
+ raise RuntimeError(f"sim service unavailable: {last}")
+
+ def spec(self) -> dict:
+ return self._spec
+
+ def status(self) -> dict:
+ r = requests.get(self.base + "/status", timeout=(5, 60))
+ r.raise_for_status()
+ return r.json()
+
+ def get_obs(self, cameras=None, seconds: float = 0.0, video_fps: int = 10, include_depth: bool = False,
+ video: str = "frames") -> Result:
+ meta = {"cameras": _names(cameras), "seconds": float(seconds), "video_fps": int(video_fps),
+ "include_depth": bool(include_depth), "video": video}
+ return self._post("/observe", meta, [], timeout=900)
+
+ def submit(self, action_type: str, actions, dt: float | None = None, interpolate: str = "linear",
+ at: int | None = None, cameras=None, video_fps: int = 10, video: str = "none",
+ include_depth: bool = False, note: str = "") -> Result:
+ a = np.asarray(actions, dtype=np.float64)
+ if a.ndim == 1:
+ a = a[None, :]
+ if not np.all(np.isfinite(a)):
+ raise ValueError("actions contain NaN or inf")
+ meta = {"type": action_type, "dt": None if dt is None else float(dt), "interpolate": interpolate,
+ "at": None if at is None else int(at), "cameras": _names(cameras), "video_fps": int(video_fps),
+ "video": video,
+ "include_depth": bool(include_depth), "note": note}
+ return self._post("/submit", meta, [rcb1.array_part("actions", a)], timeout=1800)
+
+ def reset(self, cameras=None, include_depth: bool = False) -> Result:
+ return self._post("/reset", {"cameras": _names(cameras), "include_depth": bool(include_depth)}, [], timeout=900)
+
+ def _post(self, path: str, meta: dict, parts: list, timeout: float) -> Result:
+ body = rcb1.encode(meta, parts)
+ if len(body) > MAX_BODY: # the service would close the connection on it unread
+ raise ValueError(f"{path}: {len(body)} bytes, over the protocol's {MAX_BODY}: send fewer rows")
+ r = requests.post(self.base + path, data=body, headers={"Content-Type": "application/x-rcb1"},
+ timeout=(5, timeout))
+ if r.headers.get("Content-Type", "").startswith("application/x-rcb1") and r.status_code == 200:
+ return Result(*rcb1.decode(r.content))
+ try:
+ err = r.json()
+ except ValueError:
+ err = {"error": r.text[:300]}
+ msg = f"{r.status_code} {path}: {err.get('error')}" + (f" ({err['detail']})" if err.get("detail") else "")
+ if r.status_code == 409:
+ raise EpisodeOver(msg + (f"; reason {err['reason']}" if err.get("reason") else "")
+ + (f"; resets_left {err['resets_left']}" if "resets_left" in err else ""))
+ if r.status_code in (400, 413):
+ raise ValueError(msg)
+ raise RuntimeError(msg)
+
+
+def _names(cameras) -> list[str] | None:
+ return None if cameras is None else [cameras] if isinstance(cameras, str) else list(cameras)
diff --git a/tasks/myosuite-table-tennis-i00-standard-realtime/environment/rcb_monitor/limits.json b/tasks/myosuite-table-tennis-i00-standard-realtime/environment/rcb_monitor/limits.json
new file mode 100644
index 0000000000..fa8f1b1a62
--- /dev/null
+++ b/tasks/myosuite-table-tennis-i00-standard-realtime/environment/rcb_monitor/limits.json
@@ -0,0 +1,38 @@
+{
+ "caps": {
+ "main": {
+ "cpus": 4.0,
+ "memory_gb": 10.0,
+ "pids": 8192
+ },
+ "sim": {
+ "cpus": 4.0,
+ "memory_gb": 8.0,
+ "pids": 8192
+ }
+ },
+ "disk_exclude": [],
+ "disk_paths": [
+ "/app",
+ "/tmp",
+ "/root"
+ ],
+ "disk_tick_sec": 60,
+ "format": "rcb-limits/1",
+ "gpu": false,
+ "max_disk_gb": 20.0,
+ "max_gpu_memory_gb": 0.0,
+ "max_memory_gb": 8.0,
+ "max_processes": 512,
+ "mode": "standard-realtime",
+ "scope": "agent",
+ "scope_services": [
+ "main"
+ ],
+ "sustain_sec": 60.0,
+ "task": "myosuite-table-tennis-i00-standard-realtime",
+ "tick_sec": 5,
+ "trial_max_gpu_memory_gb": 4.0,
+ "trial_max_memory_gb": 14.0,
+ "window_sec": 120.0
+}
diff --git a/tasks/myosuite-table-tennis-i00-standard-realtime/environment/rcb_monitor/monitor.py b/tasks/myosuite-table-tennis-i00-standard-realtime/environment/rcb_monitor/monitor.py
new file mode 100644
index 0000000000..81fdc91877
--- /dev/null
+++ b/tasks/myosuite-table-tennis-i00-standard-realtime/environment/rcb_monitor/monitor.py
@@ -0,0 +1,966 @@
+#!/usr/bin/env python3
+"""rcb_monitor: the trusted resource-limits monitor sidecar of every task (docs/PROTOCOL.md §2.10, record rcb-monitor/1).
+
+Runs as the compose service `rcb_monitor` (generated by scripts/rcb_monitor/inject.py): `pid: host`, `cgroup: host`,
+`network_mode: none`, CAP_SYS_PTRACE, the task's GPU reservation, its code read-only from the task directory. The agent
+cannot see it (another PID namespace), reach it (no network) or write its record (its own file system).
+
+Which containers it watches: the services that mount the trial's beacon volume `rcb_limits` (read-only at
+/run/rcb-limits) are in scope, the ones that mount `rcb_observe` are only recorded (the benchmark's own services: the
+simulator sidecar of standard modes, RoboPaint's privileged `scorer`). The compose project names both volumes, so they are unique to the trial; the monitor finds the containers as
+the mount namespaces whose /run/rcb-limits is the same directory as its own (docker compose exec'd processes included).
+A namespace that does not show the beacon is asked again at every tick (a container that is still being set up has the
+host's root until its pivot_root: "no" is never remembered); a "yes" is kept only while the process that proved it lives.
+A process whose first thread has exited is found through a thread that lives; a zombie counts in its parent's container.
+
+Every tick (5 s; disk every 60 s) it samples, per container and summed over the scope: GPU memory (NVML, per process,
+compute and graphics), memory (cgroup anon+shmem when the container has its own cgroup, else PSS anon+shmem), processes
+(TGIDs, zombies included), disk growth under the limits' disk_paths in the agent's container (complete walks only, in
+slices between the samples), CPU (recorded only). It appends to /var/rcb-monitor/samples.jsonl and
+events.jsonl and writes the agent-visible /run/rcb-limits/usage.json.
+
+Kernel OOM kills (`oom` events) are read from each watched container's cgroup `memory.events` by a thread of its own
+(OomWatch), many times a second: a kill that takes the container's main process with it ends the container, and compose
+restarts it with a new cgroup whose counters start at 0, within milliseconds -- a reading once per tick never sees it.
+Every event says whom it hit (`scope`: a container the limits bind, or a benchmark-owned service that is only
+recorded) and why (`cause`): "cap" when the container's own `oom` counter rose with it (the kernel says its memory
+reached its own cap), else "host" (the host's OOM killer, or an enclosing cgroup's: not this container's doing).
+Only a "cap" kill in the scope is a violation (rcb_limits.violations); the verifier step reports the others.
+
+RCB_LIMITS_ENFORCE=1 (run-time switch; compose default 0; 0/1/true/false, anything else is a fatal record): once a
+violation is confirmed (rcb_limits.violations, the verifier's own rule) the monitor SIGKILLs every in-scope process
+except each container's PID 1, so the agent stops at once and Harbor goes on to collection and the verifier (an agent
+killed this way ends like one that exited non-zero). It also protects the host (`emergency_kill` events, never a
+violation by themselves): above EMERGENCY_GPU_FACTOR x the GPU-memory limit at one sample it kills the scope's process
+that holds the most GPU memory; where the runtime applies no kernel memory cap, it stands in for it and kills the
+largest process once the scope is above its declared memory caps.
+0: it only records.
+
+ python3 monitor.py run # the service's command
+ python3 monitor.py seal --wait 90 # the first sidecar collect hook: seal the record (summary.json), print it
+"""
+import ctypes
+import json
+import os
+import select
+import signal
+import stat as stat_mod
+import sys
+import threading
+import time
+
+HERE = os.path.dirname(os.path.abspath(__file__))
+sys.path.insert(0, HERE)
+import rcb_limits # noqa: E402
+import selftest # noqa: E402
+
+REC = os.environ.get("RCB_MONITOR_RECORD", "/var/rcb-monitor")
+PROC, CGROUP = "/proc", "/sys/fs/cgroup" # module globals: scripts/rcb_monitor/test_monitor.py points them at a fake tree
+SCOPE_BEACON, OBSERVE_BEACON = "/run/rcb-limits", "/run/rcb-observe"
+LIMITS = os.path.join(HERE, "limits.json")
+HZ = os.sysconf("SC_CLK_TCK")
+PAGE = os.sysconf("SC_PAGE_SIZE")
+GIB = float(2 ** 30)
+EMERGENCY_GPU_FACTOR = 1.5 # host protection (enforced runs): above this x the GPU-memory limit at one sample
+DISK_SLICE_SEC = 1.5 # of every tick a running disk walk may take (the monitor's CPU cap is half a core)
+DISK_WALK_MAX_SEC = 300.0 # a walk not finished by then is partial: a gap in the record, never a baseline
+OOM_COLD_SEC = 0.02 # OomWatch: every memory.events is read this often (or at once, when the kernel notifies)
+OOM_HOT_SEC = 0.002 # ... and this often for OOM_HOT_FOR_SEC after any of them changed (memory pressure)
+OOM_HOT_FOR_SEC = 2.0
+OOM_CAP_WINDOW_SEC = 10.0 # a kill is the container's own cap's when its `oom` counter rose with it or this shortly before
+OOM_FORGET_SEC = 600.0 # a cgroup not seen by the sampler for this long is no longer watched
+
+
+# --------------------------------------------------------------------------------------------------- small helpers
+def now():
+ return time.time()
+
+
+def rl(p):
+ try:
+ return os.readlink(p)
+ except OSError:
+ return None
+
+
+def read(p):
+ try:
+ with open(p) as f:
+ return f.read()
+ except OSError:
+ return None
+
+
+def stat_id(p):
+ try:
+ st = os.stat(p)
+ return (st.st_dev, st.st_ino)
+ except OSError:
+ return None
+
+
+def stat_of(base):
+ """(starttime, utime+stime) of the process or thread at /proc path `base`, or None."""
+ s = read(base + "/stat")
+ if not s:
+ return None
+ try:
+ f = s.rsplit(")", 1)[1].split()
+ return int(f[19]), int(f[11]) + int(f[12])
+ except (IndexError, ValueError):
+ return None
+
+
+def proc_stat(pid):
+ return stat_of(f"{PROC}/{pid}")
+
+
+def status(pid):
+ s = read(f"{PROC}/{pid}/status")
+ if not s:
+ return {}
+ d = {}
+ for line in s.splitlines():
+ k, _, v = line.partition(":")
+ d[k] = v.strip()
+ return d
+
+
+def kb(v):
+ try:
+ return int((v or "0").split()[0])
+ except (ValueError, IndexError):
+ return 0
+
+
+def append(name, obj):
+ with open(os.path.join(REC, name), "a") as f:
+ f.write(json.dumps(obj, separators=(",", ":")) + "\n")
+
+
+def write_json(path, obj):
+ tmp = path + ".tmp"
+ with open(tmp, "w") as f:
+ json.dump(obj, f, indent=1)
+ os.replace(tmp, path)
+
+
+def event(kind, **kw):
+ kw.update(t=round(now(), 2), event=kind)
+ append("events.jsonl", kw)
+
+
+def no_dump():
+ """Not dumpable: without CAP_SYS_PTRACE nobody (not even the same uid) can open /proc/<monitor>/{root,mem,fd}."""
+ try:
+ ctypes.CDLL(None, use_errno=True).prctl(4, 0, 0, 0, 0) # PR_SET_DUMPABLE = 4
+ except Exception:
+ pass
+
+
+# --------------------------------------------------------------------------------------------------- NVML
+class Nvml:
+ """Per-process GPU memory (host PIDs: the monitor runs in the host PID namespace) and per-card use, via libnvidia-ml
+ (ctypes; no nvidia-smi child processes)."""
+
+ class Proc(ctypes.Structure):
+ _fields_ = [("pid", ctypes.c_uint), ("usedGpuMemory", ctypes.c_ulonglong),
+ ("gpuInstanceId", ctypes.c_uint), ("computeInstanceId", ctypes.c_uint)]
+
+ class Proc1(ctypes.Structure):
+ _fields_ = [("pid", ctypes.c_uint), ("usedGpuMemory", ctypes.c_ulonglong)]
+
+ class Mem(ctypes.Structure):
+ _fields_ = [("total", ctypes.c_ulonglong), ("free", ctypes.c_ulonglong), ("used", ctypes.c_ulonglong)]
+
+ def __init__(self):
+ self.lib = ctypes.CDLL("libnvidia-ml.so.1")
+ if self.lib.nvmlInit_v2() != 0:
+ raise RuntimeError("nvmlInit failed")
+ n = ctypes.c_uint(0)
+ self.lib.nvmlDeviceGetCount_v2(ctypes.byref(n))
+ self.handles = []
+ for i in range(n.value):
+ h = ctypes.c_void_p()
+ if self.lib.nvmlDeviceGetHandleByIndex_v2(i, ctypes.byref(h)) == 0:
+ self.handles.append(h)
+ self.fns = []
+ for name, st in (("nvmlDeviceGetComputeRunningProcesses_v3", self.Proc),
+ ("nvmlDeviceGetGraphicsRunningProcesses_v3", self.Proc),
+ ("nvmlDeviceGetComputeRunningProcesses_v2", self.Proc),
+ ("nvmlDeviceGetGraphicsRunningProcesses_v2", self.Proc),
+ ("nvmlDeviceGetComputeRunningProcesses", self.Proc1),
+ ("nvmlDeviceGetGraphicsRunningProcesses", self.Proc1)):
+ kind = "compute" if "Compute" in name else "graphics"
+ if any(k == kind for k, _, _ in self.fns):
+ continue
+ try:
+ self.fns.append((kind, getattr(self.lib, name), st))
+ except AttributeError:
+ pass
+
+ def sample(self):
+ """({host pid: MiB}, {card index: (used MiB, total MiB)}, {host pids whose memory NVML does not tell})."""
+ per, cards, unknown = {}, {}, set()
+ for i, h in enumerate(self.handles):
+ m = self.Mem()
+ if self.lib.nvmlDeviceGetMemoryInfo(h, ctypes.byref(m)) == 0:
+ cards[i] = (m.used // 2 ** 20, m.total // 2 ** 20)
+ seen = {}
+ for _kind, fn, st in self.fns:
+ cnt = ctypes.c_uint(0)
+ rc = fn(h, ctypes.byref(cnt), None)
+ if rc not in (0, 7): # 7 = NVML_ERROR_INSUFFICIENT_SIZE: cnt now holds the count
+ continue
+ cnt = ctypes.c_uint(cnt.value + 16)
+ arr = (st * cnt.value)()
+ if fn(h, ctypes.byref(cnt), arr) != 0:
+ continue
+ for k in range(cnt.value):
+ mem = arr[k].usedGpuMemory
+ if mem >= 2 ** 63: # NVML_VALUE_NOT_AVAILABLE (MIG / vGPU / WSL): not a zero
+ unknown.add(arr[k].pid)
+ mib = 0
+ else:
+ mib = mem // 2 ** 20
+ seen[arr[k].pid] = max(seen.get(arr[k].pid, 0), mib) # a C+G process is in both lists
+ for p, v in seen.items():
+ per[p] = per.get(p, 0) + v
+ return per, cards, unknown
+
+
+# --------------------------------------------------------------------------------------------------- disk
+class DiskWalk:
+ """One measurement of the files under `paths` below `root` (the agent container's root, seen through /proc): the
+ sum of st_blocks, every hard-linked file once, never descending into another mount (a directory on another device;
+ files are not asked for theirs: on an overlay root whose layers sit on different file systems they report their
+ layer's), `exclude` skipped. It runs in
+ slices (step) between the monitor's samples, so a large tree delays no sample. `done` only when every directory was
+ read: only then is `bytes` a measurement; before that it is a lower bound."""
+
+ def __init__(self, root, paths, exclude):
+ self.root, self.t0 = root, now()
+ self.bytes = self.entries = 0
+ self.done, self.why = False, None
+ self._it = self._walk(list(paths), {root + x.rstrip("/") for x in exclude})
+
+ def _walk(self, paths, excl):
+ linked = set() # (dev, ino) of files with several names: counted once
+ for base in paths:
+ top = self.root + base
+ try:
+ dev = os.lstat(top).st_dev
+ except FileNotFoundError:
+ continue # the path does not exist in this container: nothing under it
+ except OSError as e:
+ self.why = f"{base}: {type(e).__name__}"
+ return
+ stack = [top]
+ while stack:
+ d = stack.pop()
+ try:
+ it = os.scandir(d)
+ except FileNotFoundError:
+ continue # removed since its parent was read
+ except OSError as e:
+ self.why = f"{d[len(self.root):]}: {type(e).__name__}"
+ return
+ with it:
+ for e in it:
+ try:
+ st = e.stat(follow_symlinks=False)
+ except OSError:
+ continue # removed meanwhile
+ if stat_mod.S_ISDIR(st.st_mode):
+ if st.st_dev != dev or e.path in excl:
+ continue # another mount (never crossed: not the container's own files)
+ stack.append(e.path)
+ elif st.st_nlink > 1:
+ k = (st.st_dev, st.st_ino)
+ if k in linked:
+ continue
+ linked.add(k)
+ self.bytes += st.st_blocks * 512
+ self.entries += 1
+ if self.entries % 256 == 0:
+ yield
+ yield
+ self.done = True
+
+ def step(self, budget_sec):
+ """Walk on for at most `budget_sec`; True when the walk has ended (done, or failed with `why`)."""
+ end = now() + budget_sec
+ for _ in self._it:
+ if now() >= end:
+ return False
+ return True
+
+
+# --------------------------------------------------------------------------------------------------- OOM kills
+class OomWatch:
+ """Kernel OOM kills of the watched containers, from each one's cgroup v2 `memory.events`:
+ oom times the cgroup's memory reached its own limit (memory.max) and an allocation was about to fail
+ oom_kill processes of the cgroup killed by any OOM killer (its own limit's, an enclosing cgroup's, the host's)
+ A kill whose `oom` counter rose with it (or within OOM_CAP_WINDOW_SEC before) happened at the container's own cap
+ (`cause: "cap"`); one without is the host's or an enclosing cgroup's OOM killer (`cause: "host"`), which the
+ container did not cause by its own use (PROTOCOL §2.10: nothing another trial does can fail a trial).
+
+ The counters live and die with the cgroup, and a container whose main process is killed is gone (and restarted by
+ compose with fresh counters) within milliseconds. So the files are read by this object's own thread, every
+ OOM_COLD_SEC, at once when the kernel signals a change (poll: memory.events notifies), and every OOM_HOT_SEC for a
+ while after any change (the `max` counter rises under memory pressure, before a kill). A kill that removes the
+ cgroup within about a millisecond can still be missed: such a kill is then in no record.
+
+ scan() reads every file once; the sampler calls it at each tick when the thread does not run (the tests). The
+ thread ends when the record is sealed."""
+
+ def __init__(self):
+ self.lock = threading.Lock()
+ self.w = {} # cgroup dir -> {"role", "scope", "fd", "kill", "oom", "sig", "cap_t", "seen"}
+ self.hot_until = 0.0
+ self.thread = None
+ self.stopped = False # set at the seal: the record is closed, like the samples
+
+ def watch(self, cg, role, scope, t):
+ """The sampler saw this container's cgroup at time t."""
+ with self.lock:
+ e = self.w.get(cg)
+ if e is None:
+ e = self.w[cg] = {"fd": None, "kill": None, "oom": None, "sig": None, "cap_t": None}
+ e.update(role=role, scope=bool(scope), seen=t)
+
+ @staticmethod
+ def _close(e):
+ if e["fd"] is not None:
+ try:
+ os.close(e["fd"])
+ except OSError:
+ pass
+ e["fd"] = None
+
+ def _read(self, cg, e):
+ """The counters of the cgroup's memory.events ({} when the file cannot be read: the container is not up).
+ A file that stops answering belonged to a cgroup that is gone; the same path opened again is a new cgroup (the
+ container restarted: another directory, by its inode and creation time), whose counters all belong to the
+ trial: they are compared with 0. The same cgroup opened again keeps what was already counted."""
+ for _ in (0, 1):
+ try:
+ if e["fd"] is None:
+ e["fd"] = os.open(cg + "/memory.events", os.O_RDONLY)
+ st = os.stat(cg)
+ ident = (st.st_dev, st.st_ino, st.st_ctime_ns)
+ if e["sig"] is not None and ident != e.get("ident"):
+ e["kill"] = e["oom"] = 0 # read before, gone, here again: another instance of the cgroup
+ e["ident"] = ident
+ os.lseek(e["fd"], 0, os.SEEK_SET)
+ out = {}
+ for line in os.read(e["fd"], 8192).decode("ascii", "replace").splitlines():
+ k, _, v = line.partition(" ")
+ if v.strip().isdigit():
+ out[k] = int(v)
+ return out
+ except OSError:
+ self._close(e)
+ return {}
+
+ def scan(self):
+ """Read every watched file once and record the kills since the last reading. True when a counter changed."""
+ changed = False
+ with self.lock:
+ t = now()
+ for cg, e in list(self.w.items()):
+ if t - e["seen"] > OOM_FORGET_SEC:
+ self._close(e)
+ del self.w[cg]
+ continue
+ c = self._read(cg, e)
+ if "oom_kill" not in c:
+ continue
+ kill, oom = c["oom_kill"], c.get("oom", 0)
+ sig = (kill, oom, c.get("max", 0))
+ if e["kill"] is None or kill < e["kill"] or oom < e["oom"]:
+ # the first reading is the baseline; counters that went back are a new cgroup under the same path
+ e["kill"], e["oom"] = (kill, oom) if e["kill"] is None else (0, 0)
+ changed |= sig != e["sig"]
+ if oom > e["oom"]:
+ e["cap_t"] = t
+ if kill > e["kill"]:
+ at_cap = e["cap_t"] is not None and t - e["cap_t"] <= OOM_CAP_WINDOW_SEC
+ kw = {"role": e["role"], "scope": e["scope"], "count": kill - e["kill"],
+ "cause": "cap" if at_cap else "host", "cap_hits": oom - e["oom"]}
+ for name, key in (("memory.max", "memory_cap_gb"), ("memory.current", "memory_current_gb")):
+ v = (read(cg + "/" + name) or "").strip() # evidence only: the cause is the counters'
+ if v.isdigit():
+ kw[key] = round(int(v) / GIB, 3)
+ event("oom", **kw)
+ e["kill"], e["oom"], e["sig"] = kill, oom, sig
+ return changed
+
+ def start(self):
+ self.thread = threading.Thread(target=self._loop, name="rcb-oom-watch", daemon=True)
+ self.thread.start()
+
+ def _loop(self):
+ said = False
+ while not self.stopped:
+ try:
+ if self.scan():
+ self.hot_until = now() + OOM_HOT_FOR_SEC
+ if now() < self.hot_until:
+ time.sleep(OOM_HOT_SEC)
+ continue
+ with self.lock:
+ fds = [e["fd"] for e in self.w.values() if e["fd"] is not None]
+ if not fds:
+ time.sleep(OOM_COLD_SEC)
+ continue
+ p = select.poll() # memory.events notifies its readers (POLLPRI) when it changes
+ for fd in fds:
+ p.register(fd, select.POLLPRI | select.POLLERR)
+ if p.poll(OOM_COLD_SEC * 1000):
+ self.hot_until = now() + OOM_HOT_FOR_SEC
+ except Exception as e: # never stop watching
+ if not said:
+ said = True
+ try:
+ event("error", why=f"oom watch: {type(e).__name__}: {e}"[:300])
+ except Exception:
+ pass
+ time.sleep(OOM_COLD_SEC)
+
+
+# --------------------------------------------------------------------------------------------------- the monitor
+class Monitor:
+ def __init__(self):
+ self.lim = rcb_limits.read_json(LIMITS) or {}
+ self.tick = float(self.lim.get("tick_sec", 5))
+ self.disk_tick = float(self.lim.get("disk_tick_sec", 60))
+ self.enforce, self.switch_error = rcb_limits.parse_switch(os.environ.get("RCB_LIMITS_ENFORCE"))
+ self.me_mnt = rl(f"{PROC}/self/ns/mnt")
+ self.beacon = {"scope": stat_id(SCOPE_BEACON), "observe": stat_id(OBSERVE_BEACON)}
+ self.cls = {} # mnt ns -> {"kind", "role", "witness": (proc path, starttime)}: only what IS watched
+ self.neg, self.ok = set(), set() # this tick: namespaces that said no / whose witness still lives
+ self.via = {} # pid -> /proc path of a living thread, for a process whose first thread has exited
+ self.prev_cpu = {} # pid -> (starttime, ticks, t)
+ self.oomw = OomWatch() # kernel OOM kills of the watched containers (its own thread in run())
+ self.disk_base = None # GiB under disk_paths at the first complete walk (kept in state.json across restarts)
+ self.disk_last = 0.0
+ self.disk_gb = None
+ self.walk = None # (DiskWalk, mnt ns, proc path, starttime) of the measurement under way
+ self.series = [] # [(t, {metric: value})] of the last window, for enforcement decisions
+ self.confirmed = {} # metric -> time
+ self.kill_until = 0.0
+ self.killed = 0
+ self.emergency = [] # the emergency_kill events of this run
+ self.scope_seen = False
+ self.gpu_unknown = False
+ self.caps = selftest.capabilities()
+ self.nvml = None
+ try:
+ self.nvml = Nvml()
+ except Exception as e:
+ self.caps["nvml_error"] = f"{type(e).__name__}: {e}"[:200]
+ self.caps["nvml"] = int(self.nvml is not None)
+ self.caps["nvml_devices"] = len(self.nvml.handles) if self.nvml else 0
+ self.t0 = now()
+ self._last_measures, self._gpu_per = [], {}
+ self.fatal_scope = False
+
+ # ---- scope discovery
+ def base(self, pid):
+ """The /proc path to read a process's namespaces, root and memory from."""
+ return self.via.get(pid) or f"{PROC}/{pid}"
+
+ def classify(self, mnt, base):
+ """{"kind": "scope" | "observe", "role": "main" | "sidecar"} when the mount namespace `mnt` (of the process or
+ thread at /proc path `base`) is one of the trial's containers, else None.
+
+ "No" is never remembered beyond this tick: while a container is set up, its init process is already in a new
+ mount namespace but still has the host's root, and the beacon appears only with its pivot_root. "Yes" is
+ remembered while the process that proved it lives in the namespace (the kernel may give the number of a
+ namespace that ended to a new one)."""
+ if mnt in self.ok:
+ return self.cls[mnt]
+ if mnt in self.neg:
+ return None
+ c = self.cls.get(mnt)
+ if c is not None:
+ wbase, wstart = c["witness"]
+ ws = stat_of(wbase)
+ if ws and ws[0] == wstart and rl(wbase + "/ns/mnt") == mnt:
+ self.ok.add(mnt)
+ return c
+ del self.cls[mnt]
+ root = base + "/root"
+ kind = None
+ if self.beacon["scope"] and stat_id(root + SCOPE_BEACON) == self.beacon["scope"]:
+ kind = "scope"
+ elif self.beacon["observe"] and stat_id(root + OBSERVE_BEACON) == self.beacon["observe"]:
+ kind = "observe"
+ st = stat_of(base)
+ here = st is not None and rl(base + "/ns/mnt") == mnt # a vanished process tells nothing: ask the next
+ if kind and here:
+ c = {"kind": kind, "role": "main" if os.path.isdir(root + "/logs/agent") else "sidecar",
+ "witness": (base, st[0])}
+ self.cls[mnt] = c
+ self.ok.add(mnt)
+ return c
+ if here:
+ self.neg.add(mnt)
+ return None
+
+ def live_thread(self, pid):
+ """/proc path of a thread of `pid` that still runs (its first thread has exited: /proc/<pid>/ns is empty)."""
+ try:
+ tids = os.listdir(f"{PROC}/{pid}/task")
+ except OSError:
+ return None
+ for tid in tids:
+ b = f"{PROC}/{pid}/task/{tid}"
+ if tid != str(pid) and rl(b + "/ns/mnt"):
+ return b
+ return None
+
+ def groups(self):
+ """{mnt ns: {"kind", "role", "pids": [...], "zombies": n}} of the watched containers."""
+ self.neg, self.ok, self.via = set(), set(), {}
+ out, where, lost = {}, {}, []
+ try:
+ names = os.listdir(PROC)
+ except OSError:
+ names = []
+ for d in names:
+ if not d.isdigit():
+ continue
+ pid = int(d)
+ mnt = rl(f"{PROC}/{pid}/ns/mnt")
+ if mnt is None:
+ lost.append(pid)
+ continue
+ where[pid] = mnt
+ if mnt == self.me_mnt:
+ continue
+ c = self.classify(mnt, f"{PROC}/{pid}")
+ if c:
+ out.setdefault(mnt, {"kind": c["kind"], "role": c["role"], "pids": [], "zombies": 0})["pids"].append(pid)
+ zombies = []
+ for pid in lost:
+ b = self.live_thread(pid)
+ if not b:
+ zombies.append(pid)
+ continue
+ mnt = rl(b + "/ns/mnt") # its first thread exited, the others run on: a process like any other
+ if mnt is None:
+ continue
+ where[pid] = mnt
+ if mnt == self.me_mnt:
+ continue
+ c = self.classify(mnt, b)
+ if c:
+ self.via[pid] = b
+ out.setdefault(mnt, {"kind": c["kind"], "role": c["role"], "pids": [], "zombies": 0})["pids"].append(pid)
+ for pid in zombies:
+ st = status(pid)
+ if st.get("State", "").startswith("Z"): # a zombie holds a PID until its parent reaps it: counted there
+ try:
+ g = out.get(where.get(int(st.get("PPid", "0"))))
+ except ValueError:
+ g = None
+ if g:
+ g["zombies"] += 1
+ self.cls = {m: c for m, c in self.cls.items() if m in self.ok}
+ return out
+
+ # ---- per-container measurements
+ def cgroup_of(self, pids):
+ """The container's own cgroup v2 directory (under the host tree, `cgroup: host`) when every process in it is this
+ container's (rootless podman without cgroup delegation leaves containers in a shared one), else None."""
+ paths = set()
+ for p in pids[:8]:
+ s = read(f"{PROC}/{p}/cgroup") or ""
+ for line in s.splitlines():
+ if line.startswith("0::"):
+ paths.add(line[3:].strip())
+ if len(paths) != 1:
+ return None
+ d = CGROUP + paths.pop()
+ procs = read(d + "/cgroup.procs")
+ if procs is None:
+ return None
+ members = {int(x) for x in procs.split() if x.isdigit()}
+ return d if members and members <= set(pids) else None
+
+ def mem_pss(self, pids):
+ tot = 0
+ for p in pids:
+ s = read(self.base(p) + "/smaps_rollup")
+ if s is None:
+ st = status(p)
+ tot += kb(st.get("RssAnon")) + kb(st.get("RssShmem"))
+ continue
+ d = {}
+ for line in s.splitlines():
+ k, _, v = line.partition(":")
+ d[k] = v
+ tot += (kb(d.get("Pss_Anon")) + kb(d.get("Pss_Shmem"))) if "Pss_Anon" in d else kb(d.get("Pss"))
+ return tot / 1024.0 / 1024.0 # GiB
+
+ def measure(self, g, gpu_per, t, gpu_unknown=()):
+ pids = g["pids"]
+ threads = cpu = 0.0
+ for p in pids:
+ threads += int(status(p).get("Threads", "0") or 0)
+ ps = proc_stat(p)
+ if ps:
+ pv = self.prev_cpu.get(p)
+ if pv and pv[0] == ps[0] and t > pv[2]:
+ cpu += max(0, ps[1] - pv[1]) / HZ / (t - pv[2])
+ self.prev_cpu[p] = (ps[0], ps[1], t)
+ cg = self.cgroup_of(pids)
+ m = {"kind": g["kind"], "role": g["role"], "processes": len(pids) + g.get("zombies", 0),
+ "zombies": g.get("zombies", 0), "threads": int(threads), "cpu": round(cpu, 2),
+ "gpu_memory_gb": round(sum(gpu_per.get(p, 0) for p in pids) / 1024.0, 3), "cgroup": int(cg is not None)}
+ unknown = sum(1 for p in pids if p in gpu_unknown)
+ if unknown:
+ m["gpu_unmeasured"] = unknown # processes on the GPU whose memory NVML does not tell
+ if cg:
+ st = {}
+ for line in (read(cg + "/memory.stat") or "").splitlines():
+ k, _, v = line.partition(" ")
+ st[k] = v
+ try:
+ m["memory_gb"] = round((int(st.get("anon", 0)) + int(st.get("shmem", 0))) / GIB, 3)
+ except ValueError:
+ m["memory_gb"] = round(self.mem_pss(pids), 3)
+ mx = (read(cg + "/memory.max") or "").strip()
+ m["memory_cap_gb"] = None if mx in ("", "max") else round(int(mx) / GIB, 2)
+ self.oomw.watch(cg, g["role"], g["kind"] == "scope", now())
+ else:
+ m["memory_gb"] = round(self.mem_pss(pids), 3)
+ return m
+
+ # ---- disk
+ def main_proc(self, groups):
+ """(mnt ns, pid) to reach the agent container's files through: its PID 1 when visible (it outlives the others)."""
+ for mnt, g in groups.items():
+ if g["kind"] == "scope" and g["role"] == "main" and g["pids"]:
+ init = [p for p in g["pids"] if (status(p).get("NSpid") or "").split()[-1:] == ["1"]]
+ return mnt, (init[0] if init else min(g["pids"]))
+ return None, None
+
+ def disk_step(self, t, groups):
+ """The disk value of this sample: GiB the files under disk_paths grew since the baseline, or None (not measured
+ at this tick). Only a complete walk of a container that is still there is a measurement; only such a walk sets
+ the baseline. A walk that did not finish is a lower bound: it can show that the limit is exceeded, never that it
+ is not, and it never becomes the baseline."""
+ if self.walk is None:
+ mnt, pid = self.main_proc(groups)
+ if pid is None or t - self.disk_last < self.disk_tick:
+ return None
+ b = self.base(pid)
+ st = stat_of(b)
+ if st is None:
+ return None
+ self.disk_last = t
+ self.walk = (DiskWalk(b + "/root", self.lim.get("disk_paths") or ["/app"], self.lim.get("disk_exclude") or []),
+ mnt, b, st[0])
+ w, mnt, b, start = self.walk
+ ended = w.step(DISK_SLICE_SEC)
+ if not ended and now() - w.t0 <= DISK_WALK_MAX_SEC:
+ return None # goes on at the next tick
+ self.walk = None
+ st = stat_of(b)
+ there = st is not None and st[0] == start and rl(b + "/ns/mnt") == mnt # the same container, still up
+ gb = w.bytes / GIB
+ if ended and w.done and there:
+ if self.disk_base is None:
+ self.disk_base = gb
+ self.save_state()
+ event("disk_baseline", gb=round(gb, 3), entries=w.entries)
+ self.disk_gb = round(max(0.0, gb - self.disk_base), 3)
+ return self.disk_gb
+ why = "the container's process ended during the walk" if not there else (w.why or "not finished in time")
+ event("disk_partial", why=why, gb=round(gb, 3), entries=w.entries, sec=round(now() - w.t0, 1))
+ lim = float(self.lim.get("max_disk_gb") or 0)
+ if there and self.disk_base is not None and lim and gb - self.disk_base > lim:
+ self.disk_gb = round(gb - self.disk_base, 3) # already over the limit, whatever the rest holds
+ return self.disk_gb
+ return None
+
+ # ---- state that must outlive a restart of the monitor (restart: on-failure)
+ def save_state(self):
+ try:
+ write_json(os.path.join(REC, "state.json"), {"disk_base": self.disk_base})
+ except OSError:
+ pass
+
+ def restore(self):
+ """After a restart: the disk baseline, the confirmed violations and the last window of samples come back from
+ the record, so that a crash neither resets the baseline nor forgets time already spent over a limit."""
+ st = rcb_limits.read_json(os.path.join(REC, "state.json")) or {}
+ if isinstance(st.get("disk_base"), (int, float)):
+ self.disk_base = float(st["disk_base"])
+ _, samples, events = rcb_limits.load_record(REC)
+ window = float(self.lim.get("window_sec", 2 * float(self.lim.get("sustain_sec", 60))))
+ self.series = [(s["t"], s.get("scope") or {}) for s in samples
+ if isinstance(s.get("t"), (int, float)) and s["t"] >= self.t0 - window - 3 * self.tick]
+ for e in events:
+ k = e.get("event")
+ if k == "violation" and e.get("metric"):
+ self.confirmed.setdefault(e["metric"], e.get("t") or self.t0)
+ elif k in ("kill", "emergency_kill"):
+ self.killed += int(e.get("processes") or 0)
+ if k == "emergency_kill":
+ self.emergency.append(e)
+ elif k == "scope_found":
+ self.scope_seen = True
+ if self.enforce and self.confirmed and self.t0 - max(self.confirmed.values()) < 60:
+ self.kill_until = max(self.confirmed.values()) + 60
+
+ # ---- enforcement
+ @staticmethod
+ def spared(pids):
+ """The container's own processes that must outlive a kill: its PID 1 and what PID 1 started with it (Harbor's
+ `sh -c "sleep infinity"` keepalive). Agent processes are exec'd into the container (their parent is outside
+ it) or, when orphaned, adopted by PID 1 later: those are never spared."""
+ st = {p: status(p) for p in pids}
+ init = [p for p in pids if (st[p].get("NSpid") or "").split()[-1:] == ["1"]]
+ if not init:
+ return set()
+ t1 = (proc_stat(init[0]) or (0, 0))[0]
+ keep = set(init)
+ for p in pids:
+ ns_ppid = None
+ ns = (st[p].get("NSpid") or "").split()
+ if ns and st[p].get("PPid") is not None:
+ # PPid in /proc/<pid>/status is in the reader's namespace (host); compare with PID 1's host PID
+ ns_ppid = int(st[p]["PPid"])
+ ps = proc_stat(p)
+ if ns_ppid == init[0] and ps and ps[0] - t1 <= 2 * HZ: # started with the container, not adopted later
+ keep.add(p)
+ return keep
+
+ def kill_scope(self, groups, why):
+ n = 0
+ for mnt, g in groups.items():
+ if g["kind"] != "scope":
+ continue
+ snap = {p: proc_stat(p) for p in g["pids"]}
+ keep = self.spared(g["pids"])
+ for p in g["pids"]:
+ if p in keep:
+ continue # PID 1 and its keepalive: the container stays up for collection
+ if rl(self.base(p) + "/ns/mnt") != mnt or not snap[p] or proc_stat(p) is None \
+ or proc_stat(p)[0] != snap[p][0]:
+ continue # gone, or its PID now names another process
+ try:
+ os.kill(p, signal.SIGKILL)
+ n += 1
+ except OSError:
+ pass
+ if n:
+ self.killed += n
+ event("kill", why=why, processes=n)
+ return n
+
+ def scope_memory_cap(self):
+ """GiB: the declared memory caps of the services the limits bind, summed (None: not declared)."""
+ caps = self.lim.get("caps") or {}
+ names = self.lim.get("scope_services") or (
+ [n for n in caps if n != "*"] if self.lim.get("scope") == "trial" else ["main"])
+ vals = [float(caps[n]["memory_gb"]) for n in names if isinstance(caps.get(n), dict) and caps[n].get("memory_gb")]
+ return sum(vals) if vals else None
+
+ def check(self, t, scope_vals, groups):
+ self.series.append((t, scope_vals))
+ window = float(self.lim.get("window_sec", 2 * float(self.lim.get("sustain_sec", 60))))
+ self.series = [x for x in self.series if x[0] >= t - window - 3 * self.tick]
+ samples = [{"t": a, "scope": b} for a, b in self.series]
+ for v in rcb_limits.violations(samples, self.lim):
+ if v["metric"] not in self.confirmed:
+ self.confirmed[v["metric"]] = t
+ event("violation", **v)
+ if self.enforce:
+ self.kill_until = t + 60
+ if self.enforce and t <= self.kill_until:
+ self.kill_scope(groups, ",".join(sorted(self.confirmed)))
+ # Host protection (enforced runs only; PROTOCOL §2.10). An emergency kill is never a violation by itself: the
+ # trial fails only by the sustained rule above or a kernel OOM kill at its own memory cap. Both conditions are
+ # the trial's own doing (its scope's use against its own limit or caps), never the state of a card or host it
+ # shares with other trials.
+ if self.enforce:
+ cap = self.scope_memory_cap()
+ lim_gpu = float(self.lim.get("max_gpu_memory_gb") or 0)
+ if cap and scope_vals["memory_gb"] > cap and not any(
+ m.get("memory_cap_gb") for m in self._last_measures if m["kind"] == "scope"):
+ # no kernel memory cap holds here (e.g. rootless podman without cgroups): stand in for it
+ self._emergency(groups, "memory", cap, scope_vals["memory_gb"], key="mem")
+ if lim_gpu and scope_vals["gpu_memory_gb"] > EMERGENCY_GPU_FACTOR * lim_gpu:
+ self._emergency(groups, "gpu_memory", lim_gpu, scope_vals["gpu_memory_gb"], key="gpu")
+
+ def _emergency(self, groups, why, limit, observed, key):
+ best, bv = None, 0 # only a process that holds some: never one with none of it
+ for g in groups.values():
+ if g["kind"] != "scope":
+ continue
+ keep = self.spared(g["pids"])
+ for p in g["pids"]:
+ if p in keep:
+ continue
+ v = self._gpu_per.get(p, 0) if key == "gpu" else self.rss_kb(p)
+ if v > bv:
+ best, bv = p, v
+ if best:
+ try:
+ os.kill(best, signal.SIGKILL)
+ self.killed += 1
+ e = {"why": why, "limit": limit, "observed": round(observed, 3), "processes": 1}
+ event("emergency_kill", **e)
+ self.emergency.append(dict(e, t=round(now(), 2)))
+ except OSError:
+ pass
+
+ def rss_kb(self, p):
+ v = kb(status(p).get("VmRSS"))
+ if not v and p in self.via: # the first thread has exited: its status has no memory lines
+ for line in (read(self.via[p] + "/status") or "").splitlines():
+ if line.startswith("VmRSS:"):
+ v = kb(line.partition(":")[2])
+ return v
+
+ # ---- main loop
+ def run(self):
+ os.makedirs(REC, exist_ok=True)
+ restarted = os.path.exists(os.path.join(REC, "meta.json"))
+ meta = {"format": rcb_limits.FORMAT, "started": round(self.t0, 2), "enforce": self.enforce, "limits": self.lim,
+ "capabilities": self.caps, "beacons": {k: bool(v) for k, v in self.beacon.items()}}
+ if self.switch_error:
+ meta["switch_error"] = self.switch_error
+ if restarted:
+ self.restore()
+ event("restart", capabilities=self.caps)
+ else:
+ write_json(os.path.join(REC, "meta.json"), meta)
+ event("start", enforce=self.enforce, capabilities=self.caps)
+ if self.switch_error:
+ event("fatal", why=self.switch_error + " (nothing is enforced; the trial is not a result)")
+ if not self.beacon["scope"]:
+ event("fatal", why="no scope beacon mounted at /run/rcb-limits")
+ if self.lim.get("gpu") and not self.nvml:
+ event("fatal", why="GPU task but NVML is unavailable: " + str(self.caps.get("nvml_error")))
+ sys.setswitchinterval(0.001) # the OOM watch must get its turn within a millisecond
+ self.oomw.start()
+ while not os.path.exists(os.path.join(REC, "seal.request")):
+ t = now()
+ try:
+ self.tick_once(t)
+ except Exception as e: # a vanished process mid-read must never stop the record
+ import traceback
+ traceback.print_exc()
+ event("error", why=f"{type(e).__name__}: {e}"[:300])
+ for p in list(self.prev_cpu):
+ if t - self.prev_cpu[p][2] > 60:
+ del self.prev_cpu[p]
+ time.sleep(max(0.2, self.tick - (now() - t)))
+ self.seal()
+ while True: # stay up: Harbor copies the record after the hook
+ time.sleep(3600)
+
+ def tick_once(self, t):
+ groups = self.groups()
+ gpu_per, cards, gpu_unknown = self.nvml.sample() if self.nvml else ({}, {}, set())
+ self._gpu_per = gpu_per
+ ms = [self.measure(g, gpu_per, t, gpu_unknown) for g in groups.values()]
+ self._last_measures = ms
+ if self.oomw.thread is None:
+ self.oomw.scan() # no thread (tests): read the OOM counters with the sample
+ scope = [m for m in ms if m["kind"] == "scope"]
+ if scope and not self.scope_seen:
+ self.scope_seen = True
+ event("scope_found", containers=len(scope), roles=sorted(m["role"] for m in scope))
+ if not self.scope_seen and not self.fatal_scope and t - self.t0 > 120:
+ self.fatal_scope = True
+ event("fatal", why="no container mounting the trial's beacon found within 120 s")
+ disk_now = self.disk_step(t, groups)
+ vals = {"gpu_memory_gb": round(sum(m["gpu_memory_gb"] for m in scope), 3),
+ "memory_gb": round(sum(m["memory_gb"] for m in scope), 3),
+ "processes": sum(m["processes"] for m in scope), "disk_gb": disk_now}
+ unknown = sum(m.get("gpu_unmeasured", 0) for m in scope)
+ if unknown:
+ vals["gpu_unmeasured"] = unknown
+ if not self.gpu_unknown:
+ self.gpu_unknown = True
+ event("gpu_unmeasured", processes=unknown,
+ why="NVML does not tell the GPU memory of processes in scope (MIG / vGPU?): counted as 0")
+ obs = [m for m in ms if m["kind"] == "observe"]
+ append("samples.jsonl", {"t": round(t, 2), "scope": vals,
+ "observe": {"gpu_memory_gb": round(sum(m["gpu_memory_gb"] for m in obs), 3),
+ "memory_gb": round(sum(m["memory_gb"] for m in obs), 3),
+ "processes": sum(m["processes"] for m in obs)},
+ "cards": {str(k): v for k, v in cards.items()}, "containers": ms})
+ self.check(t, vals, groups)
+ try:
+ write_json(SCOPE_BEACON + "/usage.json", {
+ "t": round(t, 2), "gpu_memory_gb": vals["gpu_memory_gb"], "memory_gb": vals["memory_gb"],
+ "processes": vals["processes"], "disk_gb": self.disk_gb, "enforced": self.enforce,
+ "limits": {k: self.lim.get(k) for k in ("max_gpu_memory_gb", "max_memory_gb", "max_processes",
+ "max_disk_gb", "sustain_sec", "window_sec")},
+ "violations": sorted(self.confirmed),
+ # processes the monitor killed at once to protect the host (not violations by themselves)
+ "emergency_kills": len(self.emergency), "last_emergency_kill": (self.emergency or [None])[-1]})
+ except OSError:
+ pass
+
+ def seal(self):
+ try:
+ self.oomw.scan() # the counters once more, whatever the watch thread is doing
+ except Exception:
+ pass
+ # The record ends here, for OOM kills as for the samples: what happens to a sidecar in the collect hooks that
+ # follow (its replays) is the grader's to judge, from the hook's own evidence.
+ self.oomw.stopped = True
+ summary, samples, events = rcb_limits.load_record(REC)
+ viol = rcb_limits.violations(samples, self.lim, events)
+ meta = rcb_limits.read_json(os.path.join(REC, "meta.json")) or {}
+ out = {
+ "format": rcb_limits.FORMAT, "sealed": True, "enforce": meta.get("enforce", self.enforce),
+ "started": meta.get("started"), "ended": round(now(), 2), "n_samples": len(samples),
+ "scope_seen": self.scope_seen or any((s.get("scope") or {}).get("processes") for s in samples),
+ "peaks": rcb_limits.peaks(samples), "violations": viol,
+ "killed": sum(int(e.get("processes") or 0) for e in events if e.get("event") in ("kill", "emergency_kill")),
+ "emergency_kills": sum(1 for e in events if e.get("event") == "emergency_kill"),
+ "disk_baseline_gb": self.disk_base,
+ "disk_partial_walks": sum(1 for e in events if e.get("event") == "disk_partial"),
+ "restarts": sum(1 for e in events if e.get("event") == "restart"),
+ "capabilities": meta.get("capabilities", self.caps), "limits": self.lim}
+ if meta.get("switch_error") or self.switch_error:
+ out["switch_error"] = meta.get("switch_error") or self.switch_error
+ write_json(os.path.join(REC, "summary.json"), out)
+ event("seal", violations=len(viol))
+
+
+def seal_cmd(wait):
+ open(os.path.join(REC, "seal.request"), "a").close()
+ path = os.path.join(REC, "summary.json")
+ t_end = now() + wait
+ while now() < t_end and not os.path.exists(path):
+ time.sleep(0.5)
+ s = rcb_limits.read_json(path)
+ print(json.dumps({k: (s or {}).get(k) for k in ("sealed", "enforce", "n_samples", "peaks", "violations", "killed")}))
+ return 0
+
+
+def main(argv):
+ if argv[:1] == ["seal"]:
+ wait = float(argv[argv.index("--wait") + 1]) if "--wait" in argv else 90.0
+ return seal_cmd(wait)
+ if argv[:1] in (["run"], []):
+ no_dump()
+ Monitor().run()
+ return 0
+ print(__doc__)
+ return 2
+
+
+if __name__ == "__main__":
+ sys.exit(main(sys.argv[1:]))
diff --git a/tasks/myosuite-table-tennis-i00-standard-realtime/environment/rcb_monitor/rcb_limits.py b/tasks/myosuite-table-tennis-i00-standard-realtime/environment/rcb_monitor/rcb_limits.py
new file mode 100644
index 0000000000..a979bcdaff
--- /dev/null
+++ b/tasks/myosuite-table-tennis-i00-standard-realtime/environment/rcb_monitor/rcb_limits.py
@@ -0,0 +1,457 @@
+#!/usr/bin/env python3
+"""Resource limits (docs/PROTOCOL.md §2.10): the sustained-violation rule, shared by the monitor sidecar (monitor.py) and
+the verifier step, and the verifier step itself.
+
+Copies of this file live in every task (environment/rcb_monitor/ for the monitor, tests/ for the verifier), byte-identical
+to scripts/rcb_monitor/rcb_limits.py (ci_checks/check-monitor.py). Python 3.8+, standard library only: it runs inside
+benchmark verifier images.
+
+ python3 /tests/rcb_limits.py verify # test.sh runs it on exit (trap), after the grader
+ python3 /tests/rcb_limits.py sanitize /app/output # test.sh runs it before the grader (see sanitize())
+
+The verifier step reads the monitor's sealed record (/var/rcb-monitor/, collected from the rcb_monitor sidecar, a
+channel the agent cannot write), recomputes every violation from the raw samples with its own copy of the limits
+(/tests/rcb_limits.json), and adds to /logs/verifier/reward.json:
+ always: resource_enforced, limits_violation_observed, resource_monitor_missing, peak_gpu_memory_gb, peak_memory_gb,
+ peak_processes, peak_disk_gb, resource_emergency_kills; over_<metric>=1 for every violation; the kernel's
+ OOM kills in the record: resource_oom_kills (in the scope, at the container's own memory cap: the
+ violation), resource_host_oom_kills (in the scope, not at its cap: the host's OOM killer or an enclosing
+ cgroup's), resource_service_oom_kills (in a benchmark-owned service that is only recorded: the simulator
+ sidecar of standard modes, RoboPaint's scorer)
+ enforced (the record's mode, RCB_LIMITS_ENFORCE=1) and violated: resource_limit_exceeded=1, success=0 (reward.txt 0),
+ success_before_limits = the grader's success
+ enforced and no usable record: infra_error=1 (not a result). Not enforced (the default): a missing or unusable
+ record is only reported (resource_monitor_missing=1); the grader's verdict stands.
+ enforced, no violation, and a process in the scope was killed by an OOM killer that is not its own cap's:
+ infra_error=1, resource_host_oom=1 (not a result: another trial's or the host's doing, never the agent's)
+ RCB_LIMITS_ENFORCE neither 0/1/true/false: infra_error=1, resource_switch_error=1 (nobody knows what was asked for)
+ in both modes: the kernel OOM-killed a benchmark-owned service and the grader's verdict is not a success:
+ infra_error=1, resource_service_oom=1 (PROTOCOL §2.10: such a service is recorded, never a failure; an
+ episode that ended without success after its simulator was killed for memory is not the agent's 0). A
+ success stands: the service recovered and the grader accepted the episode.
+Wherever this step says "not a result" (infra_error=1) it also sets success=0 and reward.txt 0 and keeps the grader's
+verdict in success_before_limits, so that no reader sees a success next to an infrastructure error.
+With the default (not enforced, a readable switch) the step changes nothing of the grader's verdict (reward.txt,
+success, reward) and sets infra_error only in the one case above (a benchmark-owned service OOM-killed, no success).
+A violation is a limit exceeded for sustain_sec within window_sec (disk: two measurements in a row), or a kernel OOM kill
+in the scope at the container's own memory cap (the monitor's `oom` event with cause "cap"). The monitor's emergency
+kills (host protection) are counted, never violations by themselves.
+The reasons, in words, go to /logs/verifier/resource_limits.json. If this step itself fails, it says so in reward.json
+(resource_step_error=1, resource_monitor_missing=1; infra_error=1 when the switch asks for enforcement).
+"""
+import json
+import os
+import stat
+import sys
+
+FORMAT = "rcb-monitor/1"
+RECORD = os.environ.get("RCB_MONITOR_RECORD", "/var/rcb-monitor")
+METRICS = ("gpu_memory_gb", "memory_gb", "processes", "disk_gb")
+LIMIT_KEY = {"gpu_memory_gb": "max_gpu_memory_gb", "memory_gb": "max_memory_gb", "processes": "max_processes",
+ "disk_gb": "max_disk_gb"}
+MAX_GAP_FRACTION = 0.10 # more of the agent phase than this unobserved: the record is not usable
+
+
+def parse_switch(value):
+ """RCB_LIMITS_ENFORCE -> (1 or 0, None), or (0, why) for a value that is neither on nor off (never a silent off)."""
+ s = str("" if value is None else value).strip().lower()
+ if s in ("", "0", "false"):
+ return 0, None
+ if s in ("1", "true"):
+ return 1, None
+ return 0, f"RCB_LIMITS_ENFORCE={value!r} is not one of 0, 1, true, false"
+
+
+def sustained(series, limit, sustain_sec, window_sec, tick_sec):
+ """First time a sustained violation is confirmed, or None.
+
+ series: [(t, value)] in time order (value None = not measured: never over). Each sample stands for the time to the
+ next one (the last for none), at most 3 ticks (a gap in the record never counts against the agent). A violation is confirmed at the
+ first moment when the time over `limit` within the last `window_sec` reaches `sustain_sec` (so going under for a
+ moment does not reset it)."""
+ if limit is None or limit <= 0:
+ return None
+ iv = [] # (start, end) of the time over the limit
+ for i, (t, v) in enumerate(series):
+ if i + 1 == len(series):
+ break # the last sample has not lasted yet
+ d = max(0.0, min(series[i + 1][0] - t, 3 * tick_sec))
+ if v is not None and v > limit and d > 0:
+ if iv and abs(iv[-1][1] - t) < 1e-9:
+ iv[-1] = (iv[-1][0], t + d)
+ else:
+ iv.append((t, t + d))
+ if not iv:
+ return None
+ # the over-time in a window (e - W, e] only grows while e moves through an interval: check each interval's end,
+ # and inside the first one that reaches it, bisect for the exact moment
+ def over_in(e):
+ """Over-time inside (e - window, e]."""
+ s = e - window_sec
+ tot = 0.0
+ for a, b in iv:
+ if b <= s:
+ continue
+ if a >= e:
+ break
+ tot += min(b, e) - max(a, s)
+ return tot
+
+ for a, b in iv:
+ if over_in(b) + 1e-9 >= sustain_sec:
+ # bisect for the first moment in [a, b] at which it is reached
+ x, y = a, b
+ for _ in range(40):
+ m = (x + y) / 2
+ if over_in(m) + 1e-9 >= sustain_sec:
+ y = m
+ else:
+ x = m
+ return y
+ return None
+
+
+def consecutive(series, limit, n=2):
+ """First time `n` consecutive measured samples are over `limit` (disk: measured once a minute), or None."""
+ if limit is None or limit <= 0:
+ return None
+ run = 0
+ for t, v in series:
+ if v is None:
+ continue
+ run = run + 1 if v > limit else 0
+ if run >= n:
+ return t
+ return None
+
+
+def violations(samples, limits, events=()):
+ """Every violation in a record: [{metric, limit, at, peak}] (empty: none). samples: the record's sample dicts.
+ A limit exceeded for sustain_sec within window_sec (disk: two measurements in a row), or a kernel OOM kill in the
+ scope at the container's own memory cap (cause "cap": the cgroup's own `oom` counter rose with the kill). A kill
+ by the host's OOM killer (any other cause) is not the trial's doing and never a violation: oom_kills() counts it.
+ The monitor's own `emergency_kill` events protect the host and are not violations (PROTOCOL §2.10)."""
+ tick = float(limits.get("tick_sec", 5))
+ sustain = float(limits.get("sustain_sec", 60))
+ window = float(limits.get("window_sec", 2 * sustain))
+ out = []
+ for m in METRICS:
+ lim = limits.get(LIMIT_KEY[m])
+ if lim is None:
+ continue
+ lim = float(lim)
+ ser = [(s["t"], (s.get("scope") or {}).get(m)) for s in samples]
+ at = consecutive(ser, lim) if m == "disk_gb" else sustained(ser, lim, sustain, window, tick)
+ if at is not None:
+ peak = max((v for _, v in ser if v is not None), default=None)
+ out.append({"metric": m, "limit": lim, "at": round(at, 2), "peak": peak})
+ for e in events:
+ if isinstance(e, dict) and e.get("event") == "oom" and e.get("scope") and e.get("cause") == "cap":
+ out.append({"metric": "oom", "limit": None, "at": e.get("t"), "peak": e.get("count"),
+ "service": e.get("role")})
+ return sorted(out, key=lambda v: v["at"] if isinstance(v["at"], (int, float)) else 0)
+
+
+def oom_kills(events):
+ """The kernel's OOM kills in a record: {"cap": in the scope at the container's own memory cap (the violation),
+ "host": in the scope by another OOM killer (the host's, an enclosing cgroup's; also an event that does not say),
+ "service": in a benchmark-owned service that is only recorded (whatever the cause)}."""
+ out = {"cap": 0, "host": 0, "service": 0}
+ for e in events:
+ if not isinstance(e, dict) or e.get("event") != "oom":
+ continue
+ n = e.get("count")
+ n = int(n) if isinstance(n, (int, float)) and not isinstance(n, bool) and n == n and 1 <= n < 1e9 else 1
+ out["service" if not e.get("scope") else "cap" if e.get("cause") == "cap" else "host"] += n
+ return out
+
+
+def peaks(samples):
+ p = {}
+ for m in METRICS:
+ vals = [(s.get("scope") or {}).get(m) for s in samples]
+ vals = [v for v in vals if v is not None]
+ p[m] = round(max(vals), 3) if vals else 0.0
+ return p
+
+
+def read_jsonl(path):
+ out = []
+ try:
+ with open(path) as f:
+ for line in f:
+ line = line.strip()
+ if line:
+ try:
+ out.append(json.loads(line))
+ except ValueError:
+ pass # a torn last line (the monitor was killed mid-write)
+ except OSError:
+ pass
+ return out
+
+
+def read_json(path):
+ try:
+ with open(path) as f:
+ return json.load(f)
+ except (OSError, ValueError):
+ return None
+
+
+def load_record(record=None):
+ """(summary or None, samples, events)."""
+ record = record or os.environ.get("RCB_MONITOR_RECORD", RECORD)
+ return (read_json(os.path.join(record, "summary.json")), read_jsonl(os.path.join(record, "samples.jsonl")),
+ [e for e in read_jsonl(os.path.join(record, "events.jsonl")) if isinstance(e, dict)])
+
+
+def record_problems(summary, samples, events, limits):
+ """Why a record cannot be used ([] = usable)."""
+ why = []
+ if not summary:
+ why.append("no sealed record (/var/rcb-monitor/summary.json missing: the monitor did not run, crashed, or was not collected)")
+ elif summary.get("format") != FORMAT:
+ why.append(f"record format {summary.get('format')!r}, expected {FORMAT}")
+ if not samples:
+ why.append("no samples")
+ fatal = [e for e in events if e.get("event") == "fatal"]
+ if fatal:
+ why.append("monitor fatal: " + "; ".join(str(e.get("why")) for e in fatal)[:300])
+ if samples and summary and not summary.get("scope_seen"):
+ why.append("the monitor never found the containers it watches (no beacon seen)")
+ if float(limits.get("max_gpu_memory_gb") or 0) > 0 and samples:
+ blind = sum(1 for s in samples if (s.get("scope") or {}).get("gpu_unmeasured"))
+ if blind / len(samples) > MAX_GAP_FRACTION:
+ why.append(f"NVML did not tell the GPU memory of processes in scope in {blind} of {len(samples)} samples "
+ "(MIG / vGPU?): the GPU-memory limit was not measured")
+ if len(samples) >= 2:
+ tick = float(limits.get("tick_sec", 5))
+ span = samples[-1]["t"] - samples[0]["t"]
+ gaps = sum(max(0.0, b["t"] - a["t"] - 3 * tick) for a, b in zip(samples, samples[1:]))
+ if span > 0 and gaps / span > MAX_GAP_FRACTION:
+ why.append(f"{gaps:.0f} s of {span:.0f} s unobserved")
+ return why
+
+
+def switch_of(summary):
+ """(enforce, why the switch cannot be read or None): the record's mode when there is a record, else this step's
+ own environment (the same run-time variable, task.toml [verifier.env])."""
+ if summary and summary.get("switch_error"):
+ return 0, str(summary["switch_error"])
+ if summary and "enforce" in summary:
+ return parse_switch(summary.get("enforce"))
+ return parse_switch(os.environ.get("RCB_LIMITS_ENFORCE"))
+
+
+def not_a_result(reward, add):
+ """This step found that the trial is not a result (infra_error=1): no success is left beside it. The grader's
+ verdict stays readable in success_before_limits."""
+ add["infra_error"] = 1
+ add["success_before_limits"] = int(bool(reward.get("success", 0)))
+ add["success"] = 0
+ if "reward" in reward:
+ add["reward"] = 0
+
+
+def verify(argv=()):
+ log = os.environ.get("RCB_VERIFIER_LOGS", "/logs/verifier")
+ tests = os.path.dirname(os.path.abspath(__file__))
+ limits_file = read_json(os.path.join(tests, "rcb_limits.json"))
+ limits = limits_file or {}
+ summary, samples, events = load_record()
+ enforce, switch_error = switch_of(summary)
+ rj_path, rt_path = os.path.join(log, "reward.json"), os.path.join(log, "reward.txt")
+ reward = read_json(rj_path)
+ have_grade = isinstance(reward, dict)
+ reward = reward if have_grade else {}
+ problems = record_problems(summary, samples, events, limits)
+ if limits_file is None:
+ problems.append("the verifier's copy of the limits (tests/rcb_limits.json) is missing or unreadable")
+ if switch_error and not any(switch_error in p for p in problems):
+ problems.append(switch_error)
+ viol = violations(samples, limits, events) if samples else []
+ pk = peaks(samples)
+ emergency = [e for e in events if e.get("event") == "emergency_kill"]
+ oom = oom_kills(events)
+ detail = {"format": FORMAT, "enforced": enforce, "limits": {k: v for k, v in limits.items()},
+ "peaks": pk, "violations": viol, "record_problems": problems, "emergency_kills": emergency,
+ "oom_kills": [e for e in events if e.get("event") == "oom"],
+ "monitor": {k: (summary or {}).get(k) for k in ("started", "ended", "n_samples", "capabilities", "killed",
+ "restarts", "disk_baseline_gb", "disk_partial_walks")}}
+ add = {"resource_enforced": enforce, "limits_violation_observed": int(bool(viol)),
+ "resource_monitor_missing": int(bool(problems)),
+ "peak_gpu_memory_gb": pk["gpu_memory_gb"], "peak_memory_gb": pk["memory_gb"],
+ "peak_processes": pk["processes"], "peak_disk_gb": pk["disk_gb"],
+ "resource_emergency_kills": len(emergency), "resource_oom_kills": oom["cap"],
+ "resource_host_oom_kills": oom["host"], "resource_service_oom_kills": oom["service"]}
+ for v in viol:
+ add["over_" + v["metric"]] = 1
+ failed = void = False # void: this step says "not a result" (infra_error=1) and takes the success away
+ if switch_error:
+ # nobody knows whether enforcement was asked for: not a result in either mode
+ add["resource_switch_error"] = 1
+ void = True
+ detail["verdict"] = "not a result: " + switch_error
+ elif enforce and viol:
+ # a confirmed violation fails the trial whatever the deliverable (and wins over a grader's infra flag: the
+ # agent's processes were stopped for it)
+ failed = True
+ add["success_before_limits"] = int(reward.get("success", 0) or 0)
+ add["resource_limit_exceeded"] = 1
+ add["success"] = 0
+ if "reward" in reward:
+ add["reward"] = 0
+ if reward.get("infra_error"):
+ add["grader_infra_error"] = 1
+ add["infra_error"] = 0
+ detail["verdict"] = "failed: resource limit exceeded (" + ", ".join(v["metric"] for v in viol) + ")"
+ elif enforce and problems:
+ void = True
+ detail["verdict"] = "not a result: the resource monitor's record is not usable"
+ elif enforce and oom["host"]:
+ # a process the limits bind was killed for memory, but not at its container's own cap: the host ran out of
+ # memory (other trials), or an enclosing cgroup did. Never the agent's failure, and not a clean run either.
+ add["resource_host_oom"] = 1
+ void = True
+ detail["verdict"] = ("not a result: an out-of-memory killer other than the container's own cap (the host's) "
+ f"killed {oom['host']} process(es) of the trial")
+ elif oom["service"] and have_grade and not reward.get("success"):
+ # in both modes: the kernel killed a benchmark-owned service (the simulator sidecar, a scorer) for memory and
+ # the episode did not end in a success. Such a service is recorded, never a failure (PROTOCOL §2.10): whatever
+ # the grader made of the crash, this is not the agent's 0.
+ add["resource_service_oom"] = 1
+ add["infra_error"] = 1
+ detail["verdict"] = (f"not a result: the kernel OOM-killed a benchmark-owned service ({oom['service']} kill(s)) "
+ "and the episode did not end in a success")
+ else:
+ # not enforced: observe only. A missing or unusable record is reported (resource_monitor_missing), never infra.
+ detail["verdict"] = ("observed only (RCB_LIMITS_ENFORCE=0): " + ", ".join(v["metric"] for v in viol)) if viol \
+ else "within limits" if not problems else "observed only (RCB_LIMITS_ENFORCE=0): no usable record"
+ if void:
+ not_a_result(reward, add)
+ os.makedirs(log, exist_ok=True)
+ if have_grade or failed:
+ reward.update(add)
+ reward.setdefault("success", 0)
+ tmp = rj_path + ".tmp"
+ with open(tmp, "w") as f:
+ json.dump(reward, f, indent=1)
+ os.replace(tmp, rj_path)
+ if failed or void:
+ with open(rt_path, "w") as f:
+ f.write("0\n")
+ with open(os.path.join(log, "resource_limits.json"), "w") as f:
+ json.dump(detail, f, indent=1)
+ print("resource limits: " + detail["verdict"], file=sys.stderr)
+ return 0
+
+
+def step_failed(err):
+ """This step raised: say so in the grader's reward.json instead of leaving the trial without its resource fields
+ (and, when the switch asks for enforcement or cannot be read, as an infrastructure error: nothing was judged)."""
+ log = os.environ.get("RCB_VERIFIER_LOGS", "/logs/verifier")
+ rj_path = os.path.join(log, "reward.json")
+ reward = read_json(rj_path)
+ enforce, switch_error = parse_switch(os.environ.get("RCB_LIMITS_ENFORCE"))
+ add = {"resource_step_error": 1, "resource_monitor_missing": 1, "resource_enforced": enforce}
+ if isinstance(reward, dict):
+ if enforce or switch_error:
+ not_a_result(reward, add)
+ reward.update(add)
+ tmp = rj_path + ".tmp"
+ with open(tmp, "w") as f:
+ json.dump(reward, f, indent=1)
+ os.replace(tmp, rj_path)
+ if add.get("infra_error"):
+ with open(os.path.join(log, "reward.txt"), "w") as f:
+ f.write("0\n")
+ with open(os.path.join(log, "resource_limits.json"), "w") as f:
+ json.dump({"format": FORMAT, "enforced": enforce, "verdict": "the verifier's resource step failed: " + err,
+ "record_problems": [err]}, f, indent=1)
+
+
+MAX_SESSION_LOG = 64 * 2 ** 20 # bytes of /app/output/rcb_session_log.jsonl a grader is given to read
+
+
+def sanitize(roots=("/app/output",), max_log=MAX_SESSION_LOG):
+ """Before the grader: the files the agent left (PROTOCOL §2.6: collected from its container as they are, links and
+ all) are untrusted input, and a grader opens some of them by name. Remove what is not a file of the agent's own
+ under the root: a symbolic link that leaves it or does not end at a regular file (`rcb_session_log.jsonl ->
+ /dev/zero` made a grader read until the verifier ran out of memory or time: no verdict), and every FIFO, device and
+ socket (opening a FIFO blocks). A session log larger than `max_log` is cut there (graders read it line by line,
+ and one endless line is as bad). Regular files, directories and links to regular files inside the root stay as
+ they are: nothing a grader could have read safely changes. Returns what it removed or cut (also printed)."""
+ done = []
+ for root in roots:
+ real_root = os.path.realpath(root)
+ if not os.path.isdir(real_root) or os.path.islink(root):
+ if os.path.islink(root): # the root itself must be a directory of its own
+ try:
+ os.unlink(root)
+ done.append(f"removed {root}: a symbolic link")
+ except OSError:
+ pass
+ continue
+ for dirpath, dirnames, filenames in os.walk(real_root, followlinks=False):
+ for name in list(dirnames) + filenames:
+ path = os.path.join(dirpath, name)
+ try:
+ st = os.lstat(path)
+ if stat.S_ISLNK(st.st_mode):
+ target = os.path.realpath(path)
+ inside = target.startswith(real_root + os.sep)
+ if inside and stat.S_ISREG(os.stat(target).st_mode):
+ continue
+ why = "a symbolic link that leaves the directory or does not end at a regular file"
+ elif stat.S_ISREG(st.st_mode) or stat.S_ISDIR(st.st_mode):
+ continue
+ else:
+ why = "not a regular file (FIFO, device or socket)"
+ os.unlink(path)
+ if name in dirnames:
+ dirnames.remove(name)
+ done.append(f"removed {path}: {why}")
+ except OSError as e: # a dangling link, a file that is gone: remove what can be removed
+ try:
+ if stat.S_ISLNK(os.lstat(path).st_mode):
+ os.unlink(path)
+ done.append(f"removed {path}: a symbolic link that cannot be resolved ({type(e).__name__})")
+ except OSError:
+ pass
+ log = os.path.join(real_root, "rcb_session_log.jsonl")
+ try:
+ if os.path.isfile(log) and not os.path.islink(log) and os.path.getsize(log) > max_log:
+ os.truncate(log, max_log)
+ done.append(f"cut {log} to {max_log} bytes")
+ except OSError:
+ pass
+ for d in done:
+ print("verifier input: " + d, file=sys.stderr)
+ return done
+
+
+def run_step(argv=()):
+ """verify(), and whatever it raises reported instead of breaking the grader's own verdict. Always 0."""
+ try:
+ verify(argv)
+ except Exception as e:
+ msg = f"{type(e).__name__}: {e}"
+ print(f"resource limits: verifier step failed: {msg}", file=sys.stderr)
+ try:
+ step_failed(msg)
+ except Exception:
+ pass
+ return 0
+
+
+if __name__ == "__main__":
+ if sys.argv[1:2] == ["verify"]:
+ sys.exit(run_step(sys.argv[2:]))
+ if sys.argv[1:2] == ["sanitize"]:
+ try:
+ sanitize(sys.argv[2:] or ("/app/output",))
+ except Exception as e: # never keep the grader from running
+ print(f"verifier input: sanitize failed: {type(e).__name__}: {e}", file=sys.stderr)
+ sys.exit(0)
+ print(__doc__)
diff --git a/tasks/myosuite-table-tennis-i00-standard-realtime/environment/rcb_monitor/selftest.py b/tasks/myosuite-table-tennis-i00-standard-realtime/environment/rcb_monitor/selftest.py
new file mode 100644
index 0000000000..89d553743f
--- /dev/null
+++ b/tasks/myosuite-table-tennis-i00-standard-realtime/environment/rcb_monitor/selftest.py
@@ -0,0 +1,137 @@
+#!/usr/bin/env python3
+"""What the resource monitor can measure where it runs (capabilities(), written into its record), and a self-test of
+the shared violation rule and the verifier step (no containers needed; ci_checks/check-monitor.py runs it):
+
+ python3 scripts/rcb_monitor/selftest.py
+"""
+import json
+import os
+import sys
+import tempfile
+
+HERE = os.path.dirname(os.path.abspath(__file__))
+
+
+def _rl(p):
+ try:
+ return os.readlink(p)
+ except OSError:
+ return None
+
+
+def capabilities():
+ """Facts the verifier and the results footnotes need: does the monitor see the host's processes (pid: host), the
+ host's cgroup tree (cgroup: host; per-container memory and OOM counts), """
+ c = {}
+ try:
+ pids = [d for d in os.listdir("/proc") if d.isdigit()]
+ except OSError:
+ pids = []
+ try:
+ c["pid_host"] = int(len(pids) > 1 and open("/proc/1/comm").read().strip() != open("/proc/self/comm").read().strip())
+ except OSError:
+ c["pid_host"] = 0
+ own = ""
+ try:
+ own = [ln[3:].strip() for ln in open("/proc/self/cgroup") if ln.startswith("0::")][0]
+ except (OSError, IndexError):
+ pass
+ c["cgroup_v2"] = int(os.path.exists("/sys/fs/cgroup/cgroup.controllers"))
+ c["cgroup_host"] = int(bool(own) and own != "/" and os.path.isdir("/sys/fs/cgroup" + own))
+ return c
+
+
+# --------------------------------------------------------------------------------------------------- self-test
+def _check(cond, msg):
+ if not cond:
+ raise AssertionError(msg)
+
+
+def test_rule():
+ sys.path.insert(0, HERE)
+ import rcb_limits as r
+ tick = 5.0
+ over = [(i * tick, 10.0 if 20 <= i < 33 else 1.0) for i in range(60)] # 65 s over the limit
+ _check(abs(r.sustained(over, 5, 60, 120, tick) - 160.0) < 0.01, "65 s over: confirmed at 60 s")
+ short = [(i * tick, 10.0 if 20 <= i < 31 else 1.0) for i in range(60)] # 55 s
+ _check(r.sustained(short, 5, 60, 120, tick) is None, "55 s over: no violation")
+ flap = [(i * tick, 10.0 if (i % 2 == 0 and i >= 20) else 1.0) for i in range(80)] # half the time, never 60 s straight
+ _check(r.sustained(flap, 5, 60, 120, tick) is not None, "59 s on / 1 s off does not escape: 60 s in any 120 s")
+ sparse = [(i * 25.0, 10.0) for i in range(4)] # 25 s gaps count 15 s each
+ _check(r.sustained(sparse, 5, 60, 120, tick) is None, "gaps never count against the agent")
+ _check(r.sustained(over, 0, 60, 120, tick) is None, "limit 0 = not limited")
+ disk = [(0, 1.0), (60, 30.0), (120, 1.0), (180, 30.0), (240, 31.0)]
+ _check(r.consecutive(disk, 20) == 240, "disk: two consecutive measurements over")
+
+
+def test_verify():
+ sys.path.insert(0, HERE)
+ import rcb_limits as r
+ lim = {"max_gpu_memory_gb": 4.0, "max_memory_gb": 8.0, "max_processes": 100, "max_disk_gb": 10.0,
+ "sustain_sec": 60, "window_sec": 120, "tick_sec": 5}
+ for enforce, gpu, want in ((1, 6.0, "fail"), (0, 6.0, "observe"), (1, 3.0, "pass")):
+ with tempfile.TemporaryDirectory() as d:
+ rec, tests, logs = (os.path.join(d, x) for x in ("rec", "tests", "logs"))
+ for x in (rec, tests, logs):
+ os.makedirs(x)
+ json.dump(lim, open(os.path.join(tests, "rcb_limits.json"), "w"))
+ with open(os.path.join(rec, "samples.jsonl"), "w") as f:
+ for i in range(40):
+ f.write(json.dumps({"t": 1000 + 5 * i, "scope": {"gpu_memory_gb": gpu if i > 5 else 0.5,
+ "memory_gb": 2.0, "processes": 10,
+ "disk_gb": 0.1 if i % 12 == 0 else None}}) + "\n")
+ open(os.path.join(rec, "events.jsonl"), "w").close()
+ json.dump({"format": r.FORMAT, "sealed": True, "enforce": enforce, "scope_seen": True},
+ open(os.path.join(rec, "summary.json"), "w"))
+ json.dump({"success": 1, "infra_error": 0}, open(os.path.join(logs, "reward.json"), "w"))
+ open(os.path.join(logs, "reward.txt"), "w").write("1\n")
+ import shutil
+ shutil.copy(os.path.join(HERE, "rcb_limits.py"), tests)
+ env = dict(os.environ, RCB_MONITOR_RECORD=rec, RCB_VERIFIER_LOGS=logs)
+ import subprocess
+ subprocess.run([sys.executable, os.path.join(tests, "rcb_limits.py"), "verify"], env=env, check=True,
+ capture_output=True)
+ out = json.load(open(os.path.join(logs, "reward.json")))
+ txt = open(os.path.join(logs, "reward.txt")).read().strip()
+ if want == "fail":
+ _check(out["success"] == 0 and out["resource_limit_exceeded"] == 1 and out["success_before_limits"] == 1
+ and txt == "0" and out["over_gpu_memory_gb"] == 1, f"enforced violation fails: {out}")
+ elif want == "observe":
+ _check(out["success"] == 1 and out["limits_violation_observed"] == 1 and txt == "1"
+ and "resource_limit_exceeded" not in out, f"observe mode only records: {out}")
+ else:
+ _check(out["success"] == 1 and out["limits_violation_observed"] == 0 and out["resource_monitor_missing"] == 0,
+ f"within limits: {out}")
+ _check(out["peak_gpu_memory_gb"] == gpu, "peak recorded")
+ # no record at all: enforced = infra, observed = flagged only
+ for enforce in (1, 0):
+ with tempfile.TemporaryDirectory() as d:
+ logs, tests = os.path.join(d, "logs"), os.path.join(d, "tests")
+ os.makedirs(logs)
+ os.makedirs(tests)
+ json.dump(lim, open(os.path.join(tests, "rcb_limits.json"), "w"))
+ json.dump({"success": 1}, open(os.path.join(logs, "reward.json"), "w"))
+ import shutil
+ import subprocess
+ shutil.copy(os.path.join(HERE, "rcb_limits.py"), tests)
+ env = dict(os.environ, RCB_MONITOR_RECORD=os.path.join(d, "none"), RCB_VERIFIER_LOGS=logs,
+ RCB_LIMITS_ENFORCE=str(enforce))
+ subprocess.run([sys.executable, os.path.join(tests, "rcb_limits.py"), "verify"], env=env, check=True,
+ capture_output=True)
+ out = json.load(open(os.path.join(logs, "reward.json")))
+ _check(out["resource_monitor_missing"] == 1 and out.get("infra_error", 0) == enforce,
+ f"missing record (enforce={enforce}): {out}")
+
+
+def main():
+ if sys.argv[1:2] == ["caps"]:
+ print(json.dumps(capabilities()))
+ return 0
+ test_rule()
+ test_verify()
+ print("rcb_monitor selftest: ok")
+ return 0
+
+
+if __name__ == "__main__":
+ sys.exit(main())
diff --git a/tasks/myosuite-table-tennis-i00-standard-realtime/environment/sim/instance.npz b/tasks/myosuite-table-tennis-i00-standard-realtime/environment/sim/instance.npz
new file mode 100644
index 0000000000..f48e25b480
Binary files /dev/null and b/tasks/myosuite-table-tennis-i00-standard-realtime/environment/sim/instance.npz differ
diff --git a/tasks/myosuite-table-tennis-i00-standard-realtime/environment/sim/rcb1.py b/tasks/myosuite-table-tennis-i00-standard-realtime/environment/sim/rcb1.py
new file mode 100644
index 0000000000..6938c57e38
--- /dev/null
+++ b/tasks/myosuite-table-tennis-i00-standard-realtime/environment/sim/rcb1.py
@@ -0,0 +1,83 @@
+"""RCB1 — a tiny binary frame for one JSON header plus N raw blobs (videos, image arrays, depth, proprio).
+
+ frame = b"RCB1" | header_len uint32 LE | header JSON (UTF-8) | blob_0 | blob_1 | ...
+ header = {"v": 1, "meta": {...}, "parts": [{"name", "kind", "len", ...}, ...]} # blobs in the same order
+
+kinds: "mp4" (fps, n), "jpeg", "array" (dtype, shape; C-order raw bytes), "json" (utf-8), "bytes".
+Zero dependencies for bytes/json parts; numpy only when you ask for arrays. Used on both sides (server and client),
+over HTTP bodies (Content-Type: application/x-rcb1) and equally usable over a raw socket. `python -m rcb1 unpack f dir`.
+"""
+from __future__ import annotations
+import json, struct, sys, pathlib
+
+MAGIC = b"RCB1"
+
+
+def encode(meta: dict, parts: list[tuple[str, str, bytes, dict]]) -> bytes:
+ """parts: [(name, kind, blob, extra), ...]; extra carries dtype/shape/fps/n."""
+ hdr_parts = []
+ for name, kind, blob, extra in parts:
+ d = {"name": name, "kind": kind, "len": len(blob)}
+ d.update({k: v for k, v in (extra or {}).items() if v is not None})
+ hdr_parts.append(d)
+ header = json.dumps({"v": 1, "meta": meta, "parts": hdr_parts}, separators=(",", ":")).encode()
+ return b"".join([MAGIC, struct.pack("<I", len(header)), header] + [blob for _, _, blob, _ in parts])
+
+
+def array_part(name: str, arr, extra: dict | None = None) -> tuple[str, str, bytes, dict]:
+ import numpy as np
+ a = np.ascontiguousarray(arr)
+ return (name, "array", a.tobytes(), {"dtype": str(a.dtype), "shape": list(a.shape), **(extra or {})})
+
+
+def decode(buf: bytes, arrays: bool = True) -> tuple[dict, dict]:
+ """-> (meta, parts) where parts[name] = {"kind", ..., "data"}; arrays become numpy arrays when arrays=True."""
+ if buf[:4] != MAGIC:
+ raise ValueError("not an RCB1 frame")
+ n = struct.unpack("<I", buf[4:8])[0]
+ header = json.loads(buf[8:8 + n].decode()); off = 8 + n
+ out = {}
+ for p in header["parts"]:
+ blob = buf[off:off + p["len"]]; off += p["len"]
+ d = dict(p)
+ if p["kind"] == "array" and arrays:
+ import numpy as np
+ d["data"] = np.frombuffer(blob, dtype=np.dtype(p["dtype"])).reshape(p["shape"])
+ elif p["kind"] == "json":
+ d["data"] = json.loads(blob.decode())
+ else:
+ d["data"] = blob
+ out[p["name"]] = d
+ return header.get("meta", {}), out
+
+
+def unpack_to_dir(buf: bytes, out_dir: str) -> list[str]:
+ """Write a frame's parts as files: mp4/jpeg/bytes verbatim, arrays as .npy, meta as meta.json."""
+ import numpy as np
+ out = pathlib.Path(out_dir); out.mkdir(parents=True, exist_ok=True)
+ meta, parts = decode(buf, arrays=True); written = []
+ (out / "meta.json").write_text(json.dumps(meta, indent=1)); written.append("meta.json")
+ for name, p in parts.items():
+ safe = name.replace("/", "__")
+ if p["kind"] == "array":
+ f = out / f"{safe}.npy"; np.save(f, p["data"])
+ elif p["kind"] == "mp4":
+ f = out / f"{safe}.mp4"; f.write_bytes(p["data"])
+ elif p["kind"] == "jpeg":
+ f = out / f"{safe}.jpg"; f.write_bytes(p["data"])
+ elif p["kind"] == "json":
+ f = out / f"{safe}.json"; f.write_text(json.dumps(p["data"], indent=1))
+ else:
+ f = out / f"{safe}.bin"; f.write_bytes(p["data"])
+ written.append(f.name)
+ return written
+
+
+if __name__ == "__main__":
+ if len(sys.argv) == 4 and sys.argv[1] == "unpack":
+ print("\n".join(unpack_to_dir(open(sys.argv[2], "rb").read(), sys.argv[3])))
+ elif len(sys.argv) == 3 and sys.argv[1] == "info":
+ meta, parts = decode(open(sys.argv[2], "rb").read(), arrays=False)
+ print(json.dumps(meta, indent=1)); [print(n, {k: v for k, v in p.items() if k != "data"}) for n, p in parts.items()]
+ else:
+ print(__doc__)
diff --git a/tasks/myosuite-table-tennis-i00-standard-realtime/environment/sim/rcb_myosuite.py b/tasks/myosuite-table-tennis-i00-standard-realtime/environment/sim/rcb_myosuite.py
new file mode 100644
index 0000000000..930c0c73ba
--- /dev/null
+++ b/tasks/myosuite-table-tennis-i00-standard-realtime/environment/sim/rcb_myosuite.py
@@ -0,0 +1,722 @@
+"""
+rcb_myosuite.py — task-local tooling for MyoSuite (2.12.2, MuJoCo): the MyoHand (a musculoskeletal model of the human
+forearm and hand, 39 muscles) and the MyoArm (shoulder, arm and hand, 34 or 63 muscles). Frozen scene instance / privileged state / muscle-level actions / recording /
+fresh-process deterministic replay.
+
+Concepts
+--------
+instance : the task's env after `reset(seed=0)` (the `Fixed` / `P1` variants: fixed goal). Instance.npz stores the full
+ MuJoCo integration state (solver warm start included), every numeric model array (resets edit the model:
+ goal sites, object sizes) and the env's Python-side state; Session(instance=...) builds the env, resets it
+ and loads that snapshot. Every episode (reset, replay) starts from it. `state_sha` covers state + model.
+action : one row per control step = 39 real numbers, one per muscle. MyoSuite maps each to the muscle's excitation
+ u = 1 / (1 + exp(-5 (a - 0.5))) in (0, 1) and does not clip a: a = -1 gives u ~ 0.0006, a = 0.5 gives 0.5,
+ a = 1 gives 0.92, a = 2 gives 0.9994. Muscles only pull; activation follows excitation with a first-order
+ delay (~10 ms up, ~40 ms down); one muscle can cross several joints; force also depends on muscle length and
+ velocity. There is no joint-level command: turning excitations into motion is the agent's job.
+episode : ends in success once the task's own `solved` flag has held for HOLD_S (1 s; baoding 5 s) of consecutive control steps,
+ or in failure at the task's own `done` (object dropped, fingers far from the key, ...). MyoSuite's step
+ budget (50-200 steps) is lifted; MAX_STEPS only stops runaway trajectories. MuJoCo on the CPU is
+ deterministic: the same rows from the same snapshot give the same final state bit for bit.
+
+CLI
+---
+ python rcb_myosuite.py make-instance --task key_turn --out instance.npz
+ python rcb_myosuite.py replay --task key_turn --instance instance.npz --traj traj.npz [--repeat N] [--out r.json] [--video replay.mp4]
+
+Python (what the agent gets)
+----------------------------
+ from rcb_myosuite import Session
+ s = Session("key_turn", instance="/app/instance.npz")
+ s.reset() # back to the frozen instance state (unlimited)
+ s.state() # the task's observation dict (privileged): hand joints, object / goal poses, errors, ...
+ s.muscles(); s.joints() # muscle and joint names (+ ranges); s.moment_arms() -> (39, nv) at the current pose
+ s.model, s.data # the MuJoCo model / data, read-only by convention (only actions are replayed)
+ s.step(a) # one control step: 39 muscle commands (excitation = sigmoid(5 (a-0.5))); False once over
+ s.status() # {'success', 'done', 'reason', 'steps', 'resets', 'solved_run', 'state_sha', 'info'}
+ s.record_start(); ...; A = s.record_stop(); s.replay(A); s.save_trajectory(A, "/app/output/trajectory.npz")
+ s.render_frame("front_view"); s.render_video(A, "/app/output/check.mp4")
+"""
+from __future__ import annotations
+
+import os
+os.environ.setdefault("MUJOCO_GL", "osmesa")
+os.environ.setdefault("PYOPENGL_PLATFORM", "osmesa")
+
+import argparse
+import copy
+import hashlib
+import json
+import sys
+import time
+from pathlib import Path
+from typing import Any
+
+import numpy as np
+
+MAX_STEPS = 10_000_000 # runaway guard only; the wall clock is the budget
+# the task's `solved` must hold this long (consecutive control steps) to count. 1 s for a fixed goal, so passing
+# through it does not count; baoding's targets circle the palm once every 5 s, so there it is one full orbit (the balls
+# must really rotate: parking them where the moving targets sweep past was enough at 0.25 s).
+HOLD_S = {"key_turn": 1.0, "obj_hold": 1.0, "pen_twirl": 1.0, "die_reorient": 1.0, "baoding": 5.0,
+ "hand_reach": 1.0, "hand_pose": 1.0, "obj_reorient": 1.0, "arm_reach": 1.0, "relocate": 1.0}
+TASKS = {
+ "key_turn": "myoHandKeyTurnFixed-v0",
+ "obj_hold": "myoHandObjHoldFixed-v0",
+ "pen_twirl": "myoHandPenTwirlFixed-v0",
+ "die_reorient": "myoChallengeDieReorientP1-v0",
+ "baoding": "myoChallengeBaodingP1-v1",
+ "hand_reach": "myoHandReachFixed-v0", # all five fingertips to fixed targets
+ "hand_pose": "myoHandPoseFixed-v0", # the hand's 23 joints to a fixed target pose
+ "obj_reorient": "myoHandReorient8-v0", # MyoDex / SAR: reorient one of 8 objects (one frozen draw)
+ "arm_reach": "myoArmReachFixed-v0", # MyoArm (34 muscles): index fingertip to a target in space
+ "relocate": "myoChallengeRelocateP1-v0", # MyoChallenge 2023: MyoArm (63 muscles) moves an object into a box
+ # set 3 (2026-10-06)
+ "sarc_key_turn": "myoSarcHandKeyTurnFixed-v0", # key_turn with sarcopenic (weaker) muscles
+ "table_tennis": "myoChallengeTableTennisP1-v0", # MyoChallenge 2025: return a served ball with a paddle
+ "bimanual": "myoChallengeBimanual-v0", # MyoChallenge 2024: MyoArm hands an object to a robotic arm
+ "soccer": "myoChallengeSoccerP1-v0", # MyoChallenge 2025: MyoLegs kick the ball into the goal
+ "chase_tag": "myoChallengeChaseTagP1-v0", # MyoChallenge 2023: MyoLegs catch a moving opponent
+ "osl_run": "myoChallengeOslRunFixed-v0", # MyoChallenge 2024: MyoLeg + a prosthetic leg run a track
+ "dex_apple_lift": "MyoHandAppleLift-v0", # MyoDex: track a recorded object motion (GRAB / TCDM)
+ "dex_cylinder_pass": "MyoHandCylindersmallPass-v0",
+ "dex_bunny_inspect": "MyoHandStanfordbunnyInspect-v0",
+ "dex_hammer_use": "MyoHandHammerUse-v0",
+ "dex_cup_pour": "MyoHandCupPour-v0",
+ "dex_airplane_fly": "MyoHandAirplaneFly-v0",
+}
+BODY = {t: "arm" if t in ("arm_reach", "relocate") else "hand" for t in TASKS}
+BODY.update(table_tennis="body", bimanual="bimanual", soccer="legs", chase_tag="legs", osl_run="osl")
+# How an episode is judged (set 3 adds two rules; the first ten families keep "hold"):
+# hold : MyoSuite's `solved` held HOLD_S seconds in a row; its `done` is a failure.
+# event : a MyoChallenge's own `solved` is its success event and ends the episode (the challenge's official
+# definition: goal scored, opponent tagged, ball returned, object placed, track finished); `done` without it
+# (a fall, out of bounds, the challenge's own time limit) is a failure.
+# track : MyoDex gives no binary success (its `solved` is always 0). The episode is the whole recorded motion; it
+# succeeds iff MyoDex's termination never fires, the object's MOTION follows the reference's (its displacement
+# from where it started within TRACK_POS_M of the reference's displacement from where the reference started,
+# and, where orientation matters, its rotation since the start within TRACK_ROT_RAD of the reference's) at
+# >= TRACK_FRAC of the control steps and at the last one, and the displacement error never exceeds
+# TRACK_MAX_M. Relative, because a reference may start a few centimetres off the resting object (the apple's
+# starts 2.9 cm below it: with absolute errors, leaving the apple on the table passed); the cap, because a
+# small motion (the apple is lifted 6.8 cm) could otherwise be skipped within the 10 % allowance.
+RULE = {t: "hold" for t in TASKS}
+RULE.update({t: "event" for t in ("table_tennis", "bimanual", "soccer", "chase_tag", "osl_run")})
+RULE.update({t: "track" for t in TASKS if t.startswith("dex_")})
+HOLD_S["sarc_key_turn"] = 1.0
+TRACK_POS_M, TRACK_ROT_RAD, TRACK_FRAC, TRACK_MAX_M = 0.03, 0.5, 0.9, 0.05
+TRACK_ROT = {"dex_apple_lift": False, "dex_cylinder_pass": False, "dex_bunny_inspect": True, "dex_hammer_use": True,
+ "dex_cup_pour": True, "dex_airplane_fly": True}
+# Set 3 families restore every piece of episode state on reset: the env's random generator (opponents, goalkeepers and
+# terrains draw from it during reset or the episode) and its Python-side objects (OSL controller, reference cursor,
+# opponent policy, ...), deep-copied once after the instance is loaded. The first ten keep their original reset.
+STATE_SET3 = {t for t in TASKS if t not in ("key_turn", "obj_hold", "pen_twirl", "die_reorient", "baoding", "hand_reach",
+ "hand_pose", "obj_reorient", "arm_reach", "relocate")}
+_NO_COPY = ("mj_", "obsd_", "robot", "action_space", "observation_space", "spec", "mj_renderer", "_np_random", "np_random",
+ "sim", "sim_obsd", "viewer", "renderer")
+_SIMPLE = (bool, int, float, str, type(None), np.integer, np.floating, np.bool_, np.ndarray)
+_MODEL_SKIP = ("mesh_", "tex_", "hfield_", "skin_", "name_", "names", "paths", "text_", "flex_", "bvh_", "oct_")
+
+SESSION_LOG = os.environ.get("RCB_SESSION_LOG", "/app/output/rcb_session_log.jsonl")
+SESSION_LOG_KEEP = os.environ.get("RCB_SESSION_LOG_KEEP", "/logs/agent/rcb_session_log.jsonl")
+
+
+def _slog(ev: str, **kw) -> None:
+ if not SESSION_LOG:
+ return
+ line = json.dumps({"ts": time.time(), "pid": os.getpid(), "ev": ev, **kw}) + "\n"
+ for path, must_exist in ((SESSION_LOG, False), (SESSION_LOG_KEEP, True)):
+ if not path or (must_exist and not os.path.isdir(os.path.dirname(path))):
+ continue
+ try:
+ os.makedirs(os.path.dirname(path), exist_ok=True)
+ with open(path, "a") as f:
+ f.write(line)
+ except OSError:
+ pass
+
+
+def _log(*a: Any) -> None:
+ print("[rcb_myosuite]", *a, file=sys.stderr, flush=True)
+
+
+class PrimitiveError(RuntimeError):
+ pass
+
+
+def _plain(v: Any) -> bool:
+ if isinstance(v, _SIMPLE):
+ return True
+ if isinstance(v, (list, tuple)):
+ return all(_plain(x) for x in v)
+ if isinstance(v, dict):
+ return all(isinstance(k, (str, int)) and _plain(x) for k, x in v.items())
+ return False
+
+
+def _model_arrays(model) -> dict[str, np.ndarray]:
+ out = {}
+ for k in dir(model):
+ if k.startswith("_") or k.startswith(_MODEL_SKIP):
+ continue
+ try:
+ v = getattr(model, k)
+ except Exception: # noqa: BLE001
+ continue
+ if isinstance(v, np.ndarray) and v.dtype.kind in "fiub" and v.size:
+ out[k] = v
+ return out
+
+
+def _jsonable(v: Any) -> Any:
+ if isinstance(v, np.ndarray):
+ return v.round(6).tolist()
+ if isinstance(v, (np.floating,)):
+ return round(float(v), 6)
+ if isinstance(v, (np.integer,)):
+ return int(v)
+ if isinstance(v, (np.bool_,)):
+ return bool(v)
+ if isinstance(v, (list, tuple)):
+ return [_jsonable(x) for x in v]
+ if isinstance(v, dict):
+ return {str(k): _jsonable(x) for k, x in v.items()}
+ return v
+
+
+def _types(attrs: dict) -> dict:
+ def t(v):
+ if isinstance(v, np.ndarray):
+ return {"nd": str(v.dtype), "shape": list(v.shape)}
+ if isinstance(v, (np.bool_, np.integer, np.floating)):
+ return {"np": str(np.asarray(v).dtype)}
+ if isinstance(v, bool):
+ return "bool"
+ if isinstance(v, int):
+ return "int"
+ if isinstance(v, float):
+ return "float"
+ if isinstance(v, tuple):
+ return {"tuple": [t(x) for x in v]}
+ if isinstance(v, list):
+ return {"list": [t(x) for x in v]}
+ if isinstance(v, dict):
+ return {"dict": {str(k): t(x) for k, x in v.items()}, "keys": [type(k).__name__ for k in v]}
+ return type(v).__name__
+ return {k: t(v) for k, v in attrs.items()}
+
+
+def _to_json(attrs: dict) -> dict:
+ def j(v):
+ if isinstance(v, np.ndarray):
+ return v.tolist()
+ if isinstance(v, (np.bool_,)):
+ return bool(v)
+ if isinstance(v, np.integer):
+ return int(v)
+ if isinstance(v, np.floating):
+ return float(v)
+ if isinstance(v, (list, tuple)):
+ return [j(x) for x in v]
+ if isinstance(v, dict):
+ return {str(k): j(x) for k, x in v.items()}
+ return v
+ return {k: j(v) for k, v in attrs.items()}
+
+
+def _from_json(vals: dict, types: dict) -> dict:
+ def f(v, t):
+ if isinstance(t, dict) and "np" in t:
+ return np.asarray(v, dtype=t["np"])[()]
+ if isinstance(t, dict) and "nd" in t:
+ return np.asarray(v, dtype=t["nd"]).reshape(t["shape"])
+ if isinstance(t, dict) and "tuple" in t:
+ return tuple(f(x, tt) for x, tt in zip(v, t["tuple"]))
+ if isinstance(t, dict) and "list" in t:
+ return [f(x, tt) for x, tt in zip(v, t["list"])]
+ if isinstance(t, dict) and "dict" in t:
+ out = {}
+ for (k, x), kt in zip(v.items(), t["keys"]):
+ out[int(k) if kt == "int" else k] = f(x, t["dict"][k])
+ return out
+ if t == "bool":
+ return bool(v)
+ if t == "int":
+ return int(v)
+ if t == "float":
+ return float(v)
+ return v
+ return {k: f(v, types[k]) for k, v in vals.items()}
+
+
+class Session:
+ """One MyoSuite env on a frozen instance. Every episode starts from the instance snapshot."""
+
+ def __init__(self, task: str, instance: str | Path | None = None, *, max_steps: int = MAX_STEPS,
+ image_size: tuple[int, int] = (480, 640)):
+ import mujoco
+ from myosuite.utils import gym as mgym
+ if task not in TASKS:
+ raise KeyError(f"unknown task {task!r}; known: {sorted(TASKS)}")
+ self._mj = mujoco
+ self.task, self.env_id = task, TASKS[task]
+ self.max_steps = int(max_steps)
+ self.env = mgym.make(self.env_id)
+ self.base = self.env.unwrapped # no TimeLimit wrapper: MyoSuite's step budget is lifted
+ self.base.seed(0) # the challenge envs sample goal / mass / size in reset() before reset(seed=) reseeds
+ self._set3 = task in STATE_SET3
+ self.rule = RULE[task]
+ self._rng0 = copy.deepcopy(self._rng().bit_generator.state) if self._set3 else None
+ self.base.reset(seed=0)
+ self.model, self.data = self.base.mj_model, self.base.mj_data
+ self.nu = int(self.model.nu)
+ self.dt = float(self.base.dt)
+ self.hold_steps = max(1, int(round(HOLD_S[task] / self.dt))) if task in HOLD_S else 1
+ self._renderer = None
+ self._image_size = image_size
+ self._rec: list[np.ndarray] | None = None
+ self.resets = 0
+ self.instance_meta: dict[str, Any] = {}
+ if instance is not None:
+ self.load_instance(instance)
+ else:
+ self._snapshot = self.snapshot()
+ self._extras = self._copy_extras() if self._set3 else {}
+ self._rng_after = copy.deepcopy(self._rng().bit_generator.state) if self._set3 else None
+ self.track_steps = self._track_steps() if self.rule == "track" else None
+ self._episode_reset()
+ self._total_steps = 0
+ _slog("session_start", task=task)
+ import atexit
+ atexit.register(lambda: _slog("session_end", task=task, resets=self.resets, sim_steps=self._total_steps + self.steps))
+
+ # ------------------------------------------------------------------ snapshots
+ # MyoSuite steps the physics in its robot wrapper's own MjData (base.robot.mj_data); base.mj_data / obsd_mj_data are
+ # mirrors it refreshes from sensor readings. The snapshot is the robot's data; restore writes it back there and to both
+ # mirrors, and base.forward() then re-derives the observations from it.
+ def _phys(self) -> np.ndarray:
+ mj = self._mj
+ sz = mj.mj_stateSize(self.model, mj.mjtState.mjSTATE_INTEGRATION)
+ s = np.empty(sz); mj.mj_getState(self.model, self.base.robot.mj_data, s, mj.mjtState.mjSTATE_INTEGRATION)
+ return s
+
+ def _rng(self):
+ g = self.base.np_random
+ return getattr(g, "_np_random", g) # MyoSuite wraps the generator (NPRandomVersionWrapper)
+
+ def _copy_extras(self) -> dict[str, Any]:
+ """Set 3: deep copies of the env's Python-side episode state (controllers, opponent policies, the reference
+ cursor), taken once after the instance is loaded and put back on every reset. Objects that cannot be copied
+ (they hold MuJoCo views) are rebuilt by the env's own reset from the restored random generator."""
+ out = {}
+ for k, v in vars(self.base).items():
+ if k.startswith("__") or _plain(v) or k.startswith(_NO_COPY) or k in ("mujoco_render_frames",):
+ continue
+ try:
+ out[k] = copy.deepcopy(v)
+ except Exception: # noqa: BLE001
+ pass
+ return out
+
+ def _track_steps(self) -> int:
+ t = np.asarray(self.base.ref.reference["time"], dtype=np.float64).reshape(-1)
+ return max(1, int(round((t[-1] - t[0]) / self.dt)))
+
+ def snapshot(self) -> dict[str, Any]:
+ attrs = {k: copy.deepcopy(v) for k, v in vars(self.base).items()
+ if not k.startswith("__") and _plain(v) and k not in ("mujoco_render_frames",)}
+ snap = {"physics": self._phys(), "attrs": attrs, "model": {k: v.copy() for k, v in _model_arrays(self.model).items()}}
+ if self._set3 and self.model.nmocap:
+ snap["mocap"] = self._mocap()
+ return snap
+
+ def _mocap(self) -> np.ndarray:
+ # set 3: opponents, goalkeepers and the like are mocap bodies that MyoSuite moves in base.mj_data (the robot's own
+ # data keeps them at the origin): part of the state the observations and the success test read
+ return np.concatenate([self.data.mocap_pos.ravel(), self.data.mocap_quat.ravel()]).copy()
+
+ def restore(self, snap: dict[str, Any]) -> None:
+ mj = self._mj
+ live = _model_arrays(self.model)
+ for k, v in snap.get("model", {}).items():
+ if k in live and live[k].shape == v.shape:
+ live[k][...] = v
+ phys = np.asarray(snap["physics"])
+ datas = [self.base.robot.mj_data, self.data] + ([self.base.obsd_mj_data] if self.base.obsd_mj_data is not self.data else [])
+ for d in datas:
+ mj.mj_setState(self.model, d, phys, mj.mjtState.mjSTATE_INTEGRATION)
+ mj.mj_forward(self.model, d)
+ if "mocap" in snap:
+ m = np.asarray(snap["mocap"], dtype=np.float64)
+ n = self.model.nmocap
+ for d in datas[1:]:
+ d.mocap_pos[:] = m[: 3 * n].reshape(n, 3)
+ d.mocap_quat[:] = m[3 * n:].reshape(n, 4)
+ mj.mj_forward(self.model, d)
+ for k, v in snap["attrs"].items():
+ setattr(self.base, k, copy.deepcopy(v))
+ self.base.robot._sensor_cache.clear() # readings of the abandoned branch
+ self.base.forward() # refresh obs_dict / rwd_dict from the restored state
+
+ def state_sha(self) -> str:
+ h = hashlib.sha1(self._phys().tobytes())
+ for k, v in sorted(_model_arrays(self.model).items()):
+ h.update(k.encode()); h.update(np.ascontiguousarray(v).tobytes())
+ if self._set3 and self.model.nmocap:
+ h.update(self._mocap().tobytes())
+ if self._set3 and self.model.nhfielddata: # terrains are drawn at reset: part of the instance
+ h.update(np.ascontiguousarray(self.model.hfield_data).tobytes())
+ return h.hexdigest()[:16]
+
+ def save_instance(self, path: str | Path, meta: dict | None = None) -> None:
+ snap = self.snapshot()
+ np.savez_compressed(path, task=self.task, env_id=self.env_id, physics=snap["physics"],
+ **{"model__" + k: v for k, v in snap["model"].items()},
+ attrs=json.dumps(_to_json(snap["attrs"])), attr_types=json.dumps(_types(snap["attrs"])),
+ sha=self.state_sha(), meta=json.dumps(meta or {}),
+ **({"mocap": snap["mocap"]} if "mocap" in snap else {}))
+
+ def load_instance(self, path: str | Path) -> None:
+ d = np.load(path, allow_pickle=False)
+ if str(d["task"]) != self.task:
+ raise ValueError(f"instance is for task {str(d['task'])!r}, not {self.task!r}")
+ self._snapshot = {"physics": d["physics"], "attrs": _from_json(json.loads(str(d["attrs"])), json.loads(str(d["attr_types"]))),
+ "model": {k[len("model__"):]: d[k] for k in d.files if k.startswith("model__")}}
+ if "mocap" in d.files:
+ self._snapshot["mocap"] = d["mocap"]
+ self.instance_meta = json.loads(str(d["meta"]))
+ self.restore(self._snapshot)
+ self.instance_sha = str(d["sha"])
+ self.instance_exact = self.state_sha() == self.instance_sha
+ if not self.instance_exact:
+ _log(f"WARNING: restored state sha {self.state_sha()} != instance sha {self.instance_sha}")
+
+ # ------------------------------------------------------------------ episode
+ def _episode_reset(self) -> None:
+ # a full env reset first: MyoSuite's robot wrapper caches sensor readings outside mjData (not in the snapshot),
+ # so restoring the snapshot alone left the previous episode's key angle / object pose in the observations
+ if self._set3: # the same draws as the instance's own reset, every episode
+ self._rng().bit_generator.state = copy.deepcopy(self._rng0)
+ self.base.reset(seed=0)
+ self.restore(self._snapshot)
+ if self._set3:
+ for k, v in self._extras.items():
+ setattr(self.base, k, copy.deepcopy(v))
+ self._rng().bit_generator.state = copy.deepcopy(self._rng_after)
+ self.base.forward()
+ self.track_ok = 0
+ self.track_max = 0.0
+ if self.rule == "track":
+ od = self.base.obs_dict
+ self._track0 = {k: np.asarray(od[k], dtype=np.float64).reshape(-1).copy()
+ for k in ("curr_obj_com", "targ_obj_com", "curr_obj_rot", "targ_obj_rot")}
+ self.steps = 0
+ self.done = False
+ self.success = False
+ self.reason: str | None = None
+ self.solved_run = 0
+ self.last_info: dict[str, Any] = {}
+
+ def reset(self) -> None:
+ """Back to the frozen instance state (unlimited). Starts a new episode."""
+ _slog("reset", task=self.task, n=self.resets + 1, sim_steps_before=self.steps)
+ self._total_steps += self.steps
+ self._episode_reset()
+ self.resets += 1
+
+ def step(self, action) -> bool:
+ """One control step with 39 muscle commands a (excitation = sigmoid(5 (a - 0.5))). True while the episode runs."""
+ a = np.asarray(action, dtype=np.float64).reshape(-1)
+ if a.shape != (self.nu,):
+ raise ValueError(f"action must have {self.nu} muscle commands, got {a.shape}")
+ if not np.all(np.isfinite(a)):
+ raise ValueError("action contains NaN or inf")
+ if self.done:
+ raise PrimitiveError(f"episode is over ({self.reason}); call reset()")
+ if self.steps >= self.max_steps:
+ raise PrimitiveError(f"runaway guard: {self.max_steps} steps")
+ _, _, terminated, _, info = self.base.step(a)
+ self.steps += 1
+ if self._rec is not None:
+ self._rec.append(a.copy())
+ solved = bool(np.asarray(info.get("solved", False)).any())
+ self.solved_run = self.solved_run + 1 if solved else 0
+ self.last_info = {k: _jsonable(v) for k, v in dict(info).items() if k in ("solved", "done", "rwd_sparse", "rwd_dense")}
+ self.last_info["solved_run"] = self.solved_run
+ failed = bool(np.asarray(info.get("done", False)).any())
+ if self.rule == "track":
+ pos, rot = self.track_errors()
+ ok = pos <= TRACK_POS_M and (not TRACK_ROT[self.task] or rot <= TRACK_ROT_RAD)
+ self.track_ok += int(ok)
+ self.track_max = max(self.track_max, pos)
+ self.last_info.update(track_pos_err=round(pos, 5), track_rot_err=round(rot, 4), track_ok=self.track_ok)
+ if failed:
+ self.done, self.reason = True, "failure: " + _FAILURE.get(self.task, "task terminated")
+ elif self.steps >= self.track_steps:
+ self.done = True
+ self.success = ok and self.track_ok >= TRACK_FRAC * self.track_steps and self.track_max <= TRACK_MAX_M
+ self.reason = "success" if self.success else (
+ f"failure: on track at {self.track_ok} of {self.track_steps} steps"
+ + ("" if ok else ", off track at the end")
+ + (f", {self.track_max:.3f} m off at worst" if self.track_max > TRACK_MAX_M else ""))
+ elif self.solved_run >= self.hold_steps and self.task == "bimanual" and not self._bimanual_contacts_ok():
+ self.done, self.reason = True, "failure: placed without a clean hand-over (MyoChallenge's contact check)"
+ elif self.solved_run >= self.hold_steps:
+ self.done, self.success, self.reason = True, True, "success"
+ elif failed:
+ self.done, self.reason = True, "failure: " + _FAILURE.get(self.task, "task terminated")
+ # Baoding's targets come from a precomputed trajectory of 1000 points (baoding_v1.create_goal_trajectory,
+ # "assumes that it's greater than env horizon"): with MyoSuite's step budget lifted, step 1001 would index past
+ # it (IndexError, board 0086). The episode ends there instead, as a failure unless it succeeded on that step.
+ n_goal = len(getattr(self.base, "goal", ())) if self.task == "baoding" else 0
+ if not self.done and n_goal and self.steps >= n_goal:
+ self.done, self.reason = True, f"failure: the targets' trajectory ends after {n_goal} control steps"
+ return not self.done
+
+ def _bimanual_contacts_ok(self) -> bool:
+ """MyoChallenge 2024's own evaluation of the object's contact history: the MyoHand held it, then the prosthesis,
+ then only the goal pillar, and it never touched the table or the floor (bimanual_v0.evaluate_contact_trajectory)."""
+ from myosuite.envs.myo.myochallenge.bimanual_v0 import evaluate_contact_trajectory
+ return evaluate_contact_trajectory(self.base.touch_history) is None
+
+ def track_errors(self) -> tuple[float, float]:
+ """MyoDex families: how far the object's motion since the episode start is from the reference's: the
+ displacement error (m) of the centre of mass and the angle (rad) between the two rotations since the start."""
+ od, t0 = self.base.obs_dict, self._track0
+ v = lambda k: np.asarray(od[k], dtype=np.float64).reshape(-1) # noqa: E731
+ pos = float(np.linalg.norm((v("curr_obj_com") - t0["curr_obj_com"]) - (v("targ_obj_com") - t0["targ_obj_com"])))
+ rot = _quat_angle(_qmul(v("curr_obj_rot"), _qconj(t0["curr_obj_rot"])), _qmul(v("targ_obj_rot"), _qconj(t0["targ_obj_rot"])))
+ return pos, rot
+
+ def rest_action(self) -> np.ndarray:
+ """All muscles relaxed (a = -1: excitation ~0.0006)."""
+ return -np.ones(self.nu)
+
+ @staticmethod
+ def excitation(a):
+ """MyoSuite's command -> excitation map, u = 1 / (1 + exp(-5 (a - 0.5)))."""
+ return 1.0 / (1.0 + np.exp(-5.0 * (np.asarray(a, float) - 0.5)))
+
+ @staticmethod
+ def command_for(u):
+ """Inverse of excitation(): the command a that gives excitation u in (0, 1)."""
+ u = np.clip(np.asarray(u, float), 1e-6, 1 - 1e-6)
+ return 0.5 + np.log(u / (1 - u)) / 5.0
+
+ def status(self) -> dict[str, Any]:
+ return {"success": bool(self.success), "done": bool(self.done), "reason": self.reason, "steps": int(self.steps),
+ "rule": self.rule, "track_steps": self.track_steps, "track_ok": int(getattr(self, "track_ok", 0)),
+ "resets": int(self.resets), "solved_run": int(self.solved_run), "hold_steps": int(self.hold_steps),
+ "state_sha": self.state_sha(), "info": self.last_info}
+
+ # ------------------------------------------------------------------ recording / replay
+ def record_start(self) -> None:
+ self._rec = []
+
+ def record_stop(self) -> np.ndarray:
+ rec, self._rec = self._rec or [], None
+ _slog("record", task=self.task, n_actions=len(rec), success=bool(self.success))
+ return np.asarray(rec, dtype=np.float64).reshape(-1, self.nu)
+
+ def save_trajectory(self, actions, path: str | Path, meta: dict | None = None) -> None:
+ A = np.asarray(actions, dtype=np.float64)
+ if A.ndim != 2 or A.shape[1] != self.nu:
+ raise ValueError(f"trajectory must be (T, {self.nu}), got {A.shape}")
+ os.makedirs(os.path.dirname(os.path.abspath(path)), exist_ok=True)
+ np.savez(path, actions=A, task=self.task, instance_sha=getattr(self, "instance_sha", ""), meta=json.dumps(meta or {}))
+
+ def replay(self, actions) -> dict[str, Any]:
+ """Reset to the instance and replay open-loop until the episode ends or the rows run out."""
+ A = np.asarray(actions, dtype=np.float64)
+ self.reset()
+ n = 0
+ for row in A:
+ n += 1
+ if not self.step(row):
+ break
+ st = self.status(); st["replayed_rows"] = n; st["rows"] = int(len(A))
+ _slog("replay", task=self.task, n_actions=int(len(A)), success=st["success"])
+ return st
+
+ # ------------------------------------------------------------------ privileged state
+ def state(self) -> dict[str, list]:
+ """The task's observation dict (privileged): hand joints and velocities, object / goal / target poses and
+ their errors, muscle activations (`act`)."""
+ return {k: np.asarray(v, float).reshape(-1).round(6).tolist() for k, v in self.base.obs_dict.items() if k != "time"}
+
+ def muscles(self) -> list[dict]:
+ mj, m = self._mj, self.model
+ return [{"index": i, "name": mj.mj_id2name(m, mj.mjtObj.mjOBJ_ACTUATOR, i),
+ "ctrlrange": m.actuator_ctrlrange[i].round(3).tolist()} for i in range(m.nu)]
+
+ def actuators(self) -> list[dict]:
+ """Every actuator in action order: muscles (`kind` "muscle": the command a maps to the excitation
+ 1/(1+exp(-5(a-0.5)))) and the few non-muscle ones of some families (`kind` "position" servo or "motor"): their
+ command is the raw MuJoCo control in `ctrlrange` (MuJoCo clips it), except where `command` says otherwise."""
+ mj, m = self._mj, self.model
+ out = []
+ for i in range(m.nu):
+ name = mj.mj_id2name(m, mj.mjtObj.mjOBJ_ACTUATOR, i)
+ if m.actuator_dyntype[i] == mj.mjtDyn.mjDYN_MUSCLE:
+ out.append({"index": i, "name": name, "kind": "muscle", "command": "excitation 1/(1+exp(-5(a-0.5)))"})
+ continue
+ kp, b1 = float(m.actuator_gainprm[i, 0]), float(m.actuator_biasprm[i, 1])
+ kind = "position" if b1 != 0 and abs(b1 + kp) < 1e-9 else "motor"
+ trn = int(m.actuator_trnid[i, 0])
+ target = mj.mj_id2name(m, mj.mjtObj.mjOBJ_JOINT, trn) if m.actuator_trntype[i] == mj.mjtTrn.mjTRN_JOINT else None
+ lo, hi = (float(x) for x in m.actuator_ctrlrange[i])
+ cmd = (f"position target of joint {target} in [{lo:g}, {hi:g}]" if kind == "position" else
+ f"control in [{lo:g}, {hi:g}] (force or torque = {float(m.actuator_gear[i, 0]):g} x control)")
+ if self.task in ("bimanual", "table_tennis"): # their env's step maps [-1, 1] onto the control range
+ cmd = f"normalized: a in [-1, 1] maps linearly onto [{lo:g}, {hi:g}] ({kind} of joint {target})"
+ if self.task == "osl_run": # RunTrack.step overwrites them with the prosthesis's own controller
+ cmd = "ignored: the prosthesis's own impedance controller (MyoSuite's OSL state machine) sets it"
+ out.append({"index": i, "name": name, "kind": kind, "joint": target, "ctrlrange": [lo, hi], "command": cmd})
+ return out
+
+ def joints(self) -> list[dict]:
+ mj, m = self._mj, self.model
+ return [{"index": j, "name": mj.mj_id2name(m, mj.mjtObj.mjOBJ_JOINT, j), "qpos_adr": int(m.jnt_qposadr[j]),
+ "dof_adr": int(m.jnt_dofadr[j]), "range": m.jnt_range[j].round(3).tolist(), "qpos": float(self.data.qpos[m.jnt_qposadr[j]])}
+ for j in range(m.njnt) if m.jnt_type[j] in (2, 3)]
+
+ def moment_arms(self) -> np.ndarray:
+ """(nu, nv) muscle moment arms at the current pose (d muscle length / d q, MuJoCo's actuator_moment)."""
+ mj, m, d = self._mj, self.model, self.data
+ if hasattr(d, "actuator_moment") and d.actuator_moment.ndim == 2:
+ return d.actuator_moment.copy()
+ out = np.zeros((m.nu, m.nv))
+ mj.mju_sparse2dense(out, d.actuator_moment, d.moment_rownnz, d.moment_rowadr, d.moment_colind)
+ return out
+
+ # ------------------------------------------------------------------ rendering (on demand)
+ def cameras(self) -> list[str]:
+ mj = self._mj
+ return [mj.mj_id2name(self.model, mj.mjtObj.mjOBJ_CAMERA, i) for i in range(self.model.ncam)]
+
+ def default_camera(self) -> str:
+ return "front_view" if "front_view" in self.cameras() else self.cameras()[0]
+
+ def render_frame(self, camera: str | None = None, size: tuple[int, int] | None = None) -> np.ndarray:
+ mj = self._mj
+ h, w = size or self._image_size
+ if self._renderer is None or (self._renderer.height, self._renderer.width) != (h, w):
+ if self._renderer is not None:
+ self._renderer.close(); self._renderer = None # a second OSMesa context next to a live one reads garbage
+ self.model.vis.global_.offheight = max(self.model.vis.global_.offheight, h)
+ self.model.vis.global_.offwidth = max(self.model.vis.global_.offwidth, w)
+ self._renderer = mj.Renderer(self.model, height=h, width=w)
+ self._renderer.update_scene(self.data, camera=camera or self.default_camera())
+ return self._renderer.render().copy()
+
+ def render_video(self, actions, path: str | Path, *, cameras: list[str] | None = None, every: int = 1, fps: int = 25,
+ size: tuple[int, int] = (240, 320)) -> dict[str, Any]:
+ import imageio.v2 as imageio
+ A = np.asarray(actions, dtype=np.float64)
+ self.reset()
+ cams = cameras or [c for c in ("hand_top", "hand_side_inter") if c in self.cameras()] or [self.default_camera()]
+ w = imageio.get_writer(str(path), fps=fps, codec="libx264", quality=7, macro_block_size=8)
+ try:
+ w.append_data(np.concatenate([self.render_frame(c, size) for c in cams], axis=1))
+ for i, row in enumerate(A):
+ running = self.step(row)
+ if i % every == 0 or not running:
+ w.append_data(np.concatenate([self.render_frame(c, size) for c in cams], axis=1))
+ if not running:
+ break
+ finally:
+ w.close()
+ return self.status()
+
+ def close(self) -> None:
+ try:
+ self.env.close()
+ except Exception:
+ pass
+
+
+_FAILURE = {"key_turn": "fingertips too far from the key", "obj_hold": "object dropped", "pen_twirl": "pen dropped",
+ "die_reorient": "die dropped", "baoding": "a ball dropped or left its target", "pose": "task terminated", "hand_reach": "fingertips too far from their targets", "hand_pose": "hand pose too far from the target",
+ "obj_reorient": "object dropped", "arm_reach": "fingertip too far from the target",
+ "relocate": "object dropped (palm more than 0.5 m from it)",
+ "sarc_key_turn": "fingertips too far from the key", "table_tennis": "the rally ended without a return",
+ "bimanual": "object dropped or the time limit passed", "soccer": "fallen, or the time limit passed without a goal",
+ "chase_tag": "the opponent was not tagged in time, out of bounds or fallen",
+ "osl_run": "fallen, off the track or out of time",
+ **{t: "object more than 0.25 m from the reference (or from the wrist)" for t in TASKS if t.startswith("dex_")}}
+
+
+def _qconj(q):
+ q = np.asarray(q, dtype=np.float64)
+ return np.array([q[0], -q[1], -q[2], -q[3]]) / max(np.dot(q, q), 1e-12)
+
+
+def _qmul(a, b):
+ w1, x1, y1, z1 = a
+ w2, x2, y2, z2 = b
+ return np.array([w1 * w2 - x1 * x2 - y1 * y2 - z1 * z2, w1 * x2 + x1 * w2 + y1 * z2 - z1 * y2,
+ w1 * y2 - x1 * z2 + y1 * w2 + z1 * x2, w1 * z2 + x1 * y2 - y1 * x2 + z1 * w2])
+
+
+def _quat_angle(a, b) -> float:
+ """Angle (rad) of the rotation between two wxyz quaternions."""
+ a = a / np.linalg.norm(a); b = b / np.linalg.norm(b)
+ return float(2.0 * np.arccos(min(1.0, abs(float(np.dot(a, b))))))
+
+
+def _load_traj(path: str) -> np.ndarray:
+ return np.asarray(np.load(path, allow_pickle=False)["actions"], dtype=np.float64)
+
+
+def cmd_make_instance(a) -> None:
+ s = Session(a.task)
+ s.save_instance(a.out, {"source": f"{TASKS[a.task]} after reset(seed=0)", "myosuite": "2.12.2"})
+ _log("instance", a.out, "sha", s.state_sha())
+
+
+def cmd_replay(a) -> None:
+ """Replay a trajectory from the frozen instance, `--repeat` times in this process (the graders run it once per fresh
+ process); prints, and with --out writes, the result as JSON. --video renders the first replay (after the verdict)."""
+ A = _load_traj(a.traj)
+ a.task = a.task or str(np.load(a.instance, allow_pickle=False)["task"])
+ runs, instance_ok = [], True
+ for _ in range(a.repeat):
+ s = Session(a.task, instance=a.instance)
+ instance_ok = instance_ok and bool(s.instance_exact)
+ st = s.replay(A)
+ runs.append({k: st[k] for k in ("success", "reason", "steps", "state_sha", "replayed_rows", "rows")})
+ s.close()
+ out = {"n_actions": int(len(A)), "replays": runs, "success": all(r["success"] for r in runs),
+ "state_sha": runs[0]["state_sha"] if runs else None, "steps": runs[0]["steps"] if runs else 0,
+ "reason": runs[0]["reason"] if runs else None, "instance_ok": instance_ok,
+ "deterministic": len({r["state_sha"] for r in runs}) == 1, "video": False}
+ if a.out:
+ with open(a.out, "w") as f:
+ json.dump(out, f)
+ if a.video:
+ try:
+ s = Session(a.task, instance=a.instance)
+ s.render_video(A, a.video, every=max(1, -(-len(A) // 1500)))
+ out["video"] = os.path.isfile(a.video)
+ s.close()
+ except Exception as exc: # noqa: BLE001 -- the video never changes the verdict
+ _log("video failed:", repr(exc))
+ if a.out:
+ with open(a.out, "w") as f:
+ json.dump(out, f)
+ print(json.dumps(out))
+
+
+def main() -> None:
+ ap = argparse.ArgumentParser(description=__doc__.split("\n\n")[0])
+ sub = ap.add_subparsers(dest="cmd", required=True)
+ m = sub.add_parser("make-instance"); m.add_argument("--task", required=True); m.add_argument("--out", required=True)
+ r = sub.add_parser("replay"); r.add_argument("--task", help="default: the instance's task"); r.add_argument("--instance", required=True)
+ r.add_argument("--traj", required=True); r.add_argument("--repeat", type=int, default=1); r.add_argument("--video")
+ r.add_argument("--out", help="write the result JSON here")
+ a = ap.parse_args()
+ {"make-instance": cmd_make_instance, "replay": cmd_replay}[a.cmd](a)
+
+
+if __name__ == "__main__":
+ main()
diff --git a/tasks/myosuite-table-tennis-i00-standard-realtime/environment/sim/rcb_realtime.py b/tasks/myosuite-table-tennis-i00-standard-realtime/environment/sim/rcb_realtime.py
new file mode 100644
index 0000000000..cf4204bb26
--- /dev/null
+++ b/tasks/myosuite-table-tennis-i00-standard-realtime/environment/sim/rcb_realtime.py
@@ -0,0 +1,926 @@
+"""eai-standard/2.3, the standard-realtime mode's protocol: the sim service of HumanoidBench's `-standard-realtime` tasks
+(docs/PROTOCOL.md §4b, docs/STANDARD_MODE.md "Standard-realtime mode"; docs/benchmarks/humanoidbench.md). Entry
+point of the sidecar: `python /srv/rcb_realtime.py` (the adapter is server.py's, unchanged). 2.3 is 2.2 plus the realtime
+capability, every 2.2 addition optional: this service is rcb_service's 2.1 service with the capability, and reports
+eai-standard/2.3.
+
+It extends rcb_service.Service: the same endpoints, RCB1 frames, budgets (reset, wall clock), evidence, crash recovery
+and finalize. Only the time model differs. In standard mode time moves only inside calls; here the episode has a clock:
+
+- An episode starts paused, at the frozen start with the servos holding their starting targets. Its clock starts at its
+ first POST /submit or POST /observe with seconds > 0 (again after every reset, and after a crash recovery). From then
+ on one control step runs every control_dt x slowdown seconds of wall time, whether or not the agent sends anything,
+ until the episode ends; the clock stops there.
+- The service keeps a plan: one target row per coming control step. /submit expands its chunk (rows dt apart, `linear`
+ or `hold`, exactly as 2.1 does) into control steps and writes it into the plan from control step `at` (default: the
+ next one), replacing whatever was planned from there on, and answers at once. A chunk that arrives after its `at`
+ loses the rows whose steps are gone (`late_steps`). When the plan runs out the last target is held (zero-order
+ hold): a slow agent's robot keeps its last servo targets.
+- A simulator may have several action types (RoboOlympics: `stroke`, `toss`, `mallet`, `stop`, ...), and a plan may
+ mix them. A row is only ever combined with a row of its own type: a chunk interpolates from the target of ITS type
+ in force before `at` (the simulator's current command of that type when the plan holds another type there), and a
+ gap before `at` is filled with held steps (the simulator's hold row, as when the plan has run out) unless the type
+ is an interpolable target that is itself in force there (then that target, held). A command that is not a target
+ (not interpolable: one row is one control step) is never interpolated and never repeated into a gap.
+- /observe with seconds = 0 answers at once; with seconds > 0 it watches the running episode that long (robot time)
+ and returns what the cameras saw. It never changes the plan.
+- `cameras = []` asks for no images at all (in 2.1 an empty list means every camera): a controller's proprio-only call.
+
+Grading is unchanged: every executed control step's sim row is journaled (trajectory_live.f64) and the graded episode's
+rows are replayed in a fresh simulator at the end. What is graded is the rows the simulator applied, not what was
+submitted, so the replay equals the live run however the timing went. Evidence added (never visible to the agent):
+ticks.f64 (per journal row: wall time, and the call whose chunk gave the row, -1 for a held one), calls.jsonl (one line
+per call; no calls/NNNNNN/ directories: a controller calls 50 times a second) and submitted.f64 (every row as sent:
+call id, then the row). final.json adds realtime, slowdown, held_steps, late_steps, max_queue_wait_s (the longest a
+request waited for the simulator thread) and how well the clock was kept: clock_s (wall seconds the clock ran, over the
+whole trial), max_lag_s (the furthest a control step ran behind its time), stalls, stalled_s (the wall time the stalls
+dropped: the robot stood still that long while the clock should have run) and stalled_share = stalled_s / clock_s; and
+what the requests cost: requests_refused (answered 4xx), request_mb (request bodies read) and request_busy_s (simulator
+thread time spent on requests). The graders copy the clock's figures into reward.json; none of them decides a verdict.
+
+Threads: the simulator and its GL context belong to the main thread, which runs the clock and, between control steps,
+the part of a request that needs the simulator. Everything else about a request happens on its HTTP thread, off the
+clock: reading the body, decoding it, every check that does not depend on the episode's state, and encoding the answer.
+A request that is refused there (404, 413, a malformed body, a wrong type / shape / dt / camera ...) never reaches the
+simulator thread. What a request can cost the clock is bounded:
+- JSON text is parsed under this process's interpreter lock (which the simulator thread needs too) only up to
+ INLINE_JSON bytes; a body with more of it is parsed by a child process, so a flood of large bodies, refused or not,
+ takes no time from the clock. Bodies over LANE_BODY bytes are read, decoded and checked one at a time (one of them
+ in memory at a time) and paced (LANE_PACE); while serving, the interpreter switches threads every SWITCH_INTERVAL_S.
+- A control step that is due runs before any waiting request. Behind the clock (a long render, a simulator step slower
+ than control_dt, a busy host) the missed steps run back to back, skipping none; a waiting request is then served only
+ once the clock has caught up, or -- so that a simulator that cannot keep up at all still answers -- after
+ BEHIND_PACE times the time the last request of its kind took. Requests therefore queue behind each other, not into
+ the clock: an /observe may take longer to answer, the robot's time does not slow down. A request that renders and
+ alone takes more than MAX_LAG_S (a stall by itself: CPU rendering) keeps the next rendering request waiting
+ BEHIND_PACE times as long, clock on time or not; requests without images (cameras = [], /status) are not held
+ back by it, so a controller's loop goes on. Requests can so take at most 1 / (1 + BEHIND_PACE) of the clock's time.
+- Cameras are rendered (and the proprioception read) at most once per control step, however many requests and watches
+ ask for them: all of them get that step's frames.
+More than MAX_LAG_S behind, the service drops the debt instead of fast-forwarding the robot (a stall: counted, and its
+seconds added to stalled_s).
+
+A crash resumes the same episode from its last checkpoint (one per second of robot time while the clock runs), paused.
+A simulator crash is an infrastructure failure, never the agent's (PROTOCOL §4b, owner 2026-10-08). No call spans a
+crash here (submit answers at once), so a crash is never put down to the agent's call: past RCB_MAX_RESUMES restarts the
+trial ends `sim_crashed` (infra_error), a failed resume `resume_failed` (infra_error), and a simulator error that ends
+the graded episode (reason "error") sets infra_error in final.json too.
+"""
+from __future__ import annotations
+
+import collections
+import json
+import os
+import queue
+import struct
+import subprocess
+import sys
+import threading
+import time
+import traceback
+from http.server import BaseHTTPRequestHandler, ThreadingHTTPServer
+from urllib.parse import urlparse
+
+import numpy as np
+
+import rcb1
+import rcb_service
+from rcb_service import MAX_BODY, Recorder, Refused, Service, bad_request
+
+PROTOCOL = "eai-standard/2.3" # the standard-realtime mode's protocol (PROTOCOL §4b); rcb_service.py itself serves 2.1
+MAX_LAG_S = 0.5 # behind the clock by more than this: drop the debt (a stall) instead of catching up
+CHECKPOINT_S = 1.0 # robot time between checkpoints while the clock runs (the most a crash loses)
+HELD = (None, None, -1) # a plan entry for a step with nothing planned: the simulator's hold row, taken at that step
+# What a request may cost the clock (module docstring, "Threads"):
+INLINE_JSON = 1 << 20 # bytes of JSON text parsed in this process; more is parsed by a child process (off the lock)
+LANE_BODY = 64 << 10 # bodies above this are decoded and checked one at a time ...
+LANE_PACE = 3.0 # ... and the next one waits this many times the CPU time the last one took
+BEHIND_PACE = 4.0 # behind the clock: the next request waits this many times the time the last one took
+CHILD_TIMEOUT_S = 60.0 # the child process's time for one body
+CHILD_MEM = 1 << 30 # ... and its memory above what its interpreter needs (a body that needs more is refused)
+SWITCH_INTERVAL_S = 0.001 # the interpreter's thread switch interval while serving (default 0.005): HTTP threads
+ # hand the lock back to the simulator thread within a millisecond
+FIELDS = ("type", "dt", "interpolate", "at", "cameras", "video_fps", "video", "include_depth", "note", "seconds")
+
+# The child process that parses a body with more than INLINE_JSON bytes of JSON text (argv: the most elements a valid
+# `actions` can have, INLINE_JSON, CHILD_MEM; stdin: the body). It answers with the fields the service reads and the
+# actions as float64 -- or only their shape when no action type takes that many -- or the error a malformed body gets.
+_CHILD = r"""
+import json, struct, sys
+FIELDS = %r
+def conv(np, obj, lim):
+ try:
+ a = np.asarray(obj, dtype=np.float64)
+ except (TypeError, ValueError):
+ return {"kind": "bad"}, b""
+ if a.size > lim:
+ return {"kind": "shape", "shape": list(a.shape)}, b""
+ return {"kind": "array", "dtype": "<f8", "shape": list(a.shape)}, np.ascontiguousarray(a).astype("<f8").tobytes()
+def main():
+ lim, inline, mem = (int(x) for x in sys.argv[1:4])
+ import numpy as np
+ try:
+ import resource
+ with open("/proc/self/statm") as f:
+ vm = int(f.read().split()[0]) * resource.getpagesize()
+ resource.setrlimit(resource.RLIMIT_AS, (vm + mem, vm + mem))
+ except Exception:
+ pass
+ blob = b""
+ try:
+ buf = sys.stdin.buffer.read()
+ part = None
+ if buf[:4] == b"RCB1":
+ n = struct.unpack("<I", buf[4:8])[0]
+ header = json.loads(buf[8:8 + n].decode())
+ off = 8 + n
+ meta = header.get("meta", {})
+ for p in header["parts"]:
+ if p["name"] == "actions":
+ part = (p, off)
+ off += p["len"]
+ else:
+ meta = json.loads(buf or b"{}")
+ if not isinstance(meta, dict):
+ raise ValueError("the body must be a JSON object or an RCB1 frame")
+ desc = {"kind": "none"}
+ if part is not None:
+ p, off = part
+ raw = buf[off:off + p["len"]]
+ if p["kind"] == "array":
+ a = np.frombuffer(raw, dtype=np.dtype(p["dtype"])).reshape(p["shape"])
+ if a.size > lim:
+ desc = {"kind": "shape", "shape": list(a.shape)}
+ else:
+ desc, blob = {"kind": "array", "dtype": p["dtype"], "shape": list(a.shape)}, raw
+ elif p["kind"] == "json":
+ desc, blob = conv(np, json.loads(raw.decode()), lim)
+ else:
+ desc = {"kind": "bad"}
+ elif "actions" in meta:
+ desc, blob = conv(np, meta["actions"], lim)
+ small = {k: meta[k] for k in FIELDS if k in meta}
+ if "note" in small:
+ small["note"] = str(small["note"] or "")[:200]
+ head = json.dumps({"meta": small, "actions": desc}).encode()
+ if len(head) > inline:
+ raise ValueError("the request's fields other than the actions take more than %%d bytes" %% inline)
+ except MemoryError:
+ head, blob = json.dumps({"error": "the body takes too much memory to parse"}).encode(), b""
+ except Exception as exc:
+ head, blob = json.dumps({"error": str(exc)[:300] or type(exc).__name__}).encode(), b""
+ sys.stdout.buffer.write(struct.pack("<I", len(head)) + head + blob)
+main()
+""" % (FIELDS,)
+
+
+class _Job:
+ """One HTTP request, handed from its HTTP thread to the simulator's thread and back."""
+
+ def __init__(self, method: str, path: str, body: bytes = b"", content_length: int | None = None, req: tuple = ()):
+ self.method, self.path, self.body, self.content_length = method, path, body, content_length
+ self.req = req # (route, checked request) from RealtimeService.front(); () = not looked at yet
+ self.queued = time.monotonic()
+ self.done = threading.Event()
+ self.reply: tuple = () # ("raw", code, ctype, body) | ("rcb1", meta, parts) | ("rec", meta, Recorder)
+
+ def finish(self, *reply) -> None:
+ self.reply = reply
+ self.done.set()
+
+
+class _Watch:
+ """A pending /observe with seconds > 0: completes once its control steps have run (or the episode ended)."""
+
+ def __init__(self, job: _Job, rec: Recorder, start: int, steps: int, cameras: list[str], obs_id: int):
+ self.job, self.rec, self.cameras, self.obs_id = job, rec, cameras, obs_id
+ self.start, self.end = start, start + steps
+
+
+class _Views:
+ """The simulator as the Recorders see it: the cameras are rendered, and the proprioception read, at most once per
+ control step, however many requests and watches ask (each gets that step's arrays; nobody writes into them). The
+ state they show changes only with a control step or a reset, which is what the cache is keyed by."""
+
+ def __init__(self, svc: "RealtimeService"):
+ self._svc, self._sim = svc, svc.sim
+ self._key, self._rgb, self._dep, self._prop = None, {}, {}, None
+
+ def __getattr__(self, name):
+ return getattr(self._sim, name)
+
+ def _fresh(self) -> None:
+ key = (self._svc.journal_rows, self._svc.resets)
+ if key != self._key:
+ self._key, self._rgb, self._dep, self._prop = key, {}, {}, None
+
+ def images(self, cameras, depth):
+ self._fresh()
+ want = depth and set(self._sim.depth_cameras)
+ need = [c for c in cameras if c not in self._rgb or (want and c in want and c not in self._dep)]
+ if need:
+ rgb, dep = self._sim.images(list(dict.fromkeys(need)), depth)
+ self._rgb.update(rgb)
+ self._dep.update(dep)
+ return {c: self._rgb[c] for c in cameras}, ({c: self._dep[c] for c in cameras if c in self._dep} if depth else {})
+
+ def proprio(self):
+ self._fresh()
+ if self._prop is None:
+ self._prop = self._sim.proprio()
+ return self._prop
+
+
+class RealtimeService(Service):
+ """rcb_service.Service with the realtime clock (module docstring)."""
+
+ def __init__(self, sim, *, slowdown: float | None = None, **kw):
+ self.slowdown = float(os.environ.get("RCB_SLOWDOWN") or 1.0) if slowdown is None else float(slowdown)
+ if not self.slowdown > 0:
+ raise ValueError("RCB_SLOWDOWN must be > 0")
+ self.wall_dt = sim.control_dt * self.slowdown
+ self.max_ahead = int(round(float(kw.get("max_chunk_s", 10.0)) / sim.control_dt)) # how far ahead `at` may be
+ self.plan: collections.deque = collections.deque() # (type, row, call id) per coming control step, or HELD
+ self.running = False
+ self.anchor_t = self.anchor_n = 0 # the clock: control step anchor_n was due at anchor_t
+ self.watches: list[_Watch] = []
+ self.q: queue.Queue = queue.Queue()
+ self.max_lag = self.lag = 0.0
+ self.stalls = self.held_steps = self.late_steps = 0
+ self.stalled_s = 0.0 # wall time the stalls dropped
+ self.clock_s = 0.0 # wall time the clock ran, up to its last start
+ self._clock_t0 = 0.0 # ... and when that was
+ self.max_queue_wait = 0.0
+ self.requests_refused, self.request_bytes, self.request_busy_s = 0, 0, 0.0
+ self._ticks: list[tuple[float, float]] = [] # (wall time, call id) per journal row not yet on disk
+ super().__init__(sim, **kw)
+ self.ticks_path = os.path.join(self.data, "ticks.f64")
+ self.ckpt_every = max(1, int(round(CHECKPOINT_S / sim.control_dt)))
+ # the HTTP threads' side (front()): what they need of the simulator is read once, here
+ self.views = _Views(self)
+ self._interp = {t: bool(sim.interpolable(t)) for t in self.types}
+ self._max_rows = int((self.max_chunk_s + 1e-9) / sim.control_dt) # of any chunk that can be valid
+ self._max_elems = max([self._max_rows * int(v["dim"]) for v in self.types.values()] or [0])
+ self._spec_reply: tuple | None = None
+ self._t_seen: float | None = None # the first request's time, as its HTTP thread saw it
+ self._stat_lock, self._lane, self._child_lock = threading.Lock(), threading.Lock(), threading.Lock()
+ self._lane_free_at = 0.0
+ # the simulator thread's side (loop()): requests by kind -- True: it renders cameras, False: it does not
+ self._gate = {True: 0.0, False: 0.0} # behind the clock, none of the kind is served before this
+ self._render_gate = 0.0 # ... and no rendering request before this, behind or not
+ self._held: collections.deque = collections.deque() # requests taken from the queue before their time
+
+ # ------------------------------------------------------------------ spec, status, evidence
+ def _spec(self) -> dict:
+ spec = super()._spec()
+ spec["protocol"] = PROTOCOL
+ cdt = self.sim.control_dt
+ spec["timing"] = {
+ "realtime": True, "slowdown": self.slowdown, "control_dt": cdt,
+ "wall_s_per_control_step": cdt * self.slowdown,
+ "clock": "an episode starts paused; its clock starts at its first submit or observe with seconds > 0 (also "
+ "after a reset) and then runs one control step every wall_s_per_control_step seconds, whatever "
+ "you do, until the episode ends",
+ "submit": "answers at once: the chunk (rows dt apart, linear or hold) becomes the plan from control step "
+ "`at` (default: the next one), replacing what was planned from there; rows for steps already "
+ "gone are dropped (meta.late_steps); after the plan's last row that target is held",
+ "observe": "seconds = 0: now, at once; seconds > 0: watches the running episode that long and returns what "
+ "the cameras saw; never changes the plan",
+ "max_ahead_steps": self.max_ahead,
+ "no_images": "cameras = [] returns no images, only proprioception (the fastest call)"}
+ return spec
+
+ def _realtime_counters(self) -> dict:
+ return {"running": self.running, "planned_steps": len(self.plan), "lag_s": round(self.lag, 4)}
+
+ def status(self) -> dict:
+ return {**super().status(), **self._realtime_counters()}
+
+ def meta(self) -> dict:
+ return {**super().meta(), **self._realtime_counters()}
+
+ def _clock_seconds(self) -> float:
+ """Wall seconds the clock has run, over the whole trial."""
+ return self.clock_s + (time.monotonic() - self._clock_t0 if self.running else 0.0)
+
+ def _timing(self) -> dict:
+ """The counters a checkpoint keeps (and _resume restores)."""
+ return {"max_lag_s": self.max_lag, "stalls": self.stalls, "stalled_s": self.stalled_s,
+ "clock_s": self._clock_seconds(), "held_steps": self.held_steps, "late_steps": self.late_steps,
+ "max_queue_wait_s": self.max_queue_wait, "requests_refused": self.requests_refused,
+ "request_bytes": self.request_bytes, "request_busy_s": self.request_busy_s}
+
+ def _timing_record(self) -> dict:
+ clock = self._clock_seconds()
+ return {"protocol": PROTOCOL, "realtime": 1, "slowdown": self.slowdown, "max_lag_s": round(self.max_lag, 4),
+ "stalls": self.stalls, "stalled_s": round(self.stalled_s, 3), "clock_s": round(clock, 3),
+ "stalled_share": round(min(1.0, self.stalled_s / clock), 4) if clock > 0 else 0.0,
+ "held_steps": self.held_steps, "late_steps": self.late_steps,
+ "max_queue_wait_s": round(self.max_queue_wait, 4), "requests_refused": self.requests_refused,
+ "request_mb": round(self.request_bytes / 1e6, 3), "request_busy_s": round(self.request_busy_s, 3)}
+
+ def _write_final(self) -> None:
+ self.verdict.update(self._timing_record()) # final.json merges the verdict; none of these keys is a verdict's
+ if self.reason == "error": # a simulator error ended the graded episode: infrastructure (§4b)
+ self.verdict["infra_error"] = 1
+ else: # (an episode that erred and was reset away is not graded)
+ self.verdict.pop("infra_error", None)
+ super()._write_final()
+
+ def _checkpoint(self, ev: str, **kw) -> None:
+ if self._ticks: # the tick log first: a crash in between leaves it too long
+ with open(self.ticks_path, "ab") as f:
+ f.write(np.asarray(self._ticks, dtype=np.float64).tobytes())
+ self._ticks = []
+ super()._checkpoint(ev, **kw, timing=self._timing())
+
+ def _save_call(self, meta: dict, actions: np.ndarray | None) -> None:
+ keep = ("type", "n_rows", "dt", "interpolate", "at", "start_step", "late_steps", "planned_steps", "seconds",
+ "control_steps", "episode", "done", "reason", "n_clipped", "lag_s", "note")
+ line = {"call": self.calls, "ts": time.time(), **{k: meta[k] for k in keep if k in meta}}
+ with open(os.path.join(self.data, "calls.jsonl"), "a") as f:
+ f.write(json.dumps(line) + "\n")
+ if actions is not None:
+ rows = np.concatenate([np.full((len(actions), 1), float(self.calls)), actions], axis=1)
+ with open(os.path.join(self.data, "submitted.f64"), "ab") as f:
+ f.write(np.asarray(rows, dtype=np.float64).tobytes())
+
+ # ------------------------------------------------------------------ start, clock
+ def start(self) -> None:
+ super().start()
+ rows = np.fromfile(self.ticks_path, dtype=np.float64) if os.path.exists(self.ticks_path) else np.zeros(0)
+ n = self.journal_rows
+ if rows.size > 2 * n: # the steps a crash lost from the journal go from the tick log too
+ rows[: 2 * n].tofile(self.ticks_path)
+ elif rows.size < 2 * n: # (an older tick log, short of the journal: pad it as unknown)
+ np.concatenate([rows, np.full(2 * n - rows.size, np.nan)]).tofile(self.ticks_path)
+ self.running = False # a resumed episode waits, paused, for the agent's next call
+
+ def _resume(self, ev: list[dict]) -> None:
+ """Service._resume, with the realtime counters as of the checkpoint the episode resumes from (what the lost
+ steps had added goes with them): final.json's stalls, stalled_s, clock_s, held_steps, late_steps, max_lag_s,
+ max_queue_wait_s and the request counters cover the whole trial, not only the part after the last restart."""
+ cp = next((e for e in reversed(ev) if "state_sha" in e and "journal_rows" in e), {})
+ t = cp.get("timing") if isinstance(cp.get("timing"), dict) else {}
+ self.stalls, self.held_steps, self.late_steps, self.requests_refused, self.request_bytes = (
+ int(t.get(k) or 0) for k in ("stalls", "held_steps", "late_steps", "requests_refused", "request_bytes"))
+ self.max_lag, self.max_queue_wait, self.stalled_s, self.clock_s, self.request_busy_s = (
+ float(t.get(k) or 0.0) for k in ("max_lag_s", "max_queue_wait_s", "stalled_s", "clock_s", "request_busy_s"))
+ super()._resume(ev)
+
+ def _start_clock(self) -> None:
+ if not self.running and not self.done and not self.stopping:
+ self.running, self.anchor_t, self.anchor_n = True, time.monotonic(), len(self.episode)
+ self._clock_t0 = self.anchor_t
+ self.log("clock_start", control_steps=len(self.episode), episode=self.resets + 1)
+
+ def _stop_clock(self, why: str) -> None:
+ if self.running:
+ self.clock_s = self._clock_seconds()
+ self.running = False
+ self.log("clock_stop", why=why, control_steps=len(self.episode), max_lag_s=round(self.max_lag, 4),
+ stalls=self.stalls, stalled_s=round(self.stalled_s, 3), clock_s=round(self.clock_s, 3))
+ self.plan.clear()
+
+ def _end(self, reason: str) -> None:
+ super()._end(reason)
+ self._stop_clock(reason) # pending watches are completed by whoever ended it, after this
+
+ def due_in(self, now: float) -> float | None:
+ """Seconds until the next control step is due (<= 0: due now); None while the clock is stopped."""
+ if not self.running or self.done or self.stopping:
+ return None
+ return self.anchor_t + (len(self.episode) - self.anchor_n) * self.wall_dt - now
+
+ def tick(self) -> None:
+ """One control step of the clock: the plan's row for it, or the held target."""
+ now = time.monotonic()
+ lag = -(self.due_in(now) or 0.0)
+ self.max_lag = max(self.max_lag, lag)
+ if lag > MAX_LAG_S: # far behind: re-anchor rather than rush the robot through the debt
+ self.stalls += 1
+ self.stalled_s += lag # the robot stood still this long while its clock should have run
+ self.log("stall", behind_s=round(lag, 3), control_steps=len(self.episode))
+ self.anchor_t, self.anchor_n, lag = now, len(self.episode), 0.0
+ self.lag = max(lag, 0.0)
+ typ, row, src = self.plan.popleft() if self.plan else HELD
+ if typ is None: # nothing planned for this step (the plan ran out, or a gap in it)
+ typ, row = self.sim.hold()
+ self.held_steps += 1
+ before = self.journal_rows
+ self._ticks.append((time.time(), float(src)))
+ try:
+ self._step(typ, row)
+ except Exception as exc: # noqa: BLE001 -- it ran and failed: the episode ends, as on hardware
+ traceback.print_exc()
+ if self.journal_rows == before:
+ self._ticks.pop()
+ self.log("error", path="clock", error=repr(exc)[:500])
+ self._end("error")
+ self._release_watches()
+ return
+ for w in list(self.watches):
+ w.rec.step()
+ if self.done or len(self.episode) >= w.end:
+ self._complete_watch(w)
+ if not self.done and len(self.episode) % self.ckpt_every == 0:
+ self._checkpoint("tick")
+
+ # ------------------------------------------------------------------ requests, off the clock (HTTP threads)
+ def first_request(self) -> None:
+ if self.t0 is None:
+ self.t0 = self._t_seen or time.time() # when its HTTP thread saw it (front), logged by this thread
+ self.log("first_request", t0=self.t0)
+
+ def _count(self, refused: int = 0, read: int = 0) -> None:
+ with self._stat_lock:
+ self.requests_refused += refused
+ self.request_bytes += read
+
+ def _refusal(self, exc: Refused) -> tuple:
+ self._count(refused=1)
+ return ("reply", *self._reply_json(exc.code, exc.body))
+
+ def front(self, method: str, path: str, n: int = 0, read=None) -> tuple:
+ """Everything about a request that needs no simulator, on the caller's thread (an HTTP thread): the route, the
+ body (`read(n)`, called only for a body that will be decoded), its decoding, and every check that does not
+ depend on the episode's state. Returns ("reply", status, content type, body) for a request that ends here --
+ every refusal but the ones only the episode's state can decide (409, an `at` too far ahead, too many frames) --
+ or (route, checked request) for the simulator thread."""
+ if self._t_seen is None:
+ with self._stat_lock:
+ self._t_seen = self._t_seen or time.time()
+ path = urlparse(path).path
+ if method == "GET" and path == "/spec": # fixed for the service's life
+ if self._spec_reply is None:
+ self._spec_reply = self._reply_json(200, self.spec)
+ return ("reply", *self._spec_reply)
+ if method == "GET" and path == "/status":
+ return ("/status", None)
+ if method != "POST" or path not in ("/submit", "/observe") + (("/reset",) if self.reset_budget else ()):
+ return self._refusal(Refused(404, {"error": "no such endpoint"}))
+ if n < 0:
+ return self._refusal(Refused(400, {"error": "bad Content-Length"}))
+ if n > MAX_BODY:
+ return self._refusal(Refused(413, {"error": f"request body over {MAX_BODY} bytes"}))
+ try:
+ if n <= LANE_BODY:
+ body = read(n) if n else b""
+ self._count(read=len(body))
+ return (path, self._prepare(path, body))
+ with self._lane: # large bodies: one at a time (also in memory: the next one is
+ wait = self._lane_free_at - time.monotonic() # read when this one is done), paced by what the last took
+ if wait > 0:
+ time.sleep(wait)
+ body = read(n)
+ self._count(read=len(body))
+ t = time.thread_time()
+ try:
+ return (path, self._prepare(path, body))
+ finally:
+ self._lane_free_at = time.monotonic() + LANE_PACE * (time.thread_time() - t)
+ except Refused as exc:
+ # After a body that cannot be read at all, what the simulator thread answers before it looks at a
+ # request's fields (409) comes first here too.
+ return self._refusal(exc if getattr(exc, "malformed", False) else self._state_refusal(path) or exc)
+
+ def _state_refusal(self, path: str) -> Refused | None:
+ """The refusal a request to `path` gets whatever it says, as things stand (read without the simulator thread:
+ a request that races the episode's end may get either answer, as it always could)."""
+ if path == "/submit" and (self.done or self.stopping):
+ return self._conflict()
+ if path == "/reset":
+ try:
+ self._reset_view({})
+ except Refused as exc:
+ return exc
+ return None
+
+ def _prepare(self, path: str, body: bytes):
+ try:
+ meta, actions = self._decode_rt(body)
+ except Exception as exc: # noqa: BLE001 -- any malformed body is the caller's error
+ refused = Refused(400, {"error": f"malformed body: {exc}"[:300]})
+ refused.malformed = True
+ raise refused from None
+ if path == "/submit":
+ return self._check_submit(meta, actions)
+ if path == "/observe":
+ return self._check_observe(meta)
+ with bad_request(): # /reset: its view; the budget is the simulator thread's to check
+ self._view(meta, "frames")
+ return meta
+
+ def _decode_rt(self, body: bytes) -> tuple[dict, object]:
+ """Service._decode for this service: (meta, the actions as sent -- the RCB1 part `actions`, else meta's -- or
+ None). Only the part the service reads is decoded. More than INLINE_JSON bytes of JSON text go to the child."""
+ actions, part = None, None
+ if body[:4] == rcb1.MAGIC:
+ (n,) = struct.unpack("<I", body[4:8])
+ if n > INLINE_JSON:
+ return self._decode_child(body)
+ header = json.loads(body[8:8 + n].decode())
+ meta, off = header.get("meta", {}), 8 + n
+ for p in header["parts"]:
+ if p["name"] == "actions":
+ part = (p, off)
+ off += p["len"]
+ else:
+ if len(body) > INLINE_JSON:
+ return self._decode_child(body)
+ meta = json.loads(body or b"{}")
+ if not isinstance(meta, dict):
+ raise ValueError("the body must be a JSON object or an RCB1 frame")
+ if part is not None:
+ p, off = part
+ blob = body[off:off + p["len"]]
+ if p["kind"] == "array":
+ actions = np.frombuffer(blob, dtype=np.dtype(p["dtype"])).reshape(p["shape"])
+ elif p["kind"] == "json":
+ if len(blob) > INLINE_JSON:
+ return self._decode_child(body)
+ actions = json.loads(blob.decode())
+ else:
+ actions = blob
+ else:
+ actions = meta.get("actions")
+ return meta, actions
+
+ def _decode_child(self, body: bytes) -> tuple[dict, object]:
+ """_decode_rt by the child process (_CHILD): the same answer, without holding this process's interpreter lock
+ for the time a large JSON text takes to parse. One at a time."""
+ with self._child_lock:
+ try:
+ env = {**os.environ, "OMP_NUM_THREADS": "1", "OPENBLAS_NUM_THREADS": "1", "MKL_NUM_THREADS": "1"}
+ out = subprocess.run([sys.executable, "-c", _CHILD, str(self._max_elems), str(INLINE_JSON),
+ str(CHILD_MEM)], input=body, stdout=subprocess.PIPE, stderr=subprocess.DEVNULL,
+ timeout=CHILD_TIMEOUT_S, env=env).stdout
+ (n,) = struct.unpack("<I", out[:4])
+ head, blob = json.loads(out[4:4 + n].decode()), out[4 + n:]
+ except Exception as exc: # noqa: BLE001 -- no child, no answer: refused; never parsed on the clock's lock
+ raise ValueError(f"more than {INLINE_JSON} bytes of JSON text, which this service could not parse "
+ f"({type(exc).__name__}): send the rows as an RCB1 array part") from None
+ if "error" in head:
+ raise ValueError(head["error"])
+ d = head["actions"]
+ if d["kind"] == "array":
+ actions = np.frombuffer(blob, dtype=np.dtype(d["dtype"])).reshape(d["shape"])
+ elif d["kind"] == "shape": # more elements than any action type takes: its shape is all that
+ shape = tuple(int(x) for x in d["shape"]) # matters (it is refused for it), so none of them came back
+ actions = np.lib.stride_tricks.as_strided(np.zeros(1), shape=shape, strides=(0,) * len(shape),
+ writeable=False)
+ else:
+ actions = None if d["kind"] == "none" else b"?" # b"?": not numbers (refused as such)
+ return head["meta"], actions
+
+ def _check_submit(self, meta: dict, actions) -> dict:
+ """/submit's checks that do not depend on the episode's state; the chunk, ready for the simulator thread."""
+ with bad_request():
+ typ = meta.get("type")
+ if not isinstance(typ, str) or typ not in self.types:
+ raise ValueError(f"type must be one of {sorted(self.types)}")
+ dim = int(self.types[typ]["dim"])
+ cdt = self.sim.control_dt
+
+ def shape_ok(a) -> None: # before any pass over the rows: a refusal costs nothing
+ if a.ndim != 2 or a.shape[0] < 1 or a.shape[1] != dim:
+ raise ValueError(f"{typ} actions must be (T, {dim}) with T >= 1, got {list(a.shape)}")
+ if a.shape[0] > self._max_rows:
+ raise ValueError(f"the chunk covers {a.shape[0] * cdt:g} s: "
+ f"more than max_chunk_s = {self.max_chunk_s:g}")
+
+ if isinstance(actions, np.ndarray):
+ shape_ok(actions)
+ elif isinstance(actions, list) and len(actions) > self._max_rows:
+ raise ValueError(f"the chunk covers {len(actions) * cdt:g} s: "
+ f"more than max_chunk_s = {self.max_chunk_s:g}")
+ try:
+ A = np.asarray(actions, dtype=np.float64)
+ except (TypeError, ValueError, KeyError):
+ raise ValueError("send the actions as an RCB1 array part `actions` or a JSON list of rows") from None
+ shape_ok(A)
+ if not np.all(np.isfinite(A)):
+ raise ValueError("actions contain NaN or inf")
+ try:
+ dt = cdt if meta.get("dt") is None else float(meta["dt"])
+ except (TypeError, ValueError):
+ raise ValueError("dt must be a number of seconds") from None
+ k = round(dt / cdt) if np.isfinite(dt / cdt) else 0
+ if k < 1 or abs(k * cdt - dt) > 1e-9:
+ raise ValueError(f"dt must be a positive whole multiple of control_dt = {cdt} s (never rounded)")
+ if k > 1 and not self._interp[typ]:
+ raise ValueError(f"{typ}: one row is one control step, so dt must be control_dt = {cdt} s")
+ if len(A) * k * cdt > self.max_chunk_s + 1e-9:
+ raise ValueError(f"the chunk covers {len(A) * k * cdt:g} s: "
+ f"more than max_chunk_s = {self.max_chunk_s:g}")
+ interp = meta.get("interpolate", "linear")
+ if interp not in ("linear", "hold"):
+ raise ValueError("interpolate must be linear or hold")
+ at = meta.get("at")
+ if at is not None:
+ if isinstance(at, bool) or not isinstance(at, (int, float)) or at != int(at) or at < 0:
+ raise ValueError("at must be a control step of this episode (an integer >= 0), "
+ "or null for the next one")
+ at = int(at)
+ return {"type": typ, "actions": A, "dt": dt, "k": k, "interpolate": interp, "at": at,
+ "view": self._view_rt(meta, "none"), "note": str(meta.get("note") or "")[:200]}
+
+ def _check_observe(self, meta: dict) -> dict:
+ with bad_request():
+ try:
+ seconds = float(meta.get("seconds") or 0.0)
+ except (TypeError, ValueError):
+ raise ValueError("seconds must be a number") from None
+ if not 0.0 <= seconds <= self.max_watch_s:
+ raise ValueError(f"seconds must be in [0, {self.max_watch_s:g}]")
+ return {"seconds": seconds, "view": self._view_rt(meta, "frames")}
+
+ # ------------------------------------------------------------------ requests (simulator thread)
+ def dispatch(self, job: _Job) -> None:
+ """Serve one queued request; a watch finishes later, from tick(). A job that front() has not seen (handed over
+ whole: tests, tools) is taken through it here."""
+ self.max_queue_wait = max(self.max_queue_wait, time.monotonic() - job.queued)
+ try:
+ if not job.req:
+ n = len(job.body) if job.content_length is None else job.content_length
+ job.req = self.front(job.method, job.path, n, lambda _n: job.body)
+ self.first_request()
+ route = job.req[0]
+ if route == "reply":
+ return job.finish("raw", *job.req[1:])
+ if route == "/status":
+ return job.finish("raw", *self._reply_json(200, self.status()))
+ try:
+ {"/submit": self._rt_submit, "/observe": self._rt_observe,
+ "/reset": self._rt_reset}[route](job.req[1], job)
+ except Refused as exc: # 409, or a 400 only the episode's state decides: nothing changed
+ self._count(refused=1)
+ job.finish("raw", *self._reply_json(exc.code, exc.body))
+ except Exception as exc: # noqa: BLE001 -- it ran and failed: the episode ends, as on hardware
+ traceback.print_exc()
+ self.log("error", path=route, error=repr(exc)[:500])
+ self._end("error")
+ self._release_watches()
+ if not job.done.is_set():
+ job.finish("raw", *self._reply_json(500, {"error": "simulator error", "detail": repr(exc)[:200]}))
+ except Exception as exc: # noqa: BLE001 -- never leave an HTTP thread waiting forever
+ traceback.print_exc()
+ if not job.done.is_set():
+ job.finish("raw", *self._reply_json(500, {"error": "service error", "detail": repr(exc)[:200]}))
+
+ def _view_rt(self, meta: dict, default_video: str) -> tuple[list[str], int, str, bool]:
+ """2.1's _view, plus cameras = [] for no images."""
+ none = isinstance(meta.get("cameras"), list) and not meta["cameras"]
+ cams, fps, video, depth = self._view(meta, default_video)
+ return ([], fps, "none", False) if none else (cams, fps, video, depth)
+
+ def _in_force(self, step: int, nxt: int, typ: str) -> np.ndarray | None:
+ """The `typ` row in force at control step `step` (>= nxt - 1) as things stand: the plan's row for it, else the
+ plan's last row before it (held); with nothing planned up to there, or a step already run, the simulator's
+ current command of that type. None when the row in force there is of ANOTHER action type: rows of different
+ types have different layouts and are never combined, and what that row leaves in force for `typ` is not
+ known before it runs."""
+ for i in range(min(step - nxt, len(self.plan) - 1), -1, -1):
+ t, row, _ = self.plan[i]
+ if t is not None: # (a held step leaves in force what was before it)
+ return np.asarray(row, dtype=np.float64) if t == typ else None
+ return np.asarray(self.sim.command(typ), dtype=np.float64)
+
+ def _rt_submit(self, req: dict, job: _Job) -> None:
+ """A checked chunk (_check_submit) into the plan: what depends on the episode's state, and the planning."""
+ if self.done or self.stopping:
+ raise self._conflict()
+ nxt = len(self.episode)
+ typ, A, dt, k, interp, at = (req[x] for x in ("type", "actions", "dt", "k", "interpolate", "at"))
+ cams, fps, video, depth = req["view"]
+ if at is not None and at - nxt > self.max_ahead:
+ raise Refused(400, {"error": f"at = {at} is more than max_ahead_steps = {self.max_ahead} control steps "
+ f"ahead of the next one ({nxt})"})
+ at = nxt if at is None else at
+ rows, clipped = self.sim.clip(typ, A)
+ late = max(0, nxt - at)
+ keep = at - nxt if at > nxt else 0
+ try:
+ with bad_request(): # planning runs nothing: whatever fails here changed nothing (400)
+ rows = np.asarray(rows, dtype=np.float64)
+ target = self._interp[typ] # a target can be interpolated and held; a command cannot
+ held = self._in_force(at - 1, nxt, typ) if target else None
+ if held is not None and held.shape != rows.shape[1:]:
+ held = None # (an adapter whose command() has not the type's layout)
+ if target and interp == "linear": # linear starts from the target in force just before `at`
+ prev = np.asarray(self.sim.command(typ), dtype=np.float64) if held is None else held
+ if rows.ndim != 2 or prev.shape != rows.shape[1:]:
+ raise ValueError(f"{typ}: cannot interpolate rows of shape {list(rows.shape[1:])} from the "
+ f"command in force, of shape {list(prev.shape)}; send interpolate='hold'")
+ steps = []
+ for row in rows:
+ steps += [prev + (row - prev) * (j / k) for j in range(1, k + 1)]
+ prev = row
+ else: # hold; and a command stands for itself, one row one control step
+ steps = [row for row in rows for _ in range(k)]
+ head = list(self.plan)[:keep]
+ if len(head) < keep:
+ # A gap before `at`. A target of this type in force there is held through it (rows of call -1).
+ # Anything else -- a command (repeating it would issue it again at every step of the gap), or a
+ # plan that holds another action type there -- leaves the gap to the simulator's hold row, as
+ # when the plan has run out.
+ head += [HELD if held is None else (typ, held, -1)] * (keep - len(head))
+ except Refused as exc: # the request itself passed its checks: keep it on the record
+ self.log("plan_refused", type=typ, at=at, error=str(exc)[:300])
+ raise
+ self.calls += 1
+ self.rows += len(A)
+ self.plan = collections.deque(head + [(typ, s, self.calls) for s in steps[late:]])
+ self.late_steps += min(late, len(steps))
+ self._start_clock()
+ rec = Recorder(self.views, cams, fps, depth, "none")
+ rec.start()
+ rec.finish()
+ out = {**self.meta(), "chunk_id": self.calls, "type": typ, "n_rows": int(len(A)), "dt": dt,
+ "interpolate": interp, "at": at, "start_step": max(at, nxt), "late_steps": late,
+ "plan_end_step": nxt + len(self.plan), "clipped": clipped[:200], "n_clipped": len(clipped), "ok": True}
+ self._save_call({**out, "note": req["note"]}, A)
+ job.finish("rec", out, rec)
+
+ def _rt_observe(self, req: dict, job: _Job) -> None:
+ with bad_request():
+ seconds, (cams, fps, video, depth) = req["seconds"], req["view"]
+ n = int(round(seconds / self.sim.control_dt)) if not (self.done or self.stopping) else 0
+ rec = Recorder(self.views, cams, fps, depth, video if n or video != "mp4" else "frames") # no time: no video
+ self._frame_budget(rec, n)
+ self.calls += 1
+ rec.start()
+ if not n:
+ rec.finish()
+ out = {**self.meta(), "obs_id": self.calls, "seconds": 0.0, "control_steps_watch": 0, "ok": True}
+ self._save_call(out, None)
+ return job.finish("rec", out, rec)
+ self._start_clock()
+ self.watches.append(_Watch(job, rec, len(self.episode), n, cams, self.calls))
+
+ def _complete_watch(self, w: _Watch) -> None:
+ self.watches.remove(w)
+ w.rec.finish()
+ ran = max(0, min(len(self.episode), w.end) - w.start)
+ out = {**self.meta(), "obs_id": w.obs_id, "seconds": round(ran * self.sim.control_dt, 4),
+ "control_steps_watch": ran, "ok": True}
+ self._save_call(out, None)
+ w.job.finish("rec", out, w.rec)
+
+ def _release_watches(self) -> None:
+ for w in list(self.watches):
+ self._complete_watch(w)
+
+ def _rt_reset(self, meta: dict, job: _Job) -> None:
+ self._reset_view(meta) # a refused reset (409, 400) changes nothing: the watches go on
+ self._release_watches()
+ out, prts = self._reset(meta, {}) # 2.1's reset: abandons the episode, a new one (paused)
+ self._stop_clock("reset")
+ job.finish("rcb1", {**out, **self._realtime_counters()}, prts)
+
+ # ------------------------------------------------------------------ the loop (simulator thread) and HTTP
+ def poll(self) -> None:
+ """After SIGUSR1: end the episode, answer the pending watches, finalize (Service.poll, plus the watches)."""
+ if self.stopping and not self.finalized:
+ if not self.done:
+ self._end("timeout")
+ self._release_watches()
+ self._finalize()
+
+ @staticmethod
+ def _renders(job: _Job) -> bool:
+ """Does serving this request render cameras? (What a request without images costs is bounded by its checks.)"""
+ route, req = (job.req + (None, None))[:2]
+ if route in ("/submit", "/observe"):
+ return bool(req["view"][0])
+ return route == "/reset" and not (isinstance(req.get("cameras"), list) and not req["cameras"])
+
+ def _may_serve(self, job: _Job, behind: bool, now: float) -> bool:
+ if not self.running or self.done or self.stopping:
+ return True # no clock to keep
+ kind = self._renders(job)
+ return not (kind and now < self._render_gate) and not (behind and now < self._gate[kind])
+
+ def _take(self, behind: bool, timeout: float) -> _Job | None:
+ """The next request to serve now, waiting up to `timeout` for one: the oldest held one whose time has come,
+ else the queue's next -- which is held instead, if its time has not come."""
+ now = time.monotonic()
+ for job in self._held:
+ if self._may_serve(job, behind, now):
+ self._held.remove(job)
+ return job
+ if self._held and self.running: # do not sleep past a held request's time
+ gates = [self._gate[k] for k in (True, False)] + [self._render_gate]
+ timeout = min([timeout] + [g - now for g in gates if g > now])
+ try:
+ job = self.q.get(timeout=timeout) if timeout > 0 else self.q.get_nowait()
+ except queue.Empty:
+ return None
+ if job.req and not self._may_serve(job, behind, time.monotonic()):
+ self._held.append(job)
+ return None
+ return job
+
+ def loop(self, until: threading.Event | None = None) -> None:
+ """The simulator thread: control steps when due, requests in between, the wrap-up after SIGUSR1."""
+ while until is None or not until.is_set():
+ self.poll()
+ if self.t0 is None and self._t_seen is not None:
+ self.first_request() # a request that ended on its HTTP thread started the trial's time
+ wait = self.due_in(time.monotonic())
+ if wait is not None and wait <= 0:
+ self.tick()
+ # Behind the clock still (the next step is due at once): the steps come first. A waiting request is
+ # served once the clock has caught up -- so requests queue behind each other, not into the clock --
+ # or, for a simulator that cannot keep up by itself, BEHIND_PACE times the last request's time after
+ # it (a cheap request: at once, one per step; a long render: a fifth of the time at most).
+ wait = self.due_in(time.monotonic())
+ job = self._take(wait is not None and wait <= 0, 0.0)
+ else:
+ job = self._take(False, 0.25 if wait is None else min(wait, 0.25))
+ if job is None:
+ continue
+ kind, t = bool(job.req) and self._renders(job), time.monotonic()
+ self.dispatch(job)
+ now = time.monotonic()
+ self.request_busy_s += now - t
+ self._gate[kind] = now + BEHIND_PACE * (now - t)
+ if kind and now - t > MAX_LAG_S: # a stall by itself: the next render waits, clock on time or not
+ self._render_gate = self._gate[kind]
+
+ def submit_job(self, job: _Job) -> tuple[int, str, bytes]:
+ """From an HTTP thread: queue the request, wait for the simulator thread, encode the answer here."""
+ self.q.put(job)
+ job.done.wait()
+ kind, *rest = job.reply
+ if kind == "raw":
+ return tuple(rest)
+ meta, payload = rest # a Recorder's parts are encoded here, off the simulator's thread
+ return self._reply_rcb1(meta, payload.parts() if kind == "rec" else payload)
+
+ def http_server(self, address: tuple[str, int]) -> ThreadingHTTPServer:
+ svc = self
+
+ class Handler(BaseHTTPRequestHandler):
+ timeout = 120
+
+ def log_message(self, *a):
+ pass
+
+ def _send(self, code: int, ctype: str, body: bytes) -> None:
+ self.send_response(code)
+ self.send_header("Content-Type", ctype)
+ self.send_header("Content-Length", str(len(body)))
+ self.end_headers()
+ self.wfile.write(body)
+
+ def do_GET(self):
+ self._serve("GET", 0)
+
+ def do_POST(self):
+ try:
+ n = int(self.headers.get("Content-Length") or 0)
+ except ValueError:
+ n = -1
+ if not 0 <= n <= MAX_BODY: # refused unread: drop the connection with it
+ self.close_connection = True
+ self._serve("POST", n)
+
+ def _serve(self, method: str, n: int) -> None:
+ try:
+ req = svc.front(method, self.path, n, self.rfile.read) # all that needs no simulator: here
+ except OSError: # the client went away mid-body
+ raise
+ except Exception as exc: # noqa: BLE001 -- never an HTTP thread's traceback instead of an answer
+ traceback.print_exc()
+ req = ("reply", *svc._reply_json(500, {"error": "service error", "detail": repr(exc)[:200]}))
+ self._send(*(req[1:] if req[0] == "reply" else svc.submit_job(_Job(method, self.path, req=req))))
+
+ class Server(ThreadingHTTPServer):
+ daemon_threads = True
+
+ return Server(address, Handler)
+
+ def serve(self, port: int = 8000) -> None:
+ import signal
+ signal.signal(signal.SIGUSR1, self.on_sigusr1)
+ srv = self.bind(port)
+ sys.setswitchinterval(SWITCH_INTERVAL_S) # the HTTP threads give the simulator thread its turn quickly
+ threading.Thread(target=srv.serve_forever, name="http", daemon=True).start()
+ print(f"{rcb_service.PROTOCOL} sim service (realtime, slowdown {self.slowdown:g}) ready on :{port}", flush=True)
+ self.loop()
+
+
+def main(build, port: int = 8000) -> None:
+ """rcb_service.main -- the start count, back-off, startup watchdog, resume -- with this service in place of
+ Service (RCB_SLOWDOWN: 1 = real time), and eai-standard/2.3 in everything it logs and writes."""
+ rcb_service.Service = RealtimeService
+ rcb_service.PROTOCOL = PROTOCOL
+ rcb_service.main(build, port)
+
+
+if __name__ == "__main__":
+ sys.path.insert(0, os.path.dirname(os.path.abspath(__file__)))
+ import server # noqa: E402 -- the HumanoidBench adapter; picks the GL backend and restores the reward on import
+
+ main(server.build)
diff --git a/tasks/myosuite-table-tennis-i00-standard-realtime/environment/sim/rcb_service.py b/tasks/myosuite-table-tennis-i00-standard-realtime/environment/sim/rcb_service.py
new file mode 100644
index 0000000000..5b31ed7428
--- /dev/null
+++ b/tasks/myosuite-table-tennis-i00-standard-realtime/environment/sim/rcb_service.py
@@ -0,0 +1,835 @@
+"""eai-standard/2.1 sim service: the simulator-agnostic half of a standard-mode sidecar (docs/PROTOCOL.md §4,
+docs/STANDARD_MODE.md).
+
+The robot as a service: the agent reaches a simulator it cannot open over HTTP. GET /spec, GET /status, POST /observe
+and POST /submit; POST /reset exists only when the run grants resets (budget.reset > 0); everything else is 404.
+Responses that carry data are RCB1 frames (rcb1.py), never base64. Everything simulator-specific comes in through an
+adapter (the `Sim` contract below), so this file imports no simulator: a sidecar is its benchmark's adapter plus this.
+
+One thread does all of it -- the HTTP requests, the SIGUSR1 wrap-up, the finalize -- because a simulator's GL context
+belongs to the thread that made it (MuJoCo renders black frames from any other).
+
+The episode starts when the service is ready. It ends on success, by the task's own rule ("failed": the robot fell,
+the episode's control steps ran out), on a simulator error ("error"), or at SIGUSR1 ("timeout": the collect hook, once
+the agent's container is gone). With a reset budget, POST /reset abandons the running episode and starts another from
+the frozen instance; only the last episode is graded, and a success ends the trial.
+
+A request is checked in full before anything runs: whatever is wrong with it is a 400 and nothing ran. Once it runs,
+any failure is the simulator's: a 500, and the episode ends ("error").
+
+Evidence under DATA, never visible to the agent:
+ episode.jsonl one line per event; every state the service may resume from is a checkpoint (it carries
+ state_sha and journal_rows)
+ trajectory_live.f64 the journal: every executed control step in the simulator's own action space, float64
+ calls/NNNNNN/ each call's response meta and the actions as sent
+ startups.jsonl one line per process start
+ final.json kept current (written atomically); "finalized": true once the verdict is in -- before the
+ video is rendered ("video_pending": true while it is), so the video never decides anything
+ trajectory.npz, replay.mp4, replay.last.png the graded episode and its fresh replay
+
+Crash recovery: compose restarts a crashed service (restart: on-failure). It replays the journal up to the last
+checkpoint, checks the state hash, and continues the SAME episode -- never a reset. A call in flight is lost as a
+whole (its caller got an error, not a result, so nothing was learned). More than RCB_MAX_RESUMES restarts since the
+episode began (every process start counts, also one that died building the scene or replaying the journal) end the
+trial: sim_crashed_on_call if every crash happened while serving a call (the agent's failure), else sim_crashed
+(infrastructure, infra_error = 1). A verdict already on disk is never overwritten; a crash, a hang or a slow render
+of the video loses only the video.
+"""
+from __future__ import annotations
+
+import contextlib
+import errno
+import faulthandler
+import json
+import os
+import signal
+import tempfile
+import threading
+import time
+import traceback
+from http.server import BaseHTTPRequestHandler, HTTPServer
+from typing import Callable
+from urllib.parse import urlparse
+
+import numpy as np
+
+import rcb1
+
+PROTOCOL = "eai-standard/2.1"
+MAX_BODY = 64 * 1024 * 1024 # request bodies above this are refused (413)
+MAX_RESPONSE = 256 * 1024 * 1024 # responses above this drop frames/ and depth/ (meta.degraded says so)
+MAX_VIDEO_FPS = 30
+MAX_FRAMES_PER_CALL = 3000 # images in one response (cameras x frames, depth counted again): bounds memory
+VIDEOS = ("mp4", "frames", "both", "none")
+ON_SERVICE_ERROR = ("if a call fails with a connection error or a 5xx, the service may have crashed and resumed the "
+ "same episode (never a reset): the state is the one after the last call that completed on the "
+ "service, which you may not have received; check GET /status before continuing")
+
+
+class Sim:
+ """The adapter contract, duck-typed (this class only documents it). One object = one frozen scene.
+
+ Rows come in three forms: what the agent sends (an action of one of spec()["action_types"]), the same after
+ clip(), and a "sim row" (to_sim()), the action in the simulator's own space, which step() executes and the
+ journal and the graded trajectory record (raw_dim floats)."""
+ control_dt: float # seconds per control step; a chunk's dt is a whole multiple of it
+ max_control_steps: int # the task's episode length: the episode ends there
+ cameras: tuple[str, ...] # every camera the agent may ask for
+ depth_cameras: tuple[str, ...] # the ones that also give depth
+ proprio_layout: list[str] # the proprio row's columns
+ raw_dim: int # a sim row's width
+
+ def spec(self) -> dict: ... # simulator, robot, task_text, episode, action_types, cameras, robot facts
+ def interpolable(self, type: str) -> bool: ... # may a row of this type span several control steps?
+ def images(self, cameras: list[str], depth: bool) -> tuple[dict, dict]: ... # ({cam: uint8 HxWx3}, {cam: float32 HxW m})
+ def proprio(self) -> np.ndarray: ... # float32, one value per proprio_layout column
+ def clip(self, type: str, rows: np.ndarray) -> tuple[np.ndarray, list]: ... # into the limits; [{row, index, reason}]
+ def to_sim(self, type: str, row: np.ndarray) -> np.ndarray: ... # an agent row -> a sim row
+ def command(self, type: str) -> np.ndarray: ... # the last row executed, exactly as clip() returned it: a chunk
+ # interpolates from it, so a chunk equals its rows sent one by one
+ def hold(self) -> tuple[str, np.ndarray]: ... # the row that holds the last command (/observe with seconds > 0)
+ def step(self, row: np.ndarray) -> tuple[bool, bool, dict]: ... # one control step -> (done, success, info)
+ def reset(self) -> None: ... # back to the frozen instance
+ def state_sha(self) -> str: ...
+ def verify(self, rows: np.ndarray) -> dict: ... # replay in a fresh simulator -> {"success", "state_sha"}
+ def save_trajectory(self, rows: np.ndarray, path: str, meta: dict) -> None: ...
+ def render_video(self, rows: np.ndarray, path: str) -> None: ... # path and path.last.png
+
+
+class GiveUp(Exception):
+ """The service will not run this trial's episode again; final.json says why."""
+
+
+class Refused(Exception):
+ """A request the service turns down with `code` and the JSON `body`; nothing was executed."""
+
+ def __init__(self, code: int, body: dict):
+ super().__init__(body.get("error"))
+ self.code, self.body = code, body
+
+
+def bad_request():
+ """Checking a request, before anything runs: whatever goes wrong here is the request's fault (400)."""
+ try:
+ yield
+ except Refused:
+ raise
+ except Exception as exc: # noqa: BLE001 -- a wrong type, value or shape of any field
+ raise Refused(400, {"error": str(exc)[:300] or type(exc).__name__}) from None
+
+
+def count_start(data: str) -> int:
+ """Record one process start in startups.jsonl; returns how many there have been."""
+ path = os.path.join(data, "startups.jsonl")
+ with open(path, "a") as f:
+ f.write(json.dumps({"ts": time.time(), "pid": os.getpid()}) + "\n")
+ with open(path) as f:
+ return sum(1 for _ in f)
+
+
+def write_json_atomic(path: str, obj: dict) -> None:
+ """Write then rename: the collect hook polling the file never reads half of it."""
+ fd, tmp = tempfile.mkstemp(dir=os.path.dirname(path), suffix=".tmp")
+ with os.fdopen(fd, "w") as f:
+ json.dump(obj, f)
+ os.replace(tmp, path)
+
+
+def encode_mp4(frames: np.ndarray, fps: float) -> bytes:
+ import imageio.v2 as imageio
+ fd, path = tempfile.mkstemp(suffix=".mp4")
+ os.close(fd)
+ try:
+ with imageio.get_writer(path, format="FFMPEG", mode="I", fps=fps, codec="libx264", macro_block_size=1,
+ output_params=["-movflags", "+faststart", "-pix_fmt", "yuv420p"]) as w:
+ for f in frames:
+ w.append_data(np.ascontiguousarray(f))
+ with open(path, "rb") as f:
+ return f.read()
+ finally:
+ os.unlink(path)
+
+
+class Recorder:
+ """What one call returns: a proprio row at its start, every `every` control steps (the requested video_fps, rounded to
+ whole control steps) and at its end; with them the cameras' frames when a video or the frames are asked for, else
+ only the last frame (rendering is most of a call's cost)."""
+
+ def __init__(self, sim, cameras: list[str], fps: int, depth: bool, video: str):
+ self.sim, self.cams, self.depth, self.video = sim, cameras, depth, video
+ self.every = max(1, round(1.0 / (sim.control_dt * fps)))
+ self.fps = 1.0 / (self.every * sim.control_dt)
+ self.images = video != "none"
+ self.frames = {c: [] for c in cameras}
+ self.dep = {c: [] for c in cameras}
+ self.prop: list[np.ndarray] = []
+ self.k, self.grabbed_at = 0, None
+
+ def images_per_call(self, steps: int) -> int:
+ per_frame = len(self.cams) + (len(set(self.cams) & set(self.sim.depth_cameras)) if self.depth else 0)
+ return (steps // self.every + 2 if self.images else 1) * per_frame
+
+ def start(self) -> None:
+ self._grab(self.images)
+
+ def step(self) -> None:
+ self.k += 1
+ if self.k % self.every == 0:
+ self._grab(self.images)
+
+ def finish(self) -> None:
+ if self.grabbed_at != self.k:
+ self._grab(self.images)
+ if not self.images:
+ self._grab(True, proprio=False) # the one frame a call without video returns: the last
+
+ def _grab(self, images: bool, proprio: bool = True) -> None:
+ if images:
+ rgb, dep = self.sim.images(self.cams, self.depth)
+ for c in self.cams:
+ self.frames[c].append(rgb[c])
+ if c in dep:
+ self.dep[c].append(dep[c])
+ if proprio:
+ self.prop.append(np.asarray(self.sim.proprio(), np.float32))
+ self.grabbed_at = self.k
+
+ def parts(self) -> list:
+ out = []
+ for c in self.cams:
+ fr = np.stack(self.frames[c])
+ if self.video in ("mp4", "both"):
+ out.append((f"video/{c}", "mp4", encode_mp4(fr, self.fps), {"fps": round(self.fps, 3), "n": int(len(fr))}))
+ if self.video in ("frames", "both"):
+ out.append(rcb1.array_part(f"frames/{c}", fr))
+ out.append(rcb1.array_part(f"last/{c}", fr[-1]))
+ if self.dep[c]:
+ out.append(rcb1.array_part(f"depth/{c}", np.stack(self.dep[c]).astype(np.float16)))
+ out.append(rcb1.array_part("proprio", np.stack(self.prop)))
+ return out
+
+
+class Service:
+ """Episode bookkeeping, evidence and the HTTP endpoints around one `Sim`."""
+
+ def __init__(self, sim, *, data: str = "/data", reset_budget: int = 0, wall_clock_s: float = 3600.0,
+ max_chunk_s: float = 10.0, max_watch_s: float = 30.0, max_frames_per_call: int = MAX_FRAMES_PER_CALL,
+ max_resumes: int = 2):
+ self.sim, self.data = sim, data
+ self.reset_budget, self.wall_clock_s = int(reset_budget), float(wall_clock_s)
+ self.max_chunk_s, self.max_watch_s = float(max_chunk_s), float(max_watch_s)
+ self.max_frames, self.max_resumes = int(max_frames_per_call), int(max_resumes)
+ os.makedirs(os.path.join(data, "calls"), exist_ok=True)
+ self.journal_path = os.path.join(data, "trajectory_live.f64")
+ self.spec = self._spec()
+ self.types = {k: v for k, v in self.spec["action_types"].items() if isinstance(v, dict)}
+ self.calls = self.rows = self.resets = 0
+ self.journal_rows = 0 # control steps executed over the whole trial (the journal's length)
+ self.ep_start = 0 # where the current episode begins in the journal
+ self.episode: list[np.ndarray] = [] # the current episode's sim rows: what finalize grades
+ self.done, self.reason, self.success = False, None, False
+ self.stopping = False # SIGUSR1 seen: the agent is gone
+ self.finalized, self.verdict = False, {}
+ self.t0 = self.t_end = None # the clock starts at the agent's first request
+ self.starts = 1
+ self._unwritten: list[np.ndarray] = [] # sim rows executed since the last checkpoint
+
+ # ------------------------------------------------------------------ spec and status
+ def _spec(self) -> dict:
+ own = self.sim.spec()
+ core = {"protocol": PROTOCOL, "control_dt": self.sim.control_dt, "max_control_steps": self.sim.max_control_steps,
+ "max_chunk_s": self.max_chunk_s, "max_watch_s": self.max_watch_s, "max_video_fps": MAX_VIDEO_FPS,
+ "max_frames_per_call": self.max_frames, "proprio_layout": list(self.sim.proprio_layout),
+ "budget": {"reset": self.reset_budget, "wall_clock_s": self.wall_clock_s},
+ "clock": "time_s and elapsed_s count from your first request to the service; time_left_s is "
+ "budget.wall_clock_s minus time_s",
+ "on_service_error": ON_SERVICE_ERROR}
+ if clash := sorted(set(core) & set(own)):
+ raise ValueError(f"the adapter's spec() must not set {clash}")
+ cams = {c: {"rgb": True, "depth": c in self.sim.depth_cameras, **own["cameras"].get(c, {})}
+ for c in self.sim.cameras}
+ spec = {**core, **own, "cameras": cams}
+ if self.reset_budget:
+ spec["reset"] = ("POST /reset abandons the running episode and starts a new one from the same starting "
+ "state, budget.reset times in all; only the last episode is graded, and a success ends "
+ "the trial")
+ return spec
+
+ def clock(self) -> float:
+ return 0.0 if self.t0 is None else time.time() - self.t0
+
+ def _counters(self) -> dict:
+ out = {"calls": self.calls, "rows": self.rows, "control_steps": len(self.episode), "done": self.done,
+ "reason": self.reason}
+ if self.reset_budget:
+ out.update(episode=self.resets + 1, control_steps_total=self.journal_rows,
+ resets_left=0 if self.success or self.stopping else self.reset_budget - self.resets)
+ return out
+
+ def status(self) -> dict:
+ return {**self._counters(), "elapsed_s": round(self.clock(), 1)}
+
+ def meta(self) -> dict:
+ t = self.clock()
+ return {**self._counters(), "sim_time_s": round(len(self.episode) * self.sim.control_dt, 4),
+ "time_s": round(t, 2), "time_left_s": max(0.0, round(self.wall_clock_s - t, 1)),
+ "proprio": {"layout": list(self.sim.proprio_layout),
+ "now": np.asarray(self.sim.proprio(), np.float32).tolist()}}
+
+ # ------------------------------------------------------------------ evidence
+ def log(self, ev: str, **kw) -> None:
+ with open(os.path.join(self.data, "episode.jsonl"), "a") as f:
+ f.write(json.dumps({"ts": time.time(), "ev": ev, **kw}) + "\n")
+
+ def _events(self) -> list[dict]:
+ try:
+ with open(os.path.join(self.data, "episode.jsonl")) as f:
+ return [json.loads(line) for line in f if line.strip()]
+ except OSError:
+ return []
+
+ def _checkpoint(self, ev: str, **kw) -> None:
+ """Log a state this service may resume from, after appending everything executed since the last one to the
+ journal. Rows of a call that never reaches its checkpoint are lost with it, as the call is."""
+ if self._unwritten:
+ with open(self.journal_path, "ab") as f:
+ f.write(np.asarray(self._unwritten, dtype=np.float64).tobytes())
+ self._unwritten = []
+ self.log(ev, **kw, calls=self.calls, rows=self.rows, resets=self.resets, ep_start=self.ep_start,
+ journal_rows=self.journal_rows, done=self.done, reason=self.reason, success=self.success,
+ state_sha=self.sim.state_sha())
+ self._write_final()
+
+ def _write_final(self) -> None:
+ end = self.t_end or time.time()
+ d = {"protocol": PROTOCOL, "success": False, "live_success": self.success, "done": self.done,
+ "reason": self.reason, "calls": self.calls, "rows": self.rows, "control_steps": len(self.episode),
+ "control_steps_total": self.journal_rows, "episodes": self.resets + 1, "resets": self.resets,
+ "reset_budget": self.reset_budget, "elapsed_s": round(end - self.t0, 1) if self.t0 else 0.0,
+ "sim_startups": self.starts, "finalized": self.finalized, **self.verdict}
+ if "deterministic" in self.verdict:
+ d["success"] = bool(self.success and self.verdict["replay_success"] and self.verdict["deterministic"])
+ write_json_atomic(os.path.join(self.data, "final.json"), d)
+
+ def _save_call(self, meta: dict, actions: np.ndarray | None) -> None:
+ d = os.path.join(self.data, "calls", f"{self.calls:06d}")
+ os.makedirs(d, exist_ok=True)
+ with open(os.path.join(d, "meta.json"), "w") as f:
+ json.dump(meta, f)
+ if actions is not None:
+ np.save(os.path.join(d, "actions.npy"), actions)
+
+ # ------------------------------------------------------------------ start, resume, give up
+ def start(self) -> None:
+ """A fresh episode, or -- after a crash -- the same one again. Raises GiveUp when it must not run again."""
+ try:
+ with open(os.path.join(self.data, "startups.jsonl")) as f:
+ self.starts = max(1, sum(1 for _ in f))
+ except OSError:
+ self.starts = 1
+ ev = self._events()
+ if not any(e["ev"] == "start" for e in ev):
+ open(self.journal_path, "wb").close()
+ self._checkpoint("start", protocol=PROTOCOL, spec=self.spec, startup=self.starts)
+ else:
+ self._resume(ev)
+ with open(os.path.join(self.data, "server.pid"), "w") as f:
+ f.write(str(os.getpid()))
+
+ def _resume(self, ev: list[dict]) -> None:
+ path = os.path.join(self.data, "final.json")
+ try:
+ with open(path) as f:
+ fin = json.load(f)
+ except (OSError, ValueError):
+ fin = {}
+ if fin.get("finalized") or "deterministic" in fin:
+ # Graded already. The process died rendering the video (video_pending; before 2026-10, a verdict not yet
+ # finalized): all it lost is the video.
+ if not fin.get("finalized") or fin.get("video_pending"):
+ write_json_atomic(path, {**fin, "finalized": True, "video": False, "video_pending": False})
+ self._give_up("restarted after finalize", infra=False)
+ restarts = self.starts - int(next(e for e in ev if e["ev"] == "start")["startup"])
+ last = max(i for i, e in enumerate(ev) if e["ev"] in ("start", "resuming"))
+ in_call = False # did the process that died do so while serving a call?
+ for e in ev[last:]:
+ in_call = {"call_begin": True, "call_end": False}.get(e["ev"], in_call)
+ in_calls = sum(bool(e["crashed_in_call"]) for e in ev if e["ev"] == "resuming") + int(in_call)
+ if restarts > self.max_resumes:
+ agent = in_calls == restarts # every crash happened inside a call: the agent's commands did it
+ self._give_up("sim_crashed_on_call" if agent else "sim_crashed", infra=not agent)
+ self.log("resuming", crashed_in_call=in_call, restart=restarts)
+ cp = [e for e in ev if "state_sha" in e and "journal_rows" in e][-1]
+ n, dim = int(cp["journal_rows"]), int(self.sim.raw_dim)
+ rows = np.fromfile(self.journal_path, dtype=np.float64) if os.path.exists(self.journal_path) else np.zeros(0)
+ if rows.size < n * dim:
+ self._give_up("resume_failed: journal shorter than the log", infra=True)
+ rows = rows[: n * dim].reshape(n, dim)
+ for row in rows[cp["ep_start"]:]:
+ self.sim.step(row)
+ if self.sim.state_sha() != cp["state_sha"]:
+ self._give_up("resume_failed: replayed state differs from the logged one", infra=True)
+ rows.tofile(self.journal_path) # drop the steps of the call that was in flight
+ self.calls, self.rows, self.resets = int(cp["calls"]), int(cp["rows"]), int(cp["resets"])
+ self.ep_start, self.journal_rows = int(cp["ep_start"]), n
+ self.episode = [r.copy() for r in rows[self.ep_start:]]
+ self.done, self.reason, self.success = bool(cp["done"]), cp["reason"], bool(cp["success"])
+ self.t0 = next((e["t0"] for e in ev if e["ev"] == "first_request"), None)
+ if self.done:
+ self.t_end = next((e["ts"] for e in reversed(ev) if e["ev"] == "end"), cp["ts"])
+ self._checkpoint("resumed", downtime_s=round(time.time() - float(ev[-1]["ts"]), 1))
+
+ def _give_up(self, reason: str, infra: bool) -> None:
+ """End the trial without running its episode again: mark it failed (a finalized final.json is kept as it is),
+ leave the PID file for the collect hook, and stop."""
+ self.log("restart_refused", reason=reason)
+ path = os.path.join(self.data, "final.json")
+ try:
+ with open(path) as f:
+ fin = json.load(f)
+ except (OSError, ValueError):
+ fin = {}
+ if not fin.get("finalized"):
+ fin.update({"protocol": PROTOCOL, "success": False, "done": True, "reason": reason, "finalized": True,
+ **({"infra_error": 1} if infra else {})})
+ write_json_atomic(path, fin)
+ with open(os.path.join(self.data, "server.pid"), "w") as f:
+ f.write(str(os.getpid()))
+ raise GiveUp(reason)
+
+ # ------------------------------------------------------------------ requests
+ def first_request(self) -> None:
+ if self.t0 is None:
+ self.t0 = time.time()
+ self.log("first_request", t0=self.t0)
+
+ def handle(self, method: str, path: str, body: bytes = b"", content_length: int | None = None) -> tuple[int, str, bytes]:
+ """One request -> (status, content type, body): the whole HTTP surface, without the HTTP."""
+ self.first_request()
+ path = urlparse(path).path
+ if method == "GET" and path in ("/spec", "/status"):
+ return self._reply_json(200, self.spec if path == "/spec" else self.status())
+ routes = {"/submit": self._submit, "/observe": self._observe}
+ if self.reset_budget:
+ routes["/reset"] = self._reset
+ if method != "POST" or path not in routes:
+ return self._reply_json(404, {"error": "no such endpoint"})
+ if (len(body) if content_length is None else content_length) > MAX_BODY:
+ return self._reply_json(413, {"error": f"request body over {MAX_BODY} bytes"})
+ try:
+ meta, parts = self._decode(body)
+ except Exception as exc: # noqa: BLE001 -- any malformed body is the caller's error
+ return self._reply_json(400, {"error": f"malformed body: {exc}"[:300]})
+ self.log("call_begin", path=path, call=self.calls + 1)
+ try:
+ out, blobs = routes[path](meta, parts)
+ except Refused as exc: # 400 (bad_request), 409: nothing ran
+ return self._reply_json(exc.code, exc.body)
+ except Exception as exc: # noqa: BLE001 -- it ran and failed: the episode ends, as on hardware
+ traceback.print_exc()
+ self.log("error", path=path, error=repr(exc)[:500])
+ self._end("error")
+ return self._reply_json(500, {"error": "simulator error", "detail": repr(exc)[:200]})
+ finally:
+ self.log("call_end", path=path)
+ return self._reply_rcb1(out, blobs)
+
+ @staticmethod
+ def _decode(body: bytes) -> tuple[dict, dict]:
+ if body[:4] == rcb1.MAGIC:
+ meta, parts = rcb1.decode(body)
+ else:
+ meta, parts = json.loads(body or b"{}"), {}
+ if not isinstance(meta, dict):
+ raise ValueError("the body must be a JSON object or an RCB1 frame")
+ return meta, parts
+
+ @staticmethod
+ def _reply_json(code: int, obj: dict) -> tuple[int, str, bytes]:
+ return code, "application/json", json.dumps(obj).encode()
+
+ @staticmethod
+ def _reply_rcb1(meta: dict, parts: list) -> tuple[int, str, bytes]:
+ body = rcb1.encode(meta, parts)
+ if len(body) > MAX_RESPONSE:
+ parts = [p for p in parts if not p[0].startswith(("frames/", "depth/"))]
+ meta = {**meta, "degraded": f"response over {MAX_RESPONSE} bytes: frames/ and depth/ dropped"}
+ body = rcb1.encode(meta, parts)
+ return 200, "application/x-rcb1", body
+
+ def _view(self, meta: dict, default_video: str) -> tuple[list[str], int, str, bool]:
+ cams = meta.get("cameras") or list(self.sim.cameras)
+ if isinstance(cams, str):
+ cams = [c for c in cams.split(",") if c]
+ if not isinstance(cams, list) or (unknown := sorted(set(map(str, cams)) - set(self.sim.cameras))):
+ raise ValueError(f"unknown camera(s) {unknown if isinstance(cams, list) else cams}; "
+ f"the cameras are {list(self.sim.cameras)}")
+ # Each camera once, in the order asked for. A name repeated N times would be rendered into one list N times and
+ # returned N times (N * N images), which the frame budget counted as N: a small request could exhaust the memory.
+ cams = list(dict.fromkeys(map(str, cams)))
+ try:
+ fps = int(meta.get("video_fps", 10))
+ except (TypeError, ValueError, OverflowError):
+ raise ValueError("video_fps must be an integer") from None
+ if not 1 <= fps <= MAX_VIDEO_FPS:
+ raise ValueError(f"video_fps must be in [1, {MAX_VIDEO_FPS}]")
+ video = meta.get("video", default_video)
+ if video not in VIDEOS:
+ raise ValueError(f"video must be one of {list(VIDEOS)}")
+ depth = bool(meta.get("include_depth", False))
+ if depth and not set(cams) & set(self.sim.depth_cameras):
+ raise ValueError(f"include_depth: none of {cams} has depth; spec()['cameras'][name]['depth'] says which do")
+ return cams, fps, video, depth
+
+ def _frame_budget(self, rec: Recorder, steps: int) -> None:
+ n = rec.images_per_call(steps)
+ if n > self.max_frames:
+ raise ValueError(f"this call would return {n} images (cameras x frames, depth counted again); the limit is "
+ f"{self.max_frames}: ask for fewer cameras, a lower video_fps or a shorter chunk")
+
+ def _conflict(self) -> Refused:
+ if self.stopping:
+ return Refused(409, {"error": "trial finished: your time is up", "reason": self.reason})
+ body = {"error": "episode finished", "reason": self.reason}
+ if self.reset_budget:
+ body["resets_left"] = 0 if self.success else self.reset_budget - self.resets
+ return Refused(409, body)
+
+ def _step(self, typ: str, row: np.ndarray) -> dict:
+ """One control step: execute, journal, and end the episode if it is over."""
+ sim_row = np.asarray(self.sim.to_sim(typ, row), dtype=np.float64)
+ done, success, info = self.sim.step(sim_row)
+ self._unwritten.append(sim_row)
+ self.journal_rows += 1
+ self.episode.append(sim_row)
+ if success:
+ self._end("success")
+ elif done or len(self.episode) >= self.sim.max_control_steps:
+ self._end("failed")
+ return info
+
+ def _end(self, reason: str) -> None:
+ if self.done:
+ return
+ self.done, self.reason, self.success = True, reason, reason == "success"
+ self.t_end = time.time()
+ try:
+ self._checkpoint("end")
+ except Exception as exc: # noqa: BLE001 -- a broken simulator may not even hash: the end is still logged
+ self.log("end", reason=reason, checkpoint_error=repr(exc)[:300])
+ self._write_final()
+
+ def _submit(self, meta: dict, parts: dict) -> tuple[dict, list]:
+ if self.done or self.stopping:
+ raise self._conflict()
+ with bad_request():
+ typ = meta.get("type")
+ if not isinstance(typ, str) or typ not in self.types:
+ raise ValueError(f"type must be one of {sorted(self.types)}")
+ try:
+ A = np.asarray(parts["actions"]["data"] if "actions" in parts else meta.get("actions"), dtype=np.float64)
+ except (TypeError, ValueError, KeyError):
+ raise ValueError("send the actions as an RCB1 array part `actions` or a JSON list of rows") from None
+ dim = int(self.types[typ]["dim"])
+ if A.ndim != 2 or A.shape[0] < 1 or A.shape[1] != dim:
+ raise ValueError(f"{typ} actions must be (T, {dim}) with T >= 1, got {list(A.shape)}")
+ if not np.all(np.isfinite(A)):
+ raise ValueError("actions contain NaN or inf")
+ cdt = self.sim.control_dt
+ try:
+ dt = cdt if meta.get("dt") is None else float(meta["dt"])
+ except (TypeError, ValueError):
+ raise ValueError("dt must be a number of seconds") from None
+ k = round(dt / cdt) if np.isfinite(dt / cdt) else 0
+ if k < 1 or abs(k * cdt - dt) > 1e-9:
+ raise ValueError(f"dt must be a positive whole multiple of control_dt = {cdt} s (never rounded)")
+ if k > 1 and not self.sim.interpolable(typ):
+ raise ValueError(f"{typ}: one row is one control step, so dt must be control_dt = {cdt} s")
+ if len(A) * k * cdt > self.max_chunk_s + 1e-9:
+ raise ValueError(f"the chunk covers {len(A) * k * cdt:g} s: more than max_chunk_s = {self.max_chunk_s:g}")
+ interp = meta.get("interpolate", "linear")
+ if interp not in ("linear", "hold"):
+ raise ValueError("interpolate must be linear or hold")
+ cams, fps, video, depth = self._view(meta, "mp4")
+ rec = Recorder(self.sim, cams, fps, depth, video)
+ self._frame_budget(rec, len(A) * k)
+ rows, clipped = self.sim.clip(typ, A)
+ before, info, executed, stop = len(self.episode), {}, 0, False
+ prev = np.asarray(self.sim.command(typ), dtype=np.float64)
+ rec.start()
+ for i, row in enumerate(rows):
+ for j in range(1, k + 1):
+ info = self._step(typ, row if interp == "hold" else prev + (row - prev) * (j / k))
+ rec.step()
+ if self.done or self.stopping or info.get("blocked"):
+ stop = True
+ break
+ executed, prev = i + 1, row
+ if stop:
+ break
+ rec.finish()
+ self.calls += 1
+ self.rows += executed
+ out = {**self.meta(), "chunk_id": self.calls, "type": typ, "executed_rows": executed,
+ "control_steps_chunk": len(self.episode) - before, "clipped": clipped[:200], "n_clipped": len(clipped),
+ **info, "ok": True}
+ self._checkpoint("submit", type=typ, n_rows=len(A), dt=dt, interpolate=interp, cameras=cams,
+ note=str(meta.get("note") or "")[:500], executed_rows=executed, n_clipped=len(clipped), **info)
+ self._save_call(out, A)
+ return out, rec.parts()
+
+ def _observe(self, meta: dict, parts: dict) -> tuple[dict, list]:
+ with bad_request():
+ try:
+ seconds = float(meta.get("seconds") or 0.0)
+ except (TypeError, ValueError):
+ raise ValueError("seconds must be a number") from None
+ if not 0.0 <= seconds <= self.max_watch_s:
+ raise ValueError(f"seconds must be in [0, {self.max_watch_s:g}]")
+ cams, fps, video, depth = self._view(meta, "frames")
+ n = int(round(seconds / self.sim.control_dt)) if not (self.done or self.stopping) else 0
+ rec = Recorder(self.sim, cams, fps, depth, video if n or video != "mp4" else "frames") # no time: no video
+ self._frame_budget(rec, n)
+ before = len(self.episode)
+ rec.start()
+ if n:
+ typ, row = self.sim.hold()
+ for _ in range(n):
+ self._step(typ, row)
+ rec.step()
+ if self.done or self.stopping:
+ break
+ rec.finish()
+ self.calls += 1
+ ran = len(self.episode) - before
+ out = {**self.meta(), "obs_id": self.calls, "seconds": round(ran * self.sim.control_dt, 4),
+ "control_steps_watch": ran, "ok": True}
+ self._checkpoint("observe", seconds=out["seconds"], cameras=cams)
+ self._save_call(out, None)
+ return out, rec.parts()
+
+ def _reset_view(self, meta: dict) -> tuple[list[str], int, str, bool]:
+ """A reset's checks, all of them, before anything changes: Refused (409, 400) or the view it returns."""
+ if self.success or self.stopping:
+ raise Refused(409, {"error": "trial finished: " + ("an episode succeeded" if self.success else "your time is up"),
+ "reason": self.reason})
+ if self.resets >= self.reset_budget:
+ raise Refused(409, {"error": "reset budget exhausted", "budget": self.reset_budget})
+ with bad_request():
+ return self._view(meta, "frames")
+
+ def _reset(self, meta: dict, parts: dict) -> tuple[dict, list]:
+ cams, fps, video, depth = self._reset_view(meta)
+ if not self.done:
+ self._end("reset") # the running episode is abandoned
+ try:
+ self.sim.reset()
+ except Exception:
+ # The simulator failed to start the new episode: the trial ends on a simulator error, not on the episode
+ # the agent had every right to abandon (handle() answers 500; its _end("error") finds the episode ended).
+ self.done, self.reason, self.success = True, "error", False
+ try:
+ self._checkpoint("end", reset_failed=True)
+ except Exception as exc: # noqa: BLE001 -- a broken simulator may not even hash
+ self.log("end", reason="error", reset_failed=True, checkpoint_error=repr(exc)[:300])
+ self._write_final()
+ raise
+ self.resets += 1
+ self.ep_start, self.episode = self.journal_rows, []
+ self.done, self.reason, self.success, self.t_end = False, None, False, None
+ rec = Recorder(self.sim, cams, fps, depth, "frames" if video == "mp4" else video) # no time: no video
+ rec.start()
+ rec.finish()
+ self.calls += 1
+ out = {**self.meta(), "ok": True}
+ self._checkpoint("reset", cameras=cams)
+ self._save_call(out, None)
+ return out, rec.parts()
+
+ # ------------------------------------------------------------------ the end of the trial
+ def on_sigusr1(self, *_) -> None:
+ """The agent is gone (the [[verifier.collect]] hook). Only a flag: the handler may fire in the middle of a
+ request; a chunk stops at its next control step and poll() wraps up between requests."""
+ self.stopping = True
+
+ def poll(self) -> None:
+ """Between requests (the HTTP loop calls it every 0.5 s): after SIGUSR1, end the episode and finalize."""
+ if self.stopping and not self.finalized:
+ if not self.done:
+ self._end("timeout")
+ self._finalize()
+
+ def _finalize(self) -> None:
+ """The graded episode (every control step since the last reset) as a trajectory and its replay in a fresh
+ simulator (success, and bit-exact agreement with the live run): the verdict, final from the moment it is
+ written ("finalized": true). Then a video for humans, which can no longer change or void it: while it renders
+ final.json says "video_pending": true (the collect hook waits for it as long as its time allows)."""
+ rows = None
+ try:
+ rows = np.asarray(self.episode, dtype=np.float64).reshape(-1, self.sim.raw_dim)
+ live_sha = self.sim.state_sha()
+ self.sim.save_trajectory(rows, os.path.join(self.data, "trajectory.npz"),
+ {"source": "standard-mode episode", "protocol": PROTOCOL, "reason": self.reason,
+ "episode": self.resets + 1})
+ rep = self.sim.verify(rows)
+ self.verdict = {"n_actions": int(len(rows)), "replay_success": bool(rep["success"]),
+ "deterministic": rep["state_sha"] == live_sha, "state_sha": live_sha, "video": False}
+ self.log("replay", **self.verdict)
+ except Exception as exc: # noqa: BLE001
+ traceback.print_exc()
+ self.log("finalize_error", error=repr(exc)[:500])
+ self.verdict = {"finalize_error": repr(exc)[:300]}
+ graded = "finalize_error" not in self.verdict
+ self.verdict["video_pending"] = graded
+ self.finalized = True
+ self._write_final() # the verdict, finalized, on disk before any video is rendered
+ if not graded:
+ return
+ try:
+ self.sim.render_video(rows, os.path.join(self.data, "replay.mp4"))
+ self.verdict["video"] = True
+ except Exception as exc: # noqa: BLE001 -- the video is for humans; it never decides the verdict
+ traceback.print_exc()
+ self.log("video_error", error=repr(exc)[:500])
+ self.verdict["video_pending"] = False
+ try:
+ self._write_final()
+ except Exception: # noqa: BLE001 -- the finalized verdict is on disk already; only the video's flag is lost
+ traceback.print_exc()
+
+ # ------------------------------------------------------------------ HTTP
+ def http_server(self, address: tuple[str, int]) -> HTTPServer:
+ svc = self
+
+ class Handler(BaseHTTPRequestHandler):
+ timeout = 120 # a client that stalls mid-request is dropped: one thread serves everyone
+
+ def log_message(self, *a):
+ pass
+
+ def _send(self, code: int, ctype: str, body: bytes) -> None:
+ self.send_response(code)
+ self.send_header("Content-Type", ctype)
+ self.send_header("Content-Length", str(len(body)))
+ self.end_headers()
+ self.wfile.write(body)
+
+ def do_GET(self):
+ self._send(*svc.handle("GET", self.path))
+
+ def do_POST(self):
+ try:
+ n = int(self.headers.get("Content-Length") or 0)
+ except ValueError:
+ n = -1
+ if n < 0: # (read(-1) would read to the end of the stream, without a limit)
+ self.close_connection = True
+ return self._send(*svc._reply_json(400, {"error": "bad Content-Length"}))
+ if n > MAX_BODY: # refused unread: drop the connection with it
+ self.close_connection = True
+ return self._send(*svc.handle("POST", self.path, b"", content_length=n))
+ self._send(*svc.handle("POST", self.path, self.rfile.read(n)))
+
+ class Server(HTTPServer):
+ def service_actions(self):
+ svc.poll()
+
+ return Server(address, Handler)
+
+ def bind(self, port: int) -> HTTPServer:
+ """The HTTP server on `port`. Something else may hold the port -- in the network namespace it shares with the
+ agent's container, an agent's process can take it while this service (re)starts -- so wait, not die (a death
+ here would count as a crash); the trial can still end meanwhile (SIGUSR1)."""
+ busy = False
+ while True:
+ try:
+ return self.http_server(("0.0.0.0", port))
+ except OSError as exc:
+ if exc.errno != errno.EADDRINUSE:
+ raise
+ if not busy:
+ busy = True
+ self.log("port_busy", port=port)
+ self.poll()
+ time.sleep(1.0)
+
+ def serve(self, port: int = 8000) -> None:
+ signal.signal(signal.SIGUSR1, self.on_sigusr1)
+ srv = self.bind(port)
+ print(f"{PROTOCOL} sim service ready on :{port}", flush=True)
+ srv.serve_forever()
+
+
+def refuse_forever(reason: str, port: int = 8000) -> None:
+ """After a give-up: answer every request 409 with the reason until the container is removed. Before, nothing
+ listened once the service had given up, and a client waited out its whole connection timeout for a simulator that
+ would never answer. The verdict is on disk already: nothing here reads or changes it."""
+ body = json.dumps({"error": "the sim service gave up on this episode and will not serve it", "reason": reason,
+ "gave_up": True}).encode()
+
+ class Refusal(BaseHTTPRequestHandler):
+ timeout = 30 # a client that stops sending must not hold the only thread
+
+ def _refuse(self):
+ try:
+ n = int(self.headers.get("Content-Length") or 0)
+ if 0 < n <= MAX_BODY: # take the request first: a client still sending would see a reset
+ self.rfile.read(n)
+ self.send_response(409)
+ self.send_header("Content-Type", "application/json")
+ self.send_header("Content-Length", str(len(body)))
+ self.end_headers()
+ self.wfile.write(body)
+ except (OSError, ValueError):
+ pass
+ self.close_connection = True
+
+ do_GET = do_POST = _refuse
+
+ def log_message(self, *args):
+ pass
+
+ def answer():
+ while True:
+ try:
+ HTTPServer(("0.0.0.0", port), Refusal).serve_forever()
+ except Exception: # noqa: BLE001 -- a busy port, or anything else: the give-up stands without the answer
+ time.sleep(5.0)
+
+ threading.Thread(target=answer, daemon=True).start()
+ while True:
+ time.sleep(3600)
+
+
+def main(build: Callable[[], Sim], port: int = 8000) -> None:
+ """A sidecar's life: count the start, back off after a failed one, build the scene under a startup watchdog, resume
+ or start the episode, serve. Settings: RCB_DATA (/data), RCB_RESET_BUDGET (0), RCB_WALL_CLOCK_S (3600),
+ RCB_STARTUP_TIMEOUT_S (300), RCB_MAX_RESUMES (2)."""
+ env = os.environ.get
+ data = env("RCB_DATA", "/data")
+ os.makedirs(data, exist_ok=True)
+ starts = count_start(data)
+ # Compose restarts a failed start at once; a scene that failed to build (e.g. the GPU full of other trials) would
+ # meet the same conditions every time, so wait before rebuilding -- only before the episode exists.
+ if starts > 1 and not os.path.exists(os.path.join(data, "episode.jsonl")):
+ time.sleep(min(120, 20 * (starts - 1)))
+ pending: list[int] = []
+ signal.signal(signal.SIGUSR1, lambda *_: pending.append(1)) # its default action would kill the process
+ # A scene build that hangs (no crash, so compose never restarts it) must end the process. Neither a Python thread
+ # (a native hang may hold the GIL) nor SIGALRM's default action (PID 1 ignores it) can; faulthandler's C watchdog
+ # dumps every thread's traceback and exits. Before the scene exists no episode exists: the retry is no reset.
+ faulthandler.dump_traceback_later(float(env("RCB_STARTUP_TIMEOUT_S", "300")), exit=True)
+ sim = build()
+ faulthandler.cancel_dump_traceback_later() # a resume replay may take minutes
+ svc = Service(sim, data=data, reset_budget=int(env("RCB_RESET_BUDGET") or 0),
+ wall_clock_s=float(env("RCB_WALL_CLOCK_S") or 3600), max_resumes=int(env("RCB_MAX_RESUMES") or 2))
+ try:
+ svc.start()
+ except GiveUp as exc:
+ print(f"restart refused: {exc}", flush=True)
+ signal.signal(signal.SIGUSR1, lambda *_: None)
+ refuse_forever(str(exc), port)
+ if pending:
+ svc.on_sigusr1()
+ svc.serve(port)
diff --git a/tasks/myosuite-table-tennis-i00-standard-realtime/environment/sim/server.py b/tasks/myosuite-table-tennis-i00-standard-realtime/environment/sim/server.py
new file mode 100644
index 0000000000..2895899a20
--- /dev/null
+++ b/tasks/myosuite-table-tennis-i00-standard-realtime/environment/sim/server.py
@@ -0,0 +1,333 @@
+"""eai-standard/2.3 sim service for MyoSuite's standard-realtime tasks -- the adapter behind the shared realtime layer
+(`python /srv/rcb_realtime.py`, which imports this module and calls build(); docs/PROTOCOL.md §4b,
+docs/benchmarks/myosuite.md).
+
+The simulator-specific half only: `MyoSuite` implements rcb_service.Sim (the contract documented in rcb_service.py) for
+one frozen instance, on the benchmark's own tooling module rcb_myosuite.Session -- the same scene, action, success rule
+and replay as the standard twin's eai-standard/2.2 service (environment/sim/server.py of `-standard`). The protocol, the
+clock, the plan, the episode, crash recovery and the finalize are rcb_realtime.py's and rcb_service.py's, unchanged
+copies of the HumanoidBench sources (ci_checks/check-tooling-identity.sh).
+
+What the agent gets is what the standard twin gives, minus the 2.2 additions the realtime layer does not serve
+(`/kinematics`, per-frame camera poses, depth modes): the cameras (RGB, and depth on request from every camera that
+rides on the world or the body), the body's own proprioception (joint positions and velocities, muscle activations and
+lengths, the actuators' generalized force on each joint, and a free-standing body's root pose and velocity) and the
+actuator commands. Nothing about objects, balls, opponents or goals leaves this process; a camera that rides on a body
+of the scene gives images only (its pose would place that body).
+
+Rendering: MuJoCo on the GPU (EGL) when the container can reach the NVIDIA driver, else OSMesa on the CPU. In
+table_tennis an OSMesa frame takes 1.3 s and an EGL frame 5 ms; on a clock that does not stop, a render is time the
+simulator thread cannot step, so the realtime twins reserve a GPU. The NVIDIA container toolkit mounts libEGL_nvidia
+under NVIDIA_DRIVER_CAPABILITIES=graphics but, on some hosts, not its ICD manifest; this module writes one to /tmp and
+points libglvnd at it (inert where the library is absent: libglvnd falls back to Mesa), then tries EGL in a throwaway
+process, because MuJoCo binds its backend once per process at `import mujoco`.
+
+Settings (environment/docker-compose.yaml): RCB_INSTANCE (the frozen instance), RCB_TASK_INFO (task.json: task text,
+default camera), RCB_IMAGE_HW (the cameras' "H,W", default "480,640", the standard twin's), RCB_DATA, RCB_RESET_BUDGET,
+RCB_WALL_CLOCK_S, RCB_SLOWDOWN (rcb_realtime / rcb_service), RCB_GL (auto: EGL when it works, else OSMesa).
+"""
+from __future__ import annotations
+
+import json
+import math
+import os
+import subprocess
+import sys
+
+import numpy as np
+
+HERE = os.path.dirname(os.path.abspath(__file__))
+sys.path.insert(0, HERE)
+
+
+def _resolve_gl_backend() -> str:
+ """MuJoCo's offscreen backend: RCB_GL (egl | osmesa) if set, else EGL if a probe process can open a context on the
+ NVIDIA driver, else OSMesa. The image's own MUJOCO_GL (osmesa, for the CPU-only standard twin) is not a choice."""
+ if os.environ.get("RCB_GL", "auto") in ("egl", "osmesa"):
+ return os.environ["RCB_GL"]
+ if not os.path.exists("/dev/nvidiactl"):
+ return "osmesa"
+ vendors = [p for d in ("/usr/share/glvnd/egl_vendor.d", "/etc/glvnd/egl_vendor.d") if os.path.isdir(d)
+ for p in sorted(os.listdir(d))]
+ if not any("nvidia" in v for v in vendors) and not os.environ.get("__EGL_VENDOR_LIBRARY_FILENAMES"):
+ os.makedirs("/tmp/rcb_egl", exist_ok=True)
+ manifest = "/tmp/rcb_egl/10_nvidia.json"
+ with open(manifest, "w") as f:
+ f.write('{"file_format_version": "1.0.0", "ICD": {"library_path": "libEGL_nvidia.so.0"}}\n')
+ mesa = "/usr/share/glvnd/egl_vendor.d/50_mesa.json"
+ os.environ["__EGL_VENDOR_LIBRARY_FILENAMES"] = manifest + (f":{mesa}" if os.path.exists(mesa) else "")
+ probe = ("import os; os.environ['MUJOCO_GL'] = os.environ['PYOPENGL_PLATFORM'] = 'egl'; import mujoco; "
+ "c = mujoco.GLContext(8, 8); c.make_current(); import OpenGL.GL as g; "
+ "r = g.glGetString(g.GL_RENDERER).decode(); c.free(); print(r)")
+ try:
+ out = subprocess.run([sys.executable, "-c", probe], capture_output=True, text=True, timeout=120)
+ ok = out.returncode == 0 and "llvmpipe" not in out.stdout.lower() # Mesa's EGL renders on the CPU as well
+ except (OSError, subprocess.SubprocessError):
+ ok = False
+ return "egl" if ok else "osmesa"
+
+
+GL_BACKEND = _resolve_gl_backend()
+os.environ["MUJOCO_GL"] = os.environ["PYOPENGL_PLATFORM"] = GL_BACKEND
+os.environ["RCB_SESSION_LOG"] = ""
+import rcb_myosuite as M # noqa: E402 (after the backend: it only sets a default)
+import mujoco # noqa: E402
+
+M.SESSION_LOG = "" # the sidecar's Session is not the agent's interaction log
+from rcb_service import main # noqa: E402
+
+ROBOT = {
+ "hand": ("MyoHand", "a musculoskeletal model of a human forearm and hand (MyoSuite)"),
+ "arm": ("MyoArm", "a musculoskeletal model of a human shoulder, arm and hand (MyoSuite)"),
+ "body": ("MyoArm with torso", "a musculoskeletal model of a human torso, right arm and hand on a pelvis that slides "
+ "on the floor (MyoSuite)"),
+ "bimanual": ("MyoArm + MPL", "a musculoskeletal right arm and hand (MyoSuite's MyoArm) next to a motorized "
+ "prosthetic left arm (the Modular Prosthetic Limb)"),
+ "legs": ("MyoLegs", "a free-standing musculoskeletal model of a human torso and both legs (MyoSuite)"),
+ "osl": ("MyoLeg + OSL", "a free-standing musculoskeletal model of a human with the right leg's muscles and an "
+ "Open-Source Leg knee-ankle prosthesis on the left (MyoSuite)"),
+}
+MUSCLE_TEXT = (
+ "muscles: one real number a per muscle; excitation u = 1 / (1 + exp(-5 (a - 0.5))), a not clipped (a = -1: u ~ "
+ "0.0006, 0.5: 0.5, 1: 0.92, 2: 0.9994); activations follow excitations with a first-order delay (~10 ms up, ~40 ms "
+ "down); muscles only pull, one can cross several joints, and its force depends on its length and velocity")
+GL2CV = np.diag([1.0, -1.0, -1.0, 1.0])
+CAM_MODES = {0: "fixed", 1: "track", 2: "trackcom", 3: "targetbody", 4: "targetbodycom"}
+NO_POSE = ("not published: this camera rides on a body of the scene that is not yours, so its pose would be that "
+ "body's position; it gives images only")
+
+
+class MyoSuite:
+ """One frozen MyoSuite instance behind the rcb_service.Sim contract (see the module docstring)."""
+
+ def __init__(self, instance: str, task_info: dict, image_hw: tuple[int, int] = (480, 640)):
+ self.instance = instance
+ self.task = str(np.load(instance, allow_pickle=False)["task"])
+ self.info = task_info
+ self.s = M.Session(self.task, instance=instance)
+ m = self.m = self.s.model
+ self.H, self.W = image_hw
+ self.control_dt = float(self.s.dt)
+ self.max_control_steps = int(self.s.max_steps) # a runaway guard: the task's own rule ends an episode
+ self.nu = self.raw_dim = int(self.s.nu)
+ acts = self.s.actuators()
+ self.layout = [a["name"] for a in acts]
+ self.mus_idx = [a["index"] for a in acts if a["kind"] == "muscle"]
+ self.muscles = [acts[i]["name"] for i in self.mus_idx]
+ self.other = [a for a in acts if a["kind"] != "muscle"]
+ self.atype = "muscle" if not self.other else "actuators"
+ self.act_adr = np.array([int(m.actuator_actadr[i]) for i in self.mus_idx], dtype=int)
+ # the body's own joints (the key's hinge and the OBJ* joints are object state, never sent; free joints of balls,
+ # objects and paddles are not in S.joints(); a free-standing body's root is reported on its own)
+ limb = [j for j in self.s.joints() if j["name"] != "keyjoint" and not j["name"].startswith("OBJ")]
+ self.j_names = [j["name"] for j in limb]
+ self.j_q = np.array([j["qpos_adr"] for j in limb], dtype=int)
+ self.j_v = np.array([j["dof_adr"] for j in limb], dtype=int)
+ self.j_id = [m.joint(n).id for n in self.j_names]
+ root = mujoco.mj_name2id(m, mujoco.mjtObj.mjOBJ_JOINT, "root")
+ self.root = root if root >= 0 and int(m.jnt_type[root]) == 0 else None
+ self.own_joints = set(self.j_id) | ({self.root} if self.root is not None else set())
+ self.cameras = tuple(self.s.cameras())
+ self.scene_cams = {c for c in self.cameras if self._scene_camera(c)}
+ self.depth_cameras = tuple(c for c in self.cameras if c not in self.scene_cams)
+ self.proprio_layout = ([f"q_{n}" for n in self.j_names] + [f"qd_{n}" for n in self.j_names]
+ + [f"act_{n}" for n in self.muscles] + [f"len_{n}" for n in self.muscles]
+ + [f"effort_{n}" for n in self.j_names]
+ + (["root_x", "root_y", "root_z", "root_qw", "root_qx", "root_qy", "root_qz",
+ "root_vx", "root_vy", "root_vz", "root_wx", "root_wy", "root_wz"]
+ if self.root is not None else []))
+ self._depth_on = False
+ self.render_frame(self.cameras[0]) # warms up the GL context
+ self.cmd = self.s.rest_action()
+
+ # ------------------------------------------------------------------ what the agent may know
+ def _own_body(self, b: int) -> bool:
+ """A body fixed in the world or moved by the body's own joints alone; a mocap body (the opponent, goal markers)
+ or one behind any other joint (objects, balls, the paddle) is scene state."""
+ m = self.m
+ while b > 0:
+ if int(m.body_mocapid[b]) >= 0:
+ return False
+ first = int(m.body_jntadr[b])
+ if any(j not in self.own_joints for j in range(first, first + int(m.body_jntnum[b]))):
+ return False
+ b = int(m.body_parentid[b])
+ return True
+
+ def _scene_camera(self, cam: str) -> bool:
+ m, i = self.m, self.m.camera(cam).id
+ bodies = [int(m.cam_bodyid[i])] + ([int(m.cam_targetbodyid[i])] if int(m.cam_mode[i]) in (3, 4) else [])
+ return not all(self._own_body(b) for b in bodies if b >= 0)
+
+ def _intrinsic(self, cam: str) -> list:
+ f = 0.5 * self.H / math.tan(math.radians(float(self.m.cam_fovy[self.m.camera(cam).id])) / 2)
+ return [[f, 0.0, self.W / 2], [0.0, f, self.H / 2], [0.0, 0.0, 1.0]]
+
+ def _cam_T_gl(self, cam: str) -> np.ndarray:
+ i, d = self.m.camera(cam).id, self.s.data
+ T = np.eye(4)
+ T[:3, :3] = d.cam_xmat[i].reshape(3, 3) # MuJoCo cameras: OpenGL axes (x right, y up, z back)
+ T[:3, 3] = d.cam_xpos[i]
+ return T
+
+ def _camera_fact(self, cam: str) -> dict:
+ m, i = self.m, self.m.camera(cam).id
+ out = {"hw": [self.H, self.W], "intrinsic_cv": self._intrinsic(cam),
+ "fovy_deg": round(float(m.cam_fovy[i]), 2)}
+ if cam in self.scene_cams:
+ return {**out, "mounted": "on a body of the scene that is not yours (it moves with it): images only",
+ "pose": NO_POSE}
+ body, mode = m.body(int(m.cam_bodyid[i])).name, int(m.cam_mode[i])
+ if mode != 0:
+ return {**out, "mounted": f"tracking ({CAM_MODES.get(mode, mode)}) the body {body}: it follows that body's "
+ "position (or its subtree's centre of mass) without turning with it"}
+ if body == "world":
+ ext = np.linalg.inv(self._cam_T_gl(cam) @ GL2CV)[:3, :]
+ return {**out, "mounted": "static", "extrinsic_cv": ext.round(6).tolist(),
+ "pose_note": "extrinsic_cv: world -> camera, OpenCV axes (x right, y down, z forward)"}
+ b = m.body(body).id
+ Tb = np.eye(4)
+ Tb[:3, :3] = self.s.data.xmat[b].reshape(3, 3)
+ Tb[:3, 3] = self.s.data.xpos[b]
+ link_T_cam = np.linalg.inv(Tb) @ self._cam_T_gl(cam) @ GL2CV
+ return {**out, "mounted": f"fixed to the body {body}: it moves with it",
+ "link": body, "link_T_cam_cv": link_T_cam.round(6).tolist()}
+
+ # ------------------------------------------------------------------ the rcb_service.Sim contract
+ def spec(self) -> dict:
+ m, robot = self.m, ROBOT[M.BODY[self.task]]
+ lo = [float(m.jnt_range[j, 0]) if m.jnt_limited[j] else None for j in self.j_id]
+ hi = [float(m.jnt_range[j, 1]) if m.jnt_limited[j] else None for j in self.j_id]
+ action = {"dim": self.nu, "layout": self.layout, "lower": None, "upper": None,
+ "units": MUSCLE_TEXT + ("; the other actuators: see `actuators`" if self.other else ""),
+ "semantics": "one row per control step, every actuator in `layout` order; rows dt apart (a whole "
+ "multiple of control_dt), `linear` (default: a straight line from the command in force) "
+ "or `hold` between rows; nothing is clipped by the service (MuJoCo clips the non-muscle "
+ "controls to their ranges)"}
+ if self.other:
+ action["actuators"] = [{k: a[k] for k in ("index", "name", "kind", "command")} for a in self.other]
+ return {
+ "simulator": "MyoSuite 2.12.2 on MuJoCo 3.6.0 (CPU physics)",
+ "robot": f"{robot[0]}: {robot[1]}, {len(self.muscles)} muscles"
+ + (f" and {len(self.other)} other actuators" if self.other else ""),
+ "task_text": self.info["task_text"],
+ "episode": "an episode ends at the task's own success event or failure (see task_text), when your time is "
+ "up, or on a simulator error",
+ "action_types": {self.atype: action},
+ "cameras": {c: self._camera_fact(c) for c in self.cameras},
+ "default_camera": self.info["default_cameras"][0],
+ "proprio": {
+ "groups": {"q_*": "joint positions of the body (rad; m for sliding joints)",
+ "qd_*": "joint velocities (rad/s; m/s)",
+ "act_*": "muscle activations in [0, 1] (they follow the excitations with a first-order delay)",
+ "len_*": "muscle-tendon lengths (m)",
+ "effort_*": "generalized force of all actuators on each joint (N m; N)",
+ **({"root_*": "the free-standing root (the pelvis) in the world frame: position (m), wxyz "
+ "quaternion, linear (m/s, world axes) and angular velocity (rad/s, its own axes)"}
+ if self.root is not None else {})}},
+ "static": {"muscles": self.muscles},
+ "world": {"frame": "MyoSuite's world frame, z up (no positions of the scene are given)", "up": "+z",
+ "units": {"length": "m", "angle": "rad", "time": "s"}, "quaternion_order": "wxyz"},
+ "robot_spec": {
+ "name": robot[0],
+ "joints": [{"name": n, "type": "prismatic" if int(m.jnt_type[j]) == 2 else "revolute",
+ "position_limits": [lo[k], hi[k]] if lo[k] is not None else None}
+ for k, (n, j) in enumerate(zip(self.j_names, self.j_id))],
+ "controller": "muscle actuators (MuJoCo muscle model); no joint-level controller"
+ + ("; plus the non-muscle actuators in action_types" if self.other else ""),
+ "physics_hz": round(1.0 / self.control_dt * int(self.s.base.frame_skip), 3),
+ },
+ "depth": {"cameras": list(self.depth_cameras), "units": "z-depth in metres, pixel for pixel with the "
+ "camera's image; noise-free"} if self.depth_cameras else None,
+ "gl_backend": GL_BACKEND,
+ }
+
+ def interpolable(self, typ: str) -> bool:
+ return True
+
+ def render_frame(self, cam: str) -> np.ndarray:
+ if self._depth_on:
+ self.s._renderer.disable_depth_rendering()
+ self._depth_on = False
+ return self.s.render_frame(cam, (self.H, self.W))
+
+ def images(self, cameras: list[str], depth: bool) -> tuple[dict, dict]:
+ rgb = {c: self.render_frame(c) for c in cameras}
+ dep = {}
+ if depth:
+ r = self.s._renderer # the same renderer (a second context reads garbage
+ for c in cameras: # under OSMesa)
+ if c in self.depth_cameras:
+ r.enable_depth_rendering()
+ self._depth_on = True
+ r.update_scene(self.s.data, camera=c)
+ z = np.asarray(r.render(), dtype=np.float32).copy()
+ dep[c] = np.where(np.isfinite(z) & (z > 0), z, 0.0).astype(np.float32)
+ if self._depth_on:
+ r.disable_depth_rendering()
+ self._depth_on = False
+ return rgb, dep
+
+ def proprio(self) -> np.ndarray:
+ d, phys = self.s.data, self.s.base.robot.mj_data
+ # joint_effort from the physics, not the mirror: MyoSuite steps its robot wrapper's own MjData and refreshes
+ # S.data from qpos / qvel / act only (the standard twin's server.py, 2026-10-09)
+ parts = [d.qpos[self.j_q], d.qvel[self.j_v], d.act[self.act_adr], d.actuator_length[self.mus_idx],
+ phys.qfrc_actuator[self.j_v]]
+ if self.root is not None:
+ qa, va = int(self.m.jnt_qposadr[self.root]), int(self.m.jnt_dofadr[self.root])
+ parts += [d.qpos[qa:qa + 7], d.qvel[va:va + 6]]
+ return np.concatenate(parts).astype(np.float32)
+
+ def clip(self, typ: str, rows: np.ndarray) -> tuple[np.ndarray, list]:
+ return np.asarray(rows, dtype=np.float64), [] # as the standard twin: nothing clipped here
+
+ def to_sim(self, typ: str, row: np.ndarray) -> np.ndarray:
+ return np.asarray(row, dtype=np.float64)
+
+ def command(self, typ: str) -> np.ndarray:
+ return self.cmd
+
+ def hold(self) -> tuple[str, np.ndarray]:
+ return self.atype, self.cmd
+
+ def step(self, row: np.ndarray) -> tuple[bool, bool, dict]:
+ running = self.s.step(row)
+ self.cmd = np.asarray(row, dtype=np.float64).copy()
+ return (not running), bool(self.s.success), ({} if running else {"why": self.s.reason})
+
+ def reset(self) -> None:
+ self.s.reset()
+ self.cmd = self.s.rest_action()
+
+ def state_sha(self) -> str:
+ return self.s.state_sha()
+
+ def verify(self, rows: np.ndarray) -> dict:
+ """The recorded episode from the frozen instance in a fresh MyoSuite session (physics only)."""
+ fresh = M.Session(self.task, instance=self.instance)
+ try:
+ st = fresh.replay(rows)
+ return {"success": bool(st["success"]), "state_sha": st["state_sha"]}
+ finally:
+ fresh.close()
+
+ def save_trajectory(self, rows: np.ndarray, path: str, meta: dict) -> None:
+ self.s.save_trajectory(rows, path, meta)
+
+ def render_video(self, rows: np.ndarray, path: str) -> None:
+ import imageio.v2 as imageio
+ cam = self.info["default_cameras"][0]
+ self.s.render_video(rows, path, cameras=[cam], every=2, fps=25, size=(240, 320))
+ imageio.imwrite(os.path.splitext(path)[0] + ".last.png", self.s.render_frame(cam, (self.H, self.W)))
+
+
+def build() -> MyoSuite:
+ env = os.environ.get
+ hw = [int(v) for v in env("RCB_IMAGE_HW", "480,640").split(",")]
+ info = json.load(open(env("RCB_TASK_INFO", os.path.join(HERE, "task.json"))))
+ return MyoSuite(env("RCB_INSTANCE", os.path.join(HERE, "instance.npz")), info, (hw[0], hw[1]))
+
+
+if __name__ == "__main__":
+ main(build)
diff --git a/tasks/myosuite-table-tennis-i00-standard-realtime/environment/sim/task.json b/tasks/myosuite-table-tennis-i00-standard-realtime/environment/sim/task.json
new file mode 100644
index 0000000000..796de98421
--- /dev/null
+++ b/tasks/myosuite-table-tennis-i00-standard-realtime/environment/sim/task.json
@@ -0,0 +1,7 @@
+{
+ "task_text": "Return the served ball: hit it with the paddle so that it lands on the opponent's half of the table (MyoChallenge 2025 TableTennis P1's own success test on the ball's contact history: after the serve's bounce on your half, the paddle touches it once, then it touches the opponent's half, with no second paddle touch and no further bounce on your half). Failure (the episode ends): any other outcome of the rally, the ball falling below 0.3 m, or 3 s of simulated time.",
+ "default_cameras": [
+ "default"
+ ],
+ "reference": null
+}
diff --git a/tasks/myosuite-table-tennis-i00-standard-realtime/instruction.md b/tasks/myosuite-table-tennis-i00-standard-realtime/instruction.md
new file mode 100644
index 0000000000..8f92b131c5
--- /dev/null
+++ b/tasks/myosuite-table-tennis-i00-standard-realtime/instruction.md
@@ -0,0 +1,108 @@
+# MyoSuite · table_tennis — STANDARD-REALTIME MODE (robot as a service; time does not stop)
+
+You control the **MyoArm with torso**, a musculoskeletal model of a human torso, right arm and hand on a pelvis that slides on the floor, **as if it were real**: the simulator (MyoSuite 2.12.2 on MuJoCo) runs in a service
+you cannot open, you reach it only through a client, and like the real world it does not wait for you: once an episode
+is running, the body and the ball move while you think. You act only through its actuators (273 muscles and 2 other actuators).
+**Task:** Return the served ball: hit it with the paddle so that it lands on the opponent's half of the table (MyoChallenge 2025 TableTennis P1's own success test on the ball's contact history: after the serve's bounce on your half, the paddle touches it once, then it touches the opponent's half, with no second paddle touch and no further bounce on your half). Failure (the episode ends): any other outcome of the rally, the ball falling below 0.3 m, or 3 s of simulated time.
+
+You stand at a table-tennis table holding a paddle in the right hand; a ball is served toward you from the far side. Here the rally is fast: the serve reaches you and the rally is decided within about 1.2 s of robot time after the clock starts (an episode that does nothing ends after 121 control steps).
+
+**Slowdown: S = 1.** The world runs at 1/S of real speed: one control step (10 ms of robot time) every S x 10 ms of
+wall-clock time (real time). `spec()["timing"]["slowdown"]` states it.
+
+## Rules (real-world rules)
+- **Time does not stop.** An episode starts paused, on the frozen starting state, with every command at -1 (muscles
+ relaxed). Its clock starts at your first `submit` (or `get_obs(seconds=s)`) and from then on the simulator runs at
+ 100 Hz of robot time (S = 1), one control step every 10 ms of wall-clock time, whether or not you send anything, until
+ the episode ends. Between your commands it executes what you planned, then holds the last commands you sent; nothing
+ pauses for you, not your thinking, not your tool calls.
+- **You are far too slow to steer it turn by turn.** One turn of yours (a model response, a tool call) takes seconds;
+ the whole rally takes about a second. Drive the body with a program: a controller of your own (Python in your
+ container, with numpy and the client) that plans or streams the commands, started in the background
+ (`nohup python /app/output/gen.py > /app/output/run.log 2>&1 &`) or in the foreground if it ends with the episode. Use
+ your turns to start episodes, read what happened, and improve the program.
+- **Episodes and resets.** An episode ends at the task's own success event, at its failure conditions (see the task),
+ when your time is up, or on a simulator error. You have **10 resets** (`spec()["budget"]["reset"]`,
+ which is authoritative: if it says 0, there is no reset and the one episode is the run). `r.reset()` puts the body,
+ the ball and the scene back on the same frozen starting state and starts a new episode, paused until your next
+ `submit`, either abandoning the running episode or after one has ended. Once an episode has ended, `submit` raises
+ `EpisodeOver` (its message gives `resets_left`) until you reset. **Only the last episode is graded**: an episode you
+ reset away is never graded, a success ends the run at once (no reset after it), and when your time is up the episode
+ then running is the one graded. The wall clock is the run's and never restarts; `meta["episode"]` and
+ `meta["resets_left"]` count the episodes.
+- **No privileged state, no score.** You see what the cameras see (RGB, and depth on request) and the body's own
+ proprioception: joint positions and velocities, muscle activations and lengths, the actuators' generalized force on
+ each joint. No ball, paddle, table or goal poses, no success check: the service judges the episode and only
+ tells you that it ended and why (`meta["done"]`, `meta["reason"]`: `"success"`, `"error"` after a simulator error,
+ else `"failed"`). `spec()` gives the body's own facts: the actuators, the joints and their limits, each camera's
+ intrinsics and mounting.
+- There is nothing to hand in: the service records every control step it ran, replays the last episode in a fresh
+ simulator and grades the replay. Keep the controller that drives the body at `/app/output/gen.py` for the record:
+ save it early and keep it current.
+- Do not stop early: while the episode runs, or a reset is left, the run can still pass.
+
+## Time and streaming
+- `submit` answers at once, with the body as it is now: the rows you sent have not run yet. They become the plan from
+ control step `at` (an episode step number; by default the next step), replacing whatever was planned from `at` on and
+ keeping what is planned before it. When the plan runs out, the last commands are held. `res.meta["control_steps"]`
+ is the number of steps run so far in this episode (so the next step has that number), `res.meta["proprio"]["now"]`
+ the proprioception now, `res.meta["plan_end_step"]` where the plan now ends.
+- Open loop: plan a whole movement at once (`dt` a multiple of 0.01 s; a chunk covers at most 10 s), from step
+ `at = 0` for the episode's start: `r.submit("actuators", rows, dt=0.01, interpolate="hold", at=0, cameras=[])`
+ starts the clock and runs exactly those rows from the first step; `at` may be up to
+ `spec()["timing"]["max_ahead_steps"]` steps ahead.
+- Closed loop: send one row per control step, `r.submit("actuators", a[None], cameras=[])`, computed from the
+ proprioception the previous call returned. `cameras=[]` skips the images (the fastest call).
+- Latency: a reading is a few milliseconds old when it reaches you, and a row takes effect at the next control step at
+ the earliest. A row whose step has already run when the request arrives is dropped (`res.meta["late_steps"]` counts
+ them); the rest of its chunk applies from the next step.
+- `get_obs()` reads the body now; `get_obs(seconds=s)` (at most 30) waits while s seconds of robot time run and returns
+ what the cameras saw meanwhile (the body keeps executing its plan). `spec()["timing"]` states all this; `r.status()`
+ shows `running` and `lag_s`.
+
+## Client (`/opt/tools/rcb_client.py`, protocol `eai-standard/2.3`, standard-realtime mode)
+Your whole interface to the body. Its docstring is the full reference (`python -c "import rcb_client;
+help(rcb_client)"`); the calls and the endpoints they use:
+
+| call | endpoint | what it does |
+|---|---|---|
+| `r = Robot()` | | connects (`http://sim:8000`, then `http://localhost:8000`) and waits until the scene is ready |
+| `r.spec()` | `GET /spec` | the body's facts: the actuators' order (`action_types["actuators"]["layout"]`), `control_dt` (0.01 s), `timing`, the cameras, `proprio_layout`, `budget` |
+| `r.status()` | `GET /status` | `calls`, `rows`, `control_steps`, `done`, `reason`, `elapsed_s`, `running`, `planned_steps`, `lag_s`, and `episode`, `resets_left`, `control_steps_total` (all episodes) |
+| `r.submit("actuators", A, dt=0.01, interpolate="linear", at=None, cameras=None)` | `POST /submit` | plans `A`, rows of 275 commands `dt` seconds apart (a chunk covers at most 10 s), from step `at`; answers at once; the first one starts the episode's clock |
+| `r.get_obs(cameras=None, seconds=0, include_depth=False, video="frames")` | `POST /observe` | the body now, or what the cameras saw over the next `seconds` |
+| `r.reset(cameras=None, include_depth=False)` | `POST /reset` | a new episode from the frozen starting state, paused (at most 10 in all); returns the start's frames |
+
+The answers are binary (RCB1) and come back decoded: `res.meta` (the counters, `done`, `reason`, `proprio["now"]`),
+`res.last[camera]` (the last frame), `res.frames` or `res.video` (what the cameras saw during a watch), `res.proprio`
+(proprioception rows). A malformed request raises `ValueError` and changes nothing.
+```python
+from rcb_client import Robot
+import numpy as np
+r = Robot()
+rows = -np.ones((80, 275)) # 0.8 s of commands, one row per control step
+res = r.submit("actuators", rows, dt=0.01, interpolate="hold", at=0, cameras=[]) # starts the clock
+while not r.status()["done"]:
+ ... # or stream: r.submit("actuators", a[None], cameras=[]) each step
+```
+Cameras: `topview`, `default`, `close_up`, `hand_top`, `hand_bottom`, `hand_side_inter`, `hand_side_exter`, `plam_lookat` (640x480, default `default`), rendered on the
+service's GPU in a few milliseconds each; `cameras=None` means every camera, so name the ones you need.
+
+Each proprioception row (`meta["proprio"]["now"]`, `res.proprio`) has, in `spec()["proprio_layout"]` order: `q_<joint>`,
+`qd_<joint>`, `act_<muscle>`, `len_<muscle>` (m), `effort_<joint>` (the actuators' generalized force on the joint).
+
+### Action type `actuators` (275)
+275 real numbers per control step, one per actuator in order: 273 muscles and 2 non-muscle actuators. A muscle's command a maps to its excitation u = 1 / (1 + exp(-5 (a - 0.5))) and is not clipped (a = -1: u ~ 0.0006, a = 0.5: 0.5, a = 1: 0.92, a = 2: 0.9994); activation follows excitation with a first-order delay (~10 ms up, ~40 ms down); muscles only pull; one muscle can cross several joints; its force also depends on its length and velocity. The non-muscle actuators: `pelvis_x` (index 273): normalized: a in [-1, 1] maps linearly onto [-1, 0.05] (position of joint pelvis_x); `pelvis_y` (index 274): normalized: a in [-1, 1] maps linearly onto [-1, 1] (position of joint pelvis_y). Columns in `spec()["action_types"]["actuators"]["layout"]` order; rows `dt` apart, `interpolate="linear"`
+(default: a straight line from the commands in force) or `"hold"`; the client sends them as float32. The control rate
+is 100 Hz (5 physics substeps).
+
+## External material
+You may install packages you need (pip, apt). Please solve the task yourself: do not look up or download a solution to it, such as this benchmark's own source code, published or reference solutions, recorded demonstrations, or policies trained for it. This is an honour rule; trials are reviewed. The network reaches your model's API and the package registries (PyPI, Debian / Ubuntu archives) only.
+
+## Resource limits
+Limits for your agent environment, your own container only (the simulator in the `sim` service is the benchmark's and does not count): at most **8 GiB of memory**, **512 processes** and **20 GiB** of new files under `/app`, `/tmp` and `/root`; your container has no GPU. These limits may be enforced (from protocol v1.2 on): when enforcement is on, exceeding one for 60 s within any 2 minutes, or a process killed for lack of memory, ends the trial as a failure (reward 0), whatever the deliverable. A monitor outside your container measures every few seconds; the current values are in `/run/rcb-limits/usage.json`. CPU is capped at 4 cores for your container; the cap only slows you down.
+
+## Time limit
+You have 3600 seconds to complete this task. Do not cheat by using online solutions or hints specific to this task.
+
+This is wall-clock time from the start of your session; the time you wait for the simulator service counts. When it is up you are stopped without warning and the episode is graded as it stands at that moment. The service's `time_left_s` counts the same limit from your first request to the service, so it never runs out before your time does: your real time left is `time_left_s` minus the time you spent before that request. Plan by it. Your session started when you received this task: run `date` now, and again before you decide that time is short; work out the time left from those readings, never from a guess. Installing packages is not cheating (`## External material` above).
diff --git a/tasks/myosuite-table-tennis-i00-standard-realtime/task.toml b/tasks/myosuite-table-tennis-i00-standard-realtime/task.toml
new file mode 100644
index 0000000000..b33d5a4772
--- /dev/null
+++ b/tasks/myosuite-table-tennis-i00-standard-realtime/task.toml
@@ -0,0 +1,119 @@
+# Generated by scripts/myosuite/generate.py -- edit the generator, the shared sources or the instance facts, not this.
+schema_version = "1.3"
+artifacts = [{ source = "/var/rcb-monitor/", service = "rcb_monitor" },
+ { source = "/data/final.json", service = "sim" },
+ { source = "/data/episode.jsonl", service = "sim" },
+ { source = "/data/trajectory.npz", service = "sim" },
+ { source = "/data/trajectory_live.f64", service = "sim" },
+ { source = "/data/replay.mp4", service = "sim" },
+ { source = "/data/replay.last.png", service = "sim" },
+ { source = "/data/ticks.f64", service = "sim" },
+ { source = "/data/calls.jsonl", service = "sim" },
+ { source = "/data/submitted.f64", service = "sim" },
+ { source = "/data/startups.jsonl", service = "sim" },
+ "/app/output/",
+]
+
+[task]
+name = "robot-coding-bench/myosuite-table-tennis-i00-standard-realtime"
+description = "STANDARD-REALTIME MODE (robot as a service on a clock that does not stop, eai-standard/2.3, slowdown S = 1): MyoSuite table_tennis (MyoArm with torso, 273 muscles) on the same frozen instance as myosuite-table-tennis-i00-privileged: Return the served ball: hit it with the paddle so that it lands on the opponent's half of the table (MyoChallenge 2025 TableTennis P1's own success test on the ball's contact history: after the serve's bounce on your half, the paddle touches it once, then it touches the opponent's half, with no second paddle touch and no further bounce on your half). Once an episode starts the body and the ball move whatever the agent does, so the agent plans or streams actuator commands from a program of its own through the client (cameras, the body's joints and muscles); 10 resets, only the last episode graded; no object state; the sim service records every executed control step, replays the last episode in a fresh simulator and grades it."
+authors = [{ name = "Chi Gui" }]
+keywords = ["robotics", "myosuite", "mujoco", "musculoskeletal", "myoarm-torso", "trajectory", "table_tennis", "standard-realtime-mode"]
+
+[metadata]
+author_name = "Chi Gui"
+task_type = "trajectory"
+simulator = "myosuite" # MyoSuite 2.12.2 on MuJoCo 3.6.0 (CPU physics; EGL rendering on the GPU)
+robot = "myoarm-torso"
+family = "table_tennis"
+mode = "standard-realtime"
+dynamics = "dynamic" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
+slowdown = 1 # standard-realtime: the world runs at 1/S of real speed (docs/PROTOCOL.md §4b); RCB_SLOWDOWN's default
+protocol = "eai-standard/2.3"
+oracle = "none"
+oracle_evidence = "a successful episode of the privileged twin (GPT-6 Luna xhigh, trial `myosuite-table-tennis-i00-privil__SMNgbHk`, 82 rows) planned from step 0 in one `submit(..., at=0)` at S = 1 succeeds live and in the fresh replay, with no stall (board 0112, 2026-10-14); the standard twin's evidence holds for the same instance and rule; human review in PR #201"
+instance_seed = 0 # myoChallengeTableTennisP1-v0: seed(0), then reset(seed=0)
+expert_time_estimate_hours = 6.0
+
+[metadata.resources]
+# minimum compute to run this task (docs/PROTOCOL.md §2.9)
+min_cpus = 2.0
+min_memory_gb = 4.0
+min_gpu_memory_gb = 1.0
+gpu_required = true
+gpu_class = "any"
+recommended_gpu = "any NVIDIA GPU with EGL (RTX 4090 / L4)"
+verifier_min_cpus = 2.0
+verifier_min_memory_gb = 3.0
+verifier_min_gpu_memory_gb = 1.0
+measured = "as the standard twin (sim 1.2-1.4 GB, docker stats, board 0112); the GPU only renders the cameras over EGL (5 ms a frame; board 0112, 2026-10-14, ws-1); the sim keeps the GPU through the collect hook (fresh replay and video)"
+
+[metadata.resources.limits]
+# agent-phase limits in GiB (docs/PROTOCOL.md §2.10); scope: trial = every container, agent = the agent's container;
+# enforced from protocol v1.2 when RCB_LIMITS_ENFORCE=1 (a violation fails the trial), otherwise recorded only
+scope = "agent"
+max_gpu_memory_gb = 0.0
+max_memory_gb = 8.0
+max_processes = 512
+max_disk_gb = 20.0
+disk_paths = ["/app", "/tmp", "/root"]
+disk_exclude = []
+sustain_sec = 60.0
+window_sec = 120.0
+trial_max_gpu_memory_gb = 4.0
+trial_max_memory_gb = 14.0
+measured = "GPU: EGL rendering only, under 1 GiB (board 0112) -> peak + 4; memory: no samples of the agent's container; floor 8 GiB (at most the 8 GiB cap); whole trial: + the sim service's 6 GiB (docker peak 1.4 GiB, 5 samples); processes 512 and disk 20 GiB: design defaults, not yet measured"
+
+[metadata.resources.caps]
+# docker caps per compose service, applied from protocol v1.2 (docs/PROTOCOL.md §2.10)
+main = { cpus = 4.0, memory_gb = 10.0, pids = 8192 }
+sim = { cpus = 4.0, memory_gb = 8.0, pids = 8192 }
+
+[metadata.seeds]
+# the instances this task may run from (docs/PROTOCOL.md §2.13): the default, the ones a runner may draw
+# from (a list, { from = A, to = B } or "any"), and the ones the reference solution under solution/ is for
+default = 0
+allowed = [0]
+oracle = []
+
+[verifier]
+timeout_sec = 900.0
+environment_mode = "separate"
+network_mode = "no-network"
+
+# Resource limits (docs/PROTOCOL.md §2.10): seal the monitor's record first, before any other sidecar ends the episode.
+[[verifier.collect]]
+service = "rcb_monitor"
+command = "python3 /opt/rcb_monitor/monitor.py seal --wait 90"
+timeout_sec = 120
+
+# The agent's time is up: stop whatever it left running in its container. Harbor stops that container only after
+# collecting from it (~12 s), and a script still running there would keep driving the body past the deadline.
+[[verifier.collect]]
+service = "main"
+command = "kill -9 -1 2>/dev/null; exit 0"
+timeout_sec = 30
+
+# Then the episode ends: SIGUSR1 to the service (re-sent every 30 s, in case it was restarting), which grades it in a
+# fresh simulator. The verdict ("finalized": true, set before any video is rendered) may take until 840 s; then the
+# hook waits for the replay video ("video_pending") until 870 s -- a video that is not ready by then is lost, the
+# verdict is not.
+[[verifier.collect]]
+service = "sim"
+command = "t0=$(date +%s); while [ ! -f /data/server.pid ] && [ $(( $(date +%s) - t0 )) -lt 120 ]; do sleep 2; done; i=0; while [ $(( $(date +%s) - t0 )) -lt 840 ] && ! grep -q '\"finalized\": true' /data/final.json 2>/dev/null; do [ $((i % 15)) -eq 0 ] && kill -USR1 $(cat /data/server.pid 2>/dev/null) 2>/dev/null; i=$((i + 1)); sleep 2; done; while [ $(( $(date +%s) - t0 )) -lt 870 ] && grep -q '\"video_pending\": true' /data/final.json 2>/dev/null; do sleep 2; done; cat /data/final.json"
+timeout_sec = 900
+
+[agent]
+timeout_sec = 3600.0
+# The agent reaches the model APIs and the package registries, nothing else (scripts/set_agent_egress.py HOSTS), as in
+# every MyoSuite task (PROTOCOL §2.11).
+network_mode = "allowlist"
+allowed_hosts = ["chatgpt.com", "*.chatgpt.com", "auth.openai.com", "api.openai.com", "*.openai.com", "api.anthropic.com", "*.anthropic.com", "openrouter.ai", "*.openai.azure.com", "*.cognitiveservices.azure.com", "pypi.org", "files.pythonhosted.org", "deb.debian.org", "security.debian.org", "archive.ubuntu.com", "security.ubuntu.com", "snapshot.ubuntu.com"]
+
+[environment]
+build_timeout_sec = 1800.0
+cpus = 4
+memory_mb = 10240
+
+[verifier.env]
+RCB_LIMITS_ENFORCE = "${RCB_LIMITS_ENFORCE:-0}"
diff --git a/tasks/myosuite-table-tennis-i00-standard-realtime/tests/Dockerfile b/tasks/myosuite-table-tennis-i00-standard-realtime/tests/Dockerfile
new file mode 100644
index 0000000000..9cc4eeea88
--- /dev/null
+++ b/tasks/myosuite-table-tennis-i00-standard-realtime/tests/Dockerfile
@@ -0,0 +1,7 @@
+# Separate verifier sandbox: grades the sim service's own artifacts; needs nothing but python (the shared client base).
+FROM docker.io/library/python:3.12.14-slim@sha256:f77ac9e44ae96ef2c90b8053ea08c31f8be030f824196b0ae4db6d462c84e51f
+COPY test.sh /tests/test.sh
+COPY grade.py /tests/grade.py
+RUN chmod +x /tests/test.sh
+# resource limits: the verifier step and its copy of the limits (docs/PROTOCOL.md §2.10) # rcb-monitor
+COPY rcb_limits.py rcb_limits.json /tests/
diff --git a/tasks/myosuite-table-tennis-i00-standard-realtime/tests/grade.py b/tasks/myosuite-table-tennis-i00-standard-realtime/tests/grade.py
new file mode 100644
index 0000000000..b5766dbbea
--- /dev/null
+++ b/tasks/myosuite-table-tennis-i00-standard-realtime/tests/grade.py
@@ -0,0 +1,97 @@
+"""Standard-mode verifier (eai-standard/2.1): grades what the sim service -- which the agent could not touch -- wrote.
+
+/data/final.json holds the live run's outcome and the verdict of a fresh replay of the graded episode:
+success = live_success AND replay_success AND deterministic (the replay ends in the live run's final state).
+Writes /logs/verifier/reward.txt (1/0) and a numeric-only reward.json, and copies the evidence (replay video, last
+frame, trajectory, event log) next to them for humans; the evidence never decides the verdict.
+infra_error = 1: no verdict on the agent (the service crashed past its restarts, never finished, or failed to finalize).
+Every branch writes success, deterministic, infra_error and grader_error (PROTOCOL §2.6); sim_error = 1 when the graded
+episode ended on a simulator error (reason "error").
+A standard-realtime service's final.json (eai-standard/2.3) also says how well its clock was kept; reward.json repeats
+it (REALTIME below; PROTOCOL §4b), and none of it decides the verdict.
+"""
+import json
+import math
+import os
+import shutil
+import traceback
+
+DATA = os.environ.get("RCB_DATA", "/data")
+LOG = os.environ.get("RCB_VERIFIER_LOG_DIR", "/logs/verifier")
+os.makedirs(LOG, exist_ok=True)
+
+
+def write(ok: bool, out: dict) -> None:
+ with open(os.path.join(LOG, "reward.txt"), "w") as f:
+ f.write("1\n" if ok else "0\n")
+ with open(os.path.join(LOG, "reward.json"), "w") as f:
+ json.dump(out, f)
+ print(json.dumps(out))
+
+
+def num(x, default: float = 0.0) -> float:
+ """A finite number, or the default (reward.json takes numbers only)."""
+ try:
+ x = float(x)
+ except (TypeError, ValueError):
+ return default
+ return x if math.isfinite(x) else default
+
+
+# Standard-realtime only: slowdown S; clock_s, the wall seconds the clock ran; max_lag_s, the furthest a control step ran
+# behind its time; stalls, the times the service dropped more than 0.5 s of debt instead of fast-forwarding the robot;
+# stalled_s, the wall time dropped that way (the robot stood still while its clock should have run); stalled_share =
+# stalled_s / clock_s; held_steps and late_steps; requests_refused (answered 4xx) and request_busy_s (simulator-thread
+# time spent on requests). A field final.json does not have is 0.
+REALTIME = {"slowdown": float, "clock_s": float, "max_lag_s": float, "stalls": int, "stalled_s": float,
+ "stalled_share": float, "held_steps": int, "late_steps": int, "requests_refused": int,
+ "request_busy_s": float}
+
+
+def realtime_fields(d: dict) -> dict:
+ if not (d.get("realtime") or str(d.get("protocol")) == "eai-standard/2.3"):
+ return {}
+ return {"realtime": 1, **{k: kind(num(d.get(k))) for k, kind in REALTIME.items()}}
+
+
+def main() -> None:
+ path = os.path.join(DATA, "final.json")
+ if not os.path.exists(path):
+ write(False, {"success": 0, "deterministic": 0, "missing_log": 1, "infra_error": 1, "sim_error": 0,
+ "grader_error": 0})
+ return
+ try:
+ with open(path) as f:
+ d = json.load(f)
+ if not isinstance(d, dict):
+ raise ValueError("final.json is not a JSON object")
+ live, replay, det = (bool(d.get(k)) for k in ("live_success", "replay_success", "deterministic"))
+ failed_finalize = bool(d.get("finalize_error"))
+ unfinished = not d.get("finalized")
+ infra = bool(d.get("infra_error")) or failed_finalize or unfinished
+ ok = live and replay and det and not infra
+ out = {"success": int(ok), "live_success": int(live), "replay_success": int(replay), "deterministic": int(det),
+ "finalized": int(not unfinished), "missing_log": 0, "infra_error": int(infra),
+ "grader_error": int(failed_finalize), "sim_error": int(d.get("reason") == "error"),
+ "sim_crashed_on_call": int(d.get("reason") == "sim_crashed_on_call"),
+ "n_actions": int(d.get("n_actions") or 0), "sim_steps": int(d.get("control_steps_total") or 0),
+ "calls": int(d.get("calls") or 0), "rows": int(d.get("rows") or 0), "resets": int(d.get("resets") or 0),
+ "reset_budget": int(d.get("reset_budget") or 0),
+ "sim_startups": int(d.get("sim_startups") or 0), "elapsed_s": float(d.get("elapsed_s") or 0.0),
+ "video_rendered": int(bool(d.get("video"))), **realtime_fields(d)}
+ except Exception: # noqa: BLE001 -- an unreadable final.json is the grader's failure, not the agent's
+ traceback.print_exc()
+ write(False, {"success": 0, "deterministic": 0, "missing_log": 0, "infra_error": 1, "sim_error": 0,
+ "grader_error": 1})
+ return
+ for name in ("replay.mp4", "replay.last.png", "trajectory.npz", "episode.jsonl"):
+ try:
+ if os.path.exists(os.path.join(DATA, name)):
+ shutil.copy(os.path.join(DATA, name), os.path.join(LOG, name))
+ except OSError:
+ traceback.print_exc()
+ write(ok, out)
+
+
+if __name__ == "__main__":
+ main()
diff --git a/tasks/myosuite-table-tennis-i00-standard-realtime/tests/rcb_limits.json b/tasks/myosuite-table-tennis-i00-standard-realtime/tests/rcb_limits.json
new file mode 100644
index 0000000000..fa8f1b1a62
--- /dev/null
+++ b/tasks/myosuite-table-tennis-i00-standard-realtime/tests/rcb_limits.json
@@ -0,0 +1,38 @@
+{
+ "caps": {
+ "main": {
+ "cpus": 4.0,
+ "memory_gb": 10.0,
+ "pids": 8192
+ },
+ "sim": {
+ "cpus": 4.0,
+ "memory_gb": 8.0,
+ "pids": 8192
+ }
+ },
+ "disk_exclude": [],
+ "disk_paths": [
+ "/app",
+ "/tmp",
+ "/root"
+ ],
+ "disk_tick_sec": 60,
+ "format": "rcb-limits/1",
+ "gpu": false,
+ "max_disk_gb": 20.0,
+ "max_gpu_memory_gb": 0.0,
+ "max_memory_gb": 8.0,
+ "max_processes": 512,
+ "mode": "standard-realtime",
+ "scope": "agent",
+ "scope_services": [
+ "main"
+ ],
+ "sustain_sec": 60.0,
+ "task": "myosuite-table-tennis-i00-standard-realtime",
+ "tick_sec": 5,
+ "trial_max_gpu_memory_gb": 4.0,
+ "trial_max_memory_gb": 14.0,
+ "window_sec": 120.0
+}
diff --git a/tasks/myosuite-table-tennis-i00-standard-realtime/tests/rcb_limits.py b/tasks/myosuite-table-tennis-i00-standard-realtime/tests/rcb_limits.py
new file mode 100644
index 0000000000..a979bcdaff
--- /dev/null
+++ b/tasks/myosuite-table-tennis-i00-standard-realtime/tests/rcb_limits.py
@@ -0,0 +1,457 @@
+#!/usr/bin/env python3
+"""Resource limits (docs/PROTOCOL.md §2.10): the sustained-violation rule, shared by the monitor sidecar (monitor.py) and
+the verifier step, and the verifier step itself.
+
+Copies of this file live in every task (environment/rcb_monitor/ for the monitor, tests/ for the verifier), byte-identical
+to scripts/rcb_monitor/rcb_limits.py (ci_checks/check-monitor.py). Python 3.8+, standard library only: it runs inside
+benchmark verifier images.
+
+ python3 /tests/rcb_limits.py verify # test.sh runs it on exit (trap), after the grader
+ python3 /tests/rcb_limits.py sanitize /app/output # test.sh runs it before the grader (see sanitize())
+
+The verifier step reads the monitor's sealed record (/var/rcb-monitor/, collected from the rcb_monitor sidecar, a
+channel the agent cannot write), recomputes every violation from the raw samples with its own copy of the limits
+(/tests/rcb_limits.json), and adds to /logs/verifier/reward.json:
+ always: resource_enforced, limits_violation_observed, resource_monitor_missing, peak_gpu_memory_gb, peak_memory_gb,
+ peak_processes, peak_disk_gb, resource_emergency_kills; over_<metric>=1 for every violation; the kernel's
+ OOM kills in the record: resource_oom_kills (in the scope, at the container's own memory cap: the
+ violation), resource_host_oom_kills (in the scope, not at its cap: the host's OOM killer or an enclosing
+ cgroup's), resource_service_oom_kills (in a benchmark-owned service that is only recorded: the simulator
+ sidecar of standard modes, RoboPaint's scorer)
+ enforced (the record's mode, RCB_LIMITS_ENFORCE=1) and violated: resource_limit_exceeded=1, success=0 (reward.txt 0),
+ success_before_limits = the grader's success
+ enforced and no usable record: infra_error=1 (not a result). Not enforced (the default): a missing or unusable
+ record is only reported (resource_monitor_missing=1); the grader's verdict stands.
+ enforced, no violation, and a process in the scope was killed by an OOM killer that is not its own cap's:
+ infra_error=1, resource_host_oom=1 (not a result: another trial's or the host's doing, never the agent's)
+ RCB_LIMITS_ENFORCE neither 0/1/true/false: infra_error=1, resource_switch_error=1 (nobody knows what was asked for)
+ in both modes: the kernel OOM-killed a benchmark-owned service and the grader's verdict is not a success:
+ infra_error=1, resource_service_oom=1 (PROTOCOL §2.10: such a service is recorded, never a failure; an
+ episode that ended without success after its simulator was killed for memory is not the agent's 0). A
+ success stands: the service recovered and the grader accepted the episode.
+Wherever this step says "not a result" (infra_error=1) it also sets success=0 and reward.txt 0 and keeps the grader's
+verdict in success_before_limits, so that no reader sees a success next to an infrastructure error.
+With the default (not enforced, a readable switch) the step changes nothing of the grader's verdict (reward.txt,
+success, reward) and sets infra_error only in the one case above (a benchmark-owned service OOM-killed, no success).
+A violation is a limit exceeded for sustain_sec within window_sec (disk: two measurements in a row), or a kernel OOM kill
+in the scope at the container's own memory cap (the monitor's `oom` event with cause "cap"). The monitor's emergency
+kills (host protection) are counted, never violations by themselves.
+The reasons, in words, go to /logs/verifier/resource_limits.json. If this step itself fails, it says so in reward.json
+(resource_step_error=1, resource_monitor_missing=1; infra_error=1 when the switch asks for enforcement).
+"""
+import json
+import os
+import stat
+import sys
+
+FORMAT = "rcb-monitor/1"
+RECORD = os.environ.get("RCB_MONITOR_RECORD", "/var/rcb-monitor")
+METRICS = ("gpu_memory_gb", "memory_gb", "processes", "disk_gb")
+LIMIT_KEY = {"gpu_memory_gb": "max_gpu_memory_gb", "memory_gb": "max_memory_gb", "processes": "max_processes",
+ "disk_gb": "max_disk_gb"}
+MAX_GAP_FRACTION = 0.10 # more of the agent phase than this unobserved: the record is not usable
+
+
+def parse_switch(value):
+ """RCB_LIMITS_ENFORCE -> (1 or 0, None), or (0, why) for a value that is neither on nor off (never a silent off)."""
+ s = str("" if value is None else value).strip().lower()
+ if s in ("", "0", "false"):
+ return 0, None
+ if s in ("1", "true"):
+ return 1, None
+ return 0, f"RCB_LIMITS_ENFORCE={value!r} is not one of 0, 1, true, false"
+
+
+def sustained(series, limit, sustain_sec, window_sec, tick_sec):
+ """First time a sustained violation is confirmed, or None.
+
+ series: [(t, value)] in time order (value None = not measured: never over). Each sample stands for the time to the
+ next one (the last for none), at most 3 ticks (a gap in the record never counts against the agent). A violation is confirmed at the
+ first moment when the time over `limit` within the last `window_sec` reaches `sustain_sec` (so going under for a
+ moment does not reset it)."""
+ if limit is None or limit <= 0:
+ return None
+ iv = [] # (start, end) of the time over the limit
+ for i, (t, v) in enumerate(series):
+ if i + 1 == len(series):
+ break # the last sample has not lasted yet
+ d = max(0.0, min(series[i + 1][0] - t, 3 * tick_sec))
+ if v is not None and v > limit and d > 0:
+ if iv and abs(iv[-1][1] - t) < 1e-9:
+ iv[-1] = (iv[-1][0], t + d)
+ else:
+ iv.append((t, t + d))
+ if not iv:
+ return None
+ # the over-time in a window (e - W, e] only grows while e moves through an interval: check each interval's end,
+ # and inside the first one that reaches it, bisect for the exact moment
+ def over_in(e):
+ """Over-time inside (e - window, e]."""
+ s = e - window_sec
+ tot = 0.0
+ for a, b in iv:
+ if b <= s:
+ continue
+ if a >= e:
+ break
+ tot += min(b, e) - max(a, s)
+ return tot
+
+ for a, b in iv:
+ if over_in(b) + 1e-9 >= sustain_sec:
+ # bisect for the first moment in [a, b] at which it is reached
+ x, y = a, b
+ for _ in range(40):
+ m = (x + y) / 2
+ if over_in(m) + 1e-9 >= sustain_sec:
+ y = m
+ else:
+ x = m
+ return y
+ return None
+
+
+def consecutive(series, limit, n=2):
+ """First time `n` consecutive measured samples are over `limit` (disk: measured once a minute), or None."""
+ if limit is None or limit <= 0:
+ return None
+ run = 0
+ for t, v in series:
+ if v is None:
+ continue
+ run = run + 1 if v > limit else 0
+ if run >= n:
+ return t
+ return None
+
+
+def violations(samples, limits, events=()):
+ """Every violation in a record: [{metric, limit, at, peak}] (empty: none). samples: the record's sample dicts.
+ A limit exceeded for sustain_sec within window_sec (disk: two measurements in a row), or a kernel OOM kill in the
+ scope at the container's own memory cap (cause "cap": the cgroup's own `oom` counter rose with the kill). A kill
+ by the host's OOM killer (any other cause) is not the trial's doing and never a violation: oom_kills() counts it.
+ The monitor's own `emergency_kill` events protect the host and are not violations (PROTOCOL §2.10)."""
+ tick = float(limits.get("tick_sec", 5))
+ sustain = float(limits.get("sustain_sec", 60))
+ window = float(limits.get("window_sec", 2 * sustain))
+ out = []
+ for m in METRICS:
+ lim = limits.get(LIMIT_KEY[m])
+ if lim is None:
+ continue
+ lim = float(lim)
+ ser = [(s["t"], (s.get("scope") or {}).get(m)) for s in samples]
+ at = consecutive(ser, lim) if m == "disk_gb" else sustained(ser, lim, sustain, window, tick)
+ if at is not None:
+ peak = max((v for _, v in ser if v is not None), default=None)
+ out.append({"metric": m, "limit": lim, "at": round(at, 2), "peak": peak})
+ for e in events:
+ if isinstance(e, dict) and e.get("event") == "oom" and e.get("scope") and e.get("cause") == "cap":
+ out.append({"metric": "oom", "limit": None, "at": e.get("t"), "peak": e.get("count"),
+ "service": e.get("role")})
+ return sorted(out, key=lambda v: v["at"] if isinstance(v["at"], (int, float)) else 0)
+
+
+def oom_kills(events):
+ """The kernel's OOM kills in a record: {"cap": in the scope at the container's own memory cap (the violation),
+ "host": in the scope by another OOM killer (the host's, an enclosing cgroup's; also an event that does not say),
+ "service": in a benchmark-owned service that is only recorded (whatever the cause)}."""
+ out = {"cap": 0, "host": 0, "service": 0}
+ for e in events:
+ if not isinstance(e, dict) or e.get("event") != "oom":
+ continue
+ n = e.get("count")
+ n = int(n) if isinstance(n, (int, float)) and not isinstance(n, bool) and n == n and 1 <= n < 1e9 else 1
+ out["service" if not e.get("scope") else "cap" if e.get("cause") == "cap" else "host"] += n
+ return out
+
+
+def peaks(samples):
+ p = {}
+ for m in METRICS:
+ vals = [(s.get("scope") or {}).get(m) for s in samples]
+ vals = [v for v in vals if v is not None]
+ p[m] = round(max(vals), 3) if vals else 0.0
+ return p
+
+
+def read_jsonl(path):
+ out = []
+ try:
+ with open(path) as f:
+ for line in f:
+ line = line.strip()
+ if line:
+ try:
+ out.append(json.loads(line))
+ except ValueError:
+ pass # a torn last line (the monitor was killed mid-write)
+ except OSError:
+ pass
+ return out
+
+
+def read_json(path):
+ try:
+ with open(path) as f:
+ return json.load(f)
+ except (OSError, ValueError):
+ return None
+
+
+def load_record(record=None):
+ """(summary or None, samples, events)."""
+ record = record or os.environ.get("RCB_MONITOR_RECORD", RECORD)
+ return (read_json(os.path.join(record, "summary.json")), read_jsonl(os.path.join(record, "samples.jsonl")),
+ [e for e in read_jsonl(os.path.join(record, "events.jsonl")) if isinstance(e, dict)])
+
+
+def record_problems(summary, samples, events, limits):
+ """Why a record cannot be used ([] = usable)."""
+ why = []
+ if not summary:
+ why.append("no sealed record (/var/rcb-monitor/summary.json missing: the monitor did not run, crashed, or was not collected)")
+ elif summary.get("format") != FORMAT:
+ why.append(f"record format {summary.get('format')!r}, expected {FORMAT}")
+ if not samples:
+ why.append("no samples")
+ fatal = [e for e in events if e.get("event") == "fatal"]
+ if fatal:
+ why.append("monitor fatal: " + "; ".join(str(e.get("why")) for e in fatal)[:300])
+ if samples and summary and not summary.get("scope_seen"):
+ why.append("the monitor never found the containers it watches (no beacon seen)")
+ if float(limits.get("max_gpu_memory_gb") or 0) > 0 and samples:
+ blind = sum(1 for s in samples if (s.get("scope") or {}).get("gpu_unmeasured"))
+ if blind / len(samples) > MAX_GAP_FRACTION:
+ why.append(f"NVML did not tell the GPU memory of processes in scope in {blind} of {len(samples)} samples "
+ "(MIG / vGPU?): the GPU-memory limit was not measured")
+ if len(samples) >= 2:
+ tick = float(limits.get("tick_sec", 5))
+ span = samples[-1]["t"] - samples[0]["t"]
+ gaps = sum(max(0.0, b["t"] - a["t"] - 3 * tick) for a, b in zip(samples, samples[1:]))
+ if span > 0 and gaps / span > MAX_GAP_FRACTION:
+ why.append(f"{gaps:.0f} s of {span:.0f} s unobserved")
+ return why
+
+
+def switch_of(summary):
+ """(enforce, why the switch cannot be read or None): the record's mode when there is a record, else this step's
+ own environment (the same run-time variable, task.toml [verifier.env])."""
+ if summary and summary.get("switch_error"):
+ return 0, str(summary["switch_error"])
+ if summary and "enforce" in summary:
+ return parse_switch(summary.get("enforce"))
+ return parse_switch(os.environ.get("RCB_LIMITS_ENFORCE"))
+
+
+def not_a_result(reward, add):
+ """This step found that the trial is not a result (infra_error=1): no success is left beside it. The grader's
+ verdict stays readable in success_before_limits."""
+ add["infra_error"] = 1
+ add["success_before_limits"] = int(bool(reward.get("success", 0)))
+ add["success"] = 0
+ if "reward" in reward:
+ add["reward"] = 0
+
+
+def verify(argv=()):
+ log = os.environ.get("RCB_VERIFIER_LOGS", "/logs/verifier")
+ tests = os.path.dirname(os.path.abspath(__file__))
+ limits_file = read_json(os.path.join(tests, "rcb_limits.json"))
+ limits = limits_file or {}
+ summary, samples, events = load_record()
+ enforce, switch_error = switch_of(summary)
+ rj_path, rt_path = os.path.join(log, "reward.json"), os.path.join(log, "reward.txt")
+ reward = read_json(rj_path)
+ have_grade = isinstance(reward, dict)
+ reward = reward if have_grade else {}
+ problems = record_problems(summary, samples, events, limits)
+ if limits_file is None:
+ problems.append("the verifier's copy of the limits (tests/rcb_limits.json) is missing or unreadable")
+ if switch_error and not any(switch_error in p for p in problems):
+ problems.append(switch_error)
+ viol = violations(samples, limits, events) if samples else []
+ pk = peaks(samples)
+ emergency = [e for e in events if e.get("event") == "emergency_kill"]
+ oom = oom_kills(events)
+ detail = {"format": FORMAT, "enforced": enforce, "limits": {k: v for k, v in limits.items()},
+ "peaks": pk, "violations": viol, "record_problems": problems, "emergency_kills": emergency,
+ "oom_kills": [e for e in events if e.get("event") == "oom"],
+ "monitor": {k: (summary or {}).get(k) for k in ("started", "ended", "n_samples", "capabilities", "killed",
+ "restarts", "disk_baseline_gb", "disk_partial_walks")}}
+ add = {"resource_enforced": enforce, "limits_violation_observed": int(bool(viol)),
+ "resource_monitor_missing": int(bool(problems)),
+ "peak_gpu_memory_gb": pk["gpu_memory_gb"], "peak_memory_gb": pk["memory_gb"],
+ "peak_processes": pk["processes"], "peak_disk_gb": pk["disk_gb"],
+ "resource_emergency_kills": len(emergency), "resource_oom_kills": oom["cap"],
+ "resource_host_oom_kills": oom["host"], "resource_service_oom_kills": oom["service"]}
+ for v in viol:
+ add["over_" + v["metric"]] = 1
+ failed = void = False # void: this step says "not a result" (infra_error=1) and takes the success away
+ if switch_error:
+ # nobody knows whether enforcement was asked for: not a result in either mode
+ add["resource_switch_error"] = 1
+ void = True
+ detail["verdict"] = "not a result: " + switch_error
+ elif enforce and viol:
+ # a confirmed violation fails the trial whatever the deliverable (and wins over a grader's infra flag: the
+ # agent's processes were stopped for it)
+ failed = True
+ add["success_before_limits"] = int(reward.get("success", 0) or 0)
+ add["resource_limit_exceeded"] = 1
+ add["success"] = 0
+ if "reward" in reward:
+ add["reward"] = 0
+ if reward.get("infra_error"):
+ add["grader_infra_error"] = 1
+ add["infra_error"] = 0
+ detail["verdict"] = "failed: resource limit exceeded (" + ", ".join(v["metric"] for v in viol) + ")"
+ elif enforce and problems:
+ void = True
+ detail["verdict"] = "not a result: the resource monitor's record is not usable"
+ elif enforce and oom["host"]:
+ # a process the limits bind was killed for memory, but not at its container's own cap: the host ran out of
+ # memory (other trials), or an enclosing cgroup did. Never the agent's failure, and not a clean run either.
+ add["resource_host_oom"] = 1
+ void = True
+ detail["verdict"] = ("not a result: an out-of-memory killer other than the container's own cap (the host's) "
+ f"killed {oom['host']} process(es) of the trial")
+ elif oom["service"] and have_grade and not reward.get("success"):
+ # in both modes: the kernel killed a benchmark-owned service (the simulator sidecar, a scorer) for memory and
+ # the episode did not end in a success. Such a service is recorded, never a failure (PROTOCOL §2.10): whatever
+ # the grader made of the crash, this is not the agent's 0.
+ add["resource_service_oom"] = 1
+ add["infra_error"] = 1
+ detail["verdict"] = (f"not a result: the kernel OOM-killed a benchmark-owned service ({oom['service']} kill(s)) "
+ "and the episode did not end in a success")
+ else:
+ # not enforced: observe only. A missing or unusable record is reported (resource_monitor_missing), never infra.
+ detail["verdict"] = ("observed only (RCB_LIMITS_ENFORCE=0): " + ", ".join(v["metric"] for v in viol)) if viol \
+ else "within limits" if not problems else "observed only (RCB_LIMITS_ENFORCE=0): no usable record"
+ if void:
+ not_a_result(reward, add)
+ os.makedirs(log, exist_ok=True)
+ if have_grade or failed:
+ reward.update(add)
+ reward.setdefault("success", 0)
+ tmp = rj_path + ".tmp"
+ with open(tmp, "w") as f:
+ json.dump(reward, f, indent=1)
+ os.replace(tmp, rj_path)
+ if failed or void:
+ with open(rt_path, "w") as f:
+ f.write("0\n")
+ with open(os.path.join(log, "resource_limits.json"), "w") as f:
+ json.dump(detail, f, indent=1)
+ print("resource limits: " + detail["verdict"], file=sys.stderr)
+ return 0
+
+
+def step_failed(err):
+ """This step raised: say so in the grader's reward.json instead of leaving the trial without its resource fields
+ (and, when the switch asks for enforcement or cannot be read, as an infrastructure error: nothing was judged)."""
+ log = os.environ.get("RCB_VERIFIER_LOGS", "/logs/verifier")
+ rj_path = os.path.join(log, "reward.json")
+ reward = read_json(rj_path)
+ enforce, switch_error = parse_switch(os.environ.get("RCB_LIMITS_ENFORCE"))
+ add = {"resource_step_error": 1, "resource_monitor_missing": 1, "resource_enforced": enforce}
+ if isinstance(reward, dict):
+ if enforce or switch_error:
+ not_a_result(reward, add)
+ reward.update(add)
+ tmp = rj_path + ".tmp"
+ with open(tmp, "w") as f:
+ json.dump(reward, f, indent=1)
+ os.replace(tmp, rj_path)
+ if add.get("infra_error"):
+ with open(os.path.join(log, "reward.txt"), "w") as f:
+ f.write("0\n")
+ with open(os.path.join(log, "resource_limits.json"), "w") as f:
+ json.dump({"format": FORMAT, "enforced": enforce, "verdict": "the verifier's resource step failed: " + err,
+ "record_problems": [err]}, f, indent=1)
+
+
+MAX_SESSION_LOG = 64 * 2 ** 20 # bytes of /app/output/rcb_session_log.jsonl a grader is given to read
+
+
+def sanitize(roots=("/app/output",), max_log=MAX_SESSION_LOG):
+ """Before the grader: the files the agent left (PROTOCOL §2.6: collected from its container as they are, links and
+ all) are untrusted input, and a grader opens some of them by name. Remove what is not a file of the agent's own
+ under the root: a symbolic link that leaves it or does not end at a regular file (`rcb_session_log.jsonl ->
+ /dev/zero` made a grader read until the verifier ran out of memory or time: no verdict), and every FIFO, device and
+ socket (opening a FIFO blocks). A session log larger than `max_log` is cut there (graders read it line by line,
+ and one endless line is as bad). Regular files, directories and links to regular files inside the root stay as
+ they are: nothing a grader could have read safely changes. Returns what it removed or cut (also printed)."""
+ done = []
+ for root in roots:
+ real_root = os.path.realpath(root)
+ if not os.path.isdir(real_root) or os.path.islink(root):
+ if os.path.islink(root): # the root itself must be a directory of its own
+ try:
+ os.unlink(root)
+ done.append(f"removed {root}: a symbolic link")
+ except OSError:
+ pass
+ continue
+ for dirpath, dirnames, filenames in os.walk(real_root, followlinks=False):
+ for name in list(dirnames) + filenames:
+ path = os.path.join(dirpath, name)
+ try:
+ st = os.lstat(path)
+ if stat.S_ISLNK(st.st_mode):
+ target = os.path.realpath(path)
+ inside = target.startswith(real_root + os.sep)
+ if inside and stat.S_ISREG(os.stat(target).st_mode):
+ continue
+ why = "a symbolic link that leaves the directory or does not end at a regular file"
+ elif stat.S_ISREG(st.st_mode) or stat.S_ISDIR(st.st_mode):
+ continue
+ else:
+ why = "not a regular file (FIFO, device or socket)"
+ os.unlink(path)
+ if name in dirnames:
+ dirnames.remove(name)
+ done.append(f"removed {path}: {why}")
+ except OSError as e: # a dangling link, a file that is gone: remove what can be removed
+ try:
+ if stat.S_ISLNK(os.lstat(path).st_mode):
+ os.unlink(path)
+ done.append(f"removed {path}: a symbolic link that cannot be resolved ({type(e).__name__})")
+ except OSError:
+ pass
+ log = os.path.join(real_root, "rcb_session_log.jsonl")
+ try:
+ if os.path.isfile(log) and not os.path.islink(log) and os.path.getsize(log) > max_log:
+ os.truncate(log, max_log)
+ done.append(f"cut {log} to {max_log} bytes")
+ except OSError:
+ pass
+ for d in done:
+ print("verifier input: " + d, file=sys.stderr)
+ return done
+
+
+def run_step(argv=()):
+ """verify(), and whatever it raises reported instead of breaking the grader's own verdict. Always 0."""
+ try:
+ verify(argv)
+ except Exception as e:
+ msg = f"{type(e).__name__}: {e}"
+ print(f"resource limits: verifier step failed: {msg}", file=sys.stderr)
+ try:
+ step_failed(msg)
+ except Exception:
+ pass
+ return 0
+
+
+if __name__ == "__main__":
+ if sys.argv[1:2] == ["verify"]:
+ sys.exit(run_step(sys.argv[2:]))
+ if sys.argv[1:2] == ["sanitize"]:
+ try:
+ sanitize(sys.argv[2:] or ("/app/output",))
+ except Exception as e: # never keep the grader from running
+ print(f"verifier input: sanitize failed: {type(e).__name__}: {e}", file=sys.stderr)
+ sys.exit(0)
+ print(__doc__)
diff --git a/tasks/myosuite-table-tennis-i00-standard-realtime/tests/test.sh b/tasks/myosuite-table-tennis-i00-standard-realtime/tests/test.sh
new file mode 100755
index 0000000000..7391822123
--- /dev/null
+++ b/tasks/myosuite-table-tennis-i00-standard-realtime/tests/test.sh
@@ -0,0 +1,9 @@
+#!/bin/bash
+# Verifier (separate sandbox): the sim sidecar's artifacts were uploaded to /data by Harbor.
+# Harbor also mounts /logs/verifier into the agent's container: start from an empty directory, so nothing the agent
+# left there (a planted reward, a symlink) can count. rm never follows symlinks.
+rm -rf /logs/verifier/* # verdicts from an earlier run never survive (docs/PROTOCOL.md §2.6)
+trap 'p=$(command -v python3 || command -v python); "$p" /tests/rcb_limits.py verify' EXIT # resource limits, after the grader (docs/PROTOCOL.md §2.10)
+p=$(command -v python3 || command -v python); "$p" /tests/rcb_limits.py sanitize /app/output || true # the agent's files are untrusted input: no link out of /app/output, no FIFO or device (docs/PROTOCOL.md §2.6)
+mkdir -p /logs/verifier
+python /tests/grade.py 2> /logs/verifier/grade.log
diff --git a/tasks/myosuite-table-tennis-i00-standard/environment/sim/rcb_myosuite.py b/tasks/myosuite-table-tennis-i00-standard/environment/sim/rcb_myosuite.py
index 0f2fca5acd..930c0c73ba 100644
--- a/tasks/myosuite-table-tennis-i00-standard/environment/sim/rcb_myosuite.py
+++ b/tasks/myosuite-table-tennis-i00-standard/environment/sim/rcb_myosuite.py
@@ -465,6 +465,12 @@ class Session:
self.done, self.success, self.reason = True, True, "success"
elif failed:
self.done, self.reason = True, "failure: " + _FAILURE.get(self.task, "task terminated")
+ # Baoding's targets come from a precomputed trajectory of 1000 points (baoding_v1.create_goal_trajectory,
+ # "assumes that it's greater than env horizon"): with MyoSuite's step budget lifted, step 1001 would index past
+ # it (IndexError, board 0086). The episode ends there instead, as a failure unless it succeeded on that step.
+ n_goal = len(getattr(self.base, "goal", ())) if self.task == "baoding" else 0
+ if not self.done and n_goal and self.steps >= n_goal:
+ self.done, self.reason = True, f"failure: the targets' trajectory ends after {n_goal} control steps"
return not self.done
def _bimanual_contacts_ok(self) -> bool:
diff --git a/tasks/myosuite-table-tennis-i00-standard/task.toml b/tasks/myosuite-table-tennis-i00-standard/task.toml
index 7231144eb7..2ac1d8f097 100644
--- a/tasks/myosuite-table-tennis-i00-standard/task.toml
+++ b/tasks/myosuite-table-tennis-i00-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "myosuite" # MyoSuite 2.12.2 on MuJoCo 3.6.0 (CPU)
robot = "myoarm-torso"
family = "table_tennis"
mode = "standard"
+dynamics = "dynamic" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "the privileged twin's positive example (Claude Code + Claude Opus 5.5 (run `claude_code-claude_opus_5_5-default`, 2026-10-06), trial `myosuite-table-tennis-i00-privil__kdQdy89`, 111 rows) rounded to float32, as the client sends it, and streamed through the client: the service scores it 1 live and in both replays, in two fresh trials (docs/benchmarks/myosuite.md); human review in PR #65"
diff --git a/tasks/pandadyn-d01-reciprocal-cup-relay-i00-privileged/task.toml b/tasks/pandadyn-d01-reciprocal-cup-relay-i00-privileged/task.toml
index d6aa54731f..3ae2465783 100644
--- a/tasks/pandadyn-d01-reciprocal-cup-relay-i00-privileged/task.toml
+++ b/tasks/pandadyn-d01-reciprocal-cup-relay-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "maniskill" # Panda Dynamic Scenes on ManiSkill 3.0.1 / SA
robot = "franka-panda-bimanual"
family = "d01_reciprocal_cup_relay"
mode = "privileged"
+dynamics = "dynamic" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 0
expert_time_estimate_hours = 3.0
diff --git a/tasks/pandadyn-d01-reciprocal-cup-relay-i00-standard/task.toml b/tasks/pandadyn-d01-reciprocal-cup-relay-i00-standard/task.toml
index b7b09cab06..20613e8ff9 100644
--- a/tasks/pandadyn-d01-reciprocal-cup-relay-i00-standard/task.toml
+++ b/tasks/pandadyn-d01-reciprocal-cup-relay-i00-standard/task.toml
@@ -28,6 +28,7 @@ simulator = "maniskill"
robot = "franka-panda-bimanual"
family = "d01_reciprocal_cup_relay"
mode = "standard" # robot as a service: RGB, depth + proprioception only; no object poses, no success predicate, no reset
+dynamics = "dynamic" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "full"
instance_seed = 0
diff --git a/tasks/pandadyn-d02-strike-catch-merge-i00-privileged/task.toml b/tasks/pandadyn-d02-strike-catch-merge-i00-privileged/task.toml
index 609e502d03..df87dbddf7 100644
--- a/tasks/pandadyn-d02-strike-catch-merge-i00-privileged/task.toml
+++ b/tasks/pandadyn-d02-strike-catch-merge-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "maniskill" # Panda Dynamic Scenes on ManiSkill 3.0.1 / SA
robot = "franka-panda-bimanual"
family = "d02_strike_catch_merge"
mode = "privileged"
+dynamics = "dynamic" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 0
expert_time_estimate_hours = 3.0
diff --git a/tasks/pandadyn-d02-strike-catch-merge-i00-standard/task.toml b/tasks/pandadyn-d02-strike-catch-merge-i00-standard/task.toml
index 4bada5e1e0..313a1051e2 100644
--- a/tasks/pandadyn-d02-strike-catch-merge-i00-standard/task.toml
+++ b/tasks/pandadyn-d02-strike-catch-merge-i00-standard/task.toml
@@ -28,6 +28,7 @@ simulator = "maniskill"
robot = "franka-panda-bimanual"
family = "d02_strike_catch_merge"
mode = "standard" # robot as a service: RGB, depth + proprioception only; no object poses, no success predicate, no reset
+dynamics = "dynamic" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "full"
instance_seed = 0
diff --git a/tasks/pandadyn-d03-cooperative-puck-course-i00-privileged/task.toml b/tasks/pandadyn-d03-cooperative-puck-course-i00-privileged/task.toml
index c7d291aa84..d64fc15a9a 100644
--- a/tasks/pandadyn-d03-cooperative-puck-course-i00-privileged/task.toml
+++ b/tasks/pandadyn-d03-cooperative-puck-course-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "maniskill" # Panda Dynamic Scenes on ManiSkill 3.0.1 / SA
robot = "franka-panda-bimanual"
family = "d03_cooperative_puck_course"
mode = "privileged"
+dynamics = "dynamic" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 0
expert_time_estimate_hours = 3.0
diff --git a/tasks/pandadyn-d03-cooperative-puck-course-i00-standard/task.toml b/tasks/pandadyn-d03-cooperative-puck-course-i00-standard/task.toml
index 94ee9e9843..674365e0d1 100644
--- a/tasks/pandadyn-d03-cooperative-puck-course-i00-standard/task.toml
+++ b/tasks/pandadyn-d03-cooperative-puck-course-i00-standard/task.toml
@@ -28,6 +28,7 @@ simulator = "maniskill"
robot = "franka-panda-bimanual"
family = "d03_cooperative_puck_course"
mode = "standard" # robot as a service: RGB, depth + proprioception only; no object poses, no success predicate, no reset
+dynamics = "dynamic" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "full"
instance_seed = 0
diff --git a/tasks/pandadyn-d04-shared-tray-transport-discharge-i00-privileged/task.toml b/tasks/pandadyn-d04-shared-tray-transport-discharge-i00-privileged/task.toml
index f7af414693..b64762a894 100644
--- a/tasks/pandadyn-d04-shared-tray-transport-discharge-i00-privileged/task.toml
+++ b/tasks/pandadyn-d04-shared-tray-transport-discharge-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "maniskill" # Panda Dynamic Scenes on ManiSkill 3.0.1 / SA
robot = "franka-panda-bimanual"
family = "d04_shared_tray_transport_discharge"
mode = "privileged"
+dynamics = "dynamic" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 1
expert_time_estimate_hours = 3.0
diff --git a/tasks/pandadyn-d04-shared-tray-transport-discharge-i00-standard/task.toml b/tasks/pandadyn-d04-shared-tray-transport-discharge-i00-standard/task.toml
index a0ce85d441..3c128f797d 100644
--- a/tasks/pandadyn-d04-shared-tray-transport-discharge-i00-standard/task.toml
+++ b/tasks/pandadyn-d04-shared-tray-transport-discharge-i00-standard/task.toml
@@ -28,6 +28,7 @@ simulator = "maniskill"
robot = "franka-panda-bimanual"
family = "d04_shared_tray_transport_discharge"
mode = "standard" # robot as a service: RGB, depth + proprioception only; no object poses, no success predicate, no reset
+dynamics = "dynamic" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "full"
instance_seed = 1
diff --git a/tasks/pandadyn-d05-moving-ramp-relay-i00-privileged/task.toml b/tasks/pandadyn-d05-moving-ramp-relay-i00-privileged/task.toml
index b61d66d084..1ff6c96553 100644
--- a/tasks/pandadyn-d05-moving-ramp-relay-i00-privileged/task.toml
+++ b/tasks/pandadyn-d05-moving-ramp-relay-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "maniskill" # Panda Dynamic Scenes on ManiSkill 3.0.1 / SA
robot = "franka-panda-bimanual"
family = "d05_moving_ramp_relay"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 0
expert_time_estimate_hours = 3.0
diff --git a/tasks/pandadyn-d05-moving-ramp-relay-i00-standard/task.toml b/tasks/pandadyn-d05-moving-ramp-relay-i00-standard/task.toml
index 0d0eefe46f..fce7417c0d 100644
--- a/tasks/pandadyn-d05-moving-ramp-relay-i00-standard/task.toml
+++ b/tasks/pandadyn-d05-moving-ramp-relay-i00-standard/task.toml
@@ -28,6 +28,7 @@ simulator = "maniskill"
robot = "franka-panda-bimanual"
family = "d05_moving_ramp_relay"
mode = "standard" # robot as a service: RGB, depth + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "full"
instance_seed = 0
diff --git a/tasks/pandadyn-d06-moving-bank-relay-i00-privileged/task.toml b/tasks/pandadyn-d06-moving-bank-relay-i00-privileged/task.toml
index b360e15238..c20bceac85 100644
--- a/tasks/pandadyn-d06-moving-bank-relay-i00-privileged/task.toml
+++ b/tasks/pandadyn-d06-moving-bank-relay-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "maniskill" # Panda Dynamic Scenes on ManiSkill 3.0.1 / SA
robot = "franka-panda-bimanual"
family = "d06_moving_bank_relay"
mode = "privileged"
+dynamics = "dynamic" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 0
expert_time_estimate_hours = 3.0
diff --git a/tasks/pandadyn-d06-moving-bank-relay-i00-standard/task.toml b/tasks/pandadyn-d06-moving-bank-relay-i00-standard/task.toml
index a4466ff658..d2cc102c62 100644
--- a/tasks/pandadyn-d06-moving-bank-relay-i00-standard/task.toml
+++ b/tasks/pandadyn-d06-moving-bank-relay-i00-standard/task.toml
@@ -28,6 +28,7 @@ simulator = "maniskill"
robot = "franka-panda-bimanual"
family = "d06_moving_bank_relay"
mode = "standard" # robot as a service: RGB, depth + proprioception only; no object poses, no success predicate, no reset
+dynamics = "dynamic" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "full"
instance_seed = 0
diff --git a/tasks/pandadyn-d07-spring-gate-switchyard-i00-privileged/task.toml b/tasks/pandadyn-d07-spring-gate-switchyard-i00-privileged/task.toml
index 12eedae4ce..ffa0e65343 100644
--- a/tasks/pandadyn-d07-spring-gate-switchyard-i00-privileged/task.toml
+++ b/tasks/pandadyn-d07-spring-gate-switchyard-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "maniskill" # Panda Dynamic Scenes on ManiSkill 3.0.1 / SA
robot = "franka-panda-bimanual"
family = "d07_spring_gate_switchyard"
mode = "privileged"
+dynamics = "dynamic" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 0
expert_time_estimate_hours = 3.0
diff --git a/tasks/pandadyn-d07-spring-gate-switchyard-i00-standard/task.toml b/tasks/pandadyn-d07-spring-gate-switchyard-i00-standard/task.toml
index e4a9f29177..ff5c6b5e61 100644
--- a/tasks/pandadyn-d07-spring-gate-switchyard-i00-standard/task.toml
+++ b/tasks/pandadyn-d07-spring-gate-switchyard-i00-standard/task.toml
@@ -28,6 +28,7 @@ simulator = "maniskill"
robot = "franka-panda-bimanual"
family = "d07_spring_gate_switchyard"
mode = "standard" # robot as a service: RGB, depth + proprioception only; no object poses, no success predicate, no reset
+dynamics = "dynamic" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "full"
instance_seed = 0
diff --git a/tasks/pandadyn-d08-pendulum-energy-relay-i00-privileged/task.toml b/tasks/pandadyn-d08-pendulum-energy-relay-i00-privileged/task.toml
index 1a7f267c5a..84dc38d628 100644
--- a/tasks/pandadyn-d08-pendulum-energy-relay-i00-privileged/task.toml
+++ b/tasks/pandadyn-d08-pendulum-energy-relay-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "maniskill" # Panda Dynamic Scenes on ManiSkill 3.0.1 / SA
robot = "franka-panda-bimanual"
family = "d08_pendulum_energy_relay"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 0
expert_time_estimate_hours = 3.0
diff --git a/tasks/pandadyn-d08-pendulum-energy-relay-i00-standard/task.toml b/tasks/pandadyn-d08-pendulum-energy-relay-i00-standard/task.toml
index af16931ca8..60ae03fcb1 100644
--- a/tasks/pandadyn-d08-pendulum-energy-relay-i00-standard/task.toml
+++ b/tasks/pandadyn-d08-pendulum-energy-relay-i00-standard/task.toml
@@ -28,6 +28,7 @@ simulator = "maniskill"
robot = "franka-panda-bimanual"
family = "d08_pendulum_energy_relay"
mode = "standard" # robot as a service: RGB, depth + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "full"
instance_seed = 0
diff --git a/tasks/pandadyn-d09-lever-launch-catch-reload-i00-privileged/task.toml b/tasks/pandadyn-d09-lever-launch-catch-reload-i00-privileged/task.toml
index db13d4e57d..501820e243 100644
--- a/tasks/pandadyn-d09-lever-launch-catch-reload-i00-privileged/task.toml
+++ b/tasks/pandadyn-d09-lever-launch-catch-reload-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "maniskill" # Panda Dynamic Scenes on ManiSkill 3.0.1 / SA
robot = "franka-panda-bimanual"
family = "d09_lever_launch_catch_reload"
mode = "privileged"
+dynamics = "dynamic" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 0
expert_time_estimate_hours = 3.0
diff --git a/tasks/pandadyn-d09-lever-launch-catch-reload-i00-standard/task.toml b/tasks/pandadyn-d09-lever-launch-catch-reload-i00-standard/task.toml
index e23713b398..9abb7d9da3 100644
--- a/tasks/pandadyn-d09-lever-launch-catch-reload-i00-standard/task.toml
+++ b/tasks/pandadyn-d09-lever-launch-catch-reload-i00-standard/task.toml
@@ -28,6 +28,7 @@ simulator = "maniskill"
robot = "franka-panda-bimanual"
family = "d09_lever_launch_catch_reload"
mode = "standard" # robot as a service: RGB, depth + proprioception only; no object poses, no success predicate, no reset
+dynamics = "dynamic" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "full"
instance_seed = 0
diff --git a/tasks/pandadyn-d10-moving-grasp-handover-i00-privileged/task.toml b/tasks/pandadyn-d10-moving-grasp-handover-i00-privileged/task.toml
index bda28d6e93..33513f9e71 100644
--- a/tasks/pandadyn-d10-moving-grasp-handover-i00-privileged/task.toml
+++ b/tasks/pandadyn-d10-moving-grasp-handover-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "maniskill" # Panda Dynamic Scenes on ManiSkill 3.0.1 / SA
robot = "franka-panda-bimanual"
family = "d10_moving_grasp_handover"
mode = "privileged"
+dynamics = "dynamic" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 0
expert_time_estimate_hours = 3.0
diff --git a/tasks/pandadyn-d10-moving-grasp-handover-i00-standard/task.toml b/tasks/pandadyn-d10-moving-grasp-handover-i00-standard/task.toml
index 279fbab4ad..f34160d005 100644
--- a/tasks/pandadyn-d10-moving-grasp-handover-i00-standard/task.toml
+++ b/tasks/pandadyn-d10-moving-grasp-handover-i00-standard/task.toml
@@ -28,6 +28,7 @@ simulator = "maniskill"
robot = "franka-panda-bimanual"
family = "d10_moving_grasp_handover"
mode = "standard" # robot as a service: RGB, depth + proprioception only; no object poses, no success predicate, no reset
+dynamics = "dynamic" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "full"
instance_seed = 0
diff --git a/tasks/pandadyn-d11-moving-fixture-unlock-i00-privileged/task.toml b/tasks/pandadyn-d11-moving-fixture-unlock-i00-privileged/task.toml
index 1092cc3f8e..e7bba50665 100644
--- a/tasks/pandadyn-d11-moving-fixture-unlock-i00-privileged/task.toml
+++ b/tasks/pandadyn-d11-moving-fixture-unlock-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "maniskill" # Panda Dynamic Scenes on ManiSkill 3.0.1 / SA
robot = "franka-panda-bimanual"
family = "d11_moving_fixture_unlock"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 0
expert_time_estimate_hours = 3.0
diff --git a/tasks/pandadyn-d11-moving-fixture-unlock-i00-standard/task.toml b/tasks/pandadyn-d11-moving-fixture-unlock-i00-standard/task.toml
index b7ff8a2768..223c7d71cc 100644
--- a/tasks/pandadyn-d11-moving-fixture-unlock-i00-standard/task.toml
+++ b/tasks/pandadyn-d11-moving-fixture-unlock-i00-standard/task.toml
@@ -28,6 +28,7 @@ simulator = "maniskill"
robot = "franka-panda-bimanual"
family = "d11_moving_fixture_unlock"
mode = "standard" # robot as a service: RGB, depth + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "full"
instance_seed = 0
diff --git a/tasks/pandadyn-d12-reciprocal-sliding-transfer-i00-privileged/task.toml b/tasks/pandadyn-d12-reciprocal-sliding-transfer-i00-privileged/task.toml
index 550d38bc29..bc64aee97a 100644
--- a/tasks/pandadyn-d12-reciprocal-sliding-transfer-i00-privileged/task.toml
+++ b/tasks/pandadyn-d12-reciprocal-sliding-transfer-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "maniskill" # Panda Dynamic Scenes on ManiSkill 3.0.1 / SA
robot = "franka-panda-bimanual"
family = "d12_reciprocal_sliding_transfer"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 0
expert_time_estimate_hours = 3.0
diff --git a/tasks/pandadyn-d12-reciprocal-sliding-transfer-i00-standard/task.toml b/tasks/pandadyn-d12-reciprocal-sliding-transfer-i00-standard/task.toml
index 127aa2ef45..98bce6d713 100644
--- a/tasks/pandadyn-d12-reciprocal-sliding-transfer-i00-standard/task.toml
+++ b/tasks/pandadyn-d12-reciprocal-sliding-transfer-i00-standard/task.toml
@@ -28,6 +28,7 @@ simulator = "maniskill"
robot = "franka-panda-bimanual"
family = "d12_reciprocal_sliding_transfer"
mode = "standard" # robot as a service: RGB, depth + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "full"
instance_seed = 0
diff --git a/tasks/pandadyn-d13-supported-cube-flip-relay-i00-privileged/task.toml b/tasks/pandadyn-d13-supported-cube-flip-relay-i00-privileged/task.toml
index b34b4961e1..80515ea218 100644
--- a/tasks/pandadyn-d13-supported-cube-flip-relay-i00-privileged/task.toml
+++ b/tasks/pandadyn-d13-supported-cube-flip-relay-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "maniskill" # Panda Dynamic Scenes on ManiSkill 3.0.1 / SA
robot = "franka-panda-bimanual"
family = "d13_supported_cube_flip_relay"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 0
expert_time_estimate_hours = 3.0
diff --git a/tasks/pandadyn-d13-supported-cube-flip-relay-i00-standard/task.toml b/tasks/pandadyn-d13-supported-cube-flip-relay-i00-standard/task.toml
index 45a2a78886..65ec3af9c5 100644
--- a/tasks/pandadyn-d13-supported-cube-flip-relay-i00-standard/task.toml
+++ b/tasks/pandadyn-d13-supported-cube-flip-relay-i00-standard/task.toml
@@ -28,6 +28,7 @@ simulator = "maniskill"
robot = "franka-panda-bimanual"
family = "d13_supported_cube_flip_relay"
mode = "standard" # robot as a service: RGB, depth + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "full"
instance_seed = 0
diff --git a/tasks/pandadyn-s01-cup-route-delivery-i00-privileged/task.toml b/tasks/pandadyn-s01-cup-route-delivery-i00-privileged/task.toml
index 12da107db2..d849c24042 100644
--- a/tasks/pandadyn-s01-cup-route-delivery-i00-privileged/task.toml
+++ b/tasks/pandadyn-s01-cup-route-delivery-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "maniskill" # Panda Dynamic Scenes on ManiSkill 3.0.1 / SA
robot = "franka-panda"
family = "s01_cup_route_delivery"
mode = "privileged"
+dynamics = "dynamic" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 0
expert_time_estimate_hours = 3.0
diff --git a/tasks/pandadyn-s01-cup-route-delivery-i00-standard/task.toml b/tasks/pandadyn-s01-cup-route-delivery-i00-standard/task.toml
index 733d95af59..e2b6456274 100644
--- a/tasks/pandadyn-s01-cup-route-delivery-i00-standard/task.toml
+++ b/tasks/pandadyn-s01-cup-route-delivery-i00-standard/task.toml
@@ -28,6 +28,7 @@ simulator = "maniskill"
robot = "franka-panda"
family = "s01_cup_route_delivery"
mode = "standard" # robot as a service: RGB, depth + proprioception only; no object poses, no success predicate, no reset
+dynamics = "dynamic" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "full"
instance_seed = 0
diff --git a/tasks/pandadyn-s02-plate-catch-transport-release-i00-privileged/task.toml b/tasks/pandadyn-s02-plate-catch-transport-release-i00-privileged/task.toml
index 1514097193..0a5137f111 100644
--- a/tasks/pandadyn-s02-plate-catch-transport-release-i00-privileged/task.toml
+++ b/tasks/pandadyn-s02-plate-catch-transport-release-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "maniskill" # Panda Dynamic Scenes on ManiSkill 3.0.1 / SA
robot = "franka-panda"
family = "s02_plate_catch_transport_release"
mode = "privileged"
+dynamics = "dynamic" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 0
expert_time_estimate_hours = 3.0
diff --git a/tasks/pandadyn-s02-plate-catch-transport-release-i00-standard/task.toml b/tasks/pandadyn-s02-plate-catch-transport-release-i00-standard/task.toml
index a7ea80f41b..a23baf70d8 100644
--- a/tasks/pandadyn-s02-plate-catch-transport-release-i00-standard/task.toml
+++ b/tasks/pandadyn-s02-plate-catch-transport-release-i00-standard/task.toml
@@ -28,6 +28,7 @@ simulator = "maniskill"
robot = "franka-panda"
family = "s02_plate_catch_transport_release"
mode = "standard" # robot as a service: RGB, depth + proprioception only; no object poses, no success predicate, no reset
+dynamics = "dynamic" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "full"
instance_seed = 0
diff --git a/tasks/pandadyn-s03-clear-gate-hockey-i00-privileged/task.toml b/tasks/pandadyn-s03-clear-gate-hockey-i00-privileged/task.toml
index 05919833a2..ddf2e62c3e 100644
--- a/tasks/pandadyn-s03-clear-gate-hockey-i00-privileged/task.toml
+++ b/tasks/pandadyn-s03-clear-gate-hockey-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "maniskill" # Panda Dynamic Scenes on ManiSkill 3.0.1 / SA
robot = "franka-panda"
family = "s03_clear_gate_hockey"
mode = "privileged"
+dynamics = "dynamic" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 0
expert_time_estimate_hours = 3.0
diff --git a/tasks/pandadyn-s03-clear-gate-hockey-i00-standard/task.toml b/tasks/pandadyn-s03-clear-gate-hockey-i00-standard/task.toml
index d0b5a18fb6..cf24bbfd07 100644
--- a/tasks/pandadyn-s03-clear-gate-hockey-i00-standard/task.toml
+++ b/tasks/pandadyn-s03-clear-gate-hockey-i00-standard/task.toml
@@ -28,6 +28,7 @@ simulator = "maniskill"
robot = "franka-panda"
family = "s03_clear_gate_hockey"
mode = "standard" # robot as a service: RGB, depth + proprioception only; no object poses, no success predicate, no reset
+dynamics = "dynamic" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "full"
instance_seed = 0
diff --git a/tasks/pandadyn-s04-impact-release-relay-i00-privileged/task.toml b/tasks/pandadyn-s04-impact-release-relay-i00-privileged/task.toml
index cccd673b77..23a3871d70 100644
--- a/tasks/pandadyn-s04-impact-release-relay-i00-privileged/task.toml
+++ b/tasks/pandadyn-s04-impact-release-relay-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "maniskill" # Panda Dynamic Scenes on ManiSkill 3.0.1 / SA
robot = "franka-panda"
family = "s04_impact_release_relay"
mode = "privileged"
+dynamics = "dynamic" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 0
expert_time_estimate_hours = 3.0
diff --git a/tasks/pandadyn-s04-impact-release-relay-i00-standard/task.toml b/tasks/pandadyn-s04-impact-release-relay-i00-standard/task.toml
index 5f2e112747..73cf52532f 100644
--- a/tasks/pandadyn-s04-impact-release-relay-i00-standard/task.toml
+++ b/tasks/pandadyn-s04-impact-release-relay-i00-standard/task.toml
@@ -28,6 +28,7 @@ simulator = "maniskill"
robot = "franka-panda"
family = "s04_impact_release_relay"
mode = "standard" # robot as a service: RGB, depth + proprioception only; no object poses, no success predicate, no reset
+dynamics = "dynamic" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "full"
instance_seed = 0
diff --git a/tasks/pandadyn-s05-sliding-receive-align-deliver-i00-privileged/task.toml b/tasks/pandadyn-s05-sliding-receive-align-deliver-i00-privileged/task.toml
index ab26a3b2c2..36bd250fb2 100644
--- a/tasks/pandadyn-s05-sliding-receive-align-deliver-i00-privileged/task.toml
+++ b/tasks/pandadyn-s05-sliding-receive-align-deliver-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "maniskill" # Panda Dynamic Scenes on ManiSkill 3.0.1 / SA
robot = "franka-panda"
family = "s05_sliding_receive_align_deliver"
mode = "privileged"
+dynamics = "dynamic" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 0
expert_time_estimate_hours = 3.0
diff --git a/tasks/pandadyn-s05-sliding-receive-align-deliver-i00-standard/task.toml b/tasks/pandadyn-s05-sliding-receive-align-deliver-i00-standard/task.toml
index 10b02fc7e5..ec637ee10a 100644
--- a/tasks/pandadyn-s05-sliding-receive-align-deliver-i00-standard/task.toml
+++ b/tasks/pandadyn-s05-sliding-receive-align-deliver-i00-standard/task.toml
@@ -28,6 +28,7 @@ simulator = "maniskill"
robot = "franka-panda"
family = "s05_sliding_receive_align_deliver"
mode = "standard" # robot as a service: RGB, depth + proprioception only; no object poses, no success predicate, no reset
+dynamics = "dynamic" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "full"
instance_seed = 0
diff --git a/tasks/pandadyn-s06-cube-flip-route-recover-i00-privileged/task.toml b/tasks/pandadyn-s06-cube-flip-route-recover-i00-privileged/task.toml
index 734963fede..4662bc653a 100644
--- a/tasks/pandadyn-s06-cube-flip-route-recover-i00-privileged/task.toml
+++ b/tasks/pandadyn-s06-cube-flip-route-recover-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "maniskill" # Panda Dynamic Scenes on ManiSkill 3.0.1 / SA
robot = "franka-panda"
family = "s06_cube_flip_route_recover"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 0
expert_time_estimate_hours = 3.0
diff --git a/tasks/pandadyn-s06-cube-flip-route-recover-i00-standard/task.toml b/tasks/pandadyn-s06-cube-flip-route-recover-i00-standard/task.toml
index 1e7ca23d37..6f84c4bf4b 100644
--- a/tasks/pandadyn-s06-cube-flip-route-recover-i00-standard/task.toml
+++ b/tasks/pandadyn-s06-cube-flip-route-recover-i00-standard/task.toml
@@ -28,6 +28,7 @@ simulator = "maniskill"
robot = "franka-panda"
family = "s06_cube_flip_route_recover"
mode = "standard" # robot as a service: RGB, depth + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "full"
instance_seed = 0
diff --git a/tasks/pandadyn-s07-conveyor-load-brake-unload-i00-privileged/task.toml b/tasks/pandadyn-s07-conveyor-load-brake-unload-i00-privileged/task.toml
index 75c8a1d497..752e6c94dc 100644
--- a/tasks/pandadyn-s07-conveyor-load-brake-unload-i00-privileged/task.toml
+++ b/tasks/pandadyn-s07-conveyor-load-brake-unload-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "maniskill" # Panda Dynamic Scenes on ManiSkill 3.0.1 / SA
robot = "franka-panda"
family = "s07_conveyor_load_brake_unload"
mode = "privileged"
+dynamics = "dynamic" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 0
expert_time_estimate_hours = 3.0
diff --git a/tasks/pandadyn-s07-conveyor-load-brake-unload-i00-standard/task.toml b/tasks/pandadyn-s07-conveyor-load-brake-unload-i00-standard/task.toml
index 8b4bd7c8a9..62defb1e71 100644
--- a/tasks/pandadyn-s07-conveyor-load-brake-unload-i00-standard/task.toml
+++ b/tasks/pandadyn-s07-conveyor-load-brake-unload-i00-standard/task.toml
@@ -28,6 +28,7 @@ simulator = "maniskill"
robot = "franka-panda"
family = "s07_conveyor_load_brake_unload"
mode = "standard" # robot as a service: RGB, depth + proprioception only; no object poses, no success predicate, no reset
+dynamics = "dynamic" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "full"
instance_seed = 0
diff --git a/tasks/pandadyn-s08-load-switch-recover-i00-privileged/task.toml b/tasks/pandadyn-s08-load-switch-recover-i00-privileged/task.toml
index 845e91767b..e2584d0a3c 100644
--- a/tasks/pandadyn-s08-load-switch-recover-i00-privileged/task.toml
+++ b/tasks/pandadyn-s08-load-switch-recover-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "maniskill" # Panda Dynamic Scenes on ManiSkill 3.0.1 / SA
robot = "franka-panda"
family = "s08_load_switch_recover"
mode = "privileged"
+dynamics = "dynamic" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 0
expert_time_estimate_hours = 3.0
diff --git a/tasks/pandadyn-s08-load-switch-recover-i00-standard/task.toml b/tasks/pandadyn-s08-load-switch-recover-i00-standard/task.toml
index d410bc6b0a..9fc510946b 100644
--- a/tasks/pandadyn-s08-load-switch-recover-i00-standard/task.toml
+++ b/tasks/pandadyn-s08-load-switch-recover-i00-standard/task.toml
@@ -28,6 +28,7 @@ simulator = "maniskill"
robot = "franka-panda"
family = "s08_load_switch_recover"
mode = "standard" # robot as a service: RGB, depth + proprioception only; no object poses, no success predicate, no reset
+dynamics = "dynamic" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "full"
instance_seed = 0
diff --git a/tasks/pandadyn-s09-toss-recatch-delivery-i00-privileged/task.toml b/tasks/pandadyn-s09-toss-recatch-delivery-i00-privileged/task.toml
index f73ea0c6e7..59511ff688 100644
--- a/tasks/pandadyn-s09-toss-recatch-delivery-i00-privileged/task.toml
+++ b/tasks/pandadyn-s09-toss-recatch-delivery-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "maniskill" # Panda Dynamic Scenes on ManiSkill 3.0.1 / SA
robot = "franka-panda"
family = "s09_toss_recatch_delivery"
mode = "privileged"
+dynamics = "dynamic" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 0
expert_time_estimate_hours = 3.0
diff --git a/tasks/pandadyn-s09-toss-recatch-delivery-i00-standard/task.toml b/tasks/pandadyn-s09-toss-recatch-delivery-i00-standard/task.toml
index c37b9a0b44..12d45ba8f4 100644
--- a/tasks/pandadyn-s09-toss-recatch-delivery-i00-standard/task.toml
+++ b/tasks/pandadyn-s09-toss-recatch-delivery-i00-standard/task.toml
@@ -28,6 +28,7 @@ simulator = "maniskill"
robot = "franka-panda"
family = "s09_toss_recatch_delivery"
mode = "standard" # robot as a service: RGB, depth + proprioception only; no object poses, no success predicate, no reset
+dynamics = "dynamic" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "full"
instance_seed = 0
diff --git a/tasks/pandadyn-s10-swinging-load-transit-dock-i00-privileged/task.toml b/tasks/pandadyn-s10-swinging-load-transit-dock-i00-privileged/task.toml
index 623621bea4..1da8bce00d 100644
--- a/tasks/pandadyn-s10-swinging-load-transit-dock-i00-privileged/task.toml
+++ b/tasks/pandadyn-s10-swinging-load-transit-dock-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "maniskill" # Panda Dynamic Scenes on ManiSkill 3.0.1 / SA
robot = "franka-panda"
family = "s10_swinging_load_transit_dock"
mode = "privileged"
+dynamics = "dynamic" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 0
expert_time_estimate_hours = 3.0
diff --git a/tasks/pandadyn-s10-swinging-load-transit-dock-i00-standard/task.toml b/tasks/pandadyn-s10-swinging-load-transit-dock-i00-standard/task.toml
index 3d98cb8d20..b23459710d 100644
--- a/tasks/pandadyn-s10-swinging-load-transit-dock-i00-standard/task.toml
+++ b/tasks/pandadyn-s10-swinging-load-transit-dock-i00-standard/task.toml
@@ -28,6 +28,7 @@ simulator = "maniskill"
robot = "franka-panda"
family = "s10_swinging_load_transit_dock"
mode = "standard" # robot as a service: RGB, depth + proprioception only; no object poses, no success predicate, no reset
+dynamics = "dynamic" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "full"
instance_seed = 0
diff --git a/tasks/pandadyn-s11-cylinder-rescue-route-deliver-i00-privileged/task.toml b/tasks/pandadyn-s11-cylinder-rescue-route-deliver-i00-privileged/task.toml
index 148232a017..04dfacee4d 100644
--- a/tasks/pandadyn-s11-cylinder-rescue-route-deliver-i00-privileged/task.toml
+++ b/tasks/pandadyn-s11-cylinder-rescue-route-deliver-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "maniskill" # Panda Dynamic Scenes on ManiSkill 3.0.1 / SA
robot = "franka-panda"
family = "s11_cylinder_rescue_route_deliver"
mode = "privileged"
+dynamics = "dynamic" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 0
expert_time_estimate_hours = 3.0
diff --git a/tasks/pandadyn-s11-cylinder-rescue-route-deliver-i00-standard/task.toml b/tasks/pandadyn-s11-cylinder-rescue-route-deliver-i00-standard/task.toml
index f3d65a8abf..22726e2d33 100644
--- a/tasks/pandadyn-s11-cylinder-rescue-route-deliver-i00-standard/task.toml
+++ b/tasks/pandadyn-s11-cylinder-rescue-route-deliver-i00-standard/task.toml
@@ -28,6 +28,7 @@ simulator = "maniskill"
robot = "franka-panda"
family = "s11_cylinder_rescue_route_deliver"
mode = "standard" # robot as a service: RGB, depth + proprioception only; no object poses, no success predicate, no reset
+dynamics = "dynamic" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "full"
instance_seed = 0
diff --git a/tasks/pandadyn-s12-moving-bridge-puck-relay-i00-privileged/task.toml b/tasks/pandadyn-s12-moving-bridge-puck-relay-i00-privileged/task.toml
index a3b84ecb0b..e40a040e1c 100644
--- a/tasks/pandadyn-s12-moving-bridge-puck-relay-i00-privileged/task.toml
+++ b/tasks/pandadyn-s12-moving-bridge-puck-relay-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "maniskill" # Panda Dynamic Scenes on ManiSkill 3.0.1 / SA
robot = "franka-panda"
family = "s12_moving_bridge_puck_relay"
mode = "privileged"
+dynamics = "dynamic" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 1
expert_time_estimate_hours = 3.0
diff --git a/tasks/pandadyn-s12-moving-bridge-puck-relay-i00-standard/task.toml b/tasks/pandadyn-s12-moving-bridge-puck-relay-i00-standard/task.toml
index 797f502554..0c9adadac0 100644
--- a/tasks/pandadyn-s12-moving-bridge-puck-relay-i00-standard/task.toml
+++ b/tasks/pandadyn-s12-moving-bridge-puck-relay-i00-standard/task.toml
@@ -28,6 +28,7 @@ simulator = "maniskill"
robot = "franka-panda"
family = "s12_moving_bridge_puck_relay"
mode = "standard" # robot as a service: RGB, depth + proprioception only; no object poses, no success predicate, no reset
+dynamics = "dynamic" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "full"
instance_seed = 1
diff --git a/tasks/roboarc-blanket-flatten-01-i01-privileged/task.toml b/tasks/roboarc-blanket-flatten-01-i01-privileged/task.toml
index e9ec8e69c4..d352b7441d 100644
--- a/tasks/roboarc-blanket-flatten-01-i01-privileged/task.toml
+++ b/tasks/roboarc-blanket-flatten-01-i01-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "roboarc" # RoboARC on MuJoCo 3.14.0 (CPU physics, CUDA
robot = "aloha-2"
family = "blanket_flatten"
mode = "privileged" # privileged (state-based tooling, resets) | standard (robot as a service)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
difficulty = "very_easy" # the variation's tier (docs/TAXONOMY.md)
grasp = "assisted (the pbd_local_patch_v1 grasp regions on blanket)"
diff --git a/tasks/roboarc-blanket-flatten-01-i01-standard/task.toml b/tasks/roboarc-blanket-flatten-01-i01-standard/task.toml
index a755052809..b282864b8a 100644
--- a/tasks/roboarc-blanket-flatten-01-i01-standard/task.toml
+++ b/tasks/roboarc-blanket-flatten-01-i01-standard/task.toml
@@ -31,6 +31,7 @@ simulator = "roboarc"
robot = "aloha-2"
family = "blanket_flatten"
mode = "standard" # RGB + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.4"
oracle = "full"
difficulty = "very_easy"
diff --git a/tasks/roboarc-blanket-flatten-02-i01-privileged/task.toml b/tasks/roboarc-blanket-flatten-02-i01-privileged/task.toml
index ea0f038211..62e568d98b 100644
--- a/tasks/roboarc-blanket-flatten-02-i01-privileged/task.toml
+++ b/tasks/roboarc-blanket-flatten-02-i01-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "roboarc" # RoboARC on MuJoCo 3.14.0 (CPU physics, CUDA
robot = "aloha-2"
family = "blanket_flatten"
mode = "privileged" # privileged (state-based tooling, resets) | standard (robot as a service)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
difficulty = "very_easy" # the variation's tier (docs/rcb/TAXONOMY.md)
grasp = "assisted (the pbd_local_patch_v1 grasp regions on blanket)"
diff --git a/tasks/roboarc-blanket-flatten-02-i01-standard/task.toml b/tasks/roboarc-blanket-flatten-02-i01-standard/task.toml
index a1e304ae02..e19b256ee5 100644
--- a/tasks/roboarc-blanket-flatten-02-i01-standard/task.toml
+++ b/tasks/roboarc-blanket-flatten-02-i01-standard/task.toml
@@ -31,6 +31,7 @@ simulator = "roboarc"
robot = "aloha-2"
family = "blanket_flatten"
mode = "standard" # RGB + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.4"
oracle = "full"
difficulty = "very_easy"
diff --git a/tasks/roboarc-blanket-flatten-03-i01-privileged/task.toml b/tasks/roboarc-blanket-flatten-03-i01-privileged/task.toml
index 6677585c12..d28cb3d464 100644
--- a/tasks/roboarc-blanket-flatten-03-i01-privileged/task.toml
+++ b/tasks/roboarc-blanket-flatten-03-i01-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "roboarc" # RoboARC on MuJoCo 3.14.0 (CPU physics, CUDA
robot = "aloha-2"
family = "blanket_flatten"
mode = "privileged" # privileged (state-based tooling, resets) | standard (robot as a service)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
difficulty = "easier" # the variation's tier (docs/rcb/TAXONOMY.md)
grasp = "assisted (the pbd_local_patch_v1 grasp regions on blanket)"
diff --git a/tasks/roboarc-blanket-flatten-03-i01-standard/task.toml b/tasks/roboarc-blanket-flatten-03-i01-standard/task.toml
index ce3aecbb3a..e32803a82e 100644
--- a/tasks/roboarc-blanket-flatten-03-i01-standard/task.toml
+++ b/tasks/roboarc-blanket-flatten-03-i01-standard/task.toml
@@ -31,6 +31,7 @@ simulator = "roboarc"
robot = "aloha-2"
family = "blanket_flatten"
mode = "standard" # RGB + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.4"
oracle = "full"
difficulty = "easier"
diff --git a/tasks/roboarc-blanket-flatten-04-i01-privileged/task.toml b/tasks/roboarc-blanket-flatten-04-i01-privileged/task.toml
index 632380ada6..c736015ef2 100644
--- a/tasks/roboarc-blanket-flatten-04-i01-privileged/task.toml
+++ b/tasks/roboarc-blanket-flatten-04-i01-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "roboarc" # RoboARC on MuJoCo 3.14.0 (CPU physics, CUDA
robot = "aloha-2"
family = "blanket_flatten"
mode = "privileged" # privileged (state-based tooling, resets) | standard (robot as a service)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
difficulty = "easy" # the variation's tier (docs/rcb/TAXONOMY.md)
grasp = "assisted (the pbd_local_patch_v1 grasp regions on blanket)"
diff --git a/tasks/roboarc-blanket-flatten-04-i01-standard/task.toml b/tasks/roboarc-blanket-flatten-04-i01-standard/task.toml
index 0c4dba9ac9..04bf9b7ccb 100644
--- a/tasks/roboarc-blanket-flatten-04-i01-standard/task.toml
+++ b/tasks/roboarc-blanket-flatten-04-i01-standard/task.toml
@@ -31,6 +31,7 @@ simulator = "roboarc"
robot = "aloha-2"
family = "blanket_flatten"
mode = "standard" # RGB + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.4"
oracle = "full"
difficulty = "easy"
diff --git a/tasks/roboarc-blanket-flatten-05-i01-privileged/task.toml b/tasks/roboarc-blanket-flatten-05-i01-privileged/task.toml
index 28cd232db5..3ff6ada48d 100644
--- a/tasks/roboarc-blanket-flatten-05-i01-privileged/task.toml
+++ b/tasks/roboarc-blanket-flatten-05-i01-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "roboarc" # RoboARC on MuJoCo 3.14.0 (CPU physics, CUDA
robot = "aloha-2"
family = "blanket_flatten"
mode = "privileged" # privileged (state-based tooling, resets) | standard (robot as a service)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
difficulty = "medium" # the variation's tier (docs/rcb/TAXONOMY.md)
grasp = "assisted (the pbd_local_patch_v1 grasp regions on blanket)"
diff --git a/tasks/roboarc-blanket-flatten-05-i01-standard/task.toml b/tasks/roboarc-blanket-flatten-05-i01-standard/task.toml
index 2e146d23c7..327ac6120b 100644
--- a/tasks/roboarc-blanket-flatten-05-i01-standard/task.toml
+++ b/tasks/roboarc-blanket-flatten-05-i01-standard/task.toml
@@ -31,6 +31,7 @@ simulator = "roboarc"
robot = "aloha-2"
family = "blanket_flatten"
mode = "standard" # RGB + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.4"
oracle = "full"
difficulty = "medium"
diff --git a/tasks/roboarc-blanket-flatten-06-i01-privileged/task.toml b/tasks/roboarc-blanket-flatten-06-i01-privileged/task.toml
index b80677da44..2ac4277510 100644
--- a/tasks/roboarc-blanket-flatten-06-i01-privileged/task.toml
+++ b/tasks/roboarc-blanket-flatten-06-i01-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "roboarc" # RoboARC on MuJoCo 3.14.0 (CPU physics, CUDA
robot = "aloha-2"
family = "blanket_flatten"
mode = "privileged" # privileged (state-based tooling, resets) | standard (robot as a service)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
difficulty = "hard" # the variation's tier (docs/rcb/TAXONOMY.md)
grasp = "assisted (the pbd_local_patch_v1 grasp regions on blanket)"
diff --git a/tasks/roboarc-blanket-flatten-06-i01-standard/task.toml b/tasks/roboarc-blanket-flatten-06-i01-standard/task.toml
index e8379a5dd7..85525b1491 100644
--- a/tasks/roboarc-blanket-flatten-06-i01-standard/task.toml
+++ b/tasks/roboarc-blanket-flatten-06-i01-standard/task.toml
@@ -31,6 +31,7 @@ simulator = "roboarc"
robot = "aloha-2"
family = "blanket_flatten"
mode = "standard" # RGB + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.4"
oracle = "full"
difficulty = "hard"
diff --git a/tasks/roboarc-blanket-fold-01-i01-privileged/task.toml b/tasks/roboarc-blanket-fold-01-i01-privileged/task.toml
index 21f904260f..5a583dea59 100644
--- a/tasks/roboarc-blanket-fold-01-i01-privileged/task.toml
+++ b/tasks/roboarc-blanket-fold-01-i01-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "roboarc" # RoboARC on MuJoCo 3.14.0 (CPU physics, CUDA
robot = "aloha-2"
family = "blanket_fold"
mode = "privileged" # privileged (state-based tooling, resets) | standard (robot as a service)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
difficulty = "very_easy" # the variation's tier (docs/rcb/TAXONOMY.md)
grasp = "assisted (the pbd_local_patch_v1 grasp regions on blanket)"
diff --git a/tasks/roboarc-blanket-fold-01-i01-standard/task.toml b/tasks/roboarc-blanket-fold-01-i01-standard/task.toml
index 3a1a55e1d9..598b35b7a9 100644
--- a/tasks/roboarc-blanket-fold-01-i01-standard/task.toml
+++ b/tasks/roboarc-blanket-fold-01-i01-standard/task.toml
@@ -31,6 +31,7 @@ simulator = "roboarc"
robot = "aloha-2"
family = "blanket_fold"
mode = "standard" # RGB + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.4"
oracle = "full"
difficulty = "very_easy"
diff --git a/tasks/roboarc-blanket-fold-02-i01-privileged/task.toml b/tasks/roboarc-blanket-fold-02-i01-privileged/task.toml
index 3fdc097a5b..2b744c41a9 100644
--- a/tasks/roboarc-blanket-fold-02-i01-privileged/task.toml
+++ b/tasks/roboarc-blanket-fold-02-i01-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "roboarc" # RoboARC on MuJoCo 3.14.0 (CPU physics, CUDA
robot = "aloha-2"
family = "blanket_fold"
mode = "privileged" # privileged (state-based tooling, resets) | standard (robot as a service)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
difficulty = "easier" # the variation's tier (docs/rcb/TAXONOMY.md)
grasp = "assisted (the pbd_local_patch_v1 grasp regions on blanket)"
diff --git a/tasks/roboarc-blanket-fold-02-i01-standard/task.toml b/tasks/roboarc-blanket-fold-02-i01-standard/task.toml
index f6869f5f89..105688cbda 100644
--- a/tasks/roboarc-blanket-fold-02-i01-standard/task.toml
+++ b/tasks/roboarc-blanket-fold-02-i01-standard/task.toml
@@ -31,6 +31,7 @@ simulator = "roboarc"
robot = "aloha-2"
family = "blanket_fold"
mode = "standard" # RGB + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.4"
oracle = "full"
difficulty = "easier"
diff --git a/tasks/roboarc-blanket-fold-03-i01-privileged/task.toml b/tasks/roboarc-blanket-fold-03-i01-privileged/task.toml
index 2f56a1dd3c..d3d11c60e1 100644
--- a/tasks/roboarc-blanket-fold-03-i01-privileged/task.toml
+++ b/tasks/roboarc-blanket-fold-03-i01-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "roboarc" # RoboARC on MuJoCo 3.14.0 (CPU physics, CUDA
robot = "aloha-2"
family = "blanket_fold"
mode = "privileged" # privileged (state-based tooling, resets) | standard (robot as a service)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
difficulty = "easy" # the variation's tier (docs/rcb/TAXONOMY.md)
grasp = "assisted (the pbd_local_patch_v1 grasp regions on blanket)"
diff --git a/tasks/roboarc-blanket-fold-03-i01-standard/task.toml b/tasks/roboarc-blanket-fold-03-i01-standard/task.toml
index 29994fee6d..74d87743c7 100644
--- a/tasks/roboarc-blanket-fold-03-i01-standard/task.toml
+++ b/tasks/roboarc-blanket-fold-03-i01-standard/task.toml
@@ -31,6 +31,7 @@ simulator = "roboarc"
robot = "aloha-2"
family = "blanket_fold"
mode = "standard" # RGB + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.4"
oracle = "full"
difficulty = "easy"
diff --git a/tasks/roboarc-blanket-fold-04-i01-privileged/task.toml b/tasks/roboarc-blanket-fold-04-i01-privileged/task.toml
index be2c0780d0..7c69249467 100644
--- a/tasks/roboarc-blanket-fold-04-i01-privileged/task.toml
+++ b/tasks/roboarc-blanket-fold-04-i01-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "roboarc" # RoboARC on MuJoCo 3.14.0 (CPU physics, CUDA
robot = "aloha-2"
family = "blanket_fold"
mode = "privileged" # privileged (state-based tooling, resets) | standard (robot as a service)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
difficulty = "medium" # the variation's tier (docs/rcb/TAXONOMY.md)
grasp = "assisted (the pbd_local_patch_v1 grasp regions on blanket)"
diff --git a/tasks/roboarc-blanket-fold-04-i01-standard/task.toml b/tasks/roboarc-blanket-fold-04-i01-standard/task.toml
index 4f6928b526..5bd3864a1c 100644
--- a/tasks/roboarc-blanket-fold-04-i01-standard/task.toml
+++ b/tasks/roboarc-blanket-fold-04-i01-standard/task.toml
@@ -31,6 +31,7 @@ simulator = "roboarc"
robot = "aloha-2"
family = "blanket_fold"
mode = "standard" # RGB + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.4"
oracle = "full"
difficulty = "medium"
diff --git a/tasks/roboarc-blanket-fold-05-i01-privileged/task.toml b/tasks/roboarc-blanket-fold-05-i01-privileged/task.toml
index eedc66eefd..88d006e096 100644
--- a/tasks/roboarc-blanket-fold-05-i01-privileged/task.toml
+++ b/tasks/roboarc-blanket-fold-05-i01-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "roboarc" # RoboARC on MuJoCo 3.14.0 (CPU physics, CUDA
robot = "aloha-2"
family = "blanket_fold"
mode = "privileged" # privileged (state-based tooling, resets) | standard (robot as a service)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
difficulty = "hard" # the variation's tier (docs/rcb/TAXONOMY.md)
grasp = "assisted (the pbd_local_patch_v1 grasp regions on blanket)"
diff --git a/tasks/roboarc-blanket-fold-05-i01-standard/task.toml b/tasks/roboarc-blanket-fold-05-i01-standard/task.toml
index a9d7521ac0..e62c463afa 100644
--- a/tasks/roboarc-blanket-fold-05-i01-standard/task.toml
+++ b/tasks/roboarc-blanket-fold-05-i01-standard/task.toml
@@ -31,6 +31,7 @@ simulator = "roboarc"
robot = "aloha-2"
family = "blanket_fold"
mode = "standard" # RGB + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.4"
oracle = "full"
difficulty = "hard"
diff --git a/tasks/roboarc-buzz-wire-3d-01-i01-privileged/task.toml b/tasks/roboarc-buzz-wire-3d-01-i01-privileged/task.toml
index 4e87f4682f..64aff7315b 100644
--- a/tasks/roboarc-buzz-wire-3d-01-i01-privileged/task.toml
+++ b/tasks/roboarc-buzz-wire-3d-01-i01-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "roboarc" # RoboARC on MuJoCo 3.14.0 (CPU float64 physic
robot = "franka-robotiq-2f85"
family = "buzz_wire_3d"
mode = "privileged" # privileged (state-based tooling, resets) | standard (robot as a service)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
difficulty = "very_easy" # the variation's tier (docs/TAXONOMY.md)
grasp = "assisted (the contact_weld_v1 grasp regions on wand)"
diff --git a/tasks/roboarc-buzz-wire-3d-01-i01-standard/task.toml b/tasks/roboarc-buzz-wire-3d-01-i01-standard/task.toml
index d5abd48a11..2ae4090b6b 100644
--- a/tasks/roboarc-buzz-wire-3d-01-i01-standard/task.toml
+++ b/tasks/roboarc-buzz-wire-3d-01-i01-standard/task.toml
@@ -31,6 +31,7 @@ simulator = "roboarc"
robot = "franka-robotiq-2f85"
family = "buzz_wire_3d"
mode = "standard" # RGB + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.4"
oracle = "full"
difficulty = "very_easy"
diff --git a/tasks/roboarc-buzz-wire-3d-02-i01-privileged/task.toml b/tasks/roboarc-buzz-wire-3d-02-i01-privileged/task.toml
index 3c8d61b73a..4d43194aea 100644
--- a/tasks/roboarc-buzz-wire-3d-02-i01-privileged/task.toml
+++ b/tasks/roboarc-buzz-wire-3d-02-i01-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "roboarc" # RoboARC on MuJoCo 3.14.0 (CPU float64 physic
robot = "franka-robotiq-2f85"
family = "buzz_wire_3d"
mode = "privileged" # privileged (state-based tooling, resets) | standard (robot as a service)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
difficulty = "easier" # the variation's tier (docs/TAXONOMY.md)
grasp = "assisted (the contact_weld_v1 grasp regions on wand)"
diff --git a/tasks/roboarc-buzz-wire-3d-02-i01-standard/task.toml b/tasks/roboarc-buzz-wire-3d-02-i01-standard/task.toml
index 2572c3f40a..f2dc60bee2 100644
--- a/tasks/roboarc-buzz-wire-3d-02-i01-standard/task.toml
+++ b/tasks/roboarc-buzz-wire-3d-02-i01-standard/task.toml
@@ -31,6 +31,7 @@ simulator = "roboarc"
robot = "franka-robotiq-2f85"
family = "buzz_wire_3d"
mode = "standard" # RGB + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.4"
oracle = "full"
difficulty = "easier"
diff --git a/tasks/roboarc-buzz-wire-3d-03-i01-privileged/task.toml b/tasks/roboarc-buzz-wire-3d-03-i01-privileged/task.toml
index 1e2b8a5e65..c5219f52e5 100644
--- a/tasks/roboarc-buzz-wire-3d-03-i01-privileged/task.toml
+++ b/tasks/roboarc-buzz-wire-3d-03-i01-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "roboarc" # RoboARC on MuJoCo 3.14.0 (CPU float64 physic
robot = "franka-robotiq-2f85"
family = "buzz_wire_3d"
mode = "privileged" # privileged (state-based tooling, resets) | standard (robot as a service)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No oracle; human review in PR #N (PROTOCOL §2.7). No reference solution and no positive example yet: nothing is known to solve this task. Its physics was accepted on the family's acceptance rounds and a smoke stress round on this instance."
difficulty = "easy" # the variation's tier (docs/TAXONOMY.md)
diff --git a/tasks/roboarc-buzz-wire-3d-03-i01-standard/task.toml b/tasks/roboarc-buzz-wire-3d-03-i01-standard/task.toml
index 1a2192ffcb..9560d47a3a 100644
--- a/tasks/roboarc-buzz-wire-3d-03-i01-standard/task.toml
+++ b/tasks/roboarc-buzz-wire-3d-03-i01-standard/task.toml
@@ -31,6 +31,7 @@ simulator = "roboarc"
robot = "franka-robotiq-2f85"
family = "buzz_wire_3d"
mode = "standard" # RGB + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.4"
oracle = "none"
oracle_evidence = "No oracle; human review in PR #N (PROTOCOL §2.7). No reference solution and no positive example yet: nothing is known to solve this task. Its physics was accepted on the family's acceptance rounds and a smoke stress round on this instance."
diff --git a/tasks/roboarc-buzz-wire-3d-04-i01-privileged/task.toml b/tasks/roboarc-buzz-wire-3d-04-i01-privileged/task.toml
index 27708ffa5c..f116606ddb 100644
--- a/tasks/roboarc-buzz-wire-3d-04-i01-privileged/task.toml
+++ b/tasks/roboarc-buzz-wire-3d-04-i01-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "roboarc" # RoboARC on MuJoCo 3.14.0 (CPU float64 physic
robot = "franka-robotiq-2f85"
family = "buzz_wire_3d"
mode = "privileged" # privileged (state-based tooling, resets) | standard (robot as a service)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No oracle; human review in PR #N (PROTOCOL §2.7). No reference solution and no positive example yet: nothing is known to solve this task. Its physics was accepted on the family's acceptance rounds and a smoke stress round on this instance."
difficulty = "medium" # the variation's tier (docs/TAXONOMY.md)
diff --git a/tasks/roboarc-buzz-wire-3d-04-i01-standard/task.toml b/tasks/roboarc-buzz-wire-3d-04-i01-standard/task.toml
index adaf952645..a1f29c1486 100644
--- a/tasks/roboarc-buzz-wire-3d-04-i01-standard/task.toml
+++ b/tasks/roboarc-buzz-wire-3d-04-i01-standard/task.toml
@@ -31,6 +31,7 @@ simulator = "roboarc"
robot = "franka-robotiq-2f85"
family = "buzz_wire_3d"
mode = "standard" # RGB + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.4"
oracle = "none"
oracle_evidence = "No oracle; human review in PR #N (PROTOCOL §2.7). No reference solution and no positive example yet: nothing is known to solve this task. Its physics was accepted on the family's acceptance rounds and a smoke stress round on this instance."
diff --git a/tasks/roboarc-buzz-wire-3d-05-i01-privileged/task.toml b/tasks/roboarc-buzz-wire-3d-05-i01-privileged/task.toml
index a1eb00fdfe..aecce7e192 100644
--- a/tasks/roboarc-buzz-wire-3d-05-i01-privileged/task.toml
+++ b/tasks/roboarc-buzz-wire-3d-05-i01-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "roboarc" # RoboARC on MuJoCo 3.14.0 (CPU float64 physic
robot = "franka-robotiq-2f85"
family = "buzz_wire_3d"
mode = "privileged" # privileged (state-based tooling, resets) | standard (robot as a service)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No oracle; human review in PR #N (PROTOCOL §2.7). No reference solution and no positive example yet: nothing is known to solve this task. Its physics was accepted on the family's acceptance rounds and a smoke stress round on this instance."
difficulty = "hard" # the variation's tier (docs/TAXONOMY.md)
diff --git a/tasks/roboarc-buzz-wire-3d-05-i01-standard/task.toml b/tasks/roboarc-buzz-wire-3d-05-i01-standard/task.toml
index 6115326164..6b240978f0 100644
--- a/tasks/roboarc-buzz-wire-3d-05-i01-standard/task.toml
+++ b/tasks/roboarc-buzz-wire-3d-05-i01-standard/task.toml
@@ -31,6 +31,7 @@ simulator = "roboarc"
robot = "franka-robotiq-2f85"
family = "buzz_wire_3d"
mode = "standard" # RGB + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.4"
oracle = "none"
oracle_evidence = "No oracle; human review in PR #N (PROTOCOL §2.7). No reference solution and no positive example yet: nothing is known to solve this task. Its physics was accepted on the family's acceptance rounds and a smoke stress round on this instance."
diff --git a/tasks/roboarc-buzz-wire-3d-06-i01-privileged/task.toml b/tasks/roboarc-buzz-wire-3d-06-i01-privileged/task.toml
index cd090ff9b7..1a55dcb7b7 100644
--- a/tasks/roboarc-buzz-wire-3d-06-i01-privileged/task.toml
+++ b/tasks/roboarc-buzz-wire-3d-06-i01-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "roboarc" # RoboARC on MuJoCo 3.14.0 (CPU float64 physic
robot = "franka-robotiq-2f85"
family = "buzz_wire_3d"
mode = "privileged" # privileged (state-based tooling, resets) | standard (robot as a service)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
difficulty = "very_easy" # the variation's tier (docs/rcb/TAXONOMY.md)
grasp = "assisted (the contact_weld_v1 grasp regions on wand)"
diff --git a/tasks/roboarc-buzz-wire-3d-06-i01-standard/task.toml b/tasks/roboarc-buzz-wire-3d-06-i01-standard/task.toml
index 8bcb3903c2..97f2213e41 100644
--- a/tasks/roboarc-buzz-wire-3d-06-i01-standard/task.toml
+++ b/tasks/roboarc-buzz-wire-3d-06-i01-standard/task.toml
@@ -31,6 +31,7 @@ simulator = "roboarc"
robot = "franka-robotiq-2f85"
family = "buzz_wire_3d"
mode = "standard" # RGB + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.4"
oracle = "full"
difficulty = "very_easy"
diff --git a/tasks/roboarc-buzz-wire-3d-07-i01-privileged/task.toml b/tasks/roboarc-buzz-wire-3d-07-i01-privileged/task.toml
index 8a5cec6577..c3249cf6f7 100644
--- a/tasks/roboarc-buzz-wire-3d-07-i01-privileged/task.toml
+++ b/tasks/roboarc-buzz-wire-3d-07-i01-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "roboarc" # RoboARC on MuJoCo 3.14.0 (CPU float64 physic
robot = "franka-robotiq-2f85"
family = "buzz_wire_3d"
mode = "privileged" # privileged (state-based tooling, resets) | standard (robot as a service)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
difficulty = "easier" # the variation's tier (docs/rcb/TAXONOMY.md)
grasp = "assisted (the contact_weld_v1 grasp regions on wand)"
diff --git a/tasks/roboarc-buzz-wire-3d-07-i01-standard/task.toml b/tasks/roboarc-buzz-wire-3d-07-i01-standard/task.toml
index 59eb3764d2..16cd042507 100644
--- a/tasks/roboarc-buzz-wire-3d-07-i01-standard/task.toml
+++ b/tasks/roboarc-buzz-wire-3d-07-i01-standard/task.toml
@@ -31,6 +31,7 @@ simulator = "roboarc"
robot = "franka-robotiq-2f85"
family = "buzz_wire_3d"
mode = "standard" # RGB + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.4"
oracle = "full"
difficulty = "easier"
diff --git a/tasks/roboarc-buzz-wire-3d-08-i01-privileged/task.toml b/tasks/roboarc-buzz-wire-3d-08-i01-privileged/task.toml
index 4836d93784..458ab0d014 100644
--- a/tasks/roboarc-buzz-wire-3d-08-i01-privileged/task.toml
+++ b/tasks/roboarc-buzz-wire-3d-08-i01-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "roboarc" # RoboARC on MuJoCo 3.14.0 (CPU float64 physic
robot = "franka-robotiq-2f85"
family = "buzz_wire_3d"
mode = "privileged" # privileged (state-based tooling, resets) | standard (robot as a service)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
difficulty = "easy" # the variation's tier (docs/rcb/TAXONOMY.md)
grasp = "assisted (the contact_weld_v1 grasp regions on wand)"
diff --git a/tasks/roboarc-buzz-wire-3d-08-i01-standard/task.toml b/tasks/roboarc-buzz-wire-3d-08-i01-standard/task.toml
index 5f05c94f57..70303c9935 100644
--- a/tasks/roboarc-buzz-wire-3d-08-i01-standard/task.toml
+++ b/tasks/roboarc-buzz-wire-3d-08-i01-standard/task.toml
@@ -31,6 +31,7 @@ simulator = "roboarc"
robot = "franka-robotiq-2f85"
family = "buzz_wire_3d"
mode = "standard" # RGB + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.4"
oracle = "full"
difficulty = "easy"
diff --git a/tasks/roboarc-buzz-wire-3d-09-i01-privileged/task.toml b/tasks/roboarc-buzz-wire-3d-09-i01-privileged/task.toml
index 76fb316f37..0a471b4e1b 100644
--- a/tasks/roboarc-buzz-wire-3d-09-i01-privileged/task.toml
+++ b/tasks/roboarc-buzz-wire-3d-09-i01-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "roboarc" # RoboARC on MuJoCo 3.14.0 (CPU float64 physic
robot = "franka-robotiq-2f85"
family = "buzz_wire_3d"
mode = "privileged" # privileged (state-based tooling, resets) | standard (robot as a service)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
difficulty = "medium" # the variation's tier (docs/rcb/TAXONOMY.md)
grasp = "assisted (the contact_weld_v1 grasp regions on wand)"
diff --git a/tasks/roboarc-buzz-wire-3d-09-i01-standard/task.toml b/tasks/roboarc-buzz-wire-3d-09-i01-standard/task.toml
index f21e684533..973034ecb9 100644
--- a/tasks/roboarc-buzz-wire-3d-09-i01-standard/task.toml
+++ b/tasks/roboarc-buzz-wire-3d-09-i01-standard/task.toml
@@ -31,6 +31,7 @@ simulator = "roboarc"
robot = "franka-robotiq-2f85"
family = "buzz_wire_3d"
mode = "standard" # RGB + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.4"
oracle = "full"
difficulty = "medium"
diff --git a/tasks/roboarc-buzz-wire-3d-10-i01-privileged/task.toml b/tasks/roboarc-buzz-wire-3d-10-i01-privileged/task.toml
index 5311371b9f..5d763b0e0a 100644
--- a/tasks/roboarc-buzz-wire-3d-10-i01-privileged/task.toml
+++ b/tasks/roboarc-buzz-wire-3d-10-i01-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "roboarc" # RoboARC on MuJoCo 3.14.0 (CPU float64 physic
robot = "franka-robotiq-2f85"
family = "buzz_wire_3d"
mode = "privileged" # privileged (state-based tooling, resets) | standard (robot as a service)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
difficulty = "hard" # the variation's tier (docs/rcb/TAXONOMY.md)
grasp = "assisted (the contact_weld_v1 grasp regions on wand)"
diff --git a/tasks/roboarc-buzz-wire-3d-10-i01-standard/task.toml b/tasks/roboarc-buzz-wire-3d-10-i01-standard/task.toml
index d69e6f2358..1155b1c916 100644
--- a/tasks/roboarc-buzz-wire-3d-10-i01-standard/task.toml
+++ b/tasks/roboarc-buzz-wire-3d-10-i01-standard/task.toml
@@ -31,6 +31,7 @@ simulator = "roboarc"
robot = "franka-robotiq-2f85"
family = "buzz_wire_3d"
mode = "standard" # RGB + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.4"
oracle = "full"
difficulty = "hard"
diff --git a/tasks/roboarc-cast-coil-01-i01-privileged/task.toml b/tasks/roboarc-cast-coil-01-i01-privileged/task.toml
index a333ea6640..c05ed7e93d 100644
--- a/tasks/roboarc-cast-coil-01-i01-privileged/task.toml
+++ b/tasks/roboarc-cast-coil-01-i01-privileged/task.toml
@@ -16,6 +16,7 @@ robot = "aloha-2"
family = "cast_coil"
difficulty = "very_easy"
mode = "privileged" # privileged (state-based tooling, resets) | standard (robot as a service)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
grasp = "assisted (the contact_weld_v1 grasp regions on coil_gold, coil_silver)"
success_rule = "task-native success (goal_complete True; frozen parameters; terminal revision release_settling_v2_grasp_contact_2mm)"
diff --git a/tasks/roboarc-cast-coil-01-i01-standard/task.toml b/tasks/roboarc-cast-coil-01-i01-standard/task.toml
index 50dfea9e75..985560b39c 100644
--- a/tasks/roboarc-cast-coil-01-i01-standard/task.toml
+++ b/tasks/roboarc-cast-coil-01-i01-standard/task.toml
@@ -32,6 +32,7 @@ robot = "aloha-2"
family = "cast_coil"
difficulty = "very_easy"
mode = "standard" # RGB + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.4"
oracle = "full"
grasp = "assisted (the contact_weld_v1 grasp regions on coil_gold, coil_silver)"
diff --git a/tasks/roboarc-cast-coil-02-i01-privileged/task.toml b/tasks/roboarc-cast-coil-02-i01-privileged/task.toml
index 8d2aee05c5..8df92ad922 100644
--- a/tasks/roboarc-cast-coil-02-i01-privileged/task.toml
+++ b/tasks/roboarc-cast-coil-02-i01-privileged/task.toml
@@ -16,6 +16,7 @@ robot = "aloha-2"
family = "cast_coil"
difficulty = "easier"
mode = "privileged" # privileged (state-based tooling, resets) | standard (robot as a service)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
grasp = "assisted (the contact_weld_v1 grasp regions on coil_gold, coil_silver)"
success_rule = "task-native success (goal_complete True; frozen parameters; terminal revision release_settling_v2_grasp_contact_2mm)"
diff --git a/tasks/roboarc-cast-coil-02-i01-standard/task.toml b/tasks/roboarc-cast-coil-02-i01-standard/task.toml
index 933fde58ce..834bee1527 100644
--- a/tasks/roboarc-cast-coil-02-i01-standard/task.toml
+++ b/tasks/roboarc-cast-coil-02-i01-standard/task.toml
@@ -32,6 +32,7 @@ robot = "aloha-2"
family = "cast_coil"
difficulty = "easier"
mode = "standard" # RGB + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.4"
oracle = "full"
grasp = "assisted (the contact_weld_v1 grasp regions on coil_gold, coil_silver)"
diff --git a/tasks/roboarc-cast-coil-03-i01-privileged/task.toml b/tasks/roboarc-cast-coil-03-i01-privileged/task.toml
index 65d7b05b26..90eea31d99 100644
--- a/tasks/roboarc-cast-coil-03-i01-privileged/task.toml
+++ b/tasks/roboarc-cast-coil-03-i01-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "roboarc" # RoboARC on MuJoCo 3.14.0 (CPU float64 physic
robot = "aloha-2"
family = "cast_coil"
mode = "privileged" # privileged (state-based tooling, resets) | standard (robot as a service)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
difficulty = "easy" # the variation's tier (docs/TAXONOMY.md)
grasp = "assisted (the contact_weld_v1 grasp regions on coil_gold, coil_silver)"
diff --git a/tasks/roboarc-cast-coil-03-i01-standard/task.toml b/tasks/roboarc-cast-coil-03-i01-standard/task.toml
index cb9d327732..e407738f67 100644
--- a/tasks/roboarc-cast-coil-03-i01-standard/task.toml
+++ b/tasks/roboarc-cast-coil-03-i01-standard/task.toml
@@ -31,6 +31,7 @@ simulator = "roboarc"
robot = "aloha-2"
family = "cast_coil"
mode = "standard" # RGB + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.4"
oracle = "full"
difficulty = "easy"
diff --git a/tasks/roboarc-cast-coil-04-i01-privileged/task.toml b/tasks/roboarc-cast-coil-04-i01-privileged/task.toml
index 22fd9ec22c..f78fa0e857 100644
--- a/tasks/roboarc-cast-coil-04-i01-privileged/task.toml
+++ b/tasks/roboarc-cast-coil-04-i01-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "roboarc" # RoboARC on MuJoCo 3.14.0 (CPU float64 physic
robot = "aloha-2"
family = "cast_coil"
mode = "privileged" # privileged (state-based tooling, resets) | standard (robot as a service)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
difficulty = "medium" # the variation's tier (docs/TAXONOMY.md)
grasp = "assisted (the contact_weld_v1 grasp regions on coil_gold, coil_silver)"
diff --git a/tasks/roboarc-cast-coil-04-i01-standard/task.toml b/tasks/roboarc-cast-coil-04-i01-standard/task.toml
index c2fe496cfd..f8af0fd870 100644
--- a/tasks/roboarc-cast-coil-04-i01-standard/task.toml
+++ b/tasks/roboarc-cast-coil-04-i01-standard/task.toml
@@ -31,6 +31,7 @@ simulator = "roboarc"
robot = "aloha-2"
family = "cast_coil"
mode = "standard" # RGB + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.4"
oracle = "full"
difficulty = "medium"
diff --git a/tasks/roboarc-cast-coil-05-i01-privileged/task.toml b/tasks/roboarc-cast-coil-05-i01-privileged/task.toml
index b6cbb7c5c8..94f4c97f4b 100644
--- a/tasks/roboarc-cast-coil-05-i01-privileged/task.toml
+++ b/tasks/roboarc-cast-coil-05-i01-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "roboarc" # RoboARC on MuJoCo 3.14.0 (CPU float64 physic
robot = "aloha-2"
family = "cast_coil"
mode = "privileged" # privileged (state-based tooling, resets) | standard (robot as a service)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
difficulty = "hard" # the variation's tier (docs/TAXONOMY.md)
grasp = "assisted (the contact_weld_v1 grasp regions on coil_gold, coil_silver)"
diff --git a/tasks/roboarc-cast-coil-05-i01-standard/task.toml b/tasks/roboarc-cast-coil-05-i01-standard/task.toml
index 33d81ff005..efc1038ea0 100644
--- a/tasks/roboarc-cast-coil-05-i01-standard/task.toml
+++ b/tasks/roboarc-cast-coil-05-i01-standard/task.toml
@@ -31,6 +31,7 @@ simulator = "roboarc"
robot = "aloha-2"
family = "cast_coil"
mode = "standard" # RGB + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.4"
oracle = "full"
difficulty = "hard"
diff --git a/tasks/roboarc-cast-coil-06-i01-privileged/task.toml b/tasks/roboarc-cast-coil-06-i01-privileged/task.toml
index 89f16b8544..b2a3e7bf0b 100644
--- a/tasks/roboarc-cast-coil-06-i01-privileged/task.toml
+++ b/tasks/roboarc-cast-coil-06-i01-privileged/task.toml
@@ -16,6 +16,7 @@ robot = "aloha-2"
family = "cast_coil"
difficulty = "very_easy"
mode = "privileged" # privileged (state-based tooling, resets) | standard (robot as a service)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
grasp = "assisted (the contact_weld_v1 grasp regions on coil_gold, coil_silver)"
success_rule = "task-native success (goal_complete True; frozen parameters; terminal revision release_settling_v2_grasp_contact_2mm)"
diff --git a/tasks/roboarc-cast-coil-06-i01-standard/task.toml b/tasks/roboarc-cast-coil-06-i01-standard/task.toml
index d4657989a6..3ac3623dc7 100644
--- a/tasks/roboarc-cast-coil-06-i01-standard/task.toml
+++ b/tasks/roboarc-cast-coil-06-i01-standard/task.toml
@@ -32,6 +32,7 @@ robot = "aloha-2"
family = "cast_coil"
difficulty = "very_easy"
mode = "standard" # RGB + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.4"
oracle = "full"
grasp = "assisted (the contact_weld_v1 grasp regions on coil_gold, coil_silver)"
diff --git a/tasks/roboarc-cast-coil-07-i01-privileged/task.toml b/tasks/roboarc-cast-coil-07-i01-privileged/task.toml
index 6f54b307ef..9339e843ea 100644
--- a/tasks/roboarc-cast-coil-07-i01-privileged/task.toml
+++ b/tasks/roboarc-cast-coil-07-i01-privileged/task.toml
@@ -16,6 +16,7 @@ robot = "aloha-2"
family = "cast_coil"
difficulty = "easier"
mode = "privileged" # privileged (state-based tooling, resets) | standard (robot as a service)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
grasp = "assisted (the contact_weld_v1 grasp regions on coil_gold, coil_silver)"
success_rule = "task-native success (goal_complete True; frozen parameters; terminal revision release_settling_v2_grasp_contact_2mm)"
diff --git a/tasks/roboarc-cast-coil-07-i01-standard/task.toml b/tasks/roboarc-cast-coil-07-i01-standard/task.toml
index 2565f9f16b..1eb84888f8 100644
--- a/tasks/roboarc-cast-coil-07-i01-standard/task.toml
+++ b/tasks/roboarc-cast-coil-07-i01-standard/task.toml
@@ -32,6 +32,7 @@ robot = "aloha-2"
family = "cast_coil"
difficulty = "easier"
mode = "standard" # RGB + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.4"
oracle = "full"
grasp = "assisted (the contact_weld_v1 grasp regions on coil_gold, coil_silver)"
diff --git a/tasks/roboarc-cast-coil-08-i01-privileged/task.toml b/tasks/roboarc-cast-coil-08-i01-privileged/task.toml
index 4166969076..834f34d1ba 100644
--- a/tasks/roboarc-cast-coil-08-i01-privileged/task.toml
+++ b/tasks/roboarc-cast-coil-08-i01-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "roboarc" # RoboARC on MuJoCo 3.14.0 (CPU float64 physic
robot = "aloha-2"
family = "cast_coil"
mode = "privileged" # privileged (state-based tooling, resets) | standard (robot as a service)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
difficulty = "easy" # the variation's tier (docs/rcb/TAXONOMY.md)
grasp = "assisted (the contact_weld_v1 grasp regions on coil_gold, coil_silver)"
diff --git a/tasks/roboarc-cast-coil-08-i01-standard/task.toml b/tasks/roboarc-cast-coil-08-i01-standard/task.toml
index d81a75b446..eb972ed852 100644
--- a/tasks/roboarc-cast-coil-08-i01-standard/task.toml
+++ b/tasks/roboarc-cast-coil-08-i01-standard/task.toml
@@ -31,6 +31,7 @@ simulator = "roboarc"
robot = "aloha-2"
family = "cast_coil"
mode = "standard" # RGB + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.4"
oracle = "full"
difficulty = "easy"
diff --git a/tasks/roboarc-cast-keyhole-01-i01-privileged/task.toml b/tasks/roboarc-cast-keyhole-01-i01-privileged/task.toml
index 3887422fb1..7922a1e5e3 100644
--- a/tasks/roboarc-cast-keyhole-01-i01-privileged/task.toml
+++ b/tasks/roboarc-cast-keyhole-01-i01-privileged/task.toml
@@ -16,6 +16,7 @@ robot = "aloha-2"
family = "cast_keyhole"
difficulty = "very_easy"
mode = "privileged" # privileged (state-based tooling, resets) | standard (robot as a service)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
grasp = "assisted (the contact_weld_v1 grasp regions on keyhole_gold, keyhole_silver)"
success_rule = "task-native success (goal_complete True; frozen parameters; terminal revision cast_keyhole_placement_v4_grasp_contact_2mm)"
diff --git a/tasks/roboarc-cast-keyhole-01-i01-standard/task.toml b/tasks/roboarc-cast-keyhole-01-i01-standard/task.toml
index 192b37d2f4..313467feb8 100644
--- a/tasks/roboarc-cast-keyhole-01-i01-standard/task.toml
+++ b/tasks/roboarc-cast-keyhole-01-i01-standard/task.toml
@@ -32,6 +32,7 @@ robot = "aloha-2"
family = "cast_keyhole"
difficulty = "very_easy"
mode = "standard" # RGB + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.4"
oracle = "full"
grasp = "assisted (the contact_weld_v1 grasp regions on keyhole_gold, keyhole_silver)"
diff --git a/tasks/roboarc-cast-keyhole-02-i01-privileged/task.toml b/tasks/roboarc-cast-keyhole-02-i01-privileged/task.toml
index 277c44a123..883ef05f10 100644
--- a/tasks/roboarc-cast-keyhole-02-i01-privileged/task.toml
+++ b/tasks/roboarc-cast-keyhole-02-i01-privileged/task.toml
@@ -16,6 +16,7 @@ robot = "aloha-2"
family = "cast_keyhole"
difficulty = "easier"
mode = "privileged" # privileged (state-based tooling, resets) | standard (robot as a service)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
grasp = "assisted (the contact_weld_v1 grasp regions on keyhole_gold, keyhole_silver)"
success_rule = "task-native success (goal_complete True; frozen parameters; terminal revision cast_keyhole_placement_v4_grasp_contact_2mm)"
diff --git a/tasks/roboarc-cast-keyhole-02-i01-standard/task.toml b/tasks/roboarc-cast-keyhole-02-i01-standard/task.toml
index acdbf46661..f3d6791f0b 100644
--- a/tasks/roboarc-cast-keyhole-02-i01-standard/task.toml
+++ b/tasks/roboarc-cast-keyhole-02-i01-standard/task.toml
@@ -32,6 +32,7 @@ robot = "aloha-2"
family = "cast_keyhole"
difficulty = "easier"
mode = "standard" # RGB + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.4"
oracle = "full"
grasp = "assisted (the contact_weld_v1 grasp regions on keyhole_gold, keyhole_silver)"
diff --git a/tasks/roboarc-cast-keyhole-03-i01-privileged/task.toml b/tasks/roboarc-cast-keyhole-03-i01-privileged/task.toml
index 24d1920c75..948b5e5e56 100644
--- a/tasks/roboarc-cast-keyhole-03-i01-privileged/task.toml
+++ b/tasks/roboarc-cast-keyhole-03-i01-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "roboarc" # RoboARC on MuJoCo 3.14.0 (CPU float64 physic
robot = "aloha-2"
family = "cast_keyhole"
mode = "privileged" # privileged (state-based tooling, resets) | standard (robot as a service)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
difficulty = "easy" # the variation's tier (docs/TAXONOMY.md)
grasp = "assisted (the contact_weld_v1 grasp regions on keyhole_gold, keyhole_silver)"
diff --git a/tasks/roboarc-cast-keyhole-03-i01-standard/task.toml b/tasks/roboarc-cast-keyhole-03-i01-standard/task.toml
index c8764510a8..4bcd5df875 100644
--- a/tasks/roboarc-cast-keyhole-03-i01-standard/task.toml
+++ b/tasks/roboarc-cast-keyhole-03-i01-standard/task.toml
@@ -31,6 +31,7 @@ simulator = "roboarc"
robot = "aloha-2"
family = "cast_keyhole"
mode = "standard" # RGB + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.4"
oracle = "full"
difficulty = "easy"
diff --git a/tasks/roboarc-cast-keyhole-04-i01-privileged/task.toml b/tasks/roboarc-cast-keyhole-04-i01-privileged/task.toml
index 144944d5b3..08f68ab582 100644
--- a/tasks/roboarc-cast-keyhole-04-i01-privileged/task.toml
+++ b/tasks/roboarc-cast-keyhole-04-i01-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "roboarc" # RoboARC on MuJoCo 3.14.0 (CPU float64 physic
robot = "aloha-2"
family = "cast_keyhole"
mode = "privileged" # privileged (state-based tooling, resets) | standard (robot as a service)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
difficulty = "medium" # the variation's tier (docs/TAXONOMY.md)
grasp = "assisted (the contact_weld_v1 grasp regions on keyhole_gold, keyhole_silver)"
diff --git a/tasks/roboarc-cast-keyhole-04-i01-standard/task.toml b/tasks/roboarc-cast-keyhole-04-i01-standard/task.toml
index c340240517..3cd8dbbc03 100644
--- a/tasks/roboarc-cast-keyhole-04-i01-standard/task.toml
+++ b/tasks/roboarc-cast-keyhole-04-i01-standard/task.toml
@@ -31,6 +31,7 @@ simulator = "roboarc"
robot = "aloha-2"
family = "cast_keyhole"
mode = "standard" # RGB + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.4"
oracle = "full"
difficulty = "medium"
diff --git a/tasks/roboarc-cast-keyhole-05-i01-privileged/task.toml b/tasks/roboarc-cast-keyhole-05-i01-privileged/task.toml
index e9ef183283..c8c757e171 100644
--- a/tasks/roboarc-cast-keyhole-05-i01-privileged/task.toml
+++ b/tasks/roboarc-cast-keyhole-05-i01-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "roboarc" # RoboARC on MuJoCo 3.14.0 (CPU float64 physic
robot = "aloha-2"
family = "cast_keyhole"
mode = "privileged" # privileged (state-based tooling, resets) | standard (robot as a service)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
difficulty = "hard" # the variation's tier (docs/TAXONOMY.md)
grasp = "assisted (the contact_weld_v1 grasp regions on keyhole_gold, keyhole_silver)"
diff --git a/tasks/roboarc-cast-keyhole-05-i01-standard/task.toml b/tasks/roboarc-cast-keyhole-05-i01-standard/task.toml
index 9ff2e8bed1..5eff7a2831 100644
--- a/tasks/roboarc-cast-keyhole-05-i01-standard/task.toml
+++ b/tasks/roboarc-cast-keyhole-05-i01-standard/task.toml
@@ -31,6 +31,7 @@ simulator = "roboarc"
robot = "aloha-2"
family = "cast_keyhole"
mode = "standard" # RGB + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.4"
oracle = "full"
difficulty = "hard"
diff --git a/tasks/roboarc-cast-keyhole-06-i01-privileged/task.toml b/tasks/roboarc-cast-keyhole-06-i01-privileged/task.toml
index b1dde82b6e..047e2cb763 100644
--- a/tasks/roboarc-cast-keyhole-06-i01-privileged/task.toml
+++ b/tasks/roboarc-cast-keyhole-06-i01-privileged/task.toml
@@ -16,6 +16,7 @@ robot = "aloha-2"
family = "cast_keyhole"
difficulty = "hard"
mode = "privileged" # privileged (state-based tooling, resets) | standard (robot as a service)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
grasp = "assisted (the contact_weld_v1 pre-weld: the grippers start closed on keyhole_gold, keyhole_silver)"
success_rule = "task-native success (goal_complete True; frozen parameters; terminal revision cast_keyhole_placement_v4_grasp_contact_2mm)"
diff --git a/tasks/roboarc-cast-keyhole-06-i01-standard/task.toml b/tasks/roboarc-cast-keyhole-06-i01-standard/task.toml
index 4750a55725..eaae6f6982 100644
--- a/tasks/roboarc-cast-keyhole-06-i01-standard/task.toml
+++ b/tasks/roboarc-cast-keyhole-06-i01-standard/task.toml
@@ -32,6 +32,7 @@ robot = "aloha-2"
family = "cast_keyhole"
difficulty = "hard"
mode = "standard" # RGB + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.4"
oracle = "full"
grasp = "assisted (the contact_weld_v1 pre-weld: the grippers start closed on keyhole_gold, keyhole_silver)"
diff --git a/tasks/roboarc-fiber-bundling-01-i01-privileged/task.toml b/tasks/roboarc-fiber-bundling-01-i01-privileged/task.toml
index e118cd2a55..7dd97d1c85 100644
--- a/tasks/roboarc-fiber-bundling-01-i01-privileged/task.toml
+++ b/tasks/roboarc-fiber-bundling-01-i01-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "roboarc" # RoboARC on MuJoCo 3.14.0 (CPU float64 physic
robot = "franka-robotiq-2f85"
family = "fiber_bundling"
mode = "privileged" # privileged (state-based tooling, resets) | standard (robot as a service)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
difficulty = "very_easy" # the variation's tier (docs/TAXONOMY.md)
grasp = "assisted (the contact_weld_v1 grasp regions on fiber_bundle)"
diff --git a/tasks/roboarc-fiber-bundling-01-i01-standard/task.toml b/tasks/roboarc-fiber-bundling-01-i01-standard/task.toml
index 5d4d7b3096..4c529bb546 100644
--- a/tasks/roboarc-fiber-bundling-01-i01-standard/task.toml
+++ b/tasks/roboarc-fiber-bundling-01-i01-standard/task.toml
@@ -31,6 +31,7 @@ simulator = "roboarc"
robot = "franka-robotiq-2f85"
family = "fiber_bundling"
mode = "standard" # RGB + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.4"
oracle = "full"
difficulty = "very_easy"
diff --git a/tasks/roboarc-fiber-bundling-02-i01-privileged/task.toml b/tasks/roboarc-fiber-bundling-02-i01-privileged/task.toml
index 1a57268446..6cc8504808 100644
--- a/tasks/roboarc-fiber-bundling-02-i01-privileged/task.toml
+++ b/tasks/roboarc-fiber-bundling-02-i01-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "roboarc" # RoboARC on MuJoCo 3.14.0 (CPU float64 physic
robot = "franka-robotiq-2f85"
family = "fiber_bundling"
mode = "privileged" # privileged (state-based tooling, resets) | standard (robot as a service)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
difficulty = "easier" # the variation's tier (docs/TAXONOMY.md)
grasp = "assisted (the contact_weld_v1 grasp regions on fiber_bundle)"
diff --git a/tasks/roboarc-fiber-bundling-02-i01-standard/task.toml b/tasks/roboarc-fiber-bundling-02-i01-standard/task.toml
index 89bb1e97ec..06703dc1e7 100644
--- a/tasks/roboarc-fiber-bundling-02-i01-standard/task.toml
+++ b/tasks/roboarc-fiber-bundling-02-i01-standard/task.toml
@@ -31,6 +31,7 @@ simulator = "roboarc"
robot = "franka-robotiq-2f85"
family = "fiber_bundling"
mode = "standard" # RGB + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.4"
oracle = "full"
difficulty = "easier"
diff --git a/tasks/roboarc-fiber-bundling-03-i01-privileged/task.toml b/tasks/roboarc-fiber-bundling-03-i01-privileged/task.toml
index 7b7f853fac..cbf7688c63 100644
--- a/tasks/roboarc-fiber-bundling-03-i01-privileged/task.toml
+++ b/tasks/roboarc-fiber-bundling-03-i01-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "roboarc" # RoboARC on MuJoCo 3.14.0 (CPU float64 physic
robot = "franka-robotiq-2f85"
family = "fiber_bundling"
mode = "privileged" # privileged (state-based tooling, resets) | standard (robot as a service)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
difficulty = "easy" # the variation's tier (docs/TAXONOMY.md)
grasp = "assisted (the contact_weld_v1 grasp regions on fiber_bundle)"
diff --git a/tasks/roboarc-fiber-bundling-03-i01-standard/task.toml b/tasks/roboarc-fiber-bundling-03-i01-standard/task.toml
index 535b276256..61ab16f321 100644
--- a/tasks/roboarc-fiber-bundling-03-i01-standard/task.toml
+++ b/tasks/roboarc-fiber-bundling-03-i01-standard/task.toml
@@ -31,6 +31,7 @@ simulator = "roboarc"
robot = "franka-robotiq-2f85"
family = "fiber_bundling"
mode = "standard" # RGB + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.4"
oracle = "full"
difficulty = "easy"
diff --git a/tasks/roboarc-fiber-bundling-04-i01-privileged/task.toml b/tasks/roboarc-fiber-bundling-04-i01-privileged/task.toml
index b143dfeda5..6e5bcf58bc 100644
--- a/tasks/roboarc-fiber-bundling-04-i01-privileged/task.toml
+++ b/tasks/roboarc-fiber-bundling-04-i01-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "roboarc" # RoboARC on MuJoCo 3.14.0 (CPU float64 physic
robot = "franka-robotiq-2f85"
family = "fiber_bundling"
mode = "privileged" # privileged (state-based tooling, resets) | standard (robot as a service)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
difficulty = "medium" # the variation's tier (docs/TAXONOMY.md)
grasp = "assisted (the contact_weld_v1 grasp regions on fiber_bundle)"
diff --git a/tasks/roboarc-fiber-bundling-04-i01-standard/task.toml b/tasks/roboarc-fiber-bundling-04-i01-standard/task.toml
index 34d4ed64af..f63a87b2e1 100644
--- a/tasks/roboarc-fiber-bundling-04-i01-standard/task.toml
+++ b/tasks/roboarc-fiber-bundling-04-i01-standard/task.toml
@@ -31,6 +31,7 @@ simulator = "roboarc"
robot = "franka-robotiq-2f85"
family = "fiber_bundling"
mode = "standard" # RGB + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.4"
oracle = "full"
difficulty = "medium"
diff --git a/tasks/roboarc-fiber-bundling-05-i01-privileged/task.toml b/tasks/roboarc-fiber-bundling-05-i01-privileged/task.toml
index 31b1f72598..3fc717b47e 100644
--- a/tasks/roboarc-fiber-bundling-05-i01-privileged/task.toml
+++ b/tasks/roboarc-fiber-bundling-05-i01-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "roboarc" # RoboARC on MuJoCo 3.14.0 (CPU float64 physic
robot = "franka-robotiq-2f85"
family = "fiber_bundling"
mode = "privileged" # privileged (state-based tooling, resets) | standard (robot as a service)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No oracle; human review in PR #N (PROTOCOL §2.7). No reference solution and no positive example yet: nothing is known to solve this task. Its physics was accepted on the family's acceptance rounds and a smoke stress round on this instance."
difficulty = "hard" # the variation's tier (docs/TAXONOMY.md)
diff --git a/tasks/roboarc-fiber-bundling-05-i01-standard/task.toml b/tasks/roboarc-fiber-bundling-05-i01-standard/task.toml
index ab8239ea2d..666266b91e 100644
--- a/tasks/roboarc-fiber-bundling-05-i01-standard/task.toml
+++ b/tasks/roboarc-fiber-bundling-05-i01-standard/task.toml
@@ -31,6 +31,7 @@ simulator = "roboarc"
robot = "franka-robotiq-2f85"
family = "fiber_bundling"
mode = "standard" # RGB + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.4"
oracle = "none"
oracle_evidence = "No oracle; human review in PR #N (PROTOCOL §2.7). No reference solution and no positive example yet: nothing is known to solve this task. Its physics was accepted on the family's acceptance rounds and a smoke stress round on this instance."
diff --git a/tasks/roboarc-fiber-bundling-06-i01-privileged/task.toml b/tasks/roboarc-fiber-bundling-06-i01-privileged/task.toml
index 9e5cf03084..84babb110d 100644
--- a/tasks/roboarc-fiber-bundling-06-i01-privileged/task.toml
+++ b/tasks/roboarc-fiber-bundling-06-i01-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "roboarc" # RoboARC on MuJoCo 3.14.0 (CPU float64 physic
robot = "franka-robotiq-2f85"
family = "fiber_bundling"
mode = "privileged" # privileged (state-based tooling, resets) | standard (robot as a service)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
difficulty = "very_easy" # the variation's tier (docs/rcb/TAXONOMY.md)
grasp = "assisted (the contact_weld_v1 grasp regions on fiber_bundle)"
diff --git a/tasks/roboarc-fiber-bundling-06-i01-standard/task.toml b/tasks/roboarc-fiber-bundling-06-i01-standard/task.toml
index 5b08f3982b..5e497e8d4b 100644
--- a/tasks/roboarc-fiber-bundling-06-i01-standard/task.toml
+++ b/tasks/roboarc-fiber-bundling-06-i01-standard/task.toml
@@ -31,6 +31,7 @@ simulator = "roboarc"
robot = "franka-robotiq-2f85"
family = "fiber_bundling"
mode = "standard" # RGB + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.4"
oracle = "full"
difficulty = "very_easy"
diff --git a/tasks/roboarc-fiber-bundling-07-i01-privileged/task.toml b/tasks/roboarc-fiber-bundling-07-i01-privileged/task.toml
index 3fe387d3d2..da8d2e123a 100644
--- a/tasks/roboarc-fiber-bundling-07-i01-privileged/task.toml
+++ b/tasks/roboarc-fiber-bundling-07-i01-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "roboarc" # RoboARC on MuJoCo 3.14.0 (CPU float64 physic
robot = "franka-robotiq-2f85"
family = "fiber_bundling"
mode = "privileged" # privileged (state-based tooling, resets) | standard (robot as a service)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
difficulty = "easier" # the variation's tier (docs/rcb/TAXONOMY.md)
grasp = "assisted (the contact_weld_v1 grasp regions on fiber_bundle)"
diff --git a/tasks/roboarc-fiber-bundling-07-i01-standard/task.toml b/tasks/roboarc-fiber-bundling-07-i01-standard/task.toml
index beea26c945..e6bf1ce800 100644
--- a/tasks/roboarc-fiber-bundling-07-i01-standard/task.toml
+++ b/tasks/roboarc-fiber-bundling-07-i01-standard/task.toml
@@ -31,6 +31,7 @@ simulator = "roboarc"
robot = "franka-robotiq-2f85"
family = "fiber_bundling"
mode = "standard" # RGB + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.4"
oracle = "full"
difficulty = "easier"
diff --git a/tasks/roboarc-fiber-bundling-08-i01-privileged/task.toml b/tasks/roboarc-fiber-bundling-08-i01-privileged/task.toml
index 801d1b1459..8eee930fd8 100644
--- a/tasks/roboarc-fiber-bundling-08-i01-privileged/task.toml
+++ b/tasks/roboarc-fiber-bundling-08-i01-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "roboarc" # RoboARC on MuJoCo 3.14.0 (CPU float64 physic
robot = "franka-robotiq-2f85"
family = "fiber_bundling"
mode = "privileged" # privileged (state-based tooling, resets) | standard (robot as a service)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
difficulty = "easy" # the variation's tier (docs/rcb/TAXONOMY.md)
grasp = "assisted (the contact_weld_v1 grasp regions on fiber_bundle)"
diff --git a/tasks/roboarc-fiber-bundling-08-i01-standard/task.toml b/tasks/roboarc-fiber-bundling-08-i01-standard/task.toml
index c6abc65160..256a0722d6 100644
--- a/tasks/roboarc-fiber-bundling-08-i01-standard/task.toml
+++ b/tasks/roboarc-fiber-bundling-08-i01-standard/task.toml
@@ -31,6 +31,7 @@ simulator = "roboarc"
robot = "franka-robotiq-2f85"
family = "fiber_bundling"
mode = "standard" # RGB + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.4"
oracle = "full"
difficulty = "easy"
diff --git a/tasks/roboarc-fiber-bundling-09-i01-privileged/task.toml b/tasks/roboarc-fiber-bundling-09-i01-privileged/task.toml
index 1035061484..a679675727 100644
--- a/tasks/roboarc-fiber-bundling-09-i01-privileged/task.toml
+++ b/tasks/roboarc-fiber-bundling-09-i01-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "roboarc" # RoboARC on MuJoCo 3.14.0 (CPU float64 physic
robot = "franka-robotiq-2f85"
family = "fiber_bundling"
mode = "privileged" # privileged (state-based tooling, resets) | standard (robot as a service)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
difficulty = "medium" # the variation's tier (docs/rcb/TAXONOMY.md)
grasp = "assisted (the contact_weld_v1 grasp regions on fiber_bundle)"
diff --git a/tasks/roboarc-fiber-bundling-09-i01-standard/task.toml b/tasks/roboarc-fiber-bundling-09-i01-standard/task.toml
index 8a5efc63c3..fbd6c790c9 100644
--- a/tasks/roboarc-fiber-bundling-09-i01-standard/task.toml
+++ b/tasks/roboarc-fiber-bundling-09-i01-standard/task.toml
@@ -31,6 +31,7 @@ simulator = "roboarc"
robot = "franka-robotiq-2f85"
family = "fiber_bundling"
mode = "standard" # RGB + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.4"
oracle = "full"
difficulty = "medium"
diff --git a/tasks/roboarc-fiber-bundling-10-i01-privileged/task.toml b/tasks/roboarc-fiber-bundling-10-i01-privileged/task.toml
index 0447d312d1..41af665530 100644
--- a/tasks/roboarc-fiber-bundling-10-i01-privileged/task.toml
+++ b/tasks/roboarc-fiber-bundling-10-i01-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "roboarc" # RoboARC on MuJoCo 3.14.0 (CPU float64 physic
robot = "franka-robotiq-2f85"
family = "fiber_bundling"
mode = "privileged" # privileged (state-based tooling, resets) | standard (robot as a service)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
difficulty = "hard" # the variation's tier (docs/rcb/TAXONOMY.md)
grasp = "assisted (the contact_weld_v1 grasp regions on fiber_bundle)"
diff --git a/tasks/roboarc-fiber-bundling-10-i01-standard/task.toml b/tasks/roboarc-fiber-bundling-10-i01-standard/task.toml
index 4b480090cc..6e1075be3e 100644
--- a/tasks/roboarc-fiber-bundling-10-i01-standard/task.toml
+++ b/tasks/roboarc-fiber-bundling-10-i01-standard/task.toml
@@ -31,6 +31,7 @@ simulator = "roboarc"
robot = "franka-robotiq-2f85"
family = "fiber_bundling"
mode = "standard" # RGB + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.4"
oracle = "full"
difficulty = "hard"
diff --git a/tasks/roboarc-hedgehog-in-cage-01-i01-privileged/task.toml b/tasks/roboarc-hedgehog-in-cage-01-i01-privileged/task.toml
index 462b41c64c..c94a5d6b25 100644
--- a/tasks/roboarc-hedgehog-in-cage-01-i01-privileged/task.toml
+++ b/tasks/roboarc-hedgehog-in-cage-01-i01-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "roboarc" # RoboARC on MuJoCo 3.14.0 (CPU float64 physic
robot = "franka-robotiq-2f85"
family = "hedgehog_in_cage"
mode = "privileged" # privileged (state-based tooling, resets) | standard (robot as a service)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
difficulty = "very_easy" # the variation's tier (docs/TAXONOMY.md)
grasp = "assisted (the contact_weld_v1 grasp regions on hedgehog_ball)"
diff --git a/tasks/roboarc-hedgehog-in-cage-01-i01-standard/task.toml b/tasks/roboarc-hedgehog-in-cage-01-i01-standard/task.toml
index 8f819dfe7d..69256d0cd5 100644
--- a/tasks/roboarc-hedgehog-in-cage-01-i01-standard/task.toml
+++ b/tasks/roboarc-hedgehog-in-cage-01-i01-standard/task.toml
@@ -31,6 +31,7 @@ simulator = "roboarc"
robot = "franka-robotiq-2f85"
family = "hedgehog_in_cage"
mode = "standard" # RGB + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.4"
oracle = "full"
difficulty = "very_easy"
diff --git a/tasks/roboarc-hedgehog-in-cage-02-i01-privileged/task.toml b/tasks/roboarc-hedgehog-in-cage-02-i01-privileged/task.toml
index 60208a616e..8b1e89abb9 100644
--- a/tasks/roboarc-hedgehog-in-cage-02-i01-privileged/task.toml
+++ b/tasks/roboarc-hedgehog-in-cage-02-i01-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "roboarc" # RoboARC on MuJoCo 3.14.0 (CPU float64 physic
robot = "franka-robotiq-2f85"
family = "hedgehog_in_cage"
mode = "privileged" # privileged (state-based tooling, resets) | standard (robot as a service)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
difficulty = "easier" # the variation's tier (docs/TAXONOMY.md)
grasp = "assisted (the contact_weld_v1 grasp regions on hedgehog_ball)"
diff --git a/tasks/roboarc-hedgehog-in-cage-02-i01-standard/task.toml b/tasks/roboarc-hedgehog-in-cage-02-i01-standard/task.toml
index f137b1dd32..6939f89d3c 100644
--- a/tasks/roboarc-hedgehog-in-cage-02-i01-standard/task.toml
+++ b/tasks/roboarc-hedgehog-in-cage-02-i01-standard/task.toml
@@ -31,6 +31,7 @@ simulator = "roboarc"
robot = "franka-robotiq-2f85"
family = "hedgehog_in_cage"
mode = "standard" # RGB + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.4"
oracle = "full"
difficulty = "easier"
diff --git a/tasks/roboarc-hedgehog-in-cage-03-i01-privileged/task.toml b/tasks/roboarc-hedgehog-in-cage-03-i01-privileged/task.toml
index 1d306b2337..a064d617b6 100644
--- a/tasks/roboarc-hedgehog-in-cage-03-i01-privileged/task.toml
+++ b/tasks/roboarc-hedgehog-in-cage-03-i01-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "roboarc" # RoboARC on MuJoCo 3.14.0 (CPU float64 physic
robot = "franka-robotiq-2f85"
family = "hedgehog_in_cage"
mode = "privileged" # privileged (state-based tooling, resets) | standard (robot as a service)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
difficulty = "easy" # the variation's tier (docs/TAXONOMY.md)
grasp = "assisted (the contact_weld_v1 grasp regions on hedgehog_ball)"
diff --git a/tasks/roboarc-hedgehog-in-cage-03-i01-standard/task.toml b/tasks/roboarc-hedgehog-in-cage-03-i01-standard/task.toml
index e9de064950..0b332e6cd7 100644
--- a/tasks/roboarc-hedgehog-in-cage-03-i01-standard/task.toml
+++ b/tasks/roboarc-hedgehog-in-cage-03-i01-standard/task.toml
@@ -31,6 +31,7 @@ simulator = "roboarc"
robot = "franka-robotiq-2f85"
family = "hedgehog_in_cage"
mode = "standard" # RGB + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.4"
oracle = "full"
difficulty = "easy"
diff --git a/tasks/roboarc-hedgehog-in-cage-04-i01-privileged/task.toml b/tasks/roboarc-hedgehog-in-cage-04-i01-privileged/task.toml
index ee2e253597..5302692d64 100644
--- a/tasks/roboarc-hedgehog-in-cage-04-i01-privileged/task.toml
+++ b/tasks/roboarc-hedgehog-in-cage-04-i01-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "roboarc" # RoboARC on MuJoCo 3.14.0 (CPU float64 physic
robot = "franka-robotiq-2f85"
family = "hedgehog_in_cage"
mode = "privileged" # privileged (state-based tooling, resets) | standard (robot as a service)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
difficulty = "medium" # the variation's tier (docs/TAXONOMY.md)
grasp = "assisted (the contact_weld_v1 grasp regions on hedgehog_ball)"
diff --git a/tasks/roboarc-hedgehog-in-cage-04-i01-standard/task.toml b/tasks/roboarc-hedgehog-in-cage-04-i01-standard/task.toml
index 4c46f685a7..267f8bb289 100644
--- a/tasks/roboarc-hedgehog-in-cage-04-i01-standard/task.toml
+++ b/tasks/roboarc-hedgehog-in-cage-04-i01-standard/task.toml
@@ -31,6 +31,7 @@ simulator = "roboarc"
robot = "franka-robotiq-2f85"
family = "hedgehog_in_cage"
mode = "standard" # RGB + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.4"
oracle = "full"
difficulty = "medium"
diff --git a/tasks/roboarc-hedgehog-in-cage-05-i01-privileged/task.toml b/tasks/roboarc-hedgehog-in-cage-05-i01-privileged/task.toml
index 56a9e34da7..7a50fa817c 100644
--- a/tasks/roboarc-hedgehog-in-cage-05-i01-privileged/task.toml
+++ b/tasks/roboarc-hedgehog-in-cage-05-i01-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "roboarc" # RoboARC on MuJoCo 3.14.0 (CPU float64 physic
robot = "franka-robotiq-2f85"
family = "hedgehog_in_cage"
mode = "privileged" # privileged (state-based tooling, resets) | standard (robot as a service)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
difficulty = "hard" # the variation's tier (docs/TAXONOMY.md)
grasp = "assisted (the contact_weld_v1 grasp regions on hedgehog_ball)"
diff --git a/tasks/roboarc-hedgehog-in-cage-05-i01-standard/task.toml b/tasks/roboarc-hedgehog-in-cage-05-i01-standard/task.toml
index 2ed6c6c80f..f821b8d364 100644
--- a/tasks/roboarc-hedgehog-in-cage-05-i01-standard/task.toml
+++ b/tasks/roboarc-hedgehog-in-cage-05-i01-standard/task.toml
@@ -31,6 +31,7 @@ simulator = "roboarc"
robot = "franka-robotiq-2f85"
family = "hedgehog_in_cage"
mode = "standard" # RGB + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.4"
oracle = "full"
difficulty = "hard"
diff --git a/tasks/roboarc-hedgehog-in-cage-06-i01-privileged/task.toml b/tasks/roboarc-hedgehog-in-cage-06-i01-privileged/task.toml
index eb9f214ac0..4d4e17c5d2 100644
--- a/tasks/roboarc-hedgehog-in-cage-06-i01-privileged/task.toml
+++ b/tasks/roboarc-hedgehog-in-cage-06-i01-privileged/task.toml
@@ -16,6 +16,7 @@ robot = "franka-robotiq-2f85"
family = "hedgehog_in_cage"
difficulty = "hard"
mode = "privileged" # privileged (state-based tooling, resets) | standard (robot as a service)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
grasp = "assisted (the contact_weld_v1 grasp regions on hedgehog_ball, push_rod)"
success_rule = "task-native success (goal_complete True; frozen parameters; terminal revision hedgehog_placement_v4_grasp_contact_2mm)"
diff --git a/tasks/roboarc-hedgehog-in-cage-06-i01-standard/task.toml b/tasks/roboarc-hedgehog-in-cage-06-i01-standard/task.toml
index ba132151fc..9475ca2804 100644
--- a/tasks/roboarc-hedgehog-in-cage-06-i01-standard/task.toml
+++ b/tasks/roboarc-hedgehog-in-cage-06-i01-standard/task.toml
@@ -32,6 +32,7 @@ robot = "franka-robotiq-2f85"
family = "hedgehog_in_cage"
difficulty = "hard"
mode = "standard" # RGB + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.4"
oracle = "full"
grasp = "assisted (the contact_weld_v1 grasp regions on hedgehog_ball, push_rod)"
diff --git a/tasks/roboarc-hedgehog-in-cage-07-i01-privileged/task.toml b/tasks/roboarc-hedgehog-in-cage-07-i01-privileged/task.toml
index e0769939a6..5e133ad9fe 100644
--- a/tasks/roboarc-hedgehog-in-cage-07-i01-privileged/task.toml
+++ b/tasks/roboarc-hedgehog-in-cage-07-i01-privileged/task.toml
@@ -16,6 +16,7 @@ robot = "franka-robotiq-2f85"
family = "hedgehog_in_cage"
difficulty = "easy"
mode = "privileged" # privileged (state-based tooling, resets) | standard (robot as a service)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
grasp = "assisted (the contact_weld_v1 grasp regions on owl, rod)"
success_rule = "task-native success (goal_complete True; frozen parameters; terminal revision hedgehog_placement_v4_grasp_contact_2mm)"
diff --git a/tasks/roboarc-hedgehog-in-cage-07-i01-standard/task.toml b/tasks/roboarc-hedgehog-in-cage-07-i01-standard/task.toml
index 9de4ce094d..b211017aa2 100644
--- a/tasks/roboarc-hedgehog-in-cage-07-i01-standard/task.toml
+++ b/tasks/roboarc-hedgehog-in-cage-07-i01-standard/task.toml
@@ -32,6 +32,7 @@ robot = "franka-robotiq-2f85"
family = "hedgehog_in_cage"
difficulty = "easy"
mode = "standard" # RGB + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.4"
oracle = "full"
grasp = "assisted (the contact_weld_v1 grasp regions on owl, rod)"
diff --git a/tasks/roboarc-hedgehog-in-cage-08-i01-privileged/task.toml b/tasks/roboarc-hedgehog-in-cage-08-i01-privileged/task.toml
index e269152554..59d31c19a4 100644
--- a/tasks/roboarc-hedgehog-in-cage-08-i01-privileged/task.toml
+++ b/tasks/roboarc-hedgehog-in-cage-08-i01-privileged/task.toml
@@ -16,6 +16,7 @@ robot = "franka-robotiq-2f85"
family = "hedgehog_in_cage"
difficulty = "easier"
mode = "privileged" # privileged (state-based tooling, resets) | standard (robot as a service)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
grasp = "assisted (the contact_weld_v1 grasp regions on hedgehog_ball)"
success_rule = "task-native success (goal_complete True; frozen parameters; terminal revision hedgehog_placement_v4_grasp_contact_2mm)"
diff --git a/tasks/roboarc-hedgehog-in-cage-08-i01-standard/task.toml b/tasks/roboarc-hedgehog-in-cage-08-i01-standard/task.toml
index 26cee5fe18..79393aea80 100644
--- a/tasks/roboarc-hedgehog-in-cage-08-i01-standard/task.toml
+++ b/tasks/roboarc-hedgehog-in-cage-08-i01-standard/task.toml
@@ -32,6 +32,7 @@ robot = "franka-robotiq-2f85"
family = "hedgehog_in_cage"
difficulty = "easier"
mode = "standard" # RGB + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.4"
oracle = "full"
grasp = "assisted (the contact_weld_v1 grasp regions on hedgehog_ball)"
diff --git a/tasks/roboarc-horseshoe-ring-01-i01-privileged/task.toml b/tasks/roboarc-horseshoe-ring-01-i01-privileged/task.toml
index f4b8854127..4a8f3e3f28 100644
--- a/tasks/roboarc-horseshoe-ring-01-i01-privileged/task.toml
+++ b/tasks/roboarc-horseshoe-ring-01-i01-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "roboarc" # RoboARC on MuJoCo 3.14.0 (CPU float64 physic
robot = "aloha-2"
family = "horseshoe_ring"
mode = "privileged" # privileged (state-based tooling, resets) | standard (robot as a service)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
difficulty = "very_easy" # the variation's tier (docs/TAXONOMY.md)
grasp = "assisted (the contact_weld_v1 grasp regions on horseshoe_left, horseshoe_right, target_ring)"
diff --git a/tasks/roboarc-horseshoe-ring-01-i01-standard/task.toml b/tasks/roboarc-horseshoe-ring-01-i01-standard/task.toml
index 0c4248154d..0c185e1bbc 100644
--- a/tasks/roboarc-horseshoe-ring-01-i01-standard/task.toml
+++ b/tasks/roboarc-horseshoe-ring-01-i01-standard/task.toml
@@ -31,6 +31,7 @@ simulator = "roboarc"
robot = "aloha-2"
family = "horseshoe_ring"
mode = "standard" # RGB + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.4"
oracle = "full"
difficulty = "very_easy"
diff --git a/tasks/roboarc-horseshoe-ring-02-i01-privileged/task.toml b/tasks/roboarc-horseshoe-ring-02-i01-privileged/task.toml
index c20a1e52f2..1bbd4447e6 100644
--- a/tasks/roboarc-horseshoe-ring-02-i01-privileged/task.toml
+++ b/tasks/roboarc-horseshoe-ring-02-i01-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "roboarc" # RoboARC on MuJoCo 3.14.0 (CPU float64 physic
robot = "aloha-2"
family = "horseshoe_ring"
mode = "privileged" # privileged (state-based tooling, resets) | standard (robot as a service)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
difficulty = "easier" # the variation's tier (docs/TAXONOMY.md)
grasp = "assisted (the contact_weld_v1 grasp regions on horseshoe_left, horseshoe_right, target_ring)"
diff --git a/tasks/roboarc-horseshoe-ring-02-i01-standard/task.toml b/tasks/roboarc-horseshoe-ring-02-i01-standard/task.toml
index 7ceb6a5f99..620e5a789e 100644
--- a/tasks/roboarc-horseshoe-ring-02-i01-standard/task.toml
+++ b/tasks/roboarc-horseshoe-ring-02-i01-standard/task.toml
@@ -31,6 +31,7 @@ simulator = "roboarc"
robot = "aloha-2"
family = "horseshoe_ring"
mode = "standard" # RGB + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.4"
oracle = "full"
difficulty = "easier"
diff --git a/tasks/roboarc-horseshoe-ring-03-i01-privileged/task.toml b/tasks/roboarc-horseshoe-ring-03-i01-privileged/task.toml
index f5d1cac7ae..7468268e6a 100644
--- a/tasks/roboarc-horseshoe-ring-03-i01-privileged/task.toml
+++ b/tasks/roboarc-horseshoe-ring-03-i01-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "roboarc" # RoboARC on MuJoCo 3.14.0 (CPU float64 physic
robot = "aloha-2"
family = "horseshoe_ring"
mode = "privileged" # privileged (state-based tooling, resets) | standard (robot as a service)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
difficulty = "easy" # the variation's tier (docs/TAXONOMY.md)
grasp = "assisted (the contact_weld_v1 grasp regions on horseshoe_left, horseshoe_right, target_ring)"
diff --git a/tasks/roboarc-horseshoe-ring-03-i01-standard/task.toml b/tasks/roboarc-horseshoe-ring-03-i01-standard/task.toml
index 8e91f17d6f..909f53d07d 100644
--- a/tasks/roboarc-horseshoe-ring-03-i01-standard/task.toml
+++ b/tasks/roboarc-horseshoe-ring-03-i01-standard/task.toml
@@ -31,6 +31,7 @@ simulator = "roboarc"
robot = "aloha-2"
family = "horseshoe_ring"
mode = "standard" # RGB + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.4"
oracle = "full"
difficulty = "easy"
diff --git a/tasks/roboarc-horseshoe-ring-04-i01-privileged/task.toml b/tasks/roboarc-horseshoe-ring-04-i01-privileged/task.toml
index 579d79cbbb..b38e56b49e 100644
--- a/tasks/roboarc-horseshoe-ring-04-i01-privileged/task.toml
+++ b/tasks/roboarc-horseshoe-ring-04-i01-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "roboarc" # RoboARC on MuJoCo 3.14.0 (CPU float64 physic
robot = "aloha-2"
family = "horseshoe_ring"
mode = "privileged" # privileged (state-based tooling, resets) | standard (robot as a service)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
difficulty = "medium" # the variation's tier (docs/TAXONOMY.md)
grasp = "assisted (the contact_weld_v1 grasp regions on horseshoe_left, horseshoe_right, target_ring)"
diff --git a/tasks/roboarc-horseshoe-ring-04-i01-standard/task.toml b/tasks/roboarc-horseshoe-ring-04-i01-standard/task.toml
index a08aaad202..ef8e504e19 100644
--- a/tasks/roboarc-horseshoe-ring-04-i01-standard/task.toml
+++ b/tasks/roboarc-horseshoe-ring-04-i01-standard/task.toml
@@ -31,6 +31,7 @@ simulator = "roboarc"
robot = "aloha-2"
family = "horseshoe_ring"
mode = "standard" # RGB + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.4"
oracle = "full"
difficulty = "medium"
diff --git a/tasks/roboarc-horseshoe-ring-05-i01-privileged/task.toml b/tasks/roboarc-horseshoe-ring-05-i01-privileged/task.toml
index 0570cc76b2..00bd28a265 100644
--- a/tasks/roboarc-horseshoe-ring-05-i01-privileged/task.toml
+++ b/tasks/roboarc-horseshoe-ring-05-i01-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "roboarc" # RoboARC on MuJoCo 3.14.0 (CPU float64 physic
robot = "aloha-2"
family = "horseshoe_ring"
mode = "privileged" # privileged (state-based tooling, resets) | standard (robot as a service)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
difficulty = "hard" # the variation's tier (docs/TAXONOMY.md)
grasp = "assisted (the contact_weld_v1 grasp regions on horseshoe_left, horseshoe_right, target_ring)"
diff --git a/tasks/roboarc-horseshoe-ring-05-i01-standard/task.toml b/tasks/roboarc-horseshoe-ring-05-i01-standard/task.toml
index b60451b708..1b9d65535d 100644
--- a/tasks/roboarc-horseshoe-ring-05-i01-standard/task.toml
+++ b/tasks/roboarc-horseshoe-ring-05-i01-standard/task.toml
@@ -31,6 +31,7 @@ simulator = "roboarc"
robot = "aloha-2"
family = "horseshoe_ring"
mode = "standard" # RGB + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.4"
oracle = "full"
difficulty = "hard"
diff --git a/tasks/roboarc-interlocked-claws-01-i01-privileged/task.toml b/tasks/roboarc-interlocked-claws-01-i01-privileged/task.toml
index 797f960366..dc7cecbd2a 100644
--- a/tasks/roboarc-interlocked-claws-01-i01-privileged/task.toml
+++ b/tasks/roboarc-interlocked-claws-01-i01-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "roboarc" # RoboARC on MuJoCo 3.14.0 (CPU float64 physic
robot = "aloha-2"
family = "interlocked_claws"
mode = "privileged" # privileged (state-based tooling, resets) | standard (robot as a service)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
difficulty = "very_easy" # the variation's tier (docs/TAXONOMY.md)
grasp = "assisted (the contact_weld_v1 pre-weld: the grippers start closed on piece_0, piece_1)"
diff --git a/tasks/roboarc-interlocked-claws-01-i01-standard/task.toml b/tasks/roboarc-interlocked-claws-01-i01-standard/task.toml
index d409ba8d40..e649d3cc51 100644
--- a/tasks/roboarc-interlocked-claws-01-i01-standard/task.toml
+++ b/tasks/roboarc-interlocked-claws-01-i01-standard/task.toml
@@ -31,6 +31,7 @@ simulator = "roboarc"
robot = "aloha-2"
family = "interlocked_claws"
mode = "standard" # RGB + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.4"
oracle = "full"
difficulty = "very_easy"
diff --git a/tasks/roboarc-interlocked-claws-02-i01-privileged/task.toml b/tasks/roboarc-interlocked-claws-02-i01-privileged/task.toml
index b949fcc561..a243ac6020 100644
--- a/tasks/roboarc-interlocked-claws-02-i01-privileged/task.toml
+++ b/tasks/roboarc-interlocked-claws-02-i01-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "roboarc" # RoboARC on MuJoCo 3.14.0 (CPU float64 physic
robot = "aloha-2"
family = "interlocked_claws"
mode = "privileged" # privileged (state-based tooling, resets) | standard (robot as a service)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
difficulty = "easier" # the variation's tier (docs/TAXONOMY.md)
grasp = "assisted (the contact_weld_v1 pre-weld: the grippers start closed on piece_0, piece_1)"
diff --git a/tasks/roboarc-interlocked-claws-02-i01-standard/task.toml b/tasks/roboarc-interlocked-claws-02-i01-standard/task.toml
index e75550531a..31241d4354 100644
--- a/tasks/roboarc-interlocked-claws-02-i01-standard/task.toml
+++ b/tasks/roboarc-interlocked-claws-02-i01-standard/task.toml
@@ -31,6 +31,7 @@ simulator = "roboarc"
robot = "aloha-2"
family = "interlocked_claws"
mode = "standard" # RGB + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.4"
oracle = "full"
difficulty = "easier"
diff --git a/tasks/roboarc-interlocked-claws-03-i01-privileged/task.toml b/tasks/roboarc-interlocked-claws-03-i01-privileged/task.toml
index 5614ab70ef..572d994d4f 100644
--- a/tasks/roboarc-interlocked-claws-03-i01-privileged/task.toml
+++ b/tasks/roboarc-interlocked-claws-03-i01-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "roboarc" # RoboARC on MuJoCo 3.14.0 (CPU float64 physic
robot = "aloha-2"
family = "interlocked_claws"
mode = "privileged" # privileged (state-based tooling, resets) | standard (robot as a service)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
difficulty = "easy" # the variation's tier (docs/TAXONOMY.md)
grasp = "assisted (the contact_weld_v1 pre-weld: the grippers start closed on piece_0, piece_1)"
diff --git a/tasks/roboarc-interlocked-claws-03-i01-standard/task.toml b/tasks/roboarc-interlocked-claws-03-i01-standard/task.toml
index f45d63fa67..334a18113a 100644
--- a/tasks/roboarc-interlocked-claws-03-i01-standard/task.toml
+++ b/tasks/roboarc-interlocked-claws-03-i01-standard/task.toml
@@ -31,6 +31,7 @@ simulator = "roboarc"
robot = "aloha-2"
family = "interlocked_claws"
mode = "standard" # RGB + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.4"
oracle = "full"
difficulty = "easy"
diff --git a/tasks/roboarc-interlocked-claws-04-i01-privileged/task.toml b/tasks/roboarc-interlocked-claws-04-i01-privileged/task.toml
index a2fb040f41..21c12191d3 100644
--- a/tasks/roboarc-interlocked-claws-04-i01-privileged/task.toml
+++ b/tasks/roboarc-interlocked-claws-04-i01-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "roboarc" # RoboARC on MuJoCo 3.14.0 (CPU float64 physic
robot = "aloha-2"
family = "interlocked_claws"
mode = "privileged" # privileged (state-based tooling, resets) | standard (robot as a service)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
difficulty = "medium" # the variation's tier (docs/TAXONOMY.md)
grasp = "assisted (the contact_weld_v1 pre-weld: the grippers start closed on piece_0, piece_1)"
diff --git a/tasks/roboarc-interlocked-claws-04-i01-standard/task.toml b/tasks/roboarc-interlocked-claws-04-i01-standard/task.toml
index bbaee71f0e..2414cba550 100644
--- a/tasks/roboarc-interlocked-claws-04-i01-standard/task.toml
+++ b/tasks/roboarc-interlocked-claws-04-i01-standard/task.toml
@@ -31,6 +31,7 @@ simulator = "roboarc"
robot = "aloha-2"
family = "interlocked_claws"
mode = "standard" # RGB + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.4"
oracle = "full"
difficulty = "medium"
diff --git a/tasks/roboarc-interlocked-claws-05-i01-privileged/task.toml b/tasks/roboarc-interlocked-claws-05-i01-privileged/task.toml
index 1c2b63fc99..b102c0bc0c 100644
--- a/tasks/roboarc-interlocked-claws-05-i01-privileged/task.toml
+++ b/tasks/roboarc-interlocked-claws-05-i01-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "roboarc" # RoboARC on MuJoCo 3.14.0 (CPU float64 physic
robot = "aloha-2"
family = "interlocked_claws"
mode = "privileged" # privileged (state-based tooling, resets) | standard (robot as a service)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
difficulty = "hard" # the variation's tier (docs/TAXONOMY.md)
grasp = "assisted (the contact_weld_v1 pre-weld: the grippers start closed on claw_0, claw_1)"
diff --git a/tasks/roboarc-interlocked-claws-05-i01-standard/task.toml b/tasks/roboarc-interlocked-claws-05-i01-standard/task.toml
index ab2bddf9d9..c579ed35b9 100644
--- a/tasks/roboarc-interlocked-claws-05-i01-standard/task.toml
+++ b/tasks/roboarc-interlocked-claws-05-i01-standard/task.toml
@@ -31,6 +31,7 @@ simulator = "roboarc"
robot = "aloha-2"
family = "interlocked_claws"
mode = "standard" # RGB + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.4"
oracle = "full"
difficulty = "hard"
diff --git a/tasks/roboarc-interlocked-claws-06-i01-privileged/task.toml b/tasks/roboarc-interlocked-claws-06-i01-privileged/task.toml
index b4890c5282..ebf6acc929 100644
--- a/tasks/roboarc-interlocked-claws-06-i01-privileged/task.toml
+++ b/tasks/roboarc-interlocked-claws-06-i01-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "roboarc" # RoboARC on MuJoCo 3.14.0 (CPU float64 physic
robot = "aloha-2"
family = "interlocked_claws"
mode = "privileged" # privileged (state-based tooling, resets) | standard (robot as a service)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
difficulty = "very_easy" # the variation's tier (docs/rcb/TAXONOMY.md)
grasp = "assisted (the contact_weld_v1 pre-weld: the grippers start closed on piece_0, piece_1)"
diff --git a/tasks/roboarc-interlocked-claws-06-i01-standard/task.toml b/tasks/roboarc-interlocked-claws-06-i01-standard/task.toml
index 5ee7d7955d..75867d7e95 100644
--- a/tasks/roboarc-interlocked-claws-06-i01-standard/task.toml
+++ b/tasks/roboarc-interlocked-claws-06-i01-standard/task.toml
@@ -31,6 +31,7 @@ simulator = "roboarc"
robot = "aloha-2"
family = "interlocked_claws"
mode = "standard" # RGB + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.4"
oracle = "full"
difficulty = "very_easy"
diff --git a/tasks/roboarc-interlocked-claws-07-i01-privileged/task.toml b/tasks/roboarc-interlocked-claws-07-i01-privileged/task.toml
index 003ab0075e..b548200c6e 100644
--- a/tasks/roboarc-interlocked-claws-07-i01-privileged/task.toml
+++ b/tasks/roboarc-interlocked-claws-07-i01-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "roboarc" # RoboARC on MuJoCo 3.14.0 (CPU float64 physic
robot = "aloha-2"
family = "interlocked_claws"
mode = "privileged" # privileged (state-based tooling, resets) | standard (robot as a service)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
difficulty = "easier" # the variation's tier (docs/rcb/TAXONOMY.md)
grasp = "assisted (the contact_weld_v1 pre-weld: the grippers start closed on piece_0, piece_1)"
diff --git a/tasks/roboarc-interlocked-claws-07-i01-standard/task.toml b/tasks/roboarc-interlocked-claws-07-i01-standard/task.toml
index e333f2a3d1..87664f5896 100644
--- a/tasks/roboarc-interlocked-claws-07-i01-standard/task.toml
+++ b/tasks/roboarc-interlocked-claws-07-i01-standard/task.toml
@@ -31,6 +31,7 @@ simulator = "roboarc"
robot = "aloha-2"
family = "interlocked_claws"
mode = "standard" # RGB + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.4"
oracle = "full"
difficulty = "easier"
diff --git a/tasks/roboarc-interlocked-claws-08-i01-privileged/task.toml b/tasks/roboarc-interlocked-claws-08-i01-privileged/task.toml
index a1a5cbcefb..98fa9c1497 100644
--- a/tasks/roboarc-interlocked-claws-08-i01-privileged/task.toml
+++ b/tasks/roboarc-interlocked-claws-08-i01-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "roboarc" # RoboARC on MuJoCo 3.14.0 (CPU float64 physic
robot = "aloha-2"
family = "interlocked_claws"
mode = "privileged" # privileged (state-based tooling, resets) | standard (robot as a service)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
difficulty = "easy" # the variation's tier (docs/rcb/TAXONOMY.md)
grasp = "assisted (the contact_weld_v1 pre-weld: the grippers start closed on piece_0, piece_1)"
diff --git a/tasks/roboarc-interlocked-claws-08-i01-standard/task.toml b/tasks/roboarc-interlocked-claws-08-i01-standard/task.toml
index 19932c2968..b136effdd8 100644
--- a/tasks/roboarc-interlocked-claws-08-i01-standard/task.toml
+++ b/tasks/roboarc-interlocked-claws-08-i01-standard/task.toml
@@ -31,6 +31,7 @@ simulator = "roboarc"
robot = "aloha-2"
family = "interlocked_claws"
mode = "standard" # RGB + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.4"
oracle = "full"
difficulty = "easy"
diff --git a/tasks/roboarc-interlocked-claws-09-i01-privileged/task.toml b/tasks/roboarc-interlocked-claws-09-i01-privileged/task.toml
index fd373139b2..88155a297c 100644
--- a/tasks/roboarc-interlocked-claws-09-i01-privileged/task.toml
+++ b/tasks/roboarc-interlocked-claws-09-i01-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "roboarc" # RoboARC on MuJoCo 3.14.0 (CPU float64 physic
robot = "aloha-2"
family = "interlocked_claws"
mode = "privileged" # privileged (state-based tooling, resets) | standard (robot as a service)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
difficulty = "medium" # the variation's tier (docs/rcb/TAXONOMY.md)
grasp = "assisted (the contact_weld_v1 pre-weld: the grippers start closed on piece_0, piece_1)"
diff --git a/tasks/roboarc-interlocked-claws-09-i01-standard/task.toml b/tasks/roboarc-interlocked-claws-09-i01-standard/task.toml
index 6fac8b5e96..d930370102 100644
--- a/tasks/roboarc-interlocked-claws-09-i01-standard/task.toml
+++ b/tasks/roboarc-interlocked-claws-09-i01-standard/task.toml
@@ -31,6 +31,7 @@ simulator = "roboarc"
robot = "aloha-2"
family = "interlocked_claws"
mode = "standard" # RGB + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.4"
oracle = "full"
difficulty = "medium"
diff --git a/tasks/roboarc-interlocked-claws-10-i01-privileged/task.toml b/tasks/roboarc-interlocked-claws-10-i01-privileged/task.toml
index 17a9a826d3..e637c6b3c5 100644
--- a/tasks/roboarc-interlocked-claws-10-i01-privileged/task.toml
+++ b/tasks/roboarc-interlocked-claws-10-i01-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "roboarc" # RoboARC on MuJoCo 3.14.0 (CPU float64 physic
robot = "aloha-2"
family = "interlocked_claws"
mode = "privileged" # privileged (state-based tooling, resets) | standard (robot as a service)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
difficulty = "hard" # the variation's tier (docs/rcb/TAXONOMY.md)
grasp = "assisted (the contact_weld_v1 pre-weld: the grippers start closed on piece_0, piece_1)"
diff --git a/tasks/roboarc-interlocked-claws-10-i01-standard/task.toml b/tasks/roboarc-interlocked-claws-10-i01-standard/task.toml
index 1268d69c65..18c9708f7f 100644
--- a/tasks/roboarc-interlocked-claws-10-i01-standard/task.toml
+++ b/tasks/roboarc-interlocked-claws-10-i01-standard/task.toml
@@ -31,6 +31,7 @@ simulator = "roboarc"
robot = "aloha-2"
family = "interlocked_claws"
mode = "standard" # RGB + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.4"
oracle = "full"
difficulty = "hard"
diff --git a/tasks/roboarc-interlocked-claws-11-i01-privileged/task.toml b/tasks/roboarc-interlocked-claws-11-i01-privileged/task.toml
index bd59634df5..582e1c02f5 100644
--- a/tasks/roboarc-interlocked-claws-11-i01-privileged/task.toml
+++ b/tasks/roboarc-interlocked-claws-11-i01-privileged/task.toml
@@ -16,6 +16,7 @@ robot = "aloha-2"
family = "interlocked_claws"
difficulty = "medium"
mode = "privileged" # privileged (state-based tooling, resets) | standard (robot as a service)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
grasp = "assisted (the contact_weld_v1 pre-weld: the grippers start closed on ab_plate, c_ring)"
success_rule = "task-native success (goal_complete True; frozen parameters; terminal revision claws_separation_v1_grasp_contact_2mm)"
diff --git a/tasks/roboarc-interlocked-claws-11-i01-standard/task.toml b/tasks/roboarc-interlocked-claws-11-i01-standard/task.toml
index a8ba28a35a..2acbf892c1 100644
--- a/tasks/roboarc-interlocked-claws-11-i01-standard/task.toml
+++ b/tasks/roboarc-interlocked-claws-11-i01-standard/task.toml
@@ -32,6 +32,7 @@ robot = "aloha-2"
family = "interlocked_claws"
difficulty = "medium"
mode = "standard" # RGB + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.4"
oracle = "full"
grasp = "assisted (the contact_weld_v1 pre-weld: the grippers start closed on ab_plate, c_ring)"
diff --git a/tasks/roboarc-interlocked-claws-12-i01-privileged/task.toml b/tasks/roboarc-interlocked-claws-12-i01-privileged/task.toml
index 7ba05622e9..5b8701c5c4 100644
--- a/tasks/roboarc-interlocked-claws-12-i01-privileged/task.toml
+++ b/tasks/roboarc-interlocked-claws-12-i01-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "roboarc" # RoboARC on MuJoCo 3.14.0 (CPU float64 physic
robot = "aloha-2"
family = "interlocked_claws"
mode = "privileged" # privileged (state-based tooling, resets) | standard (robot as a service)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
difficulty = "hard" # the variation's tier (docs/rcb/TAXONOMY.md)
grasp = "assisted (the contact_weld_v1 pre-weld: the grippers start closed on piece_0, piece_1)"
diff --git a/tasks/roboarc-interlocked-claws-12-i01-standard/task.toml b/tasks/roboarc-interlocked-claws-12-i01-standard/task.toml
index 2b24d73022..1320ff96e3 100644
--- a/tasks/roboarc-interlocked-claws-12-i01-standard/task.toml
+++ b/tasks/roboarc-interlocked-claws-12-i01-standard/task.toml
@@ -31,6 +31,7 @@ simulator = "roboarc"
robot = "aloha-2"
family = "interlocked_claws"
mode = "standard" # RGB + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.4"
oracle = "full"
difficulty = "hard"
diff --git a/tasks/roboarc-knot-untying-01-i01-privileged/task.toml b/tasks/roboarc-knot-untying-01-i01-privileged/task.toml
index 37f2130bcc..a50992b8a7 100644
--- a/tasks/roboarc-knot-untying-01-i01-privileged/task.toml
+++ b/tasks/roboarc-knot-untying-01-i01-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "roboarc" # RoboARC on MuJoCo 3.14.0 (CPU physics, Nyx R
robot = "aloha-2"
family = "knot_untying"
mode = "privileged" # privileged (state-based tooling, resets) | standard (robot as a service)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
difficulty = "very_easy" # the variation's tier (docs/TAXONOMY.md)
grasp = "physical (no assisted grasping)"
diff --git a/tasks/roboarc-knot-untying-01-i01-standard/task.toml b/tasks/roboarc-knot-untying-01-i01-standard/task.toml
index a469938344..546196a651 100644
--- a/tasks/roboarc-knot-untying-01-i01-standard/task.toml
+++ b/tasks/roboarc-knot-untying-01-i01-standard/task.toml
@@ -31,6 +31,7 @@ simulator = "roboarc"
robot = "aloha-2"
family = "knot_untying"
mode = "standard" # RGB + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.4"
oracle = "full"
difficulty = "very_easy"
diff --git a/tasks/roboarc-knot-untying-02-i01-privileged/task.toml b/tasks/roboarc-knot-untying-02-i01-privileged/task.toml
index 601cb0f3d5..0ed39b2d73 100644
--- a/tasks/roboarc-knot-untying-02-i01-privileged/task.toml
+++ b/tasks/roboarc-knot-untying-02-i01-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "roboarc" # RoboARC on MuJoCo 3.14.0 (CPU physics, Nyx R
robot = "aloha-2"
family = "knot_untying"
mode = "privileged" # privileged (state-based tooling, resets) | standard (robot as a service)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
difficulty = "easier" # the variation's tier (docs/TAXONOMY.md)
grasp = "physical (no assisted grasping)"
diff --git a/tasks/roboarc-knot-untying-02-i01-standard/task.toml b/tasks/roboarc-knot-untying-02-i01-standard/task.toml
index b53082ff54..5d163daf94 100644
--- a/tasks/roboarc-knot-untying-02-i01-standard/task.toml
+++ b/tasks/roboarc-knot-untying-02-i01-standard/task.toml
@@ -31,6 +31,7 @@ simulator = "roboarc"
robot = "aloha-2"
family = "knot_untying"
mode = "standard" # RGB + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.4"
oracle = "full"
difficulty = "easier"
diff --git a/tasks/roboarc-knot-untying-03-i01-privileged/task.toml b/tasks/roboarc-knot-untying-03-i01-privileged/task.toml
index 30f925d2fd..0c6323b0ef 100644
--- a/tasks/roboarc-knot-untying-03-i01-privileged/task.toml
+++ b/tasks/roboarc-knot-untying-03-i01-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "roboarc" # RoboARC on MuJoCo 3.14.0 (CPU physics, Nyx R
robot = "aloha-2"
family = "knot_untying"
mode = "privileged" # privileged (state-based tooling, resets) | standard (robot as a service)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
difficulty = "easy" # the variation's tier (docs/TAXONOMY.md)
grasp = "physical (no assisted grasping)"
diff --git a/tasks/roboarc-knot-untying-03-i01-standard/task.toml b/tasks/roboarc-knot-untying-03-i01-standard/task.toml
index 679557354c..e3a9908ed8 100644
--- a/tasks/roboarc-knot-untying-03-i01-standard/task.toml
+++ b/tasks/roboarc-knot-untying-03-i01-standard/task.toml
@@ -31,6 +31,7 @@ simulator = "roboarc"
robot = "aloha-2"
family = "knot_untying"
mode = "standard" # RGB + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.4"
oracle = "full"
difficulty = "easy"
diff --git a/tasks/roboarc-knot-untying-04-i01-privileged/task.toml b/tasks/roboarc-knot-untying-04-i01-privileged/task.toml
index 10c101eb85..e7ee0c05f3 100644
--- a/tasks/roboarc-knot-untying-04-i01-privileged/task.toml
+++ b/tasks/roboarc-knot-untying-04-i01-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "roboarc" # RoboARC on MuJoCo 3.14.0 (CPU physics, Nyx R
robot = "aloha-2"
family = "knot_untying"
mode = "privileged" # privileged (state-based tooling, resets) | standard (robot as a service)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
difficulty = "medium" # the variation's tier (docs/TAXONOMY.md)
grasp = "physical (no assisted grasping)"
diff --git a/tasks/roboarc-knot-untying-04-i01-standard/task.toml b/tasks/roboarc-knot-untying-04-i01-standard/task.toml
index 67bcdb91d9..c36ac186a8 100644
--- a/tasks/roboarc-knot-untying-04-i01-standard/task.toml
+++ b/tasks/roboarc-knot-untying-04-i01-standard/task.toml
@@ -31,6 +31,7 @@ simulator = "roboarc"
robot = "aloha-2"
family = "knot_untying"
mode = "standard" # RGB + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.4"
oracle = "full"
difficulty = "medium"
diff --git a/tasks/roboarc-knot-untying-05-i01-privileged/task.toml b/tasks/roboarc-knot-untying-05-i01-privileged/task.toml
index e31943f0c0..053c0a1fda 100644
--- a/tasks/roboarc-knot-untying-05-i01-privileged/task.toml
+++ b/tasks/roboarc-knot-untying-05-i01-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "roboarc" # RoboARC on MuJoCo 3.14.0 (CPU physics, Nyx R
robot = "aloha-2"
family = "knot_untying"
mode = "privileged" # privileged (state-based tooling, resets) | standard (robot as a service)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
difficulty = "hard" # the variation's tier (docs/TAXONOMY.md)
grasp = "physical (no assisted grasping)"
diff --git a/tasks/roboarc-knot-untying-05-i01-standard/task.toml b/tasks/roboarc-knot-untying-05-i01-standard/task.toml
index 8dea27d102..0dba0305bf 100644
--- a/tasks/roboarc-knot-untying-05-i01-standard/task.toml
+++ b/tasks/roboarc-knot-untying-05-i01-standard/task.toml
@@ -31,6 +31,7 @@ simulator = "roboarc"
robot = "aloha-2"
family = "knot_untying"
mode = "standard" # RGB + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.4"
oracle = "full"
difficulty = "hard"
diff --git a/tasks/roboarc-lock-picking-01-i01-privileged/task.toml b/tasks/roboarc-lock-picking-01-i01-privileged/task.toml
index 6267ba60df..5e9cfe7435 100644
--- a/tasks/roboarc-lock-picking-01-i01-privileged/task.toml
+++ b/tasks/roboarc-lock-picking-01-i01-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "roboarc" # RoboARC on MuJoCo 3.14.0 (CPU physics, Nyx R
robot = "franka-robotiq-2f85"
family = "lock_picking"
mode = "privileged" # privileged (state-based tooling, resets) | standard (robot as a service)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
difficulty = "very_easy" # the variation's tier (docs/rcb/TAXONOMY.md)
grasp = "assisted (the contact_weld_v1 grasp regions on lock_key)"
diff --git a/tasks/roboarc-lock-picking-01-i01-standard/task.toml b/tasks/roboarc-lock-picking-01-i01-standard/task.toml
index 2166abddb5..73fbb3fae0 100644
--- a/tasks/roboarc-lock-picking-01-i01-standard/task.toml
+++ b/tasks/roboarc-lock-picking-01-i01-standard/task.toml
@@ -31,6 +31,7 @@ simulator = "roboarc"
robot = "franka-robotiq-2f85"
family = "lock_picking"
mode = "standard" # RGB + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.4"
oracle = "full"
difficulty = "very_easy"
diff --git a/tasks/roboarc-lock-picking-02-i01-privileged/task.toml b/tasks/roboarc-lock-picking-02-i01-privileged/task.toml
index b7c8293390..68c7ccbc0a 100644
--- a/tasks/roboarc-lock-picking-02-i01-privileged/task.toml
+++ b/tasks/roboarc-lock-picking-02-i01-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "roboarc" # RoboARC on MuJoCo 3.14.0 (CPU physics, Nyx R
robot = "franka-robotiq-2f85"
family = "lock_picking"
mode = "privileged" # privileged (state-based tooling, resets) | standard (robot as a service)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
difficulty = "easier" # the variation's tier (docs/rcb/TAXONOMY.md)
grasp = "assisted (the contact_weld_v1 grasp regions on lock_key)"
diff --git a/tasks/roboarc-lock-picking-02-i01-standard/task.toml b/tasks/roboarc-lock-picking-02-i01-standard/task.toml
index 5a8130a6d0..e5e6db3261 100644
--- a/tasks/roboarc-lock-picking-02-i01-standard/task.toml
+++ b/tasks/roboarc-lock-picking-02-i01-standard/task.toml
@@ -31,6 +31,7 @@ simulator = "roboarc"
robot = "franka-robotiq-2f85"
family = "lock_picking"
mode = "standard" # RGB + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.4"
oracle = "full"
difficulty = "easier"
diff --git a/tasks/roboarc-lock-picking-03-i01-privileged/task.toml b/tasks/roboarc-lock-picking-03-i01-privileged/task.toml
index b6af7c4093..72c3236ef8 100644
--- a/tasks/roboarc-lock-picking-03-i01-privileged/task.toml
+++ b/tasks/roboarc-lock-picking-03-i01-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "roboarc" # RoboARC on MuJoCo 3.14.0 (CPU physics, Nyx R
robot = "franka-robotiq-2f85"
family = "lock_picking"
mode = "privileged" # privileged (state-based tooling, resets) | standard (robot as a service)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
difficulty = "easy" # the variation's tier (docs/rcb/TAXONOMY.md)
grasp = "assisted (the contact_weld_v1 grasp regions on lock_key)"
diff --git a/tasks/roboarc-lock-picking-03-i01-standard/task.toml b/tasks/roboarc-lock-picking-03-i01-standard/task.toml
index 61969788f4..8ddf8c481a 100644
--- a/tasks/roboarc-lock-picking-03-i01-standard/task.toml
+++ b/tasks/roboarc-lock-picking-03-i01-standard/task.toml
@@ -31,6 +31,7 @@ simulator = "roboarc"
robot = "franka-robotiq-2f85"
family = "lock_picking"
mode = "standard" # RGB + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.4"
oracle = "full"
difficulty = "easy"
diff --git a/tasks/roboarc-lock-picking-04-i01-privileged/task.toml b/tasks/roboarc-lock-picking-04-i01-privileged/task.toml
index 5e8b76557a..0677b7888f 100644
--- a/tasks/roboarc-lock-picking-04-i01-privileged/task.toml
+++ b/tasks/roboarc-lock-picking-04-i01-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "roboarc" # RoboARC on MuJoCo 3.14.0 (CPU physics, Nyx R
robot = "franka-robotiq-2f85"
family = "lock_picking"
mode = "privileged" # privileged (state-based tooling, resets) | standard (robot as a service)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
difficulty = "medium" # the variation's tier (docs/rcb/TAXONOMY.md)
grasp = "assisted (the contact_weld_v1 grasp regions on lock_key)"
diff --git a/tasks/roboarc-lock-picking-04-i01-standard/task.toml b/tasks/roboarc-lock-picking-04-i01-standard/task.toml
index 7730481943..88051431c4 100644
--- a/tasks/roboarc-lock-picking-04-i01-standard/task.toml
+++ b/tasks/roboarc-lock-picking-04-i01-standard/task.toml
@@ -31,6 +31,7 @@ simulator = "roboarc"
robot = "franka-robotiq-2f85"
family = "lock_picking"
mode = "standard" # RGB + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.4"
oracle = "full"
difficulty = "medium"
diff --git a/tasks/roboarc-lock-picking-05-i01-privileged/task.toml b/tasks/roboarc-lock-picking-05-i01-privileged/task.toml
index cb840ab278..04c819e8d5 100644
--- a/tasks/roboarc-lock-picking-05-i01-privileged/task.toml
+++ b/tasks/roboarc-lock-picking-05-i01-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "roboarc" # RoboARC on MuJoCo 3.14.0 (CPU physics, Nyx R
robot = "franka-robotiq-2f85"
family = "lock_picking"
mode = "privileged" # privileged (state-based tooling, resets) | standard (robot as a service)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
difficulty = "hard" # the variation's tier (docs/rcb/TAXONOMY.md)
grasp = "assisted (the contact_weld_v1 grasp regions on lock_key)"
diff --git a/tasks/roboarc-lock-picking-05-i01-standard/task.toml b/tasks/roboarc-lock-picking-05-i01-standard/task.toml
index 1d0d1a1ebd..e4bce6915c 100644
--- a/tasks/roboarc-lock-picking-05-i01-standard/task.toml
+++ b/tasks/roboarc-lock-picking-05-i01-standard/task.toml
@@ -31,6 +31,7 @@ simulator = "roboarc"
robot = "franka-robotiq-2f85"
family = "lock_picking"
mode = "standard" # RGB + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.4"
oracle = "full"
difficulty = "hard"
diff --git a/tasks/roboarc-orbits-01-i01-privileged/task.toml b/tasks/roboarc-orbits-01-i01-privileged/task.toml
index 2070e647be..7501533f70 100644
--- a/tasks/roboarc-orbits-01-i01-privileged/task.toml
+++ b/tasks/roboarc-orbits-01-i01-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "roboarc" # RoboARC on MuJoCo 3.14.0 (CPU float64 physic
robot = "aloha-2"
family = "orbits"
mode = "privileged" # privileged (state-based tooling, resets) | standard (robot as a service)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
difficulty = "very_easy" # the variation's tier (docs/rcb/TAXONOMY.md)
grasp = "assisted (the contact_weld_v1 grasp regions on cord)"
diff --git a/tasks/roboarc-orbits-01-i01-standard/task.toml b/tasks/roboarc-orbits-01-i01-standard/task.toml
index 9df1969a72..311d771b07 100644
--- a/tasks/roboarc-orbits-01-i01-standard/task.toml
+++ b/tasks/roboarc-orbits-01-i01-standard/task.toml
@@ -31,6 +31,7 @@ simulator = "roboarc"
robot = "aloha-2"
family = "orbits"
mode = "standard" # RGB + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.4"
oracle = "full"
difficulty = "very_easy"
diff --git a/tasks/roboarc-rope-threading-01-i01-privileged/task.toml b/tasks/roboarc-rope-threading-01-i01-privileged/task.toml
index d258a25458..274c50416c 100644
--- a/tasks/roboarc-rope-threading-01-i01-privileged/task.toml
+++ b/tasks/roboarc-rope-threading-01-i01-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "roboarc" # RoboARC on MuJoCo 3.14.0 (CPU physics, Nyx R
robot = "aloha-2"
family = "rope_threading"
mode = "privileged" # privileged (state-based tooling, resets) | standard (robot as a service)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
difficulty = "easy" # the variation's tier (docs/TAXONOMY.md)
grasp = "assisted (the contact_weld_v1 grasp regions on rope)"
diff --git a/tasks/roboarc-rope-threading-01-i01-standard/task.toml b/tasks/roboarc-rope-threading-01-i01-standard/task.toml
index 1b430fac63..c53b912b5d 100644
--- a/tasks/roboarc-rope-threading-01-i01-standard/task.toml
+++ b/tasks/roboarc-rope-threading-01-i01-standard/task.toml
@@ -31,6 +31,7 @@ simulator = "roboarc"
robot = "aloha-2"
family = "rope_threading"
mode = "standard" # RGB + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.4"
oracle = "full"
difficulty = "easy"
diff --git a/tasks/roboarc-rope-threading-02-i01-privileged/task.toml b/tasks/roboarc-rope-threading-02-i01-privileged/task.toml
index e11deeaa2b..f8595e0134 100644
--- a/tasks/roboarc-rope-threading-02-i01-privileged/task.toml
+++ b/tasks/roboarc-rope-threading-02-i01-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "roboarc" # RoboARC on MuJoCo 3.14.0 (CPU physics, Nyx R
robot = "aloha-2"
family = "rope_threading"
mode = "privileged" # privileged (state-based tooling, resets) | standard (robot as a service)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
difficulty = "medium" # the variation's tier (docs/TAXONOMY.md)
grasp = "assisted (the contact_weld_v1 grasp regions on rope)"
diff --git a/tasks/roboarc-rope-threading-02-i01-standard/task.toml b/tasks/roboarc-rope-threading-02-i01-standard/task.toml
index 39124e384a..434d51895c 100644
--- a/tasks/roboarc-rope-threading-02-i01-standard/task.toml
+++ b/tasks/roboarc-rope-threading-02-i01-standard/task.toml
@@ -31,6 +31,7 @@ simulator = "roboarc"
robot = "aloha-2"
family = "rope_threading"
mode = "standard" # RGB + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.4"
oracle = "full"
difficulty = "medium"
diff --git a/tasks/roboarc-rope-threading-03-i01-privileged/task.toml b/tasks/roboarc-rope-threading-03-i01-privileged/task.toml
index 9bd2840c7b..30b521fab4 100644
--- a/tasks/roboarc-rope-threading-03-i01-privileged/task.toml
+++ b/tasks/roboarc-rope-threading-03-i01-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "roboarc" # RoboARC on MuJoCo 3.14.0 (CPU physics, Nyx R
robot = "aloha-2"
family = "rope_threading"
mode = "privileged" # privileged (state-based tooling, resets) | standard (robot as a service)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
difficulty = "hard" # the variation's tier (docs/TAXONOMY.md)
grasp = "assisted (the contact_weld_v1 grasp regions on rope)"
diff --git a/tasks/roboarc-rope-threading-03-i01-standard/task.toml b/tasks/roboarc-rope-threading-03-i01-standard/task.toml
index 1a6a3af65b..04208a2c81 100644
--- a/tasks/roboarc-rope-threading-03-i01-standard/task.toml
+++ b/tasks/roboarc-rope-threading-03-i01-standard/task.toml
@@ -31,6 +31,7 @@ simulator = "roboarc"
robot = "aloha-2"
family = "rope_threading"
mode = "standard" # RGB + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.4"
oracle = "full"
difficulty = "hard"
diff --git a/tasks/roboarc-rope-threading-04-i01-privileged/task.toml b/tasks/roboarc-rope-threading-04-i01-privileged/task.toml
index 8d233507df..c0801f6d07 100644
--- a/tasks/roboarc-rope-threading-04-i01-privileged/task.toml
+++ b/tasks/roboarc-rope-threading-04-i01-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "roboarc" # RoboARC on MuJoCo 3.14.0 (CPU physics, Nyx R
robot = "aloha-2"
family = "rope_threading"
mode = "privileged" # privileged (state-based tooling, resets) | standard (robot as a service)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
difficulty = "very_easy" # the variation's tier (docs/rcb/TAXONOMY.md)
grasp = "assisted (the contact_weld_v1 grasp regions on rope)"
diff --git a/tasks/roboarc-rope-threading-04-i01-standard/task.toml b/tasks/roboarc-rope-threading-04-i01-standard/task.toml
index e660fb7b36..b347db6cc3 100644
--- a/tasks/roboarc-rope-threading-04-i01-standard/task.toml
+++ b/tasks/roboarc-rope-threading-04-i01-standard/task.toml
@@ -31,6 +31,7 @@ simulator = "roboarc"
robot = "aloha-2"
family = "rope_threading"
mode = "standard" # RGB + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.4"
oracle = "full"
difficulty = "very_easy"
diff --git a/tasks/roboarc-rope-threading-05-i01-privileged/task.toml b/tasks/roboarc-rope-threading-05-i01-privileged/task.toml
index 60cac314a5..90eb0ea7d8 100644
--- a/tasks/roboarc-rope-threading-05-i01-privileged/task.toml
+++ b/tasks/roboarc-rope-threading-05-i01-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "roboarc" # RoboARC on MuJoCo 3.14.0 (CPU physics, Nyx R
robot = "aloha-2"
family = "rope_threading"
mode = "privileged" # privileged (state-based tooling, resets) | standard (robot as a service)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
difficulty = "easier" # the variation's tier (docs/rcb/TAXONOMY.md)
grasp = "assisted (the contact_weld_v1 grasp regions on rope)"
diff --git a/tasks/roboarc-rope-threading-05-i01-standard/task.toml b/tasks/roboarc-rope-threading-05-i01-standard/task.toml
index 295f0e7e38..191ea3e0c1 100644
--- a/tasks/roboarc-rope-threading-05-i01-standard/task.toml
+++ b/tasks/roboarc-rope-threading-05-i01-standard/task.toml
@@ -31,6 +31,7 @@ simulator = "roboarc"
robot = "aloha-2"
family = "rope_threading"
mode = "standard" # RGB + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.4"
oracle = "full"
difficulty = "easier"
diff --git a/tasks/roboarc-rope-threading-06-i01-privileged/task.toml b/tasks/roboarc-rope-threading-06-i01-privileged/task.toml
index 571cada5c5..34f620fefd 100644
--- a/tasks/roboarc-rope-threading-06-i01-privileged/task.toml
+++ b/tasks/roboarc-rope-threading-06-i01-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "roboarc" # RoboARC on MuJoCo 3.14.0 (CPU physics, Nyx R
robot = "aloha-2"
family = "rope_threading"
mode = "privileged" # privileged (state-based tooling, resets) | standard (robot as a service)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
difficulty = "easy" # the variation's tier (docs/rcb/TAXONOMY.md)
grasp = "assisted (the contact_weld_v1 grasp regions on rope)"
diff --git a/tasks/roboarc-rope-threading-06-i01-standard/task.toml b/tasks/roboarc-rope-threading-06-i01-standard/task.toml
index 13f8318d7d..050c1f0fce 100644
--- a/tasks/roboarc-rope-threading-06-i01-standard/task.toml
+++ b/tasks/roboarc-rope-threading-06-i01-standard/task.toml
@@ -31,6 +31,7 @@ simulator = "roboarc"
robot = "aloha-2"
family = "rope_threading"
mode = "standard" # RGB + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.4"
oracle = "full"
difficulty = "easy"
diff --git a/tasks/roboarc-rope-threading-07-i01-privileged/task.toml b/tasks/roboarc-rope-threading-07-i01-privileged/task.toml
index 6f55ceb79c..3ebc5686de 100644
--- a/tasks/roboarc-rope-threading-07-i01-privileged/task.toml
+++ b/tasks/roboarc-rope-threading-07-i01-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "roboarc" # RoboARC on MuJoCo 3.14.0 (CPU physics, Nyx R
robot = "aloha-2"
family = "rope_threading"
mode = "privileged" # privileged (state-based tooling, resets) | standard (robot as a service)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
difficulty = "medium" # the variation's tier (docs/rcb/TAXONOMY.md)
grasp = "assisted (the contact_weld_v1 grasp regions on rope)"
diff --git a/tasks/roboarc-rope-threading-07-i01-standard/task.toml b/tasks/roboarc-rope-threading-07-i01-standard/task.toml
index a7372fcaee..0db9ca5fb8 100644
--- a/tasks/roboarc-rope-threading-07-i01-standard/task.toml
+++ b/tasks/roboarc-rope-threading-07-i01-standard/task.toml
@@ -31,6 +31,7 @@ simulator = "roboarc"
robot = "aloha-2"
family = "rope_threading"
mode = "standard" # RGB + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.4"
oracle = "full"
difficulty = "medium"
diff --git a/tasks/roboarc-rope-threading-08-i01-privileged/task.toml b/tasks/roboarc-rope-threading-08-i01-privileged/task.toml
index ecf1476da5..cf0a252621 100644
--- a/tasks/roboarc-rope-threading-08-i01-privileged/task.toml
+++ b/tasks/roboarc-rope-threading-08-i01-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "roboarc" # RoboARC on MuJoCo 3.14.0 (CPU physics, Nyx R
robot = "aloha-2"
family = "rope_threading"
mode = "privileged" # privileged (state-based tooling, resets) | standard (robot as a service)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
difficulty = "hard" # the variation's tier (docs/rcb/TAXONOMY.md)
grasp = "assisted (the contact_weld_v1 grasp regions on rope)"
diff --git a/tasks/roboarc-rope-threading-08-i01-standard/task.toml b/tasks/roboarc-rope-threading-08-i01-standard/task.toml
index a8438825a9..28173024b8 100644
--- a/tasks/roboarc-rope-threading-08-i01-standard/task.toml
+++ b/tasks/roboarc-rope-threading-08-i01-standard/task.toml
@@ -31,6 +31,7 @@ simulator = "roboarc"
robot = "aloha-2"
family = "rope_threading"
mode = "standard" # RGB + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.4"
oracle = "full"
difficulty = "hard"
diff --git a/tasks/roboarc-slotted-post-ring-01-i01-privileged/task.toml b/tasks/roboarc-slotted-post-ring-01-i01-privileged/task.toml
index 9ff0c4a574..3ff7b3b78d 100644
--- a/tasks/roboarc-slotted-post-ring-01-i01-privileged/task.toml
+++ b/tasks/roboarc-slotted-post-ring-01-i01-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "roboarc" # RoboARC on MuJoCo 3.14.0 (CPU float64 physic
robot = "aloha-2"
family = "slotted_post_ring"
mode = "privileged" # privileged (state-based tooling, resets) | standard (robot as a service)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
difficulty = "very_easy" # the variation's tier (docs/rcb/TAXONOMY.md)
grasp = "assisted (the contact_weld_v1 grasp regions on slotted_chain)"
diff --git a/tasks/roboarc-slotted-post-ring-01-i01-standard/task.toml b/tasks/roboarc-slotted-post-ring-01-i01-standard/task.toml
index 2245e663b5..134621e385 100644
--- a/tasks/roboarc-slotted-post-ring-01-i01-standard/task.toml
+++ b/tasks/roboarc-slotted-post-ring-01-i01-standard/task.toml
@@ -31,6 +31,7 @@ simulator = "roboarc"
robot = "aloha-2"
family = "slotted_post_ring"
mode = "standard" # RGB + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.4"
oracle = "full"
difficulty = "very_easy"
diff --git a/tasks/roboarc-slotted-post-ring-02-i01-privileged/task.toml b/tasks/roboarc-slotted-post-ring-02-i01-privileged/task.toml
index 8525502364..18a5e8d30c 100644
--- a/tasks/roboarc-slotted-post-ring-02-i01-privileged/task.toml
+++ b/tasks/roboarc-slotted-post-ring-02-i01-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "roboarc" # RoboARC on MuJoCo 3.14.0 (CPU float64 physic
robot = "aloha-2"
family = "slotted_post_ring"
mode = "privileged" # privileged (state-based tooling, resets) | standard (robot as a service)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
difficulty = "easier" # the variation's tier (docs/rcb/TAXONOMY.md)
grasp = "assisted (the contact_weld_v1 grasp regions on slotted_chain)"
diff --git a/tasks/roboarc-slotted-post-ring-02-i01-standard/task.toml b/tasks/roboarc-slotted-post-ring-02-i01-standard/task.toml
index 36e79eaf7c..d1ce126083 100644
--- a/tasks/roboarc-slotted-post-ring-02-i01-standard/task.toml
+++ b/tasks/roboarc-slotted-post-ring-02-i01-standard/task.toml
@@ -31,6 +31,7 @@ simulator = "roboarc"
robot = "aloha-2"
family = "slotted_post_ring"
mode = "standard" # RGB + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.4"
oracle = "full"
difficulty = "easier"
diff --git a/tasks/roboarc-slotted-post-ring-03-i01-privileged/task.toml b/tasks/roboarc-slotted-post-ring-03-i01-privileged/task.toml
index 45189bdfb6..e9bc7366f9 100644
--- a/tasks/roboarc-slotted-post-ring-03-i01-privileged/task.toml
+++ b/tasks/roboarc-slotted-post-ring-03-i01-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "roboarc" # RoboARC on MuJoCo 3.14.0 (CPU float64 physic
robot = "aloha-2"
family = "slotted_post_ring"
mode = "privileged" # privileged (state-based tooling, resets) | standard (robot as a service)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
difficulty = "easy" # the variation's tier (docs/rcb/TAXONOMY.md)
grasp = "assisted (the contact_weld_v1 grasp regions on slotted_chain)"
diff --git a/tasks/roboarc-slotted-post-ring-03-i01-standard/task.toml b/tasks/roboarc-slotted-post-ring-03-i01-standard/task.toml
index 77b2953b5d..4b8af617cf 100644
--- a/tasks/roboarc-slotted-post-ring-03-i01-standard/task.toml
+++ b/tasks/roboarc-slotted-post-ring-03-i01-standard/task.toml
@@ -31,6 +31,7 @@ simulator = "roboarc"
robot = "aloha-2"
family = "slotted_post_ring"
mode = "standard" # RGB + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.4"
oracle = "full"
difficulty = "easy"
diff --git a/tasks/roboarc-slotted-post-ring-04-i01-privileged/task.toml b/tasks/roboarc-slotted-post-ring-04-i01-privileged/task.toml
index c36e5f9d31..c4646e8a9f 100644
--- a/tasks/roboarc-slotted-post-ring-04-i01-privileged/task.toml
+++ b/tasks/roboarc-slotted-post-ring-04-i01-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "roboarc" # RoboARC on MuJoCo 3.14.0 (CPU float64 physic
robot = "aloha-2"
family = "slotted_post_ring"
mode = "privileged" # privileged (state-based tooling, resets) | standard (robot as a service)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
difficulty = "medium" # the variation's tier (docs/rcb/TAXONOMY.md)
grasp = "assisted (the contact_weld_v1 grasp regions on slotted_chain)"
diff --git a/tasks/roboarc-slotted-post-ring-04-i01-standard/task.toml b/tasks/roboarc-slotted-post-ring-04-i01-standard/task.toml
index b49c5d41e4..9a07e80d8a 100644
--- a/tasks/roboarc-slotted-post-ring-04-i01-standard/task.toml
+++ b/tasks/roboarc-slotted-post-ring-04-i01-standard/task.toml
@@ -31,6 +31,7 @@ simulator = "roboarc"
robot = "aloha-2"
family = "slotted_post_ring"
mode = "standard" # RGB + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.4"
oracle = "full"
difficulty = "medium"
diff --git a/tasks/roboarc-slotted-post-ring-05-i01-privileged/task.toml b/tasks/roboarc-slotted-post-ring-05-i01-privileged/task.toml
index 86a4a59312..2e05691cc1 100644
--- a/tasks/roboarc-slotted-post-ring-05-i01-privileged/task.toml
+++ b/tasks/roboarc-slotted-post-ring-05-i01-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "roboarc" # RoboARC on MuJoCo 3.14.0 (CPU float64 physic
robot = "aloha-2"
family = "slotted_post_ring"
mode = "privileged" # privileged (state-based tooling, resets) | standard (robot as a service)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
difficulty = "hard" # the variation's tier (docs/rcb/TAXONOMY.md)
grasp = "assisted (the contact_weld_v1 grasp regions on slotted_chain)"
diff --git a/tasks/roboarc-slotted-post-ring-05-i01-standard/task.toml b/tasks/roboarc-slotted-post-ring-05-i01-standard/task.toml
index 25bd3c0882..2eadfe14de 100644
--- a/tasks/roboarc-slotted-post-ring-05-i01-standard/task.toml
+++ b/tasks/roboarc-slotted-post-ring-05-i01-standard/task.toml
@@ -31,6 +31,7 @@ simulator = "roboarc"
robot = "aloha-2"
family = "slotted_post_ring"
mode = "standard" # RGB + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.4"
oracle = "full"
difficulty = "hard"
diff --git a/tasks/robocasa-arrange-vegetables-i00-privileged/task.toml b/tasks/robocasa-arrange-vegetables-i00-privileged/task.toml
index 6c53778489..6bb260ebda 100644
--- a/tasks/robocasa-arrange-vegetables-i00-privileged/task.toml
+++ b/tasks/robocasa-arrange-vegetables-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa/ArrangeVegetables"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robocasa-arrange-vegetables-i00-standard/task.toml b/tasks/robocasa-arrange-vegetables-i00-standard/task.toml
index f33ec02208..83e5485329 100644
--- a/tasks/robocasa-arrange-vegetables-i00-standard/task.toml
+++ b/tasks/robocasa-arrange-vegetables-i00-standard/task.toml
@@ -24,6 +24,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa/ArrangeVegetables"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
expert_time_estimate_hours = 4.0
diff --git a/tasks/robocasa-close-double-door-i00-privileged/task.toml b/tasks/robocasa-close-double-door-i00-privileged/task.toml
index caf12b61c3..0c48705af2 100644
--- a/tasks/robocasa-close-double-door-i00-privileged/task.toml
+++ b/tasks/robocasa-close-double-door-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa/CloseDoubleDoor"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robocasa-close-double-door-i00-standard/task.toml b/tasks/robocasa-close-double-door-i00-standard/task.toml
index dc6e28dc7b..3c3a1be657 100644
--- a/tasks/robocasa-close-double-door-i00-standard/task.toml
+++ b/tasks/robocasa-close-double-door-i00-standard/task.toml
@@ -24,6 +24,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa/CloseDoubleDoor"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
expert_time_estimate_hours = 4.0
diff --git a/tasks/robocasa-close-drawer-i00-privileged/task.toml b/tasks/robocasa-close-drawer-i00-privileged/task.toml
index d36e3be291..1b30bf82ee 100644
--- a/tasks/robocasa-close-drawer-i00-privileged/task.toml
+++ b/tasks/robocasa-close-drawer-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa/CloseDrawer"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robocasa-close-drawer-i00-standard/task.toml b/tasks/robocasa-close-drawer-i00-standard/task.toml
index 25a75807af..6e95eb4290 100644
--- a/tasks/robocasa-close-drawer-i00-standard/task.toml
+++ b/tasks/robocasa-close-drawer-i00-standard/task.toml
@@ -24,6 +24,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa/CloseDrawer"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
expert_time_estimate_hours = 4.0
diff --git a/tasks/robocasa-close-single-door-i00-privileged/task.toml b/tasks/robocasa-close-single-door-i00-privileged/task.toml
index 15d254d565..81366c5b8e 100644
--- a/tasks/robocasa-close-single-door-i00-privileged/task.toml
+++ b/tasks/robocasa-close-single-door-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa/CloseSingleDoor"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robocasa-close-single-door-i00-standard/task.toml b/tasks/robocasa-close-single-door-i00-standard/task.toml
index 2ff7f8f9a7..1267f4dcd0 100644
--- a/tasks/robocasa-close-single-door-i00-standard/task.toml
+++ b/tasks/robocasa-close-single-door-i00-standard/task.toml
@@ -24,6 +24,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa/CloseSingleDoor"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
expert_time_estimate_hours = 4.0
diff --git a/tasks/robocasa-coffee-press-button-i00-privileged/task.toml b/tasks/robocasa-coffee-press-button-i00-privileged/task.toml
index cdc4a8653e..6080ef5eba 100644
--- a/tasks/robocasa-coffee-press-button-i00-privileged/task.toml
+++ b/tasks/robocasa-coffee-press-button-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa/CoffeePressButton"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robocasa-coffee-press-button-i00-standard/task.toml b/tasks/robocasa-coffee-press-button-i00-standard/task.toml
index 48d551040b..555d1be2dd 100644
--- a/tasks/robocasa-coffee-press-button-i00-standard/task.toml
+++ b/tasks/robocasa-coffee-press-button-i00-standard/task.toml
@@ -24,6 +24,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa/CoffeePressButton"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
expert_time_estimate_hours = 4.0
diff --git a/tasks/robocasa-coffee-serve-mug-i00-privileged/task.toml b/tasks/robocasa-coffee-serve-mug-i00-privileged/task.toml
index 5933fd1fe2..ac5b6a8fd7 100644
--- a/tasks/robocasa-coffee-serve-mug-i00-privileged/task.toml
+++ b/tasks/robocasa-coffee-serve-mug-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa/CoffeeServeMug"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robocasa-coffee-serve-mug-i00-standard/task.toml b/tasks/robocasa-coffee-serve-mug-i00-standard/task.toml
index b3c60b0f30..b856eced81 100644
--- a/tasks/robocasa-coffee-serve-mug-i00-standard/task.toml
+++ b/tasks/robocasa-coffee-serve-mug-i00-standard/task.toml
@@ -24,6 +24,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa/CoffeeServeMug"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
expert_time_estimate_hours = 4.0
diff --git a/tasks/robocasa-coffee-setup-mug-i00-privileged/task.toml b/tasks/robocasa-coffee-setup-mug-i00-privileged/task.toml
index 79ee2cdc64..600f4c041a 100644
--- a/tasks/robocasa-coffee-setup-mug-i00-privileged/task.toml
+++ b/tasks/robocasa-coffee-setup-mug-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa/CoffeeSetupMug"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robocasa-coffee-setup-mug-i00-standard/task.toml b/tasks/robocasa-coffee-setup-mug-i00-standard/task.toml
index 49d3077a06..7c3b26157d 100644
--- a/tasks/robocasa-coffee-setup-mug-i00-standard/task.toml
+++ b/tasks/robocasa-coffee-setup-mug-i00-standard/task.toml
@@ -24,6 +24,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa/CoffeeSetupMug"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
expert_time_estimate_hours = 4.0
diff --git a/tasks/robocasa-gr1-bottle-to-cabinet-close-i00-privileged/task.toml b/tasks/robocasa-gr1-bottle-to-cabinet-close-i00-privileged/task.toml
index d6e21012ba..16d627d7b3 100644
--- a/tasks/robocasa-gr1-bottle-to-cabinet-close-i00-privileged/task.toml
+++ b/tasks/robocasa-gr1-bottle-to-cabinet-close-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "robosuite"
robot = "fourier-gr1"
family = "robocasa-gr1/PnPBottleToCabinetClose"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robocasa-gr1-bottle-to-cabinet-close-i00-standard/README.md b/tasks/robocasa-gr1-bottle-to-cabinet-close-i00-standard/README.md
index 7fa255038d..e6a185f311 100644
--- a/tasks/robocasa-gr1-bottle-to-cabinet-close-i00-standard/README.md
+++ b/tasks/robocasa-gr1-bottle-to-cabinet-close-i00-standard/README.md
@@ -48,3 +48,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
- 2026-10-10 (review round 3, board 0067): the sim service's request handling (`environment/sim/server_casa.py`) and `tests/grade.py`. A negative or non-numeric `Content-Length` is answered 400 without reading the body (the service read until the client hung up, without a bound, which could exhaust its memory); a camera named more than once in `cameras` is rendered and returned once (N copies of one name made it build N x N images within the frame budget); a port that is still held when the service restarts is waited for, not counted as a crash; a JSON `true` or `false` where `dt` or `seconds` takes a number is a 400 (`true` ran as 1 second). The grader: a `final.json` whose evidence fields (`control_steps`, `calls`, `elapsed_s`, `final_reward`) are not numbers no longer stops it before it writes a reward (they are reported as 0 and decide nothing). An episode's physics and the pass rule are unchanged.
- 2026-10-13 (owner, boards 0085 and 0087): the sim service's clock (`environment/sim/server_casa.py`, the docstring of `environment/rcb_client.py`). `meta["time_s"]` and `meta["time_left_s"]` count from the agent's first request to the service (any endpoint), and a restarted service keeps that moment; before, they counted from the first start of the service's container (`time_left_s`) and from when its scene was ready (`time_s`). Where the service is ready before the agent's phase begins (the harness is installed in between: a median of 79 s over the recorded standard trials), that clock ran ahead of the agent's and `time_left_s` reached 0 while the agent still had that time. `/spec` says what the clock counts from (`budget.clock`), and the instruction's `## Time limit` tells the agent to plan by `time_left_s`. The service never ended an episode on this clock: the scene, the frozen instance, the pass rule and the grader are unchanged, so the task keeps its name and earlier results stand.
- 2026-10-13 (board 0093; found in board 0086): the camera calibration (`environment/sim/server_casa.py`). `cam2world_gl` held the pose of MuJoCo's left-eye camera (`scene.camera[0]`), 3.4 cm (half the model's 0.068 m interpupillary distance) to the camera's left of the viewpoint the mono RGB and depth images are rendered from; it now is that viewpoint (`data.cam_xpos` / `data.cam_xmat`). Checked against MuJoCo's own renders of marker spheres at known world points, every served camera, before and after the robot moved: the reprojection error was 10.5-10.7 px median (up to 28 px for near points, always to the same side), now 0.15 px median and at most 0.4 px. The intrinsics were right (every camera here is given by its `fovy`, none by a sensor size) and are unchanged, as are the images, the scene, the frozen instance, the pass rule and the grader, so the task keeps its name. **Standard-mode results from before this date used the offset calibration**: a point located from the images and the reported calibration came out about 3.4 cm to the side (agents missed a microwave button by 3.5 cm), so they are not strictly comparable with later ones.
+- 2026-10-14 (board 0113, from board 0086): the instruction describes the hands: each hand command is the position target (rad) of its finger's joints, 0 open and larger closing, within each joint's range; the grip site is the origin of the hand's mounting frame at the wrist end of the hand; the pink spheres beside each hand are visual markers of the robot model, not objects (board 0086: most standard runs took them for objects). Text only: the scene, the frozen instance, the reference solution, the success check and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/robocasa-gr1-bottle-to-cabinet-close-i00-standard/instruction.md b/tasks/robocasa-gr1-bottle-to-cabinet-close-i00-standard/instruction.md
index 36b42e7106..9617499147 100644
--- a/tasks/robocasa-gr1-bottle-to-cabinet-close-i00-standard/instruction.md
+++ b/tasks/robocasa-gr1-bottle-to-cabinet-close-i00-standard/instruction.md
@@ -51,6 +51,13 @@ res.video[cam]; res.frames[cam]; res.last[cam]; res.proprio; res.save("/app/runs
channel is an absolute target, so the grip sites stay within about a millimetre, but the IK's posture term can still
move the waist and elbows by a few hundredths of a radian while you watch.
- The cameras are mounted on the robot and move with it: use the calibration that comes with each response.
+- The hands: each hand command is the position target, in radians, of that finger's joints (`thumb_flexion` drives
+ two thumb joints, `thumb_rotation` one): 0 is the straight, open finger, larger values bend it toward the palm
+ (the hand closes), and every finger joint stops at the end of its range (1.22 to 1.74 rad), so values below 0 or
+ above about 1.7 change nothing more. The grip site that the pose targets place is the origin of the hand's mounting
+ frame (`gripper0_right_grip_site`, `gripper0_left_grip_site`), at the wrist end of the hand, not a point between
+ the fingertips. The four small pink spheres beside each hand are part of the robot model (visual markers
+ `robot0_<side>_pinch_spheres_0..3`, about 13.5 cm from its grip site, with no collision): they are not objects.
- World frame: metres, z up. In the proprio, `eef_left_*` / `eef_right_*` are the grip sites' positions in the world
frame; the action's pose targets and the `cmd_*` entries (the last command) are in the robot's base frame.
The robot's base frame (body `robot0_base`, fixed in the world) has its origin at world (2.65000, -2.17500, 0.97000) m and is turned +90.0 deg about z (quaternion w x y z (0.70711, 0, 0, 0.70711)): `p_world = R @ p_base + t` with R = [[0, -1, 0], [1, 0, 0], [0, 0, 1]].
diff --git a/tasks/robocasa-gr1-bottle-to-cabinet-close-i00-standard/task.toml b/tasks/robocasa-gr1-bottle-to-cabinet-close-i00-standard/task.toml
index 97edaf1765..403f7d1b8b 100644
--- a/tasks/robocasa-gr1-bottle-to-cabinet-close-i00-standard/task.toml
+++ b/tasks/robocasa-gr1-bottle-to-cabinet-close-i00-standard/task.toml
@@ -24,6 +24,7 @@ simulator = "robosuite"
robot = "fourier-gr1"
family = "robocasa-gr1/PnPBottleToCabinetClose"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
expert_time_estimate_hours = 4.0
diff --git a/tasks/robocasa-gr1-can-to-drawer-close-i00-privileged/task.toml b/tasks/robocasa-gr1-can-to-drawer-close-i00-privileged/task.toml
index c5f9a39a2d..9860094905 100644
--- a/tasks/robocasa-gr1-can-to-drawer-close-i00-privileged/task.toml
+++ b/tasks/robocasa-gr1-can-to-drawer-close-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "robosuite"
robot = "fourier-gr1"
family = "robocasa-gr1/PnPCanToDrawerClose"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robocasa-gr1-can-to-drawer-close-i00-standard/README.md b/tasks/robocasa-gr1-can-to-drawer-close-i00-standard/README.md
index 9875f62860..fe40f220f3 100644
--- a/tasks/robocasa-gr1-can-to-drawer-close-i00-standard/README.md
+++ b/tasks/robocasa-gr1-can-to-drawer-close-i00-standard/README.md
@@ -48,3 +48,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
- 2026-10-10 (review round 3, board 0067): the sim service's request handling (`environment/sim/server_casa.py`) and `tests/grade.py`. A negative or non-numeric `Content-Length` is answered 400 without reading the body (the service read until the client hung up, without a bound, which could exhaust its memory); a camera named more than once in `cameras` is rendered and returned once (N copies of one name made it build N x N images within the frame budget); a port that is still held when the service restarts is waited for, not counted as a crash; a JSON `true` or `false` where `dt` or `seconds` takes a number is a 400 (`true` ran as 1 second). The grader: a `final.json` whose evidence fields (`control_steps`, `calls`, `elapsed_s`, `final_reward`) are not numbers no longer stops it before it writes a reward (they are reported as 0 and decide nothing). An episode's physics and the pass rule are unchanged.
- 2026-10-13 (owner, boards 0085 and 0087): the sim service's clock (`environment/sim/server_casa.py`, the docstring of `environment/rcb_client.py`). `meta["time_s"]` and `meta["time_left_s"]` count from the agent's first request to the service (any endpoint), and a restarted service keeps that moment; before, they counted from the first start of the service's container (`time_left_s`) and from when its scene was ready (`time_s`). Where the service is ready before the agent's phase begins (the harness is installed in between: a median of 79 s over the recorded standard trials), that clock ran ahead of the agent's and `time_left_s` reached 0 while the agent still had that time. `/spec` says what the clock counts from (`budget.clock`), and the instruction's `## Time limit` tells the agent to plan by `time_left_s`. The service never ended an episode on this clock: the scene, the frozen instance, the pass rule and the grader are unchanged, so the task keeps its name and earlier results stand.
- 2026-10-13 (board 0093; found in board 0086): the camera calibration (`environment/sim/server_casa.py`). `cam2world_gl` held the pose of MuJoCo's left-eye camera (`scene.camera[0]`), 3.4 cm (half the model's 0.068 m interpupillary distance) to the camera's left of the viewpoint the mono RGB and depth images are rendered from; it now is that viewpoint (`data.cam_xpos` / `data.cam_xmat`). Checked against MuJoCo's own renders of marker spheres at known world points, every served camera, before and after the robot moved: the reprojection error was 10.5-10.7 px median (up to 28 px for near points, always to the same side), now 0.15 px median and at most 0.4 px. The intrinsics were right (every camera here is given by its `fovy`, none by a sensor size) and are unchanged, as are the images, the scene, the frozen instance, the pass rule and the grader, so the task keeps its name. **Standard-mode results from before this date used the offset calibration**: a point located from the images and the reported calibration came out about 3.4 cm to the side (agents missed a microwave button by 3.5 cm), so they are not strictly comparable with later ones.
+- 2026-10-14 (board 0113, from board 0086): the instruction describes the hands: each hand command is the position target (rad) of its finger's joints, 0 open and larger closing, within each joint's range; the grip site is the origin of the hand's mounting frame at the wrist end of the hand; the pink spheres beside each hand are visual markers of the robot model, not objects (board 0086: most standard runs took them for objects). Text only: the scene, the frozen instance, the reference solution, the success check and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/robocasa-gr1-can-to-drawer-close-i00-standard/instruction.md b/tasks/robocasa-gr1-can-to-drawer-close-i00-standard/instruction.md
index 2cc168430d..65795e662a 100644
--- a/tasks/robocasa-gr1-can-to-drawer-close-i00-standard/instruction.md
+++ b/tasks/robocasa-gr1-can-to-drawer-close-i00-standard/instruction.md
@@ -51,6 +51,13 @@ res.video[cam]; res.frames[cam]; res.last[cam]; res.proprio; res.save("/app/runs
channel is an absolute target, so the grip sites stay within about a millimetre, but the IK's posture term can still
move the waist and elbows by a few hundredths of a radian while you watch.
- The cameras are mounted on the robot and move with it: use the calibration that comes with each response.
+- The hands: each hand command is the position target, in radians, of that finger's joints (`thumb_flexion` drives
+ two thumb joints, `thumb_rotation` one): 0 is the straight, open finger, larger values bend it toward the palm
+ (the hand closes), and every finger joint stops at the end of its range (1.22 to 1.74 rad), so values below 0 or
+ above about 1.7 change nothing more. The grip site that the pose targets place is the origin of the hand's mounting
+ frame (`gripper0_right_grip_site`, `gripper0_left_grip_site`), at the wrist end of the hand, not a point between
+ the fingertips. The four small pink spheres beside each hand are part of the robot model (visual markers
+ `robot0_<side>_pinch_spheres_0..3`, about 13.5 cm from its grip site, with no collision): they are not objects.
- World frame: metres, z up. In the proprio, `eef_left_*` / `eef_right_*` are the grip sites' positions in the world
frame; the action's pose targets and the `cmd_*` entries (the last command) are in the robot's base frame.
The robot's base frame (body `robot0_base`, fixed in the world) has its origin at world (2.65000, -2.17500, 0.97000) m and is turned +90.0 deg about z (quaternion w x y z (0.70711, 0, 0, 0.70711)): `p_world = R @ p_base + t` with R = [[0, -1, 0], [1, 0, 0], [0, 0, 1]].
diff --git a/tasks/robocasa-gr1-can-to-drawer-close-i00-standard/task.toml b/tasks/robocasa-gr1-can-to-drawer-close-i00-standard/task.toml
index fa7bc3f262..1839ac829e 100644
--- a/tasks/robocasa-gr1-can-to-drawer-close-i00-standard/task.toml
+++ b/tasks/robocasa-gr1-can-to-drawer-close-i00-standard/task.toml
@@ -24,6 +24,7 @@ simulator = "robosuite"
robot = "fourier-gr1"
family = "robocasa-gr1/PnPCanToDrawerClose"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
expert_time_estimate_hours = 4.0
diff --git a/tasks/robocasa-gr1-cup-to-drawer-close-i00-privileged/task.toml b/tasks/robocasa-gr1-cup-to-drawer-close-i00-privileged/task.toml
index c5c74dbe51..e39bfd81d4 100644
--- a/tasks/robocasa-gr1-cup-to-drawer-close-i00-privileged/task.toml
+++ b/tasks/robocasa-gr1-cup-to-drawer-close-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "robosuite"
robot = "fourier-gr1"
family = "robocasa-gr1/PnPCupToDrawerClose"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robocasa-gr1-cup-to-drawer-close-i00-standard/README.md b/tasks/robocasa-gr1-cup-to-drawer-close-i00-standard/README.md
index b48ddac3de..d9c2e6b8e1 100644
--- a/tasks/robocasa-gr1-cup-to-drawer-close-i00-standard/README.md
+++ b/tasks/robocasa-gr1-cup-to-drawer-close-i00-standard/README.md
@@ -48,3 +48,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
- 2026-10-10 (review round 3, board 0067): the sim service's request handling (`environment/sim/server_casa.py`) and `tests/grade.py`. A negative or non-numeric `Content-Length` is answered 400 without reading the body (the service read until the client hung up, without a bound, which could exhaust its memory); a camera named more than once in `cameras` is rendered and returned once (N copies of one name made it build N x N images within the frame budget); a port that is still held when the service restarts is waited for, not counted as a crash; a JSON `true` or `false` where `dt` or `seconds` takes a number is a 400 (`true` ran as 1 second). The grader: a `final.json` whose evidence fields (`control_steps`, `calls`, `elapsed_s`, `final_reward`) are not numbers no longer stops it before it writes a reward (they are reported as 0 and decide nothing). An episode's physics and the pass rule are unchanged.
- 2026-10-13 (owner, boards 0085 and 0087): the sim service's clock (`environment/sim/server_casa.py`, the docstring of `environment/rcb_client.py`). `meta["time_s"]` and `meta["time_left_s"]` count from the agent's first request to the service (any endpoint), and a restarted service keeps that moment; before, they counted from the first start of the service's container (`time_left_s`) and from when its scene was ready (`time_s`). Where the service is ready before the agent's phase begins (the harness is installed in between: a median of 79 s over the recorded standard trials), that clock ran ahead of the agent's and `time_left_s` reached 0 while the agent still had that time. `/spec` says what the clock counts from (`budget.clock`), and the instruction's `## Time limit` tells the agent to plan by `time_left_s`. The service never ended an episode on this clock: the scene, the frozen instance, the pass rule and the grader are unchanged, so the task keeps its name and earlier results stand.
- 2026-10-13 (board 0093; found in board 0086): the camera calibration (`environment/sim/server_casa.py`). `cam2world_gl` held the pose of MuJoCo's left-eye camera (`scene.camera[0]`), 3.4 cm (half the model's 0.068 m interpupillary distance) to the camera's left of the viewpoint the mono RGB and depth images are rendered from; it now is that viewpoint (`data.cam_xpos` / `data.cam_xmat`). Checked against MuJoCo's own renders of marker spheres at known world points, every served camera, before and after the robot moved: the reprojection error was 10.5-10.7 px median (up to 28 px for near points, always to the same side), now 0.15 px median and at most 0.4 px. The intrinsics were right (every camera here is given by its `fovy`, none by a sensor size) and are unchanged, as are the images, the scene, the frozen instance, the pass rule and the grader, so the task keeps its name. **Standard-mode results from before this date used the offset calibration**: a point located from the images and the reported calibration came out about 3.4 cm to the side (agents missed a microwave button by 3.5 cm), so they are not strictly comparable with later ones.
+- 2026-10-14 (board 0113, from board 0086): the instruction describes the hands: each hand command is the position target (rad) of its finger's joints, 0 open and larger closing, within each joint's range; the grip site is the origin of the hand's mounting frame at the wrist end of the hand; the pink spheres beside each hand are visual markers of the robot model, not objects (board 0086: most standard runs took them for objects). Text only: the scene, the frozen instance, the reference solution, the success check and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/robocasa-gr1-cup-to-drawer-close-i00-standard/instruction.md b/tasks/robocasa-gr1-cup-to-drawer-close-i00-standard/instruction.md
index faacde49ae..fefab2cff6 100644
--- a/tasks/robocasa-gr1-cup-to-drawer-close-i00-standard/instruction.md
+++ b/tasks/robocasa-gr1-cup-to-drawer-close-i00-standard/instruction.md
@@ -51,6 +51,13 @@ res.video[cam]; res.frames[cam]; res.last[cam]; res.proprio; res.save("/app/runs
channel is an absolute target, so the grip sites stay within about a millimetre, but the IK's posture term can still
move the waist and elbows by a few hundredths of a radian while you watch.
- The cameras are mounted on the robot and move with it: use the calibration that comes with each response.
+- The hands: each hand command is the position target, in radians, of that finger's joints (`thumb_flexion` drives
+ two thumb joints, `thumb_rotation` one): 0 is the straight, open finger, larger values bend it toward the palm
+ (the hand closes), and every finger joint stops at the end of its range (1.22 to 1.74 rad), so values below 0 or
+ above about 1.7 change nothing more. The grip site that the pose targets place is the origin of the hand's mounting
+ frame (`gripper0_right_grip_site`, `gripper0_left_grip_site`), at the wrist end of the hand, not a point between
+ the fingertips. The four small pink spheres beside each hand are part of the robot model (visual markers
+ `robot0_<side>_pinch_spheres_0..3`, about 13.5 cm from its grip site, with no collision): they are not objects.
- World frame: metres, z up. In the proprio, `eef_left_*` / `eef_right_*` are the grip sites' positions in the world
frame; the action's pose targets and the `cmd_*` entries (the last command) are in the robot's base frame.
The robot's base frame (body `robot0_base`, fixed in the world) has its origin at world (2.65000, -2.17500, 0.97000) m and is turned +90.0 deg about z (quaternion w x y z (0.70711, 0, 0, 0.70711)): `p_world = R @ p_base + t` with R = [[0, -1, 0], [1, 0, 0], [0, 0, 1]].
diff --git a/tasks/robocasa-gr1-cup-to-drawer-close-i00-standard/task.toml b/tasks/robocasa-gr1-cup-to-drawer-close-i00-standard/task.toml
index 7d8b6be2d0..bb5b0eed52 100644
--- a/tasks/robocasa-gr1-cup-to-drawer-close-i00-standard/task.toml
+++ b/tasks/robocasa-gr1-cup-to-drawer-close-i00-standard/task.toml
@@ -24,6 +24,7 @@ simulator = "robosuite"
robot = "fourier-gr1"
family = "robocasa-gr1/PnPCupToDrawerClose"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
expert_time_estimate_hours = 4.0
diff --git a/tasks/robocasa-gr1-cuttingboard-to-basket-i00-privileged/task.toml b/tasks/robocasa-gr1-cuttingboard-to-basket-i00-privileged/task.toml
index 7a45946a7a..aa712fbc0c 100644
--- a/tasks/robocasa-gr1-cuttingboard-to-basket-i00-privileged/task.toml
+++ b/tasks/robocasa-gr1-cuttingboard-to-basket-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "robosuite"
robot = "fourier-gr1"
family = "robocasa-gr1/PosttrainPnPNovelFromCuttingboardToBasketSplitA"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robocasa-gr1-cuttingboard-to-basket-i00-standard/README.md b/tasks/robocasa-gr1-cuttingboard-to-basket-i00-standard/README.md
index 0c25b846ba..67600c3c59 100644
--- a/tasks/robocasa-gr1-cuttingboard-to-basket-i00-standard/README.md
+++ b/tasks/robocasa-gr1-cuttingboard-to-basket-i00-standard/README.md
@@ -48,3 +48,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
- 2026-10-10 (review round 3, board 0067): the sim service's request handling (`environment/sim/server_casa.py`) and `tests/grade.py`. A negative or non-numeric `Content-Length` is answered 400 without reading the body (the service read until the client hung up, without a bound, which could exhaust its memory); a camera named more than once in `cameras` is rendered and returned once (N copies of one name made it build N x N images within the frame budget); a port that is still held when the service restarts is waited for, not counted as a crash; a JSON `true` or `false` where `dt` or `seconds` takes a number is a 400 (`true` ran as 1 second). The grader: a `final.json` whose evidence fields (`control_steps`, `calls`, `elapsed_s`, `final_reward`) are not numbers no longer stops it before it writes a reward (they are reported as 0 and decide nothing). An episode's physics and the pass rule are unchanged.
- 2026-10-13 (owner, boards 0085 and 0087): the sim service's clock (`environment/sim/server_casa.py`, the docstring of `environment/rcb_client.py`). `meta["time_s"]` and `meta["time_left_s"]` count from the agent's first request to the service (any endpoint), and a restarted service keeps that moment; before, they counted from the first start of the service's container (`time_left_s`) and from when its scene was ready (`time_s`). Where the service is ready before the agent's phase begins (the harness is installed in between: a median of 79 s over the recorded standard trials), that clock ran ahead of the agent's and `time_left_s` reached 0 while the agent still had that time. `/spec` says what the clock counts from (`budget.clock`), and the instruction's `## Time limit` tells the agent to plan by `time_left_s`. The service never ended an episode on this clock: the scene, the frozen instance, the pass rule and the grader are unchanged, so the task keeps its name and earlier results stand.
- 2026-10-13 (board 0093; found in board 0086): the camera calibration (`environment/sim/server_casa.py`). `cam2world_gl` held the pose of MuJoCo's left-eye camera (`scene.camera[0]`), 3.4 cm (half the model's 0.068 m interpupillary distance) to the camera's left of the viewpoint the mono RGB and depth images are rendered from; it now is that viewpoint (`data.cam_xpos` / `data.cam_xmat`). Checked against MuJoCo's own renders of marker spheres at known world points, every served camera, before and after the robot moved: the reprojection error was 10.5-10.7 px median (up to 28 px for near points, always to the same side), now 0.15 px median and at most 0.4 px. The intrinsics were right (every camera here is given by its `fovy`, none by a sensor size) and are unchanged, as are the images, the scene, the frozen instance, the pass rule and the grader, so the task keeps its name. **Standard-mode results from before this date used the offset calibration**: a point located from the images and the reported calibration came out about 3.4 cm to the side (agents missed a microwave button by 3.5 cm), so they are not strictly comparable with later ones.
+- 2026-10-14 (board 0113, from board 0086): the instruction describes the hands: each hand command is the position target (rad) of its finger's joints, 0 open and larger closing, within each joint's range; the grip site is the origin of the hand's mounting frame at the wrist end of the hand; the pink spheres beside each hand are visual markers of the robot model, not objects (board 0086: most standard runs took them for objects). Text only: the scene, the frozen instance, the reference solution, the success check and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/robocasa-gr1-cuttingboard-to-basket-i00-standard/instruction.md b/tasks/robocasa-gr1-cuttingboard-to-basket-i00-standard/instruction.md
index 6f8b08a667..bf52ad6996 100644
--- a/tasks/robocasa-gr1-cuttingboard-to-basket-i00-standard/instruction.md
+++ b/tasks/robocasa-gr1-cuttingboard-to-basket-i00-standard/instruction.md
@@ -51,6 +51,13 @@ res.video[cam]; res.frames[cam]; res.last[cam]; res.proprio; res.save("/app/runs
channel is an absolute target, so the grip sites stay within about a millimetre, but the IK's posture term can still
move the waist and elbows by a few hundredths of a radian while you watch.
- The cameras are mounted on the robot and move with it: use the calibration that comes with each response.
+- The hands: each hand command is the position target, in radians, of that finger's joints (`thumb_flexion` drives
+ two thumb joints, `thumb_rotation` one): 0 is the straight, open finger, larger values bend it toward the palm
+ (the hand closes), and every finger joint stops at the end of its range (1.22 to 1.74 rad), so values below 0 or
+ above about 1.7 change nothing more. The grip site that the pose targets place is the origin of the hand's mounting
+ frame (`gripper0_right_grip_site`, `gripper0_left_grip_site`), at the wrist end of the hand, not a point between
+ the fingertips. The four small pink spheres beside each hand are part of the robot model (visual markers
+ `robot0_<side>_pinch_spheres_0..3`, about 13.5 cm from its grip site, with no collision): they are not objects.
- World frame: metres, z up. In the proprio, `eef_left_*` / `eef_right_*` are the grip sites' positions in the world
frame; the action's pose targets and the `cmd_*` entries (the last command) are in the robot's base frame.
The robot's base frame (body `robot0_base`, fixed in the world) has its origin at world (2.67072, -2.12925, 0.89836) m and is turned +90.0 deg about z (quaternion w x y z (0.70711, 0, 0, 0.70711)): `p_world = R @ p_base + t` with R = [[0, -1, 0], [1, 0, 0], [0, 0, 1]].
diff --git a/tasks/robocasa-gr1-cuttingboard-to-basket-i00-standard/task.toml b/tasks/robocasa-gr1-cuttingboard-to-basket-i00-standard/task.toml
index 332fc07737..966b0a2ec4 100644
--- a/tasks/robocasa-gr1-cuttingboard-to-basket-i00-standard/task.toml
+++ b/tasks/robocasa-gr1-cuttingboard-to-basket-i00-standard/task.toml
@@ -24,6 +24,7 @@ simulator = "robosuite"
robot = "fourier-gr1"
family = "robocasa-gr1/PosttrainPnPNovelFromCuttingboardToBasketSplitA"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
expert_time_estimate_hours = 4.0
diff --git a/tasks/robocasa-gr1-cuttingboard-to-cardboardbox-i00-privileged/task.toml b/tasks/robocasa-gr1-cuttingboard-to-cardboardbox-i00-privileged/task.toml
index d0b2120a1a..5d7d0ef484 100644
--- a/tasks/robocasa-gr1-cuttingboard-to-cardboardbox-i00-privileged/task.toml
+++ b/tasks/robocasa-gr1-cuttingboard-to-cardboardbox-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "robosuite"
robot = "fourier-gr1"
family = "robocasa-gr1/PosttrainPnPNovelFromCuttingboardToCardboardboxSplitA"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robocasa-gr1-cuttingboard-to-cardboardbox-i00-standard/README.md b/tasks/robocasa-gr1-cuttingboard-to-cardboardbox-i00-standard/README.md
index c266223531..513adbb507 100644
--- a/tasks/robocasa-gr1-cuttingboard-to-cardboardbox-i00-standard/README.md
+++ b/tasks/robocasa-gr1-cuttingboard-to-cardboardbox-i00-standard/README.md
@@ -48,3 +48,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
- 2026-10-10 (review round 3, board 0067): the sim service's request handling (`environment/sim/server_casa.py`) and `tests/grade.py`. A negative or non-numeric `Content-Length` is answered 400 without reading the body (the service read until the client hung up, without a bound, which could exhaust its memory); a camera named more than once in `cameras` is rendered and returned once (N copies of one name made it build N x N images within the frame budget); a port that is still held when the service restarts is waited for, not counted as a crash; a JSON `true` or `false` where `dt` or `seconds` takes a number is a 400 (`true` ran as 1 second). The grader: a `final.json` whose evidence fields (`control_steps`, `calls`, `elapsed_s`, `final_reward`) are not numbers no longer stops it before it writes a reward (they are reported as 0 and decide nothing). An episode's physics and the pass rule are unchanged.
- 2026-10-13 (owner, boards 0085 and 0087): the sim service's clock (`environment/sim/server_casa.py`, the docstring of `environment/rcb_client.py`). `meta["time_s"]` and `meta["time_left_s"]` count from the agent's first request to the service (any endpoint), and a restarted service keeps that moment; before, they counted from the first start of the service's container (`time_left_s`) and from when its scene was ready (`time_s`). Where the service is ready before the agent's phase begins (the harness is installed in between: a median of 79 s over the recorded standard trials), that clock ran ahead of the agent's and `time_left_s` reached 0 while the agent still had that time. `/spec` says what the clock counts from (`budget.clock`), and the instruction's `## Time limit` tells the agent to plan by `time_left_s`. The service never ended an episode on this clock: the scene, the frozen instance, the pass rule and the grader are unchanged, so the task keeps its name and earlier results stand.
- 2026-10-13 (board 0093; found in board 0086): the camera calibration (`environment/sim/server_casa.py`). `cam2world_gl` held the pose of MuJoCo's left-eye camera (`scene.camera[0]`), 3.4 cm (half the model's 0.068 m interpupillary distance) to the camera's left of the viewpoint the mono RGB and depth images are rendered from; it now is that viewpoint (`data.cam_xpos` / `data.cam_xmat`). Checked against MuJoCo's own renders of marker spheres at known world points, every served camera, before and after the robot moved: the reprojection error was 10.5-10.7 px median (up to 28 px for near points, always to the same side), now 0.15 px median and at most 0.4 px. The intrinsics were right (every camera here is given by its `fovy`, none by a sensor size) and are unchanged, as are the images, the scene, the frozen instance, the pass rule and the grader, so the task keeps its name. **Standard-mode results from before this date used the offset calibration**: a point located from the images and the reported calibration came out about 3.4 cm to the side (agents missed a microwave button by 3.5 cm), so they are not strictly comparable with later ones.
+- 2026-10-14 (board 0113, from board 0086): the instruction describes the hands: each hand command is the position target (rad) of its finger's joints, 0 open and larger closing, within each joint's range; the grip site is the origin of the hand's mounting frame at the wrist end of the hand; the pink spheres beside each hand are visual markers of the robot model, not objects (board 0086: most standard runs took them for objects). Text only: the scene, the frozen instance, the reference solution, the success check and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/robocasa-gr1-cuttingboard-to-cardboardbox-i00-standard/instruction.md b/tasks/robocasa-gr1-cuttingboard-to-cardboardbox-i00-standard/instruction.md
index 7dd1338749..26395a94f8 100644
--- a/tasks/robocasa-gr1-cuttingboard-to-cardboardbox-i00-standard/instruction.md
+++ b/tasks/robocasa-gr1-cuttingboard-to-cardboardbox-i00-standard/instruction.md
@@ -52,6 +52,13 @@ res.video[cam]; res.frames[cam]; res.last[cam]; res.proprio; res.save("/app/runs
channel is an absolute target, so the grip sites stay within about a millimetre, but the IK's posture term can still
move the waist and elbows by a few hundredths of a radian while you watch.
- The cameras are mounted on the robot and move with it: use the calibration that comes with each response.
+- The hands: each hand command is the position target, in radians, of that finger's joints (`thumb_flexion` drives
+ two thumb joints, `thumb_rotation` one): 0 is the straight, open finger, larger values bend it toward the palm
+ (the hand closes), and every finger joint stops at the end of its range (1.22 to 1.74 rad), so values below 0 or
+ above about 1.7 change nothing more. The grip site that the pose targets place is the origin of the hand's mounting
+ frame (`gripper0_right_grip_site`, `gripper0_left_grip_site`), at the wrist end of the hand, not a point between
+ the fingertips. The four small pink spheres beside each hand are part of the robot model (visual markers
+ `robot0_<side>_pinch_spheres_0..3`, about 13.5 cm from its grip site, with no collision): they are not objects.
- World frame: metres, z up. In the proprio, `eef_left_*` / `eef_right_*` are the grip sites' positions in the world
frame; the action's pose targets and the `cmd_*` entries (the last command) are in the robot's base frame.
The robot's base frame (body `robot0_base`, fixed in the world) has its origin at world (2.67858, -2.16910, 0.87305) m and is turned +90.0 deg about z (quaternion w x y z (0.70711, 0, 0, 0.70711)): `p_world = R @ p_base + t` with R = [[0, -1, 0], [1, 0, 0], [0, 0, 1]].
diff --git a/tasks/robocasa-gr1-cuttingboard-to-cardboardbox-i00-standard/task.toml b/tasks/robocasa-gr1-cuttingboard-to-cardboardbox-i00-standard/task.toml
index 95bd08d6be..afe275878c 100644
--- a/tasks/robocasa-gr1-cuttingboard-to-cardboardbox-i00-standard/task.toml
+++ b/tasks/robocasa-gr1-cuttingboard-to-cardboardbox-i00-standard/task.toml
@@ -24,6 +24,7 @@ simulator = "robosuite"
robot = "fourier-gr1"
family = "robocasa-gr1/PosttrainPnPNovelFromCuttingboardToCardboardboxSplitA"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
expert_time_estimate_hours = 4.0
diff --git a/tasks/robocasa-gr1-cuttingboard-to-pan-i00-privileged/task.toml b/tasks/robocasa-gr1-cuttingboard-to-pan-i00-privileged/task.toml
index f158cd606a..bf88e8ceb0 100644
--- a/tasks/robocasa-gr1-cuttingboard-to-pan-i00-privileged/task.toml
+++ b/tasks/robocasa-gr1-cuttingboard-to-pan-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "robosuite"
robot = "fourier-gr1"
family = "robocasa-gr1/PosttrainPnPNovelFromCuttingboardToPanSplitA"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robocasa-gr1-cuttingboard-to-pan-i00-standard/README.md b/tasks/robocasa-gr1-cuttingboard-to-pan-i00-standard/README.md
index ba06cb258f..e61078cec3 100644
--- a/tasks/robocasa-gr1-cuttingboard-to-pan-i00-standard/README.md
+++ b/tasks/robocasa-gr1-cuttingboard-to-pan-i00-standard/README.md
@@ -48,3 +48,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
- 2026-10-10 (review round 3, board 0067): the sim service's request handling (`environment/sim/server_casa.py`) and `tests/grade.py`. A negative or non-numeric `Content-Length` is answered 400 without reading the body (the service read until the client hung up, without a bound, which could exhaust its memory); a camera named more than once in `cameras` is rendered and returned once (N copies of one name made it build N x N images within the frame budget); a port that is still held when the service restarts is waited for, not counted as a crash; a JSON `true` or `false` where `dt` or `seconds` takes a number is a 400 (`true` ran as 1 second). The grader: a `final.json` whose evidence fields (`control_steps`, `calls`, `elapsed_s`, `final_reward`) are not numbers no longer stops it before it writes a reward (they are reported as 0 and decide nothing). An episode's physics and the pass rule are unchanged.
- 2026-10-13 (owner, boards 0085 and 0087): the sim service's clock (`environment/sim/server_casa.py`, the docstring of `environment/rcb_client.py`). `meta["time_s"]` and `meta["time_left_s"]` count from the agent's first request to the service (any endpoint), and a restarted service keeps that moment; before, they counted from the first start of the service's container (`time_left_s`) and from when its scene was ready (`time_s`). Where the service is ready before the agent's phase begins (the harness is installed in between: a median of 79 s over the recorded standard trials), that clock ran ahead of the agent's and `time_left_s` reached 0 while the agent still had that time. `/spec` says what the clock counts from (`budget.clock`), and the instruction's `## Time limit` tells the agent to plan by `time_left_s`. The service never ended an episode on this clock: the scene, the frozen instance, the pass rule and the grader are unchanged, so the task keeps its name and earlier results stand.
- 2026-10-13 (board 0093; found in board 0086): the camera calibration (`environment/sim/server_casa.py`). `cam2world_gl` held the pose of MuJoCo's left-eye camera (`scene.camera[0]`), 3.4 cm (half the model's 0.068 m interpupillary distance) to the camera's left of the viewpoint the mono RGB and depth images are rendered from; it now is that viewpoint (`data.cam_xpos` / `data.cam_xmat`). Checked against MuJoCo's own renders of marker spheres at known world points, every served camera, before and after the robot moved: the reprojection error was 10.5-10.7 px median (up to 28 px for near points, always to the same side), now 0.15 px median and at most 0.4 px. The intrinsics were right (every camera here is given by its `fovy`, none by a sensor size) and are unchanged, as are the images, the scene, the frozen instance, the pass rule and the grader, so the task keeps its name. **Standard-mode results from before this date used the offset calibration**: a point located from the images and the reported calibration came out about 3.4 cm to the side (agents missed a microwave button by 3.5 cm), so they are not strictly comparable with later ones.
+- 2026-10-14 (board 0113, from board 0086): the instruction describes the hands: each hand command is the position target (rad) of its finger's joints, 0 open and larger closing, within each joint's range; the grip site is the origin of the hand's mounting frame at the wrist end of the hand; the pink spheres beside each hand are visual markers of the robot model, not objects (board 0086: most standard runs took them for objects). Text only: the scene, the frozen instance, the reference solution, the success check and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/robocasa-gr1-cuttingboard-to-pan-i00-standard/instruction.md b/tasks/robocasa-gr1-cuttingboard-to-pan-i00-standard/instruction.md
index bfce81b57f..1e870a6156 100644
--- a/tasks/robocasa-gr1-cuttingboard-to-pan-i00-standard/instruction.md
+++ b/tasks/robocasa-gr1-cuttingboard-to-pan-i00-standard/instruction.md
@@ -51,6 +51,13 @@ res.video[cam]; res.frames[cam]; res.last[cam]; res.proprio; res.save("/app/runs
channel is an absolute target, so the grip sites stay within about a millimetre, but the IK's posture term can still
move the waist and elbows by a few hundredths of a radian while you watch.
- The cameras are mounted on the robot and move with it: use the calibration that comes with each response.
+- The hands: each hand command is the position target, in radians, of that finger's joints (`thumb_flexion` drives
+ two thumb joints, `thumb_rotation` one): 0 is the straight, open finger, larger values bend it toward the palm
+ (the hand closes), and every finger joint stops at the end of its range (1.22 to 1.74 rad), so values below 0 or
+ above about 1.7 change nothing more. The grip site that the pose targets place is the origin of the hand's mounting
+ frame (`gripper0_right_grip_site`, `gripper0_left_grip_site`), at the wrist end of the hand, not a point between
+ the fingertips. The four small pink spheres beside each hand are part of the robot model (visual markers
+ `robot0_<side>_pinch_spheres_0..3`, about 13.5 cm from its grip site, with no collision): they are not objects.
- World frame: metres, z up. In the proprio, `eef_left_*` / `eef_right_*` are the grip sites' positions in the world
frame; the action's pose targets and the `cmd_*` entries (the last command) are in the robot's base frame.
The robot's base frame (body `robot0_base`, fixed in the world) has its origin at world (2.69788, -2.14277, 0.90223) m and is turned +90.0 deg about z (quaternion w x y z (0.70711, 0, 0, 0.70711)): `p_world = R @ p_base + t` with R = [[0, -1, 0], [1, 0, 0], [0, 0, 1]].
diff --git a/tasks/robocasa-gr1-cuttingboard-to-pan-i00-standard/task.toml b/tasks/robocasa-gr1-cuttingboard-to-pan-i00-standard/task.toml
index deaa0bbc1c..2ccc916784 100644
--- a/tasks/robocasa-gr1-cuttingboard-to-pan-i00-standard/task.toml
+++ b/tasks/robocasa-gr1-cuttingboard-to-pan-i00-standard/task.toml
@@ -24,6 +24,7 @@ simulator = "robosuite"
robot = "fourier-gr1"
family = "robocasa-gr1/PosttrainPnPNovelFromCuttingboardToPanSplitA"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
expert_time_estimate_hours = 4.0
diff --git a/tasks/robocasa-gr1-cuttingboard-to-pot-i00-privileged/task.toml b/tasks/robocasa-gr1-cuttingboard-to-pot-i00-privileged/task.toml
index 151ddf35fc..837edbbb6c 100644
--- a/tasks/robocasa-gr1-cuttingboard-to-pot-i00-privileged/task.toml
+++ b/tasks/robocasa-gr1-cuttingboard-to-pot-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "robosuite"
robot = "fourier-gr1"
family = "robocasa-gr1/PosttrainPnPNovelFromCuttingboardToPotSplitA"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robocasa-gr1-cuttingboard-to-pot-i00-standard/README.md b/tasks/robocasa-gr1-cuttingboard-to-pot-i00-standard/README.md
index a96acd3dcd..68d917c0bd 100644
--- a/tasks/robocasa-gr1-cuttingboard-to-pot-i00-standard/README.md
+++ b/tasks/robocasa-gr1-cuttingboard-to-pot-i00-standard/README.md
@@ -48,3 +48,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
- 2026-10-10 (review round 3, board 0067): the sim service's request handling (`environment/sim/server_casa.py`) and `tests/grade.py`. A negative or non-numeric `Content-Length` is answered 400 without reading the body (the service read until the client hung up, without a bound, which could exhaust its memory); a camera named more than once in `cameras` is rendered and returned once (N copies of one name made it build N x N images within the frame budget); a port that is still held when the service restarts is waited for, not counted as a crash; a JSON `true` or `false` where `dt` or `seconds` takes a number is a 400 (`true` ran as 1 second). The grader: a `final.json` whose evidence fields (`control_steps`, `calls`, `elapsed_s`, `final_reward`) are not numbers no longer stops it before it writes a reward (they are reported as 0 and decide nothing). An episode's physics and the pass rule are unchanged.
- 2026-10-13 (owner, boards 0085 and 0087): the sim service's clock (`environment/sim/server_casa.py`, the docstring of `environment/rcb_client.py`). `meta["time_s"]` and `meta["time_left_s"]` count from the agent's first request to the service (any endpoint), and a restarted service keeps that moment; before, they counted from the first start of the service's container (`time_left_s`) and from when its scene was ready (`time_s`). Where the service is ready before the agent's phase begins (the harness is installed in between: a median of 79 s over the recorded standard trials), that clock ran ahead of the agent's and `time_left_s` reached 0 while the agent still had that time. `/spec` says what the clock counts from (`budget.clock`), and the instruction's `## Time limit` tells the agent to plan by `time_left_s`. The service never ended an episode on this clock: the scene, the frozen instance, the pass rule and the grader are unchanged, so the task keeps its name and earlier results stand.
- 2026-10-13 (board 0093; found in board 0086): the camera calibration (`environment/sim/server_casa.py`). `cam2world_gl` held the pose of MuJoCo's left-eye camera (`scene.camera[0]`), 3.4 cm (half the model's 0.068 m interpupillary distance) to the camera's left of the viewpoint the mono RGB and depth images are rendered from; it now is that viewpoint (`data.cam_xpos` / `data.cam_xmat`). Checked against MuJoCo's own renders of marker spheres at known world points, every served camera, before and after the robot moved: the reprojection error was 10.5-10.7 px median (up to 28 px for near points, always to the same side), now 0.15 px median and at most 0.4 px. The intrinsics were right (every camera here is given by its `fovy`, none by a sensor size) and are unchanged, as are the images, the scene, the frozen instance, the pass rule and the grader, so the task keeps its name. **Standard-mode results from before this date used the offset calibration**: a point located from the images and the reported calibration came out about 3.4 cm to the side (agents missed a microwave button by 3.5 cm), so they are not strictly comparable with later ones.
+- 2026-10-14 (board 0113, from board 0086): the instruction describes the hands: each hand command is the position target (rad) of its finger's joints, 0 open and larger closing, within each joint's range; the grip site is the origin of the hand's mounting frame at the wrist end of the hand; the pink spheres beside each hand are visual markers of the robot model, not objects (board 0086: most standard runs took them for objects). Text only: the scene, the frozen instance, the reference solution, the success check and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/robocasa-gr1-cuttingboard-to-pot-i00-standard/instruction.md b/tasks/robocasa-gr1-cuttingboard-to-pot-i00-standard/instruction.md
index f3fb30d2ad..3bcd28b8d7 100644
--- a/tasks/robocasa-gr1-cuttingboard-to-pot-i00-standard/instruction.md
+++ b/tasks/robocasa-gr1-cuttingboard-to-pot-i00-standard/instruction.md
@@ -51,6 +51,13 @@ res.video[cam]; res.frames[cam]; res.last[cam]; res.proprio; res.save("/app/runs
channel is an absolute target, so the grip sites stay within about a millimetre, but the IK's posture term can still
move the waist and elbows by a few hundredths of a radian while you watch.
- The cameras are mounted on the robot and move with it: use the calibration that comes with each response.
+- The hands: each hand command is the position target, in radians, of that finger's joints (`thumb_flexion` drives
+ two thumb joints, `thumb_rotation` one): 0 is the straight, open finger, larger values bend it toward the palm
+ (the hand closes), and every finger joint stops at the end of its range (1.22 to 1.74 rad), so values below 0 or
+ above about 1.7 change nothing more. The grip site that the pose targets place is the origin of the hand's mounting
+ frame (`gripper0_right_grip_site`, `gripper0_left_grip_site`), at the wrist end of the hand, not a point between
+ the fingertips. The four small pink spheres beside each hand are part of the robot model (visual markers
+ `robot0_<side>_pinch_spheres_0..3`, about 13.5 cm from its grip site, with no collision): they are not objects.
- World frame: metres, z up. In the proprio, `eef_left_*` / `eef_right_*` are the grip sites' positions in the world
frame; the action's pose targets and the `cmd_*` entries (the last command) are in the robot's base frame.
The robot's base frame (body `robot0_base`, fixed in the world) has its origin at world (2.66899, -2.16906, 0.89825) m and is turned +90.0 deg about z (quaternion w x y z (0.70711, 0, 0, 0.70711)): `p_world = R @ p_base + t` with R = [[0, -1, 0], [1, 0, 0], [0, 0, 1]].
diff --git a/tasks/robocasa-gr1-cuttingboard-to-pot-i00-standard/task.toml b/tasks/robocasa-gr1-cuttingboard-to-pot-i00-standard/task.toml
index 5c1e14a876..b862cdfc1c 100644
--- a/tasks/robocasa-gr1-cuttingboard-to-pot-i00-standard/task.toml
+++ b/tasks/robocasa-gr1-cuttingboard-to-pot-i00-standard/task.toml
@@ -24,6 +24,7 @@ simulator = "robosuite"
robot = "fourier-gr1"
family = "robocasa-gr1/PosttrainPnPNovelFromCuttingboardToPotSplitA"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
expert_time_estimate_hours = 4.0
diff --git a/tasks/robocasa-gr1-cuttingboard-to-tieredbasket-i00-privileged/task.toml b/tasks/robocasa-gr1-cuttingboard-to-tieredbasket-i00-privileged/task.toml
index 544d871617..837524b564 100644
--- a/tasks/robocasa-gr1-cuttingboard-to-tieredbasket-i00-privileged/task.toml
+++ b/tasks/robocasa-gr1-cuttingboard-to-tieredbasket-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "robosuite"
robot = "fourier-gr1"
family = "robocasa-gr1/PosttrainPnPNovelFromCuttingboardToTieredbasketSplitA"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robocasa-gr1-cuttingboard-to-tieredbasket-i00-standard/README.md b/tasks/robocasa-gr1-cuttingboard-to-tieredbasket-i00-standard/README.md
index fef563235f..baa1f23699 100644
--- a/tasks/robocasa-gr1-cuttingboard-to-tieredbasket-i00-standard/README.md
+++ b/tasks/robocasa-gr1-cuttingboard-to-tieredbasket-i00-standard/README.md
@@ -48,3 +48,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
- 2026-10-10 (review round 3, board 0067): the sim service's request handling (`environment/sim/server_casa.py`) and `tests/grade.py`. A negative or non-numeric `Content-Length` is answered 400 without reading the body (the service read until the client hung up, without a bound, which could exhaust its memory); a camera named more than once in `cameras` is rendered and returned once (N copies of one name made it build N x N images within the frame budget); a port that is still held when the service restarts is waited for, not counted as a crash; a JSON `true` or `false` where `dt` or `seconds` takes a number is a 400 (`true` ran as 1 second). The grader: a `final.json` whose evidence fields (`control_steps`, `calls`, `elapsed_s`, `final_reward`) are not numbers no longer stops it before it writes a reward (they are reported as 0 and decide nothing). An episode's physics and the pass rule are unchanged.
- 2026-10-13 (owner, boards 0085 and 0087): the sim service's clock (`environment/sim/server_casa.py`, the docstring of `environment/rcb_client.py`). `meta["time_s"]` and `meta["time_left_s"]` count from the agent's first request to the service (any endpoint), and a restarted service keeps that moment; before, they counted from the first start of the service's container (`time_left_s`) and from when its scene was ready (`time_s`). Where the service is ready before the agent's phase begins (the harness is installed in between: a median of 79 s over the recorded standard trials), that clock ran ahead of the agent's and `time_left_s` reached 0 while the agent still had that time. `/spec` says what the clock counts from (`budget.clock`), and the instruction's `## Time limit` tells the agent to plan by `time_left_s`. The service never ended an episode on this clock: the scene, the frozen instance, the pass rule and the grader are unchanged, so the task keeps its name and earlier results stand.
- 2026-10-13 (board 0093; found in board 0086): the camera calibration (`environment/sim/server_casa.py`). `cam2world_gl` held the pose of MuJoCo's left-eye camera (`scene.camera[0]`), 3.4 cm (half the model's 0.068 m interpupillary distance) to the camera's left of the viewpoint the mono RGB and depth images are rendered from; it now is that viewpoint (`data.cam_xpos` / `data.cam_xmat`). Checked against MuJoCo's own renders of marker spheres at known world points, every served camera, before and after the robot moved: the reprojection error was 10.5-10.7 px median (up to 28 px for near points, always to the same side), now 0.15 px median and at most 0.4 px. The intrinsics were right (every camera here is given by its `fovy`, none by a sensor size) and are unchanged, as are the images, the scene, the frozen instance, the pass rule and the grader, so the task keeps its name. **Standard-mode results from before this date used the offset calibration**: a point located from the images and the reported calibration came out about 3.4 cm to the side (agents missed a microwave button by 3.5 cm), so they are not strictly comparable with later ones.
+- 2026-10-14 (board 0113, from board 0086): the instruction describes the hands: each hand command is the position target (rad) of its finger's joints, 0 open and larger closing, within each joint's range; the grip site is the origin of the hand's mounting frame at the wrist end of the hand; the pink spheres beside each hand are visual markers of the robot model, not objects (board 0086: most standard runs took them for objects). Text only: the scene, the frozen instance, the reference solution, the success check and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/robocasa-gr1-cuttingboard-to-tieredbasket-i00-standard/instruction.md b/tasks/robocasa-gr1-cuttingboard-to-tieredbasket-i00-standard/instruction.md
index 56c6379044..67d4fc8b3c 100644
--- a/tasks/robocasa-gr1-cuttingboard-to-tieredbasket-i00-standard/instruction.md
+++ b/tasks/robocasa-gr1-cuttingboard-to-tieredbasket-i00-standard/instruction.md
@@ -52,6 +52,13 @@ res.video[cam]; res.frames[cam]; res.last[cam]; res.proprio; res.save("/app/runs
channel is an absolute target, so the grip sites stay within about a millimetre, but the IK's posture term can still
move the waist and elbows by a few hundredths of a radian while you watch.
- The cameras are mounted on the robot and move with it: use the calibration that comes with each response.
+- The hands: each hand command is the position target, in radians, of that finger's joints (`thumb_flexion` drives
+ two thumb joints, `thumb_rotation` one): 0 is the straight, open finger, larger values bend it toward the palm
+ (the hand closes), and every finger joint stops at the end of its range (1.22 to 1.74 rad), so values below 0 or
+ above about 1.7 change nothing more. The grip site that the pose targets place is the origin of the hand's mounting
+ frame (`gripper0_right_grip_site`, `gripper0_left_grip_site`), at the wrist end of the hand, not a point between
+ the fingertips. The four small pink spheres beside each hand are part of the robot model (visual markers
+ `robot0_<side>_pinch_spheres_0..3`, about 13.5 cm from its grip site, with no collision): they are not objects.
- World frame: metres, z up. In the proprio, `eef_left_*` / `eef_right_*` are the grip sites' positions in the world
frame; the action's pose targets and the `cmd_*` entries (the last command) are in the robot's base frame.
The robot's base frame (body `robot0_base`, fixed in the world) has its origin at world (2.68478, -2.15993, 0.88819) m and is turned +90.0 deg about z (quaternion w x y z (0.70711, 0, 0, 0.70711)): `p_world = R @ p_base + t` with R = [[0, -1, 0], [1, 0, 0], [0, 0, 1]].
diff --git a/tasks/robocasa-gr1-cuttingboard-to-tieredbasket-i00-standard/task.toml b/tasks/robocasa-gr1-cuttingboard-to-tieredbasket-i00-standard/task.toml
index 1f38db2d7c..9370184957 100644
--- a/tasks/robocasa-gr1-cuttingboard-to-tieredbasket-i00-standard/task.toml
+++ b/tasks/robocasa-gr1-cuttingboard-to-tieredbasket-i00-standard/task.toml
@@ -24,6 +24,7 @@ simulator = "robosuite"
robot = "fourier-gr1"
family = "robocasa-gr1/PosttrainPnPNovelFromCuttingboardToTieredbasketSplitA"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
expert_time_estimate_hours = 4.0
diff --git a/tasks/robocasa-gr1-milk-to-microwave-close-i00-privileged/task.toml b/tasks/robocasa-gr1-milk-to-microwave-close-i00-privileged/task.toml
index 0d9447494f..e8b1d95048 100644
--- a/tasks/robocasa-gr1-milk-to-microwave-close-i00-privileged/task.toml
+++ b/tasks/robocasa-gr1-milk-to-microwave-close-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "robosuite"
robot = "fourier-gr1"
family = "robocasa-gr1/PnPMilkToMicrowaveClose"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robocasa-gr1-milk-to-microwave-close-i00-standard/README.md b/tasks/robocasa-gr1-milk-to-microwave-close-i00-standard/README.md
index 924ba63961..168d96661f 100644
--- a/tasks/robocasa-gr1-milk-to-microwave-close-i00-standard/README.md
+++ b/tasks/robocasa-gr1-milk-to-microwave-close-i00-standard/README.md
@@ -48,3 +48,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
- 2026-10-10 (review round 3, board 0067): the sim service's request handling (`environment/sim/server_casa.py`) and `tests/grade.py`. A negative or non-numeric `Content-Length` is answered 400 without reading the body (the service read until the client hung up, without a bound, which could exhaust its memory); a camera named more than once in `cameras` is rendered and returned once (N copies of one name made it build N x N images within the frame budget); a port that is still held when the service restarts is waited for, not counted as a crash; a JSON `true` or `false` where `dt` or `seconds` takes a number is a 400 (`true` ran as 1 second). The grader: a `final.json` whose evidence fields (`control_steps`, `calls`, `elapsed_s`, `final_reward`) are not numbers no longer stops it before it writes a reward (they are reported as 0 and decide nothing). An episode's physics and the pass rule are unchanged.
- 2026-10-13 (owner, boards 0085 and 0087): the sim service's clock (`environment/sim/server_casa.py`, the docstring of `environment/rcb_client.py`). `meta["time_s"]` and `meta["time_left_s"]` count from the agent's first request to the service (any endpoint), and a restarted service keeps that moment; before, they counted from the first start of the service's container (`time_left_s`) and from when its scene was ready (`time_s`). Where the service is ready before the agent's phase begins (the harness is installed in between: a median of 79 s over the recorded standard trials), that clock ran ahead of the agent's and `time_left_s` reached 0 while the agent still had that time. `/spec` says what the clock counts from (`budget.clock`), and the instruction's `## Time limit` tells the agent to plan by `time_left_s`. The service never ended an episode on this clock: the scene, the frozen instance, the pass rule and the grader are unchanged, so the task keeps its name and earlier results stand.
- 2026-10-13 (board 0093; found in board 0086): the camera calibration (`environment/sim/server_casa.py`). `cam2world_gl` held the pose of MuJoCo's left-eye camera (`scene.camera[0]`), 3.4 cm (half the model's 0.068 m interpupillary distance) to the camera's left of the viewpoint the mono RGB and depth images are rendered from; it now is that viewpoint (`data.cam_xpos` / `data.cam_xmat`). Checked against MuJoCo's own renders of marker spheres at known world points, every served camera, before and after the robot moved: the reprojection error was 10.5-10.7 px median (up to 28 px for near points, always to the same side), now 0.15 px median and at most 0.4 px. The intrinsics were right (every camera here is given by its `fovy`, none by a sensor size) and are unchanged, as are the images, the scene, the frozen instance, the pass rule and the grader, so the task keeps its name. **Standard-mode results from before this date used the offset calibration**: a point located from the images and the reported calibration came out about 3.4 cm to the side (agents missed a microwave button by 3.5 cm), so they are not strictly comparable with later ones.
+- 2026-10-14 (board 0113, from board 0086): the instruction describes the hands: each hand command is the position target (rad) of its finger's joints, 0 open and larger closing, within each joint's range; the grip site is the origin of the hand's mounting frame at the wrist end of the hand; the pink spheres beside each hand are visual markers of the robot model, not objects (board 0086: most standard runs took them for objects). Text only: the scene, the frozen instance, the reference solution, the success check and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/robocasa-gr1-milk-to-microwave-close-i00-standard/instruction.md b/tasks/robocasa-gr1-milk-to-microwave-close-i00-standard/instruction.md
index 4c1f1970e9..7f386006a7 100644
--- a/tasks/robocasa-gr1-milk-to-microwave-close-i00-standard/instruction.md
+++ b/tasks/robocasa-gr1-milk-to-microwave-close-i00-standard/instruction.md
@@ -51,6 +51,13 @@ res.video[cam]; res.frames[cam]; res.last[cam]; res.proprio; res.save("/app/runs
channel is an absolute target, so the grip sites stay within about a millimetre, but the IK's posture term can still
move the waist and elbows by a few hundredths of a radian while you watch.
- The cameras are mounted on the robot and move with it: use the calibration that comes with each response.
+- The hands: each hand command is the position target, in radians, of that finger's joints (`thumb_flexion` drives
+ two thumb joints, `thumb_rotation` one): 0 is the straight, open finger, larger values bend it toward the palm
+ (the hand closes), and every finger joint stops at the end of its range (1.22 to 1.74 rad), so values below 0 or
+ above about 1.7 change nothing more. The grip site that the pose targets place is the origin of the hand's mounting
+ frame (`gripper0_right_grip_site`, `gripper0_left_grip_site`), at the wrist end of the hand, not a point between
+ the fingertips. The four small pink spheres beside each hand are part of the robot model (visual markers
+ `robot0_<side>_pinch_spheres_0..3`, about 13.5 cm from its grip site, with no collision): they are not objects.
- World frame: metres, z up. In the proprio, `eef_left_*` / `eef_right_*` are the grip sites' positions in the world
frame; the action's pose targets and the `cmd_*` entries (the last command) are in the robot's base frame.
The robot's base frame (body `robot0_base`, fixed in the world) has its origin at world (2.65000, -2.17500, 0.97000) m and is turned +90.0 deg about z (quaternion w x y z (0.70711, 0, 0, 0.70711)): `p_world = R @ p_base + t` with R = [[0, -1, 0], [1, 0, 0], [0, 0, 1]].
diff --git a/tasks/robocasa-gr1-milk-to-microwave-close-i00-standard/task.toml b/tasks/robocasa-gr1-milk-to-microwave-close-i00-standard/task.toml
index 3115718054..937d547779 100644
--- a/tasks/robocasa-gr1-milk-to-microwave-close-i00-standard/task.toml
+++ b/tasks/robocasa-gr1-milk-to-microwave-close-i00-standard/task.toml
@@ -24,6 +24,7 @@ simulator = "robosuite"
robot = "fourier-gr1"
family = "robocasa-gr1/PnPMilkToMicrowaveClose"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
expert_time_estimate_hours = 4.0
diff --git a/tasks/robocasa-gr1-placemat-to-basket-i00-privileged/task.toml b/tasks/robocasa-gr1-placemat-to-basket-i00-privileged/task.toml
index 1c88a4903c..f0130863e5 100644
--- a/tasks/robocasa-gr1-placemat-to-basket-i00-privileged/task.toml
+++ b/tasks/robocasa-gr1-placemat-to-basket-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "robosuite"
robot = "fourier-gr1"
family = "robocasa-gr1/PosttrainPnPNovelFromPlacematToBasketSplitA"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robocasa-gr1-placemat-to-basket-i00-standard/README.md b/tasks/robocasa-gr1-placemat-to-basket-i00-standard/README.md
index 0ce0128438..15f8b8bf51 100644
--- a/tasks/robocasa-gr1-placemat-to-basket-i00-standard/README.md
+++ b/tasks/robocasa-gr1-placemat-to-basket-i00-standard/README.md
@@ -48,3 +48,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
- 2026-10-10 (review round 3, board 0067): the sim service's request handling (`environment/sim/server_casa.py`) and `tests/grade.py`. A negative or non-numeric `Content-Length` is answered 400 without reading the body (the service read until the client hung up, without a bound, which could exhaust its memory); a camera named more than once in `cameras` is rendered and returned once (N copies of one name made it build N x N images within the frame budget); a port that is still held when the service restarts is waited for, not counted as a crash; a JSON `true` or `false` where `dt` or `seconds` takes a number is a 400 (`true` ran as 1 second). The grader: a `final.json` whose evidence fields (`control_steps`, `calls`, `elapsed_s`, `final_reward`) are not numbers no longer stops it before it writes a reward (they are reported as 0 and decide nothing). An episode's physics and the pass rule are unchanged.
- 2026-10-13 (owner, boards 0085 and 0087): the sim service's clock (`environment/sim/server_casa.py`, the docstring of `environment/rcb_client.py`). `meta["time_s"]` and `meta["time_left_s"]` count from the agent's first request to the service (any endpoint), and a restarted service keeps that moment; before, they counted from the first start of the service's container (`time_left_s`) and from when its scene was ready (`time_s`). Where the service is ready before the agent's phase begins (the harness is installed in between: a median of 79 s over the recorded standard trials), that clock ran ahead of the agent's and `time_left_s` reached 0 while the agent still had that time. `/spec` says what the clock counts from (`budget.clock`), and the instruction's `## Time limit` tells the agent to plan by `time_left_s`. The service never ended an episode on this clock: the scene, the frozen instance, the pass rule and the grader are unchanged, so the task keeps its name and earlier results stand.
- 2026-10-13 (board 0093; found in board 0086): the camera calibration (`environment/sim/server_casa.py`). `cam2world_gl` held the pose of MuJoCo's left-eye camera (`scene.camera[0]`), 3.4 cm (half the model's 0.068 m interpupillary distance) to the camera's left of the viewpoint the mono RGB and depth images are rendered from; it now is that viewpoint (`data.cam_xpos` / `data.cam_xmat`). Checked against MuJoCo's own renders of marker spheres at known world points, every served camera, before and after the robot moved: the reprojection error was 10.5-10.7 px median (up to 28 px for near points, always to the same side), now 0.15 px median and at most 0.4 px. The intrinsics were right (every camera here is given by its `fovy`, none by a sensor size) and are unchanged, as are the images, the scene, the frozen instance, the pass rule and the grader, so the task keeps its name. **Standard-mode results from before this date used the offset calibration**: a point located from the images and the reported calibration came out about 3.4 cm to the side (agents missed a microwave button by 3.5 cm), so they are not strictly comparable with later ones.
+- 2026-10-14 (board 0113, from board 0086): the instruction describes the hands: each hand command is the position target (rad) of its finger's joints, 0 open and larger closing, within each joint's range; the grip site is the origin of the hand's mounting frame at the wrist end of the hand; the pink spheres beside each hand are visual markers of the robot model, not objects (board 0086: most standard runs took them for objects). Text only: the scene, the frozen instance, the reference solution, the success check and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/robocasa-gr1-placemat-to-basket-i00-standard/instruction.md b/tasks/robocasa-gr1-placemat-to-basket-i00-standard/instruction.md
index 7fdf392a2e..50bddc3ec2 100644
--- a/tasks/robocasa-gr1-placemat-to-basket-i00-standard/instruction.md
+++ b/tasks/robocasa-gr1-placemat-to-basket-i00-standard/instruction.md
@@ -51,6 +51,13 @@ res.video[cam]; res.frames[cam]; res.last[cam]; res.proprio; res.save("/app/runs
channel is an absolute target, so the grip sites stay within about a millimetre, but the IK's posture term can still
move the waist and elbows by a few hundredths of a radian while you watch.
- The cameras are mounted on the robot and move with it: use the calibration that comes with each response.
+- The hands: each hand command is the position target, in radians, of that finger's joints (`thumb_flexion` drives
+ two thumb joints, `thumb_rotation` one): 0 is the straight, open finger, larger values bend it toward the palm
+ (the hand closes), and every finger joint stops at the end of its range (1.22 to 1.74 rad), so values below 0 or
+ above about 1.7 change nothing more. The grip site that the pose targets place is the origin of the hand's mounting
+ frame (`gripper0_right_grip_site`, `gripper0_left_grip_site`), at the wrist end of the hand, not a point between
+ the fingertips. The four small pink spheres beside each hand are part of the robot model (visual markers
+ `robot0_<side>_pinch_spheres_0..3`, about 13.5 cm from its grip site, with no collision): they are not objects.
- World frame: metres, z up. In the proprio, `eef_left_*` / `eef_right_*` are the grip sites' positions in the world
frame; the action's pose targets and the `cmd_*` entries (the last command) are in the robot's base frame.
The robot's base frame (body `robot0_base`, fixed in the world) has its origin at world (2.66969, -2.13364, 0.89794) m and is turned +90.0 deg about z (quaternion w x y z (0.70711, 0, 0, 0.70711)): `p_world = R @ p_base + t` with R = [[0, -1, 0], [1, 0, 0], [0, 0, 1]].
diff --git a/tasks/robocasa-gr1-placemat-to-basket-i00-standard/task.toml b/tasks/robocasa-gr1-placemat-to-basket-i00-standard/task.toml
index c22f7e2bb3..fed58b1e18 100644
--- a/tasks/robocasa-gr1-placemat-to-basket-i00-standard/task.toml
+++ b/tasks/robocasa-gr1-placemat-to-basket-i00-standard/task.toml
@@ -24,6 +24,7 @@ simulator = "robosuite"
robot = "fourier-gr1"
family = "robocasa-gr1/PosttrainPnPNovelFromPlacematToBasketSplitA"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
expert_time_estimate_hours = 4.0
diff --git a/tasks/robocasa-gr1-placemat-to-bowl-i00-privileged/task.toml b/tasks/robocasa-gr1-placemat-to-bowl-i00-privileged/task.toml
index c539294e16..f8eff6d71a 100644
--- a/tasks/robocasa-gr1-placemat-to-bowl-i00-privileged/task.toml
+++ b/tasks/robocasa-gr1-placemat-to-bowl-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "robosuite"
robot = "fourier-gr1"
family = "robocasa-gr1/PosttrainPnPNovelFromPlacematToBowlSplitA"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robocasa-gr1-placemat-to-bowl-i00-standard/README.md b/tasks/robocasa-gr1-placemat-to-bowl-i00-standard/README.md
index 2557c9380b..9e2fff511c 100644
--- a/tasks/robocasa-gr1-placemat-to-bowl-i00-standard/README.md
+++ b/tasks/robocasa-gr1-placemat-to-bowl-i00-standard/README.md
@@ -48,3 +48,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
- 2026-10-10 (review round 3, board 0067): the sim service's request handling (`environment/sim/server_casa.py`) and `tests/grade.py`. A negative or non-numeric `Content-Length` is answered 400 without reading the body (the service read until the client hung up, without a bound, which could exhaust its memory); a camera named more than once in `cameras` is rendered and returned once (N copies of one name made it build N x N images within the frame budget); a port that is still held when the service restarts is waited for, not counted as a crash; a JSON `true` or `false` where `dt` or `seconds` takes a number is a 400 (`true` ran as 1 second). The grader: a `final.json` whose evidence fields (`control_steps`, `calls`, `elapsed_s`, `final_reward`) are not numbers no longer stops it before it writes a reward (they are reported as 0 and decide nothing). An episode's physics and the pass rule are unchanged.
- 2026-10-13 (owner, boards 0085 and 0087): the sim service's clock (`environment/sim/server_casa.py`, the docstring of `environment/rcb_client.py`). `meta["time_s"]` and `meta["time_left_s"]` count from the agent's first request to the service (any endpoint), and a restarted service keeps that moment; before, they counted from the first start of the service's container (`time_left_s`) and from when its scene was ready (`time_s`). Where the service is ready before the agent's phase begins (the harness is installed in between: a median of 79 s over the recorded standard trials), that clock ran ahead of the agent's and `time_left_s` reached 0 while the agent still had that time. `/spec` says what the clock counts from (`budget.clock`), and the instruction's `## Time limit` tells the agent to plan by `time_left_s`. The service never ended an episode on this clock: the scene, the frozen instance, the pass rule and the grader are unchanged, so the task keeps its name and earlier results stand.
- 2026-10-13 (board 0093; found in board 0086): the camera calibration (`environment/sim/server_casa.py`). `cam2world_gl` held the pose of MuJoCo's left-eye camera (`scene.camera[0]`), 3.4 cm (half the model's 0.068 m interpupillary distance) to the camera's left of the viewpoint the mono RGB and depth images are rendered from; it now is that viewpoint (`data.cam_xpos` / `data.cam_xmat`). Checked against MuJoCo's own renders of marker spheres at known world points, every served camera, before and after the robot moved: the reprojection error was 10.5-10.7 px median (up to 28 px for near points, always to the same side), now 0.15 px median and at most 0.4 px. The intrinsics were right (every camera here is given by its `fovy`, none by a sensor size) and are unchanged, as are the images, the scene, the frozen instance, the pass rule and the grader, so the task keeps its name. **Standard-mode results from before this date used the offset calibration**: a point located from the images and the reported calibration came out about 3.4 cm to the side (agents missed a microwave button by 3.5 cm), so they are not strictly comparable with later ones.
+- 2026-10-14 (board 0113, from board 0086): the instruction describes the hands: each hand command is the position target (rad) of its finger's joints, 0 open and larger closing, within each joint's range; the grip site is the origin of the hand's mounting frame at the wrist end of the hand; the pink spheres beside each hand are visual markers of the robot model, not objects (board 0086: most standard runs took them for objects). Text only: the scene, the frozen instance, the reference solution, the success check and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/robocasa-gr1-placemat-to-bowl-i00-standard/instruction.md b/tasks/robocasa-gr1-placemat-to-bowl-i00-standard/instruction.md
index ade2b91b8d..3b271112ea 100644
--- a/tasks/robocasa-gr1-placemat-to-bowl-i00-standard/instruction.md
+++ b/tasks/robocasa-gr1-placemat-to-bowl-i00-standard/instruction.md
@@ -52,6 +52,13 @@ res.video[cam]; res.frames[cam]; res.last[cam]; res.proprio; res.save("/app/runs
channel is an absolute target, so the grip sites stay within about a millimetre, but the IK's posture term can still
move the waist and elbows by a few hundredths of a radian while you watch.
- The cameras are mounted on the robot and move with it: use the calibration that comes with each response.
+- The hands: each hand command is the position target, in radians, of that finger's joints (`thumb_flexion` drives
+ two thumb joints, `thumb_rotation` one): 0 is the straight, open finger, larger values bend it toward the palm
+ (the hand closes), and every finger joint stops at the end of its range (1.22 to 1.74 rad), so values below 0 or
+ above about 1.7 change nothing more. The grip site that the pose targets place is the origin of the hand's mounting
+ frame (`gripper0_right_grip_site`, `gripper0_left_grip_site`), at the wrist end of the hand, not a point between
+ the fingertips. The four small pink spheres beside each hand are part of the robot model (visual markers
+ `robot0_<side>_pinch_spheres_0..3`, about 13.5 cm from its grip site, with no collision): they are not objects.
- World frame: metres, z up. In the proprio, `eef_left_*` / `eef_right_*` are the grip sites' positions in the world
frame; the action's pose targets and the `cmd_*` entries (the last command) are in the robot's base frame.
The robot's base frame (body `robot0_base`, fixed in the world) has its origin at world (2.65064, -2.13092, 0.91581) m and is turned +90.0 deg about z (quaternion w x y z (0.70711, 0, 0, 0.70711)): `p_world = R @ p_base + t` with R = [[0, -1, 0], [1, 0, 0], [0, 0, 1]].
diff --git a/tasks/robocasa-gr1-placemat-to-bowl-i00-standard/task.toml b/tasks/robocasa-gr1-placemat-to-bowl-i00-standard/task.toml
index 0cc2a62eaf..5be297912a 100644
--- a/tasks/robocasa-gr1-placemat-to-bowl-i00-standard/task.toml
+++ b/tasks/robocasa-gr1-placemat-to-bowl-i00-standard/task.toml
@@ -24,6 +24,7 @@ simulator = "robosuite"
robot = "fourier-gr1"
family = "robocasa-gr1/PosttrainPnPNovelFromPlacematToBowlSplitA"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
expert_time_estimate_hours = 4.0
diff --git a/tasks/robocasa-gr1-placemat-to-plate-i00-privileged/task.toml b/tasks/robocasa-gr1-placemat-to-plate-i00-privileged/task.toml
index 52c4d2734c..51185da22e 100644
--- a/tasks/robocasa-gr1-placemat-to-plate-i00-privileged/task.toml
+++ b/tasks/robocasa-gr1-placemat-to-plate-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "robosuite"
robot = "fourier-gr1"
family = "robocasa-gr1/PosttrainPnPNovelFromPlacematToPlateSplitA"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robocasa-gr1-placemat-to-plate-i00-standard/README.md b/tasks/robocasa-gr1-placemat-to-plate-i00-standard/README.md
index a85e6be697..fac1257fcd 100644
--- a/tasks/robocasa-gr1-placemat-to-plate-i00-standard/README.md
+++ b/tasks/robocasa-gr1-placemat-to-plate-i00-standard/README.md
@@ -48,3 +48,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
- 2026-10-10 (review round 3, board 0067): the sim service's request handling (`environment/sim/server_casa.py`) and `tests/grade.py`. A negative or non-numeric `Content-Length` is answered 400 without reading the body (the service read until the client hung up, without a bound, which could exhaust its memory); a camera named more than once in `cameras` is rendered and returned once (N copies of one name made it build N x N images within the frame budget); a port that is still held when the service restarts is waited for, not counted as a crash; a JSON `true` or `false` where `dt` or `seconds` takes a number is a 400 (`true` ran as 1 second). The grader: a `final.json` whose evidence fields (`control_steps`, `calls`, `elapsed_s`, `final_reward`) are not numbers no longer stops it before it writes a reward (they are reported as 0 and decide nothing). An episode's physics and the pass rule are unchanged.
- 2026-10-13 (owner, boards 0085 and 0087): the sim service's clock (`environment/sim/server_casa.py`, the docstring of `environment/rcb_client.py`). `meta["time_s"]` and `meta["time_left_s"]` count from the agent's first request to the service (any endpoint), and a restarted service keeps that moment; before, they counted from the first start of the service's container (`time_left_s`) and from when its scene was ready (`time_s`). Where the service is ready before the agent's phase begins (the harness is installed in between: a median of 79 s over the recorded standard trials), that clock ran ahead of the agent's and `time_left_s` reached 0 while the agent still had that time. `/spec` says what the clock counts from (`budget.clock`), and the instruction's `## Time limit` tells the agent to plan by `time_left_s`. The service never ended an episode on this clock: the scene, the frozen instance, the pass rule and the grader are unchanged, so the task keeps its name and earlier results stand.
- 2026-10-13 (board 0093; found in board 0086): the camera calibration (`environment/sim/server_casa.py`). `cam2world_gl` held the pose of MuJoCo's left-eye camera (`scene.camera[0]`), 3.4 cm (half the model's 0.068 m interpupillary distance) to the camera's left of the viewpoint the mono RGB and depth images are rendered from; it now is that viewpoint (`data.cam_xpos` / `data.cam_xmat`). Checked against MuJoCo's own renders of marker spheres at known world points, every served camera, before and after the robot moved: the reprojection error was 10.5-10.7 px median (up to 28 px for near points, always to the same side), now 0.15 px median and at most 0.4 px. The intrinsics were right (every camera here is given by its `fovy`, none by a sensor size) and are unchanged, as are the images, the scene, the frozen instance, the pass rule and the grader, so the task keeps its name. **Standard-mode results from before this date used the offset calibration**: a point located from the images and the reported calibration came out about 3.4 cm to the side (agents missed a microwave button by 3.5 cm), so they are not strictly comparable with later ones.
+- 2026-10-14 (board 0113, from board 0086): the instruction describes the hands: each hand command is the position target (rad) of its finger's joints, 0 open and larger closing, within each joint's range; the grip site is the origin of the hand's mounting frame at the wrist end of the hand; the pink spheres beside each hand are visual markers of the robot model, not objects (board 0086: most standard runs took them for objects). Text only: the scene, the frozen instance, the reference solution, the success check and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/robocasa-gr1-placemat-to-plate-i00-standard/instruction.md b/tasks/robocasa-gr1-placemat-to-plate-i00-standard/instruction.md
index 77f4f11bdb..8ca7345caa 100644
--- a/tasks/robocasa-gr1-placemat-to-plate-i00-standard/instruction.md
+++ b/tasks/robocasa-gr1-placemat-to-plate-i00-standard/instruction.md
@@ -52,6 +52,13 @@ res.video[cam]; res.frames[cam]; res.last[cam]; res.proprio; res.save("/app/runs
channel is an absolute target, so the grip sites stay within about a millimetre, but the IK's posture term can still
move the waist and elbows by a few hundredths of a radian while you watch.
- The cameras are mounted on the robot and move with it: use the calibration that comes with each response.
+- The hands: each hand command is the position target, in radians, of that finger's joints (`thumb_flexion` drives
+ two thumb joints, `thumb_rotation` one): 0 is the straight, open finger, larger values bend it toward the palm
+ (the hand closes), and every finger joint stops at the end of its range (1.22 to 1.74 rad), so values below 0 or
+ above about 1.7 change nothing more. The grip site that the pose targets place is the origin of the hand's mounting
+ frame (`gripper0_right_grip_site`, `gripper0_left_grip_site`), at the wrist end of the hand, not a point between
+ the fingertips. The four small pink spheres beside each hand are part of the robot model (visual markers
+ `robot0_<side>_pinch_spheres_0..3`, about 13.5 cm from its grip site, with no collision): they are not objects.
- World frame: metres, z up. In the proprio, `eef_left_*` / `eef_right_*` are the grip sites' positions in the world
frame; the action's pose targets and the `cmd_*` entries (the last command) are in the robot's base frame.
The robot's base frame (body `robot0_base`, fixed in the world) has its origin at world (2.65289, -2.16654, 0.90010) m and is turned +90.0 deg about z (quaternion w x y z (0.70711, 0, 0, 0.70711)): `p_world = R @ p_base + t` with R = [[0, -1, 0], [1, 0, 0], [0, 0, 1]].
diff --git a/tasks/robocasa-gr1-placemat-to-plate-i00-standard/task.toml b/tasks/robocasa-gr1-placemat-to-plate-i00-standard/task.toml
index ad08508bb9..a5cabb639e 100644
--- a/tasks/robocasa-gr1-placemat-to-plate-i00-standard/task.toml
+++ b/tasks/robocasa-gr1-placemat-to-plate-i00-standard/task.toml
@@ -24,6 +24,7 @@ simulator = "robosuite"
robot = "fourier-gr1"
family = "robocasa-gr1/PosttrainPnPNovelFromPlacematToPlateSplitA"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
expert_time_estimate_hours = 4.0
diff --git a/tasks/robocasa-gr1-placemat-to-tieredshelf-i00-privileged/task.toml b/tasks/robocasa-gr1-placemat-to-tieredshelf-i00-privileged/task.toml
index 500b4d8b08..74a4d7f3fe 100644
--- a/tasks/robocasa-gr1-placemat-to-tieredshelf-i00-privileged/task.toml
+++ b/tasks/robocasa-gr1-placemat-to-tieredshelf-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "robosuite"
robot = "fourier-gr1"
family = "robocasa-gr1/PosttrainPnPNovelFromPlacematToTieredshelfSplitA"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robocasa-gr1-placemat-to-tieredshelf-i00-standard/README.md b/tasks/robocasa-gr1-placemat-to-tieredshelf-i00-standard/README.md
index ad032e2aa9..1dfa801da2 100644
--- a/tasks/robocasa-gr1-placemat-to-tieredshelf-i00-standard/README.md
+++ b/tasks/robocasa-gr1-placemat-to-tieredshelf-i00-standard/README.md
@@ -48,3 +48,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
- 2026-10-10 (review round 3, board 0067): the sim service's request handling (`environment/sim/server_casa.py`) and `tests/grade.py`. A negative or non-numeric `Content-Length` is answered 400 without reading the body (the service read until the client hung up, without a bound, which could exhaust its memory); a camera named more than once in `cameras` is rendered and returned once (N copies of one name made it build N x N images within the frame budget); a port that is still held when the service restarts is waited for, not counted as a crash; a JSON `true` or `false` where `dt` or `seconds` takes a number is a 400 (`true` ran as 1 second). The grader: a `final.json` whose evidence fields (`control_steps`, `calls`, `elapsed_s`, `final_reward`) are not numbers no longer stops it before it writes a reward (they are reported as 0 and decide nothing). An episode's physics and the pass rule are unchanged.
- 2026-10-13 (owner, boards 0085 and 0087): the sim service's clock (`environment/sim/server_casa.py`, the docstring of `environment/rcb_client.py`). `meta["time_s"]` and `meta["time_left_s"]` count from the agent's first request to the service (any endpoint), and a restarted service keeps that moment; before, they counted from the first start of the service's container (`time_left_s`) and from when its scene was ready (`time_s`). Where the service is ready before the agent's phase begins (the harness is installed in between: a median of 79 s over the recorded standard trials), that clock ran ahead of the agent's and `time_left_s` reached 0 while the agent still had that time. `/spec` says what the clock counts from (`budget.clock`), and the instruction's `## Time limit` tells the agent to plan by `time_left_s`. The service never ended an episode on this clock: the scene, the frozen instance, the pass rule and the grader are unchanged, so the task keeps its name and earlier results stand.
- 2026-10-13 (board 0093; found in board 0086): the camera calibration (`environment/sim/server_casa.py`). `cam2world_gl` held the pose of MuJoCo's left-eye camera (`scene.camera[0]`), 3.4 cm (half the model's 0.068 m interpupillary distance) to the camera's left of the viewpoint the mono RGB and depth images are rendered from; it now is that viewpoint (`data.cam_xpos` / `data.cam_xmat`). Checked against MuJoCo's own renders of marker spheres at known world points, every served camera, before and after the robot moved: the reprojection error was 10.5-10.7 px median (up to 28 px for near points, always to the same side), now 0.15 px median and at most 0.4 px. The intrinsics were right (every camera here is given by its `fovy`, none by a sensor size) and are unchanged, as are the images, the scene, the frozen instance, the pass rule and the grader, so the task keeps its name. **Standard-mode results from before this date used the offset calibration**: a point located from the images and the reported calibration came out about 3.4 cm to the side (agents missed a microwave button by 3.5 cm), so they are not strictly comparable with later ones.
+- 2026-10-14 (board 0113, from board 0086): the instruction describes the hands: each hand command is the position target (rad) of its finger's joints, 0 open and larger closing, within each joint's range; the grip site is the origin of the hand's mounting frame at the wrist end of the hand; the pink spheres beside each hand are visual markers of the robot model, not objects (board 0086: most standard runs took them for objects). Text only: the scene, the frozen instance, the reference solution, the success check and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/robocasa-gr1-placemat-to-tieredshelf-i00-standard/instruction.md b/tasks/robocasa-gr1-placemat-to-tieredshelf-i00-standard/instruction.md
index b3010be484..50f137d76f 100644
--- a/tasks/robocasa-gr1-placemat-to-tieredshelf-i00-standard/instruction.md
+++ b/tasks/robocasa-gr1-placemat-to-tieredshelf-i00-standard/instruction.md
@@ -52,6 +52,13 @@ res.video[cam]; res.frames[cam]; res.last[cam]; res.proprio; res.save("/app/runs
channel is an absolute target, so the grip sites stay within about a millimetre, but the IK's posture term can still
move the waist and elbows by a few hundredths of a radian while you watch.
- The cameras are mounted on the robot and move with it: use the calibration that comes with each response.
+- The hands: each hand command is the position target, in radians, of that finger's joints (`thumb_flexion` drives
+ two thumb joints, `thumb_rotation` one): 0 is the straight, open finger, larger values bend it toward the palm
+ (the hand closes), and every finger joint stops at the end of its range (1.22 to 1.74 rad), so values below 0 or
+ above about 1.7 change nothing more. The grip site that the pose targets place is the origin of the hand's mounting
+ frame (`gripper0_right_grip_site`, `gripper0_left_grip_site`), at the wrist end of the hand, not a point between
+ the fingertips. The four small pink spheres beside each hand are part of the robot model (visual markers
+ `robot0_<side>_pinch_spheres_0..3`, about 13.5 cm from its grip site, with no collision): they are not objects.
- World frame: metres, z up. In the proprio, `eef_left_*` / `eef_right_*` are the grip sites' positions in the world
frame; the action's pose targets and the `cmd_*` entries (the last command) are in the robot's base frame.
The robot's base frame (body `robot0_base`, fixed in the world) has its origin at world (2.66102, -2.15373, 0.90716) m and is turned +90.0 deg about z (quaternion w x y z (0.70711, 0, 0, 0.70711)): `p_world = R @ p_base + t` with R = [[0, -1, 0], [1, 0, 0], [0, 0, 1]].
diff --git a/tasks/robocasa-gr1-placemat-to-tieredshelf-i00-standard/task.toml b/tasks/robocasa-gr1-placemat-to-tieredshelf-i00-standard/task.toml
index 20773825ba..7d4e5b9872 100644
--- a/tasks/robocasa-gr1-placemat-to-tieredshelf-i00-standard/task.toml
+++ b/tasks/robocasa-gr1-placemat-to-tieredshelf-i00-standard/task.toml
@@ -24,6 +24,7 @@ simulator = "robosuite"
robot = "fourier-gr1"
family = "robocasa-gr1/PosttrainPnPNovelFromPlacematToTieredshelfSplitA"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
expert_time_estimate_hours = 4.0
diff --git a/tasks/robocasa-gr1-plate-to-bowl-i00-privileged/task.toml b/tasks/robocasa-gr1-plate-to-bowl-i00-privileged/task.toml
index 0d4c69abed..285fd0e35b 100644
--- a/tasks/robocasa-gr1-plate-to-bowl-i00-privileged/task.toml
+++ b/tasks/robocasa-gr1-plate-to-bowl-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "robosuite"
robot = "fourier-gr1"
family = "robocasa-gr1/PosttrainPnPNovelFromPlateToBowlSplitA"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robocasa-gr1-plate-to-bowl-i00-standard/README.md b/tasks/robocasa-gr1-plate-to-bowl-i00-standard/README.md
index 636f8028c1..cd73ca6b96 100644
--- a/tasks/robocasa-gr1-plate-to-bowl-i00-standard/README.md
+++ b/tasks/robocasa-gr1-plate-to-bowl-i00-standard/README.md
@@ -48,3 +48,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
- 2026-10-10 (review round 3, board 0067): the sim service's request handling (`environment/sim/server_casa.py`) and `tests/grade.py`. A negative or non-numeric `Content-Length` is answered 400 without reading the body (the service read until the client hung up, without a bound, which could exhaust its memory); a camera named more than once in `cameras` is rendered and returned once (N copies of one name made it build N x N images within the frame budget); a port that is still held when the service restarts is waited for, not counted as a crash; a JSON `true` or `false` where `dt` or `seconds` takes a number is a 400 (`true` ran as 1 second). The grader: a `final.json` whose evidence fields (`control_steps`, `calls`, `elapsed_s`, `final_reward`) are not numbers no longer stops it before it writes a reward (they are reported as 0 and decide nothing). An episode's physics and the pass rule are unchanged.
- 2026-10-13 (owner, boards 0085 and 0087): the sim service's clock (`environment/sim/server_casa.py`, the docstring of `environment/rcb_client.py`). `meta["time_s"]` and `meta["time_left_s"]` count from the agent's first request to the service (any endpoint), and a restarted service keeps that moment; before, they counted from the first start of the service's container (`time_left_s`) and from when its scene was ready (`time_s`). Where the service is ready before the agent's phase begins (the harness is installed in between: a median of 79 s over the recorded standard trials), that clock ran ahead of the agent's and `time_left_s` reached 0 while the agent still had that time. `/spec` says what the clock counts from (`budget.clock`), and the instruction's `## Time limit` tells the agent to plan by `time_left_s`. The service never ended an episode on this clock: the scene, the frozen instance, the pass rule and the grader are unchanged, so the task keeps its name and earlier results stand.
- 2026-10-13 (board 0093; found in board 0086): the camera calibration (`environment/sim/server_casa.py`). `cam2world_gl` held the pose of MuJoCo's left-eye camera (`scene.camera[0]`), 3.4 cm (half the model's 0.068 m interpupillary distance) to the camera's left of the viewpoint the mono RGB and depth images are rendered from; it now is that viewpoint (`data.cam_xpos` / `data.cam_xmat`). Checked against MuJoCo's own renders of marker spheres at known world points, every served camera, before and after the robot moved: the reprojection error was 10.5-10.7 px median (up to 28 px for near points, always to the same side), now 0.15 px median and at most 0.4 px. The intrinsics were right (every camera here is given by its `fovy`, none by a sensor size) and are unchanged, as are the images, the scene, the frozen instance, the pass rule and the grader, so the task keeps its name. **Standard-mode results from before this date used the offset calibration**: a point located from the images and the reported calibration came out about 3.4 cm to the side (agents missed a microwave button by 3.5 cm), so they are not strictly comparable with later ones.
+- 2026-10-14 (board 0113, from board 0086): the instruction describes the hands: each hand command is the position target (rad) of its finger's joints, 0 open and larger closing, within each joint's range; the grip site is the origin of the hand's mounting frame at the wrist end of the hand; the pink spheres beside each hand are visual markers of the robot model, not objects (board 0086: most standard runs took them for objects). Text only: the scene, the frozen instance, the reference solution, the success check and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/robocasa-gr1-plate-to-bowl-i00-standard/instruction.md b/tasks/robocasa-gr1-plate-to-bowl-i00-standard/instruction.md
index 55a9125895..b69029e1fd 100644
--- a/tasks/robocasa-gr1-plate-to-bowl-i00-standard/instruction.md
+++ b/tasks/robocasa-gr1-plate-to-bowl-i00-standard/instruction.md
@@ -52,6 +52,13 @@ res.video[cam]; res.frames[cam]; res.last[cam]; res.proprio; res.save("/app/runs
channel is an absolute target, so the grip sites stay within about a millimetre, but the IK's posture term can still
move the waist and elbows by a few hundredths of a radian while you watch.
- The cameras are mounted on the robot and move with it: use the calibration that comes with each response.
+- The hands: each hand command is the position target, in radians, of that finger's joints (`thumb_flexion` drives
+ two thumb joints, `thumb_rotation` one): 0 is the straight, open finger, larger values bend it toward the palm
+ (the hand closes), and every finger joint stops at the end of its range (1.22 to 1.74 rad), so values below 0 or
+ above about 1.7 change nothing more. The grip site that the pose targets place is the origin of the hand's mounting
+ frame (`gripper0_right_grip_site`, `gripper0_left_grip_site`), at the wrist end of the hand, not a point between
+ the fingertips. The four small pink spheres beside each hand are part of the robot model (visual markers
+ `robot0_<side>_pinch_spheres_0..3`, about 13.5 cm from its grip site, with no collision): they are not objects.
- World frame: metres, z up. In the proprio, `eef_left_*` / `eef_right_*` are the grip sites' positions in the world
frame; the action's pose targets and the `cmd_*` entries (the last command) are in the robot's base frame.
The robot's base frame (body `robot0_base`, fixed in the world) has its origin at world (2.66261, -2.15254, 0.88087) m and is turned +90.0 deg about z (quaternion w x y z (0.70711, 0, 0, 0.70711)): `p_world = R @ p_base + t` with R = [[0, -1, 0], [1, 0, 0], [0, 0, 1]].
diff --git a/tasks/robocasa-gr1-plate-to-bowl-i00-standard/task.toml b/tasks/robocasa-gr1-plate-to-bowl-i00-standard/task.toml
index 620e2ce8df..6692913d2c 100644
--- a/tasks/robocasa-gr1-plate-to-bowl-i00-standard/task.toml
+++ b/tasks/robocasa-gr1-plate-to-bowl-i00-standard/task.toml
@@ -24,6 +24,7 @@ simulator = "robosuite"
robot = "fourier-gr1"
family = "robocasa-gr1/PosttrainPnPNovelFromPlateToBowlSplitA"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
expert_time_estimate_hours = 4.0
diff --git a/tasks/robocasa-gr1-plate-to-cardboardbox-i00-privileged/task.toml b/tasks/robocasa-gr1-plate-to-cardboardbox-i00-privileged/task.toml
index af5506c437..3e9bf85628 100644
--- a/tasks/robocasa-gr1-plate-to-cardboardbox-i00-privileged/task.toml
+++ b/tasks/robocasa-gr1-plate-to-cardboardbox-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "robosuite"
robot = "fourier-gr1"
family = "robocasa-gr1/PosttrainPnPNovelFromPlateToCardboardboxSplitA"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robocasa-gr1-plate-to-cardboardbox-i00-standard/README.md b/tasks/robocasa-gr1-plate-to-cardboardbox-i00-standard/README.md
index 249dc13708..4e49d65cd4 100644
--- a/tasks/robocasa-gr1-plate-to-cardboardbox-i00-standard/README.md
+++ b/tasks/robocasa-gr1-plate-to-cardboardbox-i00-standard/README.md
@@ -48,3 +48,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
- 2026-10-10 (review round 3, board 0067): the sim service's request handling (`environment/sim/server_casa.py`) and `tests/grade.py`. A negative or non-numeric `Content-Length` is answered 400 without reading the body (the service read until the client hung up, without a bound, which could exhaust its memory); a camera named more than once in `cameras` is rendered and returned once (N copies of one name made it build N x N images within the frame budget); a port that is still held when the service restarts is waited for, not counted as a crash; a JSON `true` or `false` where `dt` or `seconds` takes a number is a 400 (`true` ran as 1 second). The grader: a `final.json` whose evidence fields (`control_steps`, `calls`, `elapsed_s`, `final_reward`) are not numbers no longer stops it before it writes a reward (they are reported as 0 and decide nothing). An episode's physics and the pass rule are unchanged.
- 2026-10-13 (owner, boards 0085 and 0087): the sim service's clock (`environment/sim/server_casa.py`, the docstring of `environment/rcb_client.py`). `meta["time_s"]` and `meta["time_left_s"]` count from the agent's first request to the service (any endpoint), and a restarted service keeps that moment; before, they counted from the first start of the service's container (`time_left_s`) and from when its scene was ready (`time_s`). Where the service is ready before the agent's phase begins (the harness is installed in between: a median of 79 s over the recorded standard trials), that clock ran ahead of the agent's and `time_left_s` reached 0 while the agent still had that time. `/spec` says what the clock counts from (`budget.clock`), and the instruction's `## Time limit` tells the agent to plan by `time_left_s`. The service never ended an episode on this clock: the scene, the frozen instance, the pass rule and the grader are unchanged, so the task keeps its name and earlier results stand.
- 2026-10-13 (board 0093; found in board 0086): the camera calibration (`environment/sim/server_casa.py`). `cam2world_gl` held the pose of MuJoCo's left-eye camera (`scene.camera[0]`), 3.4 cm (half the model's 0.068 m interpupillary distance) to the camera's left of the viewpoint the mono RGB and depth images are rendered from; it now is that viewpoint (`data.cam_xpos` / `data.cam_xmat`). Checked against MuJoCo's own renders of marker spheres at known world points, every served camera, before and after the robot moved: the reprojection error was 10.5-10.7 px median (up to 28 px for near points, always to the same side), now 0.15 px median and at most 0.4 px. The intrinsics were right (every camera here is given by its `fovy`, none by a sensor size) and are unchanged, as are the images, the scene, the frozen instance, the pass rule and the grader, so the task keeps its name. **Standard-mode results from before this date used the offset calibration**: a point located from the images and the reported calibration came out about 3.4 cm to the side (agents missed a microwave button by 3.5 cm), so they are not strictly comparable with later ones.
+- 2026-10-14 (board 0113, from board 0086): the instruction describes the hands: each hand command is the position target (rad) of its finger's joints, 0 open and larger closing, within each joint's range; the grip site is the origin of the hand's mounting frame at the wrist end of the hand; the pink spheres beside each hand are visual markers of the robot model, not objects (board 0086: most standard runs took them for objects). Text only: the scene, the frozen instance, the reference solution, the success check and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/robocasa-gr1-plate-to-cardboardbox-i00-standard/instruction.md b/tasks/robocasa-gr1-plate-to-cardboardbox-i00-standard/instruction.md
index 783fb5453a..5e3eeda784 100644
--- a/tasks/robocasa-gr1-plate-to-cardboardbox-i00-standard/instruction.md
+++ b/tasks/robocasa-gr1-plate-to-cardboardbox-i00-standard/instruction.md
@@ -52,6 +52,13 @@ res.video[cam]; res.frames[cam]; res.last[cam]; res.proprio; res.save("/app/runs
channel is an absolute target, so the grip sites stay within about a millimetre, but the IK's posture term can still
move the waist and elbows by a few hundredths of a radian while you watch.
- The cameras are mounted on the robot and move with it: use the calibration that comes with each response.
+- The hands: each hand command is the position target, in radians, of that finger's joints (`thumb_flexion` drives
+ two thumb joints, `thumb_rotation` one): 0 is the straight, open finger, larger values bend it toward the palm
+ (the hand closes), and every finger joint stops at the end of its range (1.22 to 1.74 rad), so values below 0 or
+ above about 1.7 change nothing more. The grip site that the pose targets place is the origin of the hand's mounting
+ frame (`gripper0_right_grip_site`, `gripper0_left_grip_site`), at the wrist end of the hand, not a point between
+ the fingertips. The four small pink spheres beside each hand are part of the robot model (visual markers
+ `robot0_<side>_pinch_spheres_0..3`, about 13.5 cm from its grip site, with no collision): they are not objects.
- World frame: metres, z up. In the proprio, `eef_left_*` / `eef_right_*` are the grip sites' positions in the world
frame; the action's pose targets and the `cmd_*` entries (the last command) are in the robot's base frame.
The robot's base frame (body `robot0_base`, fixed in the world) has its origin at world (2.66655, -2.16344, 0.91737) m and is turned +90.0 deg about z (quaternion w x y z (0.70711, 0, 0, 0.70711)): `p_world = R @ p_base + t` with R = [[0, -1, 0], [1, 0, 0], [0, 0, 1]].
diff --git a/tasks/robocasa-gr1-plate-to-cardboardbox-i00-standard/task.toml b/tasks/robocasa-gr1-plate-to-cardboardbox-i00-standard/task.toml
index 00d8b2f6b1..6076ba5378 100644
--- a/tasks/robocasa-gr1-plate-to-cardboardbox-i00-standard/task.toml
+++ b/tasks/robocasa-gr1-plate-to-cardboardbox-i00-standard/task.toml
@@ -24,6 +24,7 @@ simulator = "robosuite"
robot = "fourier-gr1"
family = "robocasa-gr1/PosttrainPnPNovelFromPlateToCardboardboxSplitA"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
expert_time_estimate_hours = 4.0
diff --git a/tasks/robocasa-gr1-plate-to-pan-i00-privileged/task.toml b/tasks/robocasa-gr1-plate-to-pan-i00-privileged/task.toml
index d8624c24b2..4e10f0a7c7 100644
--- a/tasks/robocasa-gr1-plate-to-pan-i00-privileged/task.toml
+++ b/tasks/robocasa-gr1-plate-to-pan-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "robosuite"
robot = "fourier-gr1"
family = "robocasa-gr1/PosttrainPnPNovelFromPlateToPanSplitA"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robocasa-gr1-plate-to-pan-i00-standard/README.md b/tasks/robocasa-gr1-plate-to-pan-i00-standard/README.md
index 5d56cb3e90..8287675765 100644
--- a/tasks/robocasa-gr1-plate-to-pan-i00-standard/README.md
+++ b/tasks/robocasa-gr1-plate-to-pan-i00-standard/README.md
@@ -48,3 +48,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
- 2026-10-10 (review round 3, board 0067): the sim service's request handling (`environment/sim/server_casa.py`) and `tests/grade.py`. A negative or non-numeric `Content-Length` is answered 400 without reading the body (the service read until the client hung up, without a bound, which could exhaust its memory); a camera named more than once in `cameras` is rendered and returned once (N copies of one name made it build N x N images within the frame budget); a port that is still held when the service restarts is waited for, not counted as a crash; a JSON `true` or `false` where `dt` or `seconds` takes a number is a 400 (`true` ran as 1 second). The grader: a `final.json` whose evidence fields (`control_steps`, `calls`, `elapsed_s`, `final_reward`) are not numbers no longer stops it before it writes a reward (they are reported as 0 and decide nothing). An episode's physics and the pass rule are unchanged.
- 2026-10-13 (owner, boards 0085 and 0087): the sim service's clock (`environment/sim/server_casa.py`, the docstring of `environment/rcb_client.py`). `meta["time_s"]` and `meta["time_left_s"]` count from the agent's first request to the service (any endpoint), and a restarted service keeps that moment; before, they counted from the first start of the service's container (`time_left_s`) and from when its scene was ready (`time_s`). Where the service is ready before the agent's phase begins (the harness is installed in between: a median of 79 s over the recorded standard trials), that clock ran ahead of the agent's and `time_left_s` reached 0 while the agent still had that time. `/spec` says what the clock counts from (`budget.clock`), and the instruction's `## Time limit` tells the agent to plan by `time_left_s`. The service never ended an episode on this clock: the scene, the frozen instance, the pass rule and the grader are unchanged, so the task keeps its name and earlier results stand.
- 2026-10-13 (board 0093; found in board 0086): the camera calibration (`environment/sim/server_casa.py`). `cam2world_gl` held the pose of MuJoCo's left-eye camera (`scene.camera[0]`), 3.4 cm (half the model's 0.068 m interpupillary distance) to the camera's left of the viewpoint the mono RGB and depth images are rendered from; it now is that viewpoint (`data.cam_xpos` / `data.cam_xmat`). Checked against MuJoCo's own renders of marker spheres at known world points, every served camera, before and after the robot moved: the reprojection error was 10.5-10.7 px median (up to 28 px for near points, always to the same side), now 0.15 px median and at most 0.4 px. The intrinsics were right (every camera here is given by its `fovy`, none by a sensor size) and are unchanged, as are the images, the scene, the frozen instance, the pass rule and the grader, so the task keeps its name. **Standard-mode results from before this date used the offset calibration**: a point located from the images and the reported calibration came out about 3.4 cm to the side (agents missed a microwave button by 3.5 cm), so they are not strictly comparable with later ones.
+- 2026-10-14 (board 0113, from board 0086): the instruction describes the hands: each hand command is the position target (rad) of its finger's joints, 0 open and larger closing, within each joint's range; the grip site is the origin of the hand's mounting frame at the wrist end of the hand; the pink spheres beside each hand are visual markers of the robot model, not objects (board 0086: most standard runs took them for objects). Text only: the scene, the frozen instance, the reference solution, the success check and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/robocasa-gr1-plate-to-pan-i00-standard/instruction.md b/tasks/robocasa-gr1-plate-to-pan-i00-standard/instruction.md
index 9ec4a436e9..7714b0245b 100644
--- a/tasks/robocasa-gr1-plate-to-pan-i00-standard/instruction.md
+++ b/tasks/robocasa-gr1-plate-to-pan-i00-standard/instruction.md
@@ -52,6 +52,13 @@ res.video[cam]; res.frames[cam]; res.last[cam]; res.proprio; res.save("/app/runs
channel is an absolute target, so the grip sites stay within about a millimetre, but the IK's posture term can still
move the waist and elbows by a few hundredths of a radian while you watch.
- The cameras are mounted on the robot and move with it: use the calibration that comes with each response.
+- The hands: each hand command is the position target, in radians, of that finger's joints (`thumb_flexion` drives
+ two thumb joints, `thumb_rotation` one): 0 is the straight, open finger, larger values bend it toward the palm
+ (the hand closes), and every finger joint stops at the end of its range (1.22 to 1.74 rad), so values below 0 or
+ above about 1.7 change nothing more. The grip site that the pose targets place is the origin of the hand's mounting
+ frame (`gripper0_right_grip_site`, `gripper0_left_grip_site`), at the wrist end of the hand, not a point between
+ the fingertips. The four small pink spheres beside each hand are part of the robot model (visual markers
+ `robot0_<side>_pinch_spheres_0..3`, about 13.5 cm from its grip site, with no collision): they are not objects.
- World frame: metres, z up. In the proprio, `eef_left_*` / `eef_right_*` are the grip sites' positions in the world
frame; the action's pose targets and the `cmd_*` entries (the last command) are in the robot's base frame.
The robot's base frame (body `robot0_base`, fixed in the world) has its origin at world (2.67278, -2.13159, 0.89064) m and is turned +90.0 deg about z (quaternion w x y z (0.70711, 0, 0, 0.70711)): `p_world = R @ p_base + t` with R = [[0, -1, 0], [1, 0, 0], [0, 0, 1]].
diff --git a/tasks/robocasa-gr1-plate-to-pan-i00-standard/task.toml b/tasks/robocasa-gr1-plate-to-pan-i00-standard/task.toml
index df02ab88b3..b06741f349 100644
--- a/tasks/robocasa-gr1-plate-to-pan-i00-standard/task.toml
+++ b/tasks/robocasa-gr1-plate-to-pan-i00-standard/task.toml
@@ -24,6 +24,7 @@ simulator = "robosuite"
robot = "fourier-gr1"
family = "robocasa-gr1/PosttrainPnPNovelFromPlateToPanSplitA"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
expert_time_estimate_hours = 4.0
diff --git a/tasks/robocasa-gr1-plate-to-plate-i00-privileged/task.toml b/tasks/robocasa-gr1-plate-to-plate-i00-privileged/task.toml
index dc7b50d22f..f4f4401c23 100644
--- a/tasks/robocasa-gr1-plate-to-plate-i00-privileged/task.toml
+++ b/tasks/robocasa-gr1-plate-to-plate-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "robosuite"
robot = "fourier-gr1"
family = "robocasa-gr1/PosttrainPnPNovelFromPlateToPlateSplitA"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robocasa-gr1-plate-to-plate-i00-standard/README.md b/tasks/robocasa-gr1-plate-to-plate-i00-standard/README.md
index e4076b5afa..6a17be2a99 100644
--- a/tasks/robocasa-gr1-plate-to-plate-i00-standard/README.md
+++ b/tasks/robocasa-gr1-plate-to-plate-i00-standard/README.md
@@ -48,3 +48,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
- 2026-10-10 (review round 3, board 0067): the sim service's request handling (`environment/sim/server_casa.py`) and `tests/grade.py`. A negative or non-numeric `Content-Length` is answered 400 without reading the body (the service read until the client hung up, without a bound, which could exhaust its memory); a camera named more than once in `cameras` is rendered and returned once (N copies of one name made it build N x N images within the frame budget); a port that is still held when the service restarts is waited for, not counted as a crash; a JSON `true` or `false` where `dt` or `seconds` takes a number is a 400 (`true` ran as 1 second). The grader: a `final.json` whose evidence fields (`control_steps`, `calls`, `elapsed_s`, `final_reward`) are not numbers no longer stops it before it writes a reward (they are reported as 0 and decide nothing). An episode's physics and the pass rule are unchanged.
- 2026-10-13 (owner, boards 0085 and 0087): the sim service's clock (`environment/sim/server_casa.py`, the docstring of `environment/rcb_client.py`). `meta["time_s"]` and `meta["time_left_s"]` count from the agent's first request to the service (any endpoint), and a restarted service keeps that moment; before, they counted from the first start of the service's container (`time_left_s`) and from when its scene was ready (`time_s`). Where the service is ready before the agent's phase begins (the harness is installed in between: a median of 79 s over the recorded standard trials), that clock ran ahead of the agent's and `time_left_s` reached 0 while the agent still had that time. `/spec` says what the clock counts from (`budget.clock`), and the instruction's `## Time limit` tells the agent to plan by `time_left_s`. The service never ended an episode on this clock: the scene, the frozen instance, the pass rule and the grader are unchanged, so the task keeps its name and earlier results stand.
- 2026-10-13 (board 0093; found in board 0086): the camera calibration (`environment/sim/server_casa.py`). `cam2world_gl` held the pose of MuJoCo's left-eye camera (`scene.camera[0]`), 3.4 cm (half the model's 0.068 m interpupillary distance) to the camera's left of the viewpoint the mono RGB and depth images are rendered from; it now is that viewpoint (`data.cam_xpos` / `data.cam_xmat`). Checked against MuJoCo's own renders of marker spheres at known world points, every served camera, before and after the robot moved: the reprojection error was 10.5-10.7 px median (up to 28 px for near points, always to the same side), now 0.15 px median and at most 0.4 px. The intrinsics were right (every camera here is given by its `fovy`, none by a sensor size) and are unchanged, as are the images, the scene, the frozen instance, the pass rule and the grader, so the task keeps its name. **Standard-mode results from before this date used the offset calibration**: a point located from the images and the reported calibration came out about 3.4 cm to the side (agents missed a microwave button by 3.5 cm), so they are not strictly comparable with later ones.
+- 2026-10-14 (board 0113, from board 0086): the instruction describes the hands: each hand command is the position target (rad) of its finger's joints, 0 open and larger closing, within each joint's range; the grip site is the origin of the hand's mounting frame at the wrist end of the hand; the pink spheres beside each hand are visual markers of the robot model, not objects (board 0086: most standard runs took them for objects). Text only: the scene, the frozen instance, the reference solution, the success check and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/robocasa-gr1-plate-to-plate-i00-standard/instruction.md b/tasks/robocasa-gr1-plate-to-plate-i00-standard/instruction.md
index d82b706612..4fc07a5572 100644
--- a/tasks/robocasa-gr1-plate-to-plate-i00-standard/instruction.md
+++ b/tasks/robocasa-gr1-plate-to-plate-i00-standard/instruction.md
@@ -51,6 +51,13 @@ res.video[cam]; res.frames[cam]; res.last[cam]; res.proprio; res.save("/app/runs
channel is an absolute target, so the grip sites stay within about a millimetre, but the IK's posture term can still
move the waist and elbows by a few hundredths of a radian while you watch.
- The cameras are mounted on the robot and move with it: use the calibration that comes with each response.
+- The hands: each hand command is the position target, in radians, of that finger's joints (`thumb_flexion` drives
+ two thumb joints, `thumb_rotation` one): 0 is the straight, open finger, larger values bend it toward the palm
+ (the hand closes), and every finger joint stops at the end of its range (1.22 to 1.74 rad), so values below 0 or
+ above about 1.7 change nothing more. The grip site that the pose targets place is the origin of the hand's mounting
+ frame (`gripper0_right_grip_site`, `gripper0_left_grip_site`), at the wrist end of the hand, not a point between
+ the fingertips. The four small pink spheres beside each hand are part of the robot model (visual markers
+ `robot0_<side>_pinch_spheres_0..3`, about 13.5 cm from its grip site, with no collision): they are not objects.
- World frame: metres, z up. In the proprio, `eef_left_*` / `eef_right_*` are the grip sites' positions in the world
frame; the action's pose targets and the `cmd_*` entries (the last command) are in the robot's base frame.
The robot's base frame (body `robot0_base`, fixed in the world) has its origin at world (2.66859, -2.14823, 0.90329) m and is turned +90.0 deg about z (quaternion w x y z (0.70711, 0, 0, 0.70711)): `p_world = R @ p_base + t` with R = [[0, -1, 0], [1, 0, 0], [0, 0, 1]].
diff --git a/tasks/robocasa-gr1-plate-to-plate-i00-standard/task.toml b/tasks/robocasa-gr1-plate-to-plate-i00-standard/task.toml
index 045ec481a0..851ee66a7d 100644
--- a/tasks/robocasa-gr1-plate-to-plate-i00-standard/task.toml
+++ b/tasks/robocasa-gr1-plate-to-plate-i00-standard/task.toml
@@ -24,6 +24,7 @@ simulator = "robosuite"
robot = "fourier-gr1"
family = "robocasa-gr1/PosttrainPnPNovelFromPlateToPlateSplitA"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
expert_time_estimate_hours = 4.0
diff --git a/tasks/robocasa-gr1-potato-to-microwave-close-i00-privileged/task.toml b/tasks/robocasa-gr1-potato-to-microwave-close-i00-privileged/task.toml
index db973ebb80..c5ef3f968b 100644
--- a/tasks/robocasa-gr1-potato-to-microwave-close-i00-privileged/task.toml
+++ b/tasks/robocasa-gr1-potato-to-microwave-close-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "robosuite"
robot = "fourier-gr1"
family = "robocasa-gr1/PnPPotatoToMicrowaveClose"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robocasa-gr1-potato-to-microwave-close-i00-standard/README.md b/tasks/robocasa-gr1-potato-to-microwave-close-i00-standard/README.md
index a08b6a5699..534b6f8a56 100644
--- a/tasks/robocasa-gr1-potato-to-microwave-close-i00-standard/README.md
+++ b/tasks/robocasa-gr1-potato-to-microwave-close-i00-standard/README.md
@@ -48,3 +48,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
- 2026-10-10 (review round 3, board 0067): the sim service's request handling (`environment/sim/server_casa.py`) and `tests/grade.py`. A negative or non-numeric `Content-Length` is answered 400 without reading the body (the service read until the client hung up, without a bound, which could exhaust its memory); a camera named more than once in `cameras` is rendered and returned once (N copies of one name made it build N x N images within the frame budget); a port that is still held when the service restarts is waited for, not counted as a crash; a JSON `true` or `false` where `dt` or `seconds` takes a number is a 400 (`true` ran as 1 second). The grader: a `final.json` whose evidence fields (`control_steps`, `calls`, `elapsed_s`, `final_reward`) are not numbers no longer stops it before it writes a reward (they are reported as 0 and decide nothing). An episode's physics and the pass rule are unchanged.
- 2026-10-13 (owner, boards 0085 and 0087): the sim service's clock (`environment/sim/server_casa.py`, the docstring of `environment/rcb_client.py`). `meta["time_s"]` and `meta["time_left_s"]` count from the agent's first request to the service (any endpoint), and a restarted service keeps that moment; before, they counted from the first start of the service's container (`time_left_s`) and from when its scene was ready (`time_s`). Where the service is ready before the agent's phase begins (the harness is installed in between: a median of 79 s over the recorded standard trials), that clock ran ahead of the agent's and `time_left_s` reached 0 while the agent still had that time. `/spec` says what the clock counts from (`budget.clock`), and the instruction's `## Time limit` tells the agent to plan by `time_left_s`. The service never ended an episode on this clock: the scene, the frozen instance, the pass rule and the grader are unchanged, so the task keeps its name and earlier results stand.
- 2026-10-13 (board 0093; found in board 0086): the camera calibration (`environment/sim/server_casa.py`). `cam2world_gl` held the pose of MuJoCo's left-eye camera (`scene.camera[0]`), 3.4 cm (half the model's 0.068 m interpupillary distance) to the camera's left of the viewpoint the mono RGB and depth images are rendered from; it now is that viewpoint (`data.cam_xpos` / `data.cam_xmat`). Checked against MuJoCo's own renders of marker spheres at known world points, every served camera, before and after the robot moved: the reprojection error was 10.5-10.7 px median (up to 28 px for near points, always to the same side), now 0.15 px median and at most 0.4 px. The intrinsics were right (every camera here is given by its `fovy`, none by a sensor size) and are unchanged, as are the images, the scene, the frozen instance, the pass rule and the grader, so the task keeps its name. **Standard-mode results from before this date used the offset calibration**: a point located from the images and the reported calibration came out about 3.4 cm to the side (agents missed a microwave button by 3.5 cm), so they are not strictly comparable with later ones.
+- 2026-10-14 (board 0113, from board 0086): the instruction describes the hands: each hand command is the position target (rad) of its finger's joints, 0 open and larger closing, within each joint's range; the grip site is the origin of the hand's mounting frame at the wrist end of the hand; the pink spheres beside each hand are visual markers of the robot model, not objects (board 0086: most standard runs took them for objects). Text only: the scene, the frozen instance, the reference solution, the success check and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/robocasa-gr1-potato-to-microwave-close-i00-standard/instruction.md b/tasks/robocasa-gr1-potato-to-microwave-close-i00-standard/instruction.md
index da3ed1c1d8..f66a4b866e 100644
--- a/tasks/robocasa-gr1-potato-to-microwave-close-i00-standard/instruction.md
+++ b/tasks/robocasa-gr1-potato-to-microwave-close-i00-standard/instruction.md
@@ -51,6 +51,13 @@ res.video[cam]; res.frames[cam]; res.last[cam]; res.proprio; res.save("/app/runs
channel is an absolute target, so the grip sites stay within about a millimetre, but the IK's posture term can still
move the waist and elbows by a few hundredths of a radian while you watch.
- The cameras are mounted on the robot and move with it: use the calibration that comes with each response.
+- The hands: each hand command is the position target, in radians, of that finger's joints (`thumb_flexion` drives
+ two thumb joints, `thumb_rotation` one): 0 is the straight, open finger, larger values bend it toward the palm
+ (the hand closes), and every finger joint stops at the end of its range (1.22 to 1.74 rad), so values below 0 or
+ above about 1.7 change nothing more. The grip site that the pose targets place is the origin of the hand's mounting
+ frame (`gripper0_right_grip_site`, `gripper0_left_grip_site`), at the wrist end of the hand, not a point between
+ the fingertips. The four small pink spheres beside each hand are part of the robot model (visual markers
+ `robot0_<side>_pinch_spheres_0..3`, about 13.5 cm from its grip site, with no collision): they are not objects.
- World frame: metres, z up. In the proprio, `eef_left_*` / `eef_right_*` are the grip sites' positions in the world
frame; the action's pose targets and the `cmd_*` entries (the last command) are in the robot's base frame.
The robot's base frame (body `robot0_base`, fixed in the world) has its origin at world (2.65000, -2.17500, 0.97000) m and is turned +90.0 deg about z (quaternion w x y z (0.70711, 0, 0, 0.70711)): `p_world = R @ p_base + t` with R = [[0, -1, 0], [1, 0, 0], [0, 0, 1]].
diff --git a/tasks/robocasa-gr1-potato-to-microwave-close-i00-standard/task.toml b/tasks/robocasa-gr1-potato-to-microwave-close-i00-standard/task.toml
index faae6844b5..ca07824321 100644
--- a/tasks/robocasa-gr1-potato-to-microwave-close-i00-standard/task.toml
+++ b/tasks/robocasa-gr1-potato-to-microwave-close-i00-standard/task.toml
@@ -24,6 +24,7 @@ simulator = "robosuite"
robot = "fourier-gr1"
family = "robocasa-gr1/PnPPotatoToMicrowaveClose"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
expert_time_estimate_hours = 4.0
diff --git a/tasks/robocasa-gr1-tray-to-cardboardbox-i00-privileged/task.toml b/tasks/robocasa-gr1-tray-to-cardboardbox-i00-privileged/task.toml
index bed6d61405..605255820a 100644
--- a/tasks/robocasa-gr1-tray-to-cardboardbox-i00-privileged/task.toml
+++ b/tasks/robocasa-gr1-tray-to-cardboardbox-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "robosuite"
robot = "fourier-gr1"
family = "robocasa-gr1/PosttrainPnPNovelFromTrayToCardboardboxSplitA"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robocasa-gr1-tray-to-cardboardbox-i00-standard/README.md b/tasks/robocasa-gr1-tray-to-cardboardbox-i00-standard/README.md
index 628a33ac93..5c840b9f22 100644
--- a/tasks/robocasa-gr1-tray-to-cardboardbox-i00-standard/README.md
+++ b/tasks/robocasa-gr1-tray-to-cardboardbox-i00-standard/README.md
@@ -48,3 +48,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
- 2026-10-10 (review round 3, board 0067): the sim service's request handling (`environment/sim/server_casa.py`) and `tests/grade.py`. A negative or non-numeric `Content-Length` is answered 400 without reading the body (the service read until the client hung up, without a bound, which could exhaust its memory); a camera named more than once in `cameras` is rendered and returned once (N copies of one name made it build N x N images within the frame budget); a port that is still held when the service restarts is waited for, not counted as a crash; a JSON `true` or `false` where `dt` or `seconds` takes a number is a 400 (`true` ran as 1 second). The grader: a `final.json` whose evidence fields (`control_steps`, `calls`, `elapsed_s`, `final_reward`) are not numbers no longer stops it before it writes a reward (they are reported as 0 and decide nothing). An episode's physics and the pass rule are unchanged.
- 2026-10-13 (owner, boards 0085 and 0087): the sim service's clock (`environment/sim/server_casa.py`, the docstring of `environment/rcb_client.py`). `meta["time_s"]` and `meta["time_left_s"]` count from the agent's first request to the service (any endpoint), and a restarted service keeps that moment; before, they counted from the first start of the service's container (`time_left_s`) and from when its scene was ready (`time_s`). Where the service is ready before the agent's phase begins (the harness is installed in between: a median of 79 s over the recorded standard trials), that clock ran ahead of the agent's and `time_left_s` reached 0 while the agent still had that time. `/spec` says what the clock counts from (`budget.clock`), and the instruction's `## Time limit` tells the agent to plan by `time_left_s`. The service never ended an episode on this clock: the scene, the frozen instance, the pass rule and the grader are unchanged, so the task keeps its name and earlier results stand.
- 2026-10-13 (board 0093; found in board 0086): the camera calibration (`environment/sim/server_casa.py`). `cam2world_gl` held the pose of MuJoCo's left-eye camera (`scene.camera[0]`), 3.4 cm (half the model's 0.068 m interpupillary distance) to the camera's left of the viewpoint the mono RGB and depth images are rendered from; it now is that viewpoint (`data.cam_xpos` / `data.cam_xmat`). Checked against MuJoCo's own renders of marker spheres at known world points, every served camera, before and after the robot moved: the reprojection error was 10.5-10.7 px median (up to 28 px for near points, always to the same side), now 0.15 px median and at most 0.4 px. The intrinsics were right (every camera here is given by its `fovy`, none by a sensor size) and are unchanged, as are the images, the scene, the frozen instance, the pass rule and the grader, so the task keeps its name. **Standard-mode results from before this date used the offset calibration**: a point located from the images and the reported calibration came out about 3.4 cm to the side (agents missed a microwave button by 3.5 cm), so they are not strictly comparable with later ones.
+- 2026-10-14 (board 0113, from board 0086): the instruction describes the hands: each hand command is the position target (rad) of its finger's joints, 0 open and larger closing, within each joint's range; the grip site is the origin of the hand's mounting frame at the wrist end of the hand; the pink spheres beside each hand are visual markers of the robot model, not objects (board 0086: most standard runs took them for objects). Text only: the scene, the frozen instance, the reference solution, the success check and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/robocasa-gr1-tray-to-cardboardbox-i00-standard/instruction.md b/tasks/robocasa-gr1-tray-to-cardboardbox-i00-standard/instruction.md
index ee167a49e2..02e78d6182 100644
--- a/tasks/robocasa-gr1-tray-to-cardboardbox-i00-standard/instruction.md
+++ b/tasks/robocasa-gr1-tray-to-cardboardbox-i00-standard/instruction.md
@@ -51,6 +51,13 @@ res.video[cam]; res.frames[cam]; res.last[cam]; res.proprio; res.save("/app/runs
channel is an absolute target, so the grip sites stay within about a millimetre, but the IK's posture term can still
move the waist and elbows by a few hundredths of a radian while you watch.
- The cameras are mounted on the robot and move with it: use the calibration that comes with each response.
+- The hands: each hand command is the position target, in radians, of that finger's joints (`thumb_flexion` drives
+ two thumb joints, `thumb_rotation` one): 0 is the straight, open finger, larger values bend it toward the palm
+ (the hand closes), and every finger joint stops at the end of its range (1.22 to 1.74 rad), so values below 0 or
+ above about 1.7 change nothing more. The grip site that the pose targets place is the origin of the hand's mounting
+ frame (`gripper0_right_grip_site`, `gripper0_left_grip_site`), at the wrist end of the hand, not a point between
+ the fingertips. The four small pink spheres beside each hand are part of the robot model (visual markers
+ `robot0_<side>_pinch_spheres_0..3`, about 13.5 cm from its grip site, with no collision): they are not objects.
- World frame: metres, z up. In the proprio, `eef_left_*` / `eef_right_*` are the grip sites' positions in the world
frame; the action's pose targets and the `cmd_*` entries (the last command) are in the robot's base frame.
The robot's base frame (body `robot0_base`, fixed in the world) has its origin at world (2.68282, -2.16303, 0.90053) m and is turned +90.0 deg about z (quaternion w x y z (0.70711, 0, 0, 0.70711)): `p_world = R @ p_base + t` with R = [[0, -1, 0], [1, 0, 0], [0, 0, 1]].
diff --git a/tasks/robocasa-gr1-tray-to-cardboardbox-i00-standard/task.toml b/tasks/robocasa-gr1-tray-to-cardboardbox-i00-standard/task.toml
index 4bedfb1ca9..196b59b181 100644
--- a/tasks/robocasa-gr1-tray-to-cardboardbox-i00-standard/task.toml
+++ b/tasks/robocasa-gr1-tray-to-cardboardbox-i00-standard/task.toml
@@ -24,6 +24,7 @@ simulator = "robosuite"
robot = "fourier-gr1"
family = "robocasa-gr1/PosttrainPnPNovelFromTrayToCardboardboxSplitA"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
expert_time_estimate_hours = 4.0
diff --git a/tasks/robocasa-gr1-tray-to-plate-i00-privileged/task.toml b/tasks/robocasa-gr1-tray-to-plate-i00-privileged/task.toml
index 5a3bb24e99..0d34fab29b 100644
--- a/tasks/robocasa-gr1-tray-to-plate-i00-privileged/task.toml
+++ b/tasks/robocasa-gr1-tray-to-plate-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "robosuite"
robot = "fourier-gr1"
family = "robocasa-gr1/PosttrainPnPNovelFromTrayToPlateSplitA"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robocasa-gr1-tray-to-plate-i00-standard/README.md b/tasks/robocasa-gr1-tray-to-plate-i00-standard/README.md
index 50c246ee7b..c0756c3508 100644
--- a/tasks/robocasa-gr1-tray-to-plate-i00-standard/README.md
+++ b/tasks/robocasa-gr1-tray-to-plate-i00-standard/README.md
@@ -48,3 +48,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
- 2026-10-10 (review round 3, board 0067): the sim service's request handling (`environment/sim/server_casa.py`) and `tests/grade.py`. A negative or non-numeric `Content-Length` is answered 400 without reading the body (the service read until the client hung up, without a bound, which could exhaust its memory); a camera named more than once in `cameras` is rendered and returned once (N copies of one name made it build N x N images within the frame budget); a port that is still held when the service restarts is waited for, not counted as a crash; a JSON `true` or `false` where `dt` or `seconds` takes a number is a 400 (`true` ran as 1 second). The grader: a `final.json` whose evidence fields (`control_steps`, `calls`, `elapsed_s`, `final_reward`) are not numbers no longer stops it before it writes a reward (they are reported as 0 and decide nothing). An episode's physics and the pass rule are unchanged.
- 2026-10-13 (owner, boards 0085 and 0087): the sim service's clock (`environment/sim/server_casa.py`, the docstring of `environment/rcb_client.py`). `meta["time_s"]` and `meta["time_left_s"]` count from the agent's first request to the service (any endpoint), and a restarted service keeps that moment; before, they counted from the first start of the service's container (`time_left_s`) and from when its scene was ready (`time_s`). Where the service is ready before the agent's phase begins (the harness is installed in between: a median of 79 s over the recorded standard trials), that clock ran ahead of the agent's and `time_left_s` reached 0 while the agent still had that time. `/spec` says what the clock counts from (`budget.clock`), and the instruction's `## Time limit` tells the agent to plan by `time_left_s`. The service never ended an episode on this clock: the scene, the frozen instance, the pass rule and the grader are unchanged, so the task keeps its name and earlier results stand.
- 2026-10-13 (board 0093; found in board 0086): the camera calibration (`environment/sim/server_casa.py`). `cam2world_gl` held the pose of MuJoCo's left-eye camera (`scene.camera[0]`), 3.4 cm (half the model's 0.068 m interpupillary distance) to the camera's left of the viewpoint the mono RGB and depth images are rendered from; it now is that viewpoint (`data.cam_xpos` / `data.cam_xmat`). Checked against MuJoCo's own renders of marker spheres at known world points, every served camera, before and after the robot moved: the reprojection error was 10.5-10.7 px median (up to 28 px for near points, always to the same side), now 0.15 px median and at most 0.4 px. The intrinsics were right (every camera here is given by its `fovy`, none by a sensor size) and are unchanged, as are the images, the scene, the frozen instance, the pass rule and the grader, so the task keeps its name. **Standard-mode results from before this date used the offset calibration**: a point located from the images and the reported calibration came out about 3.4 cm to the side (agents missed a microwave button by 3.5 cm), so they are not strictly comparable with later ones.
+- 2026-10-14 (board 0113, from board 0086): the instruction describes the hands: each hand command is the position target (rad) of its finger's joints, 0 open and larger closing, within each joint's range; the grip site is the origin of the hand's mounting frame at the wrist end of the hand; the pink spheres beside each hand are visual markers of the robot model, not objects (board 0086: most standard runs took them for objects). Text only: the scene, the frozen instance, the reference solution, the success check and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/robocasa-gr1-tray-to-plate-i00-standard/instruction.md b/tasks/robocasa-gr1-tray-to-plate-i00-standard/instruction.md
index f0eb8b68dc..c69e5a0cd8 100644
--- a/tasks/robocasa-gr1-tray-to-plate-i00-standard/instruction.md
+++ b/tasks/robocasa-gr1-tray-to-plate-i00-standard/instruction.md
@@ -51,6 +51,13 @@ res.video[cam]; res.frames[cam]; res.last[cam]; res.proprio; res.save("/app/runs
channel is an absolute target, so the grip sites stay within about a millimetre, but the IK's posture term can still
move the waist and elbows by a few hundredths of a radian while you watch.
- The cameras are mounted on the robot and move with it: use the calibration that comes with each response.
+- The hands: each hand command is the position target, in radians, of that finger's joints (`thumb_flexion` drives
+ two thumb joints, `thumb_rotation` one): 0 is the straight, open finger, larger values bend it toward the palm
+ (the hand closes), and every finger joint stops at the end of its range (1.22 to 1.74 rad), so values below 0 or
+ above about 1.7 change nothing more. The grip site that the pose targets place is the origin of the hand's mounting
+ frame (`gripper0_right_grip_site`, `gripper0_left_grip_site`), at the wrist end of the hand, not a point between
+ the fingertips. The four small pink spheres beside each hand are part of the robot model (visual markers
+ `robot0_<side>_pinch_spheres_0..3`, about 13.5 cm from its grip site, with no collision): they are not objects.
- World frame: metres, z up. In the proprio, `eef_left_*` / `eef_right_*` are the grip sites' positions in the world
frame; the action's pose targets and the `cmd_*` entries (the last command) are in the robot's base frame.
The robot's base frame (body `robot0_base`, fixed in the world) has its origin at world (2.65879, -2.17094, 0.88403) m and is turned +90.0 deg about z (quaternion w x y z (0.70711, 0, 0, 0.70711)): `p_world = R @ p_base + t` with R = [[0, -1, 0], [1, 0, 0], [0, 0, 1]].
diff --git a/tasks/robocasa-gr1-tray-to-plate-i00-standard/task.toml b/tasks/robocasa-gr1-tray-to-plate-i00-standard/task.toml
index b0df7c438c..1676e3cd7b 100644
--- a/tasks/robocasa-gr1-tray-to-plate-i00-standard/task.toml
+++ b/tasks/robocasa-gr1-tray-to-plate-i00-standard/task.toml
@@ -24,6 +24,7 @@ simulator = "robosuite"
robot = "fourier-gr1"
family = "robocasa-gr1/PosttrainPnPNovelFromTrayToPlateSplitA"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
expert_time_estimate_hours = 4.0
diff --git a/tasks/robocasa-gr1-tray-to-pot-i00-privileged/task.toml b/tasks/robocasa-gr1-tray-to-pot-i00-privileged/task.toml
index 173ed086c3..d12371adc6 100644
--- a/tasks/robocasa-gr1-tray-to-pot-i00-privileged/task.toml
+++ b/tasks/robocasa-gr1-tray-to-pot-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "robosuite"
robot = "fourier-gr1"
family = "robocasa-gr1/PosttrainPnPNovelFromTrayToPotSplitA"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robocasa-gr1-tray-to-pot-i00-standard/README.md b/tasks/robocasa-gr1-tray-to-pot-i00-standard/README.md
index f89762b3a4..9c22f0e35b 100644
--- a/tasks/robocasa-gr1-tray-to-pot-i00-standard/README.md
+++ b/tasks/robocasa-gr1-tray-to-pot-i00-standard/README.md
@@ -48,3 +48,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
- 2026-10-10 (review round 3, board 0067): the sim service's request handling (`environment/sim/server_casa.py`) and `tests/grade.py`. A negative or non-numeric `Content-Length` is answered 400 without reading the body (the service read until the client hung up, without a bound, which could exhaust its memory); a camera named more than once in `cameras` is rendered and returned once (N copies of one name made it build N x N images within the frame budget); a port that is still held when the service restarts is waited for, not counted as a crash; a JSON `true` or `false` where `dt` or `seconds` takes a number is a 400 (`true` ran as 1 second). The grader: a `final.json` whose evidence fields (`control_steps`, `calls`, `elapsed_s`, `final_reward`) are not numbers no longer stops it before it writes a reward (they are reported as 0 and decide nothing). An episode's physics and the pass rule are unchanged.
- 2026-10-13 (owner, boards 0085 and 0087): the sim service's clock (`environment/sim/server_casa.py`, the docstring of `environment/rcb_client.py`). `meta["time_s"]` and `meta["time_left_s"]` count from the agent's first request to the service (any endpoint), and a restarted service keeps that moment; before, they counted from the first start of the service's container (`time_left_s`) and from when its scene was ready (`time_s`). Where the service is ready before the agent's phase begins (the harness is installed in between: a median of 79 s over the recorded standard trials), that clock ran ahead of the agent's and `time_left_s` reached 0 while the agent still had that time. `/spec` says what the clock counts from (`budget.clock`), and the instruction's `## Time limit` tells the agent to plan by `time_left_s`. The service never ended an episode on this clock: the scene, the frozen instance, the pass rule and the grader are unchanged, so the task keeps its name and earlier results stand.
- 2026-10-13 (board 0093; found in board 0086): the camera calibration (`environment/sim/server_casa.py`). `cam2world_gl` held the pose of MuJoCo's left-eye camera (`scene.camera[0]`), 3.4 cm (half the model's 0.068 m interpupillary distance) to the camera's left of the viewpoint the mono RGB and depth images are rendered from; it now is that viewpoint (`data.cam_xpos` / `data.cam_xmat`). Checked against MuJoCo's own renders of marker spheres at known world points, every served camera, before and after the robot moved: the reprojection error was 10.5-10.7 px median (up to 28 px for near points, always to the same side), now 0.15 px median and at most 0.4 px. The intrinsics were right (every camera here is given by its `fovy`, none by a sensor size) and are unchanged, as are the images, the scene, the frozen instance, the pass rule and the grader, so the task keeps its name. **Standard-mode results from before this date used the offset calibration**: a point located from the images and the reported calibration came out about 3.4 cm to the side (agents missed a microwave button by 3.5 cm), so they are not strictly comparable with later ones.
+- 2026-10-14 (board 0113, from board 0086): the instruction describes the hands: each hand command is the position target (rad) of its finger's joints, 0 open and larger closing, within each joint's range; the grip site is the origin of the hand's mounting frame at the wrist end of the hand; the pink spheres beside each hand are visual markers of the robot model, not objects (board 0086: most standard runs took them for objects). Text only: the scene, the frozen instance, the reference solution, the success check and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/robocasa-gr1-tray-to-pot-i00-standard/instruction.md b/tasks/robocasa-gr1-tray-to-pot-i00-standard/instruction.md
index 4764db287f..9ec253086f 100644
--- a/tasks/robocasa-gr1-tray-to-pot-i00-standard/instruction.md
+++ b/tasks/robocasa-gr1-tray-to-pot-i00-standard/instruction.md
@@ -51,6 +51,13 @@ res.video[cam]; res.frames[cam]; res.last[cam]; res.proprio; res.save("/app/runs
channel is an absolute target, so the grip sites stay within about a millimetre, but the IK's posture term can still
move the waist and elbows by a few hundredths of a radian while you watch.
- The cameras are mounted on the robot and move with it: use the calibration that comes with each response.
+- The hands: each hand command is the position target, in radians, of that finger's joints (`thumb_flexion` drives
+ two thumb joints, `thumb_rotation` one): 0 is the straight, open finger, larger values bend it toward the palm
+ (the hand closes), and every finger joint stops at the end of its range (1.22 to 1.74 rad), so values below 0 or
+ above about 1.7 change nothing more. The grip site that the pose targets place is the origin of the hand's mounting
+ frame (`gripper0_right_grip_site`, `gripper0_left_grip_site`), at the wrist end of the hand, not a point between
+ the fingertips. The four small pink spheres beside each hand are part of the robot model (visual markers
+ `robot0_<side>_pinch_spheres_0..3`, about 13.5 cm from its grip site, with no collision): they are not objects.
- World frame: metres, z up. In the proprio, `eef_left_*` / `eef_right_*` are the grip sites' positions in the world
frame; the action's pose targets and the `cmd_*` entries (the last command) are in the robot's base frame.
The robot's base frame (body `robot0_base`, fixed in the world) has its origin at world (2.67000, -2.17004, 0.91472) m and is turned +90.0 deg about z (quaternion w x y z (0.70711, 0, 0, 0.70711)): `p_world = R @ p_base + t` with R = [[0, -1, 0], [1, 0, 0], [0, 0, 1]].
diff --git a/tasks/robocasa-gr1-tray-to-pot-i00-standard/task.toml b/tasks/robocasa-gr1-tray-to-pot-i00-standard/task.toml
index ad647d5fe1..4f05340708 100644
--- a/tasks/robocasa-gr1-tray-to-pot-i00-standard/task.toml
+++ b/tasks/robocasa-gr1-tray-to-pot-i00-standard/task.toml
@@ -24,6 +24,7 @@ simulator = "robosuite"
robot = "fourier-gr1"
family = "robocasa-gr1/PosttrainPnPNovelFromTrayToPotSplitA"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
expert_time_estimate_hours = 4.0
diff --git a/tasks/robocasa-gr1-tray-to-tieredbasket-i00-privileged/task.toml b/tasks/robocasa-gr1-tray-to-tieredbasket-i00-privileged/task.toml
index f24dd8f138..6eda070e41 100644
--- a/tasks/robocasa-gr1-tray-to-tieredbasket-i00-privileged/task.toml
+++ b/tasks/robocasa-gr1-tray-to-tieredbasket-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "robosuite"
robot = "fourier-gr1"
family = "robocasa-gr1/PosttrainPnPNovelFromTrayToTieredbasketSplitA"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robocasa-gr1-tray-to-tieredbasket-i00-standard/README.md b/tasks/robocasa-gr1-tray-to-tieredbasket-i00-standard/README.md
index b4fe07a20a..05ec7d422c 100644
--- a/tasks/robocasa-gr1-tray-to-tieredbasket-i00-standard/README.md
+++ b/tasks/robocasa-gr1-tray-to-tieredbasket-i00-standard/README.md
@@ -48,3 +48,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
- 2026-10-10 (review round 3, board 0067): the sim service's request handling (`environment/sim/server_casa.py`) and `tests/grade.py`. A negative or non-numeric `Content-Length` is answered 400 without reading the body (the service read until the client hung up, without a bound, which could exhaust its memory); a camera named more than once in `cameras` is rendered and returned once (N copies of one name made it build N x N images within the frame budget); a port that is still held when the service restarts is waited for, not counted as a crash; a JSON `true` or `false` where `dt` or `seconds` takes a number is a 400 (`true` ran as 1 second). The grader: a `final.json` whose evidence fields (`control_steps`, `calls`, `elapsed_s`, `final_reward`) are not numbers no longer stops it before it writes a reward (they are reported as 0 and decide nothing). An episode's physics and the pass rule are unchanged.
- 2026-10-13 (owner, boards 0085 and 0087): the sim service's clock (`environment/sim/server_casa.py`, the docstring of `environment/rcb_client.py`). `meta["time_s"]` and `meta["time_left_s"]` count from the agent's first request to the service (any endpoint), and a restarted service keeps that moment; before, they counted from the first start of the service's container (`time_left_s`) and from when its scene was ready (`time_s`). Where the service is ready before the agent's phase begins (the harness is installed in between: a median of 79 s over the recorded standard trials), that clock ran ahead of the agent's and `time_left_s` reached 0 while the agent still had that time. `/spec` says what the clock counts from (`budget.clock`), and the instruction's `## Time limit` tells the agent to plan by `time_left_s`. The service never ended an episode on this clock: the scene, the frozen instance, the pass rule and the grader are unchanged, so the task keeps its name and earlier results stand.
- 2026-10-13 (board 0093; found in board 0086): the camera calibration (`environment/sim/server_casa.py`). `cam2world_gl` held the pose of MuJoCo's left-eye camera (`scene.camera[0]`), 3.4 cm (half the model's 0.068 m interpupillary distance) to the camera's left of the viewpoint the mono RGB and depth images are rendered from; it now is that viewpoint (`data.cam_xpos` / `data.cam_xmat`). Checked against MuJoCo's own renders of marker spheres at known world points, every served camera, before and after the robot moved: the reprojection error was 10.5-10.7 px median (up to 28 px for near points, always to the same side), now 0.15 px median and at most 0.4 px. The intrinsics were right (every camera here is given by its `fovy`, none by a sensor size) and are unchanged, as are the images, the scene, the frozen instance, the pass rule and the grader, so the task keeps its name. **Standard-mode results from before this date used the offset calibration**: a point located from the images and the reported calibration came out about 3.4 cm to the side (agents missed a microwave button by 3.5 cm), so they are not strictly comparable with later ones.
+- 2026-10-14 (board 0113, from board 0086): the instruction describes the hands: each hand command is the position target (rad) of its finger's joints, 0 open and larger closing, within each joint's range; the grip site is the origin of the hand's mounting frame at the wrist end of the hand; the pink spheres beside each hand are visual markers of the robot model, not objects (board 0086: most standard runs took them for objects). Text only: the scene, the frozen instance, the reference solution, the success check and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/robocasa-gr1-tray-to-tieredbasket-i00-standard/instruction.md b/tasks/robocasa-gr1-tray-to-tieredbasket-i00-standard/instruction.md
index 42589ff777..29a71afb4b 100644
--- a/tasks/robocasa-gr1-tray-to-tieredbasket-i00-standard/instruction.md
+++ b/tasks/robocasa-gr1-tray-to-tieredbasket-i00-standard/instruction.md
@@ -51,6 +51,13 @@ res.video[cam]; res.frames[cam]; res.last[cam]; res.proprio; res.save("/app/runs
channel is an absolute target, so the grip sites stay within about a millimetre, but the IK's posture term can still
move the waist and elbows by a few hundredths of a radian while you watch.
- The cameras are mounted on the robot and move with it: use the calibration that comes with each response.
+- The hands: each hand command is the position target, in radians, of that finger's joints (`thumb_flexion` drives
+ two thumb joints, `thumb_rotation` one): 0 is the straight, open finger, larger values bend it toward the palm
+ (the hand closes), and every finger joint stops at the end of its range (1.22 to 1.74 rad), so values below 0 or
+ above about 1.7 change nothing more. The grip site that the pose targets place is the origin of the hand's mounting
+ frame (`gripper0_right_grip_site`, `gripper0_left_grip_site`), at the wrist end of the hand, not a point between
+ the fingertips. The four small pink spheres beside each hand are part of the robot model (visual markers
+ `robot0_<side>_pinch_spheres_0..3`, about 13.5 cm from its grip site, with no collision): they are not objects.
- World frame: metres, z up. In the proprio, `eef_left_*` / `eef_right_*` are the grip sites' positions in the world
frame; the action's pose targets and the `cmd_*` entries (the last command) are in the robot's base frame.
The robot's base frame (body `robot0_base`, fixed in the world) has its origin at world (2.66894, -2.14399, 0.87359) m and is turned +90.0 deg about z (quaternion w x y z (0.70711, 0, 0, 0.70711)): `p_world = R @ p_base + t` with R = [[0, -1, 0], [1, 0, 0], [0, 0, 1]].
diff --git a/tasks/robocasa-gr1-tray-to-tieredbasket-i00-standard/task.toml b/tasks/robocasa-gr1-tray-to-tieredbasket-i00-standard/task.toml
index 14742f41f5..a1b9cc8585 100644
--- a/tasks/robocasa-gr1-tray-to-tieredbasket-i00-standard/task.toml
+++ b/tasks/robocasa-gr1-tray-to-tieredbasket-i00-standard/task.toml
@@ -24,6 +24,7 @@ simulator = "robosuite"
robot = "fourier-gr1"
family = "robocasa-gr1/PosttrainPnPNovelFromTrayToTieredbasketSplitA"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
expert_time_estimate_hours = 4.0
diff --git a/tasks/robocasa-gr1-tray-to-tieredshelf-i00-privileged/task.toml b/tasks/robocasa-gr1-tray-to-tieredshelf-i00-privileged/task.toml
index a22ba53d84..4cf26bc0a1 100644
--- a/tasks/robocasa-gr1-tray-to-tieredshelf-i00-privileged/task.toml
+++ b/tasks/robocasa-gr1-tray-to-tieredshelf-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "robosuite"
robot = "fourier-gr1"
family = "robocasa-gr1/PosttrainPnPNovelFromTrayToTieredshelfSplitA"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robocasa-gr1-tray-to-tieredshelf-i00-standard/README.md b/tasks/robocasa-gr1-tray-to-tieredshelf-i00-standard/README.md
index a38e98024f..dcda375071 100644
--- a/tasks/robocasa-gr1-tray-to-tieredshelf-i00-standard/README.md
+++ b/tasks/robocasa-gr1-tray-to-tieredshelf-i00-standard/README.md
@@ -48,3 +48,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
- 2026-10-10 (review round 3, board 0067): the sim service's request handling (`environment/sim/server_casa.py`) and `tests/grade.py`. A negative or non-numeric `Content-Length` is answered 400 without reading the body (the service read until the client hung up, without a bound, which could exhaust its memory); a camera named more than once in `cameras` is rendered and returned once (N copies of one name made it build N x N images within the frame budget); a port that is still held when the service restarts is waited for, not counted as a crash; a JSON `true` or `false` where `dt` or `seconds` takes a number is a 400 (`true` ran as 1 second). The grader: a `final.json` whose evidence fields (`control_steps`, `calls`, `elapsed_s`, `final_reward`) are not numbers no longer stops it before it writes a reward (they are reported as 0 and decide nothing). An episode's physics and the pass rule are unchanged.
- 2026-10-13 (owner, boards 0085 and 0087): the sim service's clock (`environment/sim/server_casa.py`, the docstring of `environment/rcb_client.py`). `meta["time_s"]` and `meta["time_left_s"]` count from the agent's first request to the service (any endpoint), and a restarted service keeps that moment; before, they counted from the first start of the service's container (`time_left_s`) and from when its scene was ready (`time_s`). Where the service is ready before the agent's phase begins (the harness is installed in between: a median of 79 s over the recorded standard trials), that clock ran ahead of the agent's and `time_left_s` reached 0 while the agent still had that time. `/spec` says what the clock counts from (`budget.clock`), and the instruction's `## Time limit` tells the agent to plan by `time_left_s`. The service never ended an episode on this clock: the scene, the frozen instance, the pass rule and the grader are unchanged, so the task keeps its name and earlier results stand.
- 2026-10-13 (board 0093; found in board 0086): the camera calibration (`environment/sim/server_casa.py`). `cam2world_gl` held the pose of MuJoCo's left-eye camera (`scene.camera[0]`), 3.4 cm (half the model's 0.068 m interpupillary distance) to the camera's left of the viewpoint the mono RGB and depth images are rendered from; it now is that viewpoint (`data.cam_xpos` / `data.cam_xmat`). Checked against MuJoCo's own renders of marker spheres at known world points, every served camera, before and after the robot moved: the reprojection error was 10.5-10.7 px median (up to 28 px for near points, always to the same side), now 0.15 px median and at most 0.4 px. The intrinsics were right (every camera here is given by its `fovy`, none by a sensor size) and are unchanged, as are the images, the scene, the frozen instance, the pass rule and the grader, so the task keeps its name. **Standard-mode results from before this date used the offset calibration**: a point located from the images and the reported calibration came out about 3.4 cm to the side (agents missed a microwave button by 3.5 cm), so they are not strictly comparable with later ones.
+- 2026-10-14 (board 0113, from board 0086): the instruction describes the hands: each hand command is the position target (rad) of its finger's joints, 0 open and larger closing, within each joint's range; the grip site is the origin of the hand's mounting frame at the wrist end of the hand; the pink spheres beside each hand are visual markers of the robot model, not objects (board 0086: most standard runs took them for objects). Text only: the scene, the frozen instance, the reference solution, the success check and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/robocasa-gr1-tray-to-tieredshelf-i00-standard/instruction.md b/tasks/robocasa-gr1-tray-to-tieredshelf-i00-standard/instruction.md
index dc74890de5..5d973a2b0f 100644
--- a/tasks/robocasa-gr1-tray-to-tieredshelf-i00-standard/instruction.md
+++ b/tasks/robocasa-gr1-tray-to-tieredshelf-i00-standard/instruction.md
@@ -52,6 +52,13 @@ res.video[cam]; res.frames[cam]; res.last[cam]; res.proprio; res.save("/app/runs
channel is an absolute target, so the grip sites stay within about a millimetre, but the IK's posture term can still
move the waist and elbows by a few hundredths of a radian while you watch.
- The cameras are mounted on the robot and move with it: use the calibration that comes with each response.
+- The hands: each hand command is the position target, in radians, of that finger's joints (`thumb_flexion` drives
+ two thumb joints, `thumb_rotation` one): 0 is the straight, open finger, larger values bend it toward the palm
+ (the hand closes), and every finger joint stops at the end of its range (1.22 to 1.74 rad), so values below 0 or
+ above about 1.7 change nothing more. The grip site that the pose targets place is the origin of the hand's mounting
+ frame (`gripper0_right_grip_site`, `gripper0_left_grip_site`), at the wrist end of the hand, not a point between
+ the fingertips. The four small pink spheres beside each hand are part of the robot model (visual markers
+ `robot0_<side>_pinch_spheres_0..3`, about 13.5 cm from its grip site, with no collision): they are not objects.
- World frame: metres, z up. In the proprio, `eef_left_*` / `eef_right_*` are the grip sites' positions in the world
frame; the action's pose targets and the `cmd_*` entries (the last command) are in the robot's base frame.
The robot's base frame (body `robot0_base`, fixed in the world) has its origin at world (2.67441, -2.13856, 0.89778) m and is turned +90.0 deg about z (quaternion w x y z (0.70711, 0, 0, 0.70711)): `p_world = R @ p_base + t` with R = [[0, -1, 0], [1, 0, 0], [0, 0, 1]].
diff --git a/tasks/robocasa-gr1-tray-to-tieredshelf-i00-standard/task.toml b/tasks/robocasa-gr1-tray-to-tieredshelf-i00-standard/task.toml
index 67b64f3715..0ae63384d0 100644
--- a/tasks/robocasa-gr1-tray-to-tieredshelf-i00-standard/task.toml
+++ b/tasks/robocasa-gr1-tray-to-tieredshelf-i00-standard/task.toml
@@ -24,6 +24,7 @@ simulator = "robosuite"
robot = "fourier-gr1"
family = "robocasa-gr1/PosttrainPnPNovelFromTrayToTieredshelfSplitA"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
expert_time_estimate_hours = 4.0
diff --git a/tasks/robocasa-gr1-wine-to-cabinet-close-i00-privileged/task.toml b/tasks/robocasa-gr1-wine-to-cabinet-close-i00-privileged/task.toml
index 27f6acf1c6..837c1f710d 100644
--- a/tasks/robocasa-gr1-wine-to-cabinet-close-i00-privileged/task.toml
+++ b/tasks/robocasa-gr1-wine-to-cabinet-close-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "robosuite"
robot = "fourier-gr1"
family = "robocasa-gr1/PnPWineToCabinetClose"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robocasa-gr1-wine-to-cabinet-close-i00-standard/README.md b/tasks/robocasa-gr1-wine-to-cabinet-close-i00-standard/README.md
index 34d3c66d6b..77017c5be0 100644
--- a/tasks/robocasa-gr1-wine-to-cabinet-close-i00-standard/README.md
+++ b/tasks/robocasa-gr1-wine-to-cabinet-close-i00-standard/README.md
@@ -48,3 +48,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
- 2026-10-10 (review round 3, board 0067): the sim service's request handling (`environment/sim/server_casa.py`) and `tests/grade.py`. A negative or non-numeric `Content-Length` is answered 400 without reading the body (the service read until the client hung up, without a bound, which could exhaust its memory); a camera named more than once in `cameras` is rendered and returned once (N copies of one name made it build N x N images within the frame budget); a port that is still held when the service restarts is waited for, not counted as a crash; a JSON `true` or `false` where `dt` or `seconds` takes a number is a 400 (`true` ran as 1 second). The grader: a `final.json` whose evidence fields (`control_steps`, `calls`, `elapsed_s`, `final_reward`) are not numbers no longer stops it before it writes a reward (they are reported as 0 and decide nothing). An episode's physics and the pass rule are unchanged.
- 2026-10-13 (owner, boards 0085 and 0087): the sim service's clock (`environment/sim/server_casa.py`, the docstring of `environment/rcb_client.py`). `meta["time_s"]` and `meta["time_left_s"]` count from the agent's first request to the service (any endpoint), and a restarted service keeps that moment; before, they counted from the first start of the service's container (`time_left_s`) and from when its scene was ready (`time_s`). Where the service is ready before the agent's phase begins (the harness is installed in between: a median of 79 s over the recorded standard trials), that clock ran ahead of the agent's and `time_left_s` reached 0 while the agent still had that time. `/spec` says what the clock counts from (`budget.clock`), and the instruction's `## Time limit` tells the agent to plan by `time_left_s`. The service never ended an episode on this clock: the scene, the frozen instance, the pass rule and the grader are unchanged, so the task keeps its name and earlier results stand.
- 2026-10-13 (board 0093; found in board 0086): the camera calibration (`environment/sim/server_casa.py`). `cam2world_gl` held the pose of MuJoCo's left-eye camera (`scene.camera[0]`), 3.4 cm (half the model's 0.068 m interpupillary distance) to the camera's left of the viewpoint the mono RGB and depth images are rendered from; it now is that viewpoint (`data.cam_xpos` / `data.cam_xmat`). Checked against MuJoCo's own renders of marker spheres at known world points, every served camera, before and after the robot moved: the reprojection error was 10.5-10.7 px median (up to 28 px for near points, always to the same side), now 0.15 px median and at most 0.4 px. The intrinsics were right (every camera here is given by its `fovy`, none by a sensor size) and are unchanged, as are the images, the scene, the frozen instance, the pass rule and the grader, so the task keeps its name. **Standard-mode results from before this date used the offset calibration**: a point located from the images and the reported calibration came out about 3.4 cm to the side (agents missed a microwave button by 3.5 cm), so they are not strictly comparable with later ones.
+- 2026-10-14 (board 0113, from board 0086): the instruction describes the hands: each hand command is the position target (rad) of its finger's joints, 0 open and larger closing, within each joint's range; the grip site is the origin of the hand's mounting frame at the wrist end of the hand; the pink spheres beside each hand are visual markers of the robot model, not objects (board 0086: most standard runs took them for objects). Text only: the scene, the frozen instance, the reference solution, the success check and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/robocasa-gr1-wine-to-cabinet-close-i00-standard/instruction.md b/tasks/robocasa-gr1-wine-to-cabinet-close-i00-standard/instruction.md
index 698ecd6975..8d12ca11ec 100644
--- a/tasks/robocasa-gr1-wine-to-cabinet-close-i00-standard/instruction.md
+++ b/tasks/robocasa-gr1-wine-to-cabinet-close-i00-standard/instruction.md
@@ -51,6 +51,13 @@ res.video[cam]; res.frames[cam]; res.last[cam]; res.proprio; res.save("/app/runs
channel is an absolute target, so the grip sites stay within about a millimetre, but the IK's posture term can still
move the waist and elbows by a few hundredths of a radian while you watch.
- The cameras are mounted on the robot and move with it: use the calibration that comes with each response.
+- The hands: each hand command is the position target, in radians, of that finger's joints (`thumb_flexion` drives
+ two thumb joints, `thumb_rotation` one): 0 is the straight, open finger, larger values bend it toward the palm
+ (the hand closes), and every finger joint stops at the end of its range (1.22 to 1.74 rad), so values below 0 or
+ above about 1.7 change nothing more. The grip site that the pose targets place is the origin of the hand's mounting
+ frame (`gripper0_right_grip_site`, `gripper0_left_grip_site`), at the wrist end of the hand, not a point between
+ the fingertips. The four small pink spheres beside each hand are part of the robot model (visual markers
+ `robot0_<side>_pinch_spheres_0..3`, about 13.5 cm from its grip site, with no collision): they are not objects.
- World frame: metres, z up. In the proprio, `eef_left_*` / `eef_right_*` are the grip sites' positions in the world
frame; the action's pose targets and the `cmd_*` entries (the last command) are in the robot's base frame.
The robot's base frame (body `robot0_base`, fixed in the world) has its origin at world (2.65000, -2.17500, 0.97000) m and is turned +90.0 deg about z (quaternion w x y z (0.70711, 0, 0, 0.70711)): `p_world = R @ p_base + t` with R = [[0, -1, 0], [1, 0, 0], [0, 0, 1]].
diff --git a/tasks/robocasa-gr1-wine-to-cabinet-close-i00-standard/task.toml b/tasks/robocasa-gr1-wine-to-cabinet-close-i00-standard/task.toml
index 654d5496dc..6b7bac9791 100644
--- a/tasks/robocasa-gr1-wine-to-cabinet-close-i00-standard/task.toml
+++ b/tasks/robocasa-gr1-wine-to-cabinet-close-i00-standard/task.toml
@@ -24,6 +24,7 @@ simulator = "robosuite"
robot = "fourier-gr1"
family = "robocasa-gr1/PnPWineToCabinetClose"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
expert_time_estimate_hours = 4.0
diff --git a/tasks/robocasa-microwave-thawing-i00-privileged/task.toml b/tasks/robocasa-microwave-thawing-i00-privileged/task.toml
index 0f7e61c9f0..7a720ef25c 100644
--- a/tasks/robocasa-microwave-thawing-i00-privileged/task.toml
+++ b/tasks/robocasa-microwave-thawing-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa/MicrowaveThawing"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robocasa-microwave-thawing-i00-standard/task.toml b/tasks/robocasa-microwave-thawing-i00-standard/task.toml
index d4b1437b59..40e6486f57 100644
--- a/tasks/robocasa-microwave-thawing-i00-standard/task.toml
+++ b/tasks/robocasa-microwave-thawing-i00-standard/task.toml
@@ -24,6 +24,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa/MicrowaveThawing"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
expert_time_estimate_hours = 4.0
diff --git a/tasks/robocasa-open-double-door-i00-privileged/task.toml b/tasks/robocasa-open-double-door-i00-privileged/task.toml
index 084757999c..cd8764971d 100644
--- a/tasks/robocasa-open-double-door-i00-privileged/task.toml
+++ b/tasks/robocasa-open-double-door-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa/OpenDoubleDoor"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robocasa-open-double-door-i00-standard/task.toml b/tasks/robocasa-open-double-door-i00-standard/task.toml
index 60b0db64bc..3d0c980cbb 100644
--- a/tasks/robocasa-open-double-door-i00-standard/task.toml
+++ b/tasks/robocasa-open-double-door-i00-standard/task.toml
@@ -24,6 +24,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa/OpenDoubleDoor"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
expert_time_estimate_hours = 4.0
diff --git a/tasks/robocasa-open-drawer-i00-privileged/task.toml b/tasks/robocasa-open-drawer-i00-privileged/task.toml
index 64595d8848..3641ebdc58 100644
--- a/tasks/robocasa-open-drawer-i00-privileged/task.toml
+++ b/tasks/robocasa-open-drawer-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa/OpenDrawer"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robocasa-open-drawer-i00-standard/task.toml b/tasks/robocasa-open-drawer-i00-standard/task.toml
index fc6822c6f4..05be76f57b 100644
--- a/tasks/robocasa-open-drawer-i00-standard/task.toml
+++ b/tasks/robocasa-open-drawer-i00-standard/task.toml
@@ -24,6 +24,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa/OpenDrawer"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
expert_time_estimate_hours = 4.0
diff --git a/tasks/robocasa-open-single-door-i00-privileged/task.toml b/tasks/robocasa-open-single-door-i00-privileged/task.toml
index 72984aaa5c..0e8bbfd534 100644
--- a/tasks/robocasa-open-single-door-i00-privileged/task.toml
+++ b/tasks/robocasa-open-single-door-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa/OpenSingleDoor"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robocasa-open-single-door-i00-standard/task.toml b/tasks/robocasa-open-single-door-i00-standard/task.toml
index ee99e79091..06e628baa4 100644
--- a/tasks/robocasa-open-single-door-i00-standard/task.toml
+++ b/tasks/robocasa-open-single-door-i00-standard/task.toml
@@ -24,6 +24,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa/OpenSingleDoor"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
expert_time_estimate_hours = 4.0
diff --git a/tasks/robocasa-pnp-cab-to-counter-i00-privileged/task.toml b/tasks/robocasa-pnp-cab-to-counter-i00-privileged/task.toml
index ddd99cfc51..d64faca678 100644
--- a/tasks/robocasa-pnp-cab-to-counter-i00-privileged/task.toml
+++ b/tasks/robocasa-pnp-cab-to-counter-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa/PnPCabToCounter"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robocasa-pnp-cab-to-counter-i00-standard/task.toml b/tasks/robocasa-pnp-cab-to-counter-i00-standard/task.toml
index b2a2b76135..635b9388ec 100644
--- a/tasks/robocasa-pnp-cab-to-counter-i00-standard/task.toml
+++ b/tasks/robocasa-pnp-cab-to-counter-i00-standard/task.toml
@@ -24,6 +24,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa/PnPCabToCounter"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
expert_time_estimate_hours = 4.0
diff --git a/tasks/robocasa-pnp-counter-to-cab-i00-privileged/task.toml b/tasks/robocasa-pnp-counter-to-cab-i00-privileged/task.toml
index 6e5d50951c..826d163cc0 100644
--- a/tasks/robocasa-pnp-counter-to-cab-i00-privileged/task.toml
+++ b/tasks/robocasa-pnp-counter-to-cab-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa/PnPCounterToCab"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robocasa-pnp-counter-to-cab-i00-standard/task.toml b/tasks/robocasa-pnp-counter-to-cab-i00-standard/task.toml
index 553c7b03c1..134c84829e 100644
--- a/tasks/robocasa-pnp-counter-to-cab-i00-standard/task.toml
+++ b/tasks/robocasa-pnp-counter-to-cab-i00-standard/task.toml
@@ -24,6 +24,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa/PnPCounterToCab"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
expert_time_estimate_hours = 4.0
diff --git a/tasks/robocasa-pnp-counter-to-microwave-i00-privileged/task.toml b/tasks/robocasa-pnp-counter-to-microwave-i00-privileged/task.toml
index e5c6af1e5d..f1a0b2cf68 100644
--- a/tasks/robocasa-pnp-counter-to-microwave-i00-privileged/task.toml
+++ b/tasks/robocasa-pnp-counter-to-microwave-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa/PnPCounterToMicrowave"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robocasa-pnp-counter-to-microwave-i00-standard/task.toml b/tasks/robocasa-pnp-counter-to-microwave-i00-standard/task.toml
index a0648d03cb..c487bd8e7a 100644
--- a/tasks/robocasa-pnp-counter-to-microwave-i00-standard/task.toml
+++ b/tasks/robocasa-pnp-counter-to-microwave-i00-standard/task.toml
@@ -24,6 +24,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa/PnPCounterToMicrowave"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
expert_time_estimate_hours = 4.0
diff --git a/tasks/robocasa-pnp-counter-to-sink-i00-privileged/task.toml b/tasks/robocasa-pnp-counter-to-sink-i00-privileged/task.toml
index b528661b84..cb23d5b0a5 100644
--- a/tasks/robocasa-pnp-counter-to-sink-i00-privileged/task.toml
+++ b/tasks/robocasa-pnp-counter-to-sink-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa/PnPCounterToSink"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robocasa-pnp-counter-to-sink-i00-standard/task.toml b/tasks/robocasa-pnp-counter-to-sink-i00-standard/task.toml
index 1e8c310b2d..880dc70d04 100644
--- a/tasks/robocasa-pnp-counter-to-sink-i00-standard/task.toml
+++ b/tasks/robocasa-pnp-counter-to-sink-i00-standard/task.toml
@@ -24,6 +24,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa/PnPCounterToSink"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
expert_time_estimate_hours = 4.0
diff --git a/tasks/robocasa-pnp-counter-to-stove-i00-privileged/task.toml b/tasks/robocasa-pnp-counter-to-stove-i00-privileged/task.toml
index 08f613e4f1..217af334b8 100644
--- a/tasks/robocasa-pnp-counter-to-stove-i00-privileged/task.toml
+++ b/tasks/robocasa-pnp-counter-to-stove-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa/PnPCounterToStove"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robocasa-pnp-counter-to-stove-i00-standard/task.toml b/tasks/robocasa-pnp-counter-to-stove-i00-standard/task.toml
index 8e4b027410..1bb3e7a422 100644
--- a/tasks/robocasa-pnp-counter-to-stove-i00-standard/task.toml
+++ b/tasks/robocasa-pnp-counter-to-stove-i00-standard/task.toml
@@ -24,6 +24,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa/PnPCounterToStove"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
expert_time_estimate_hours = 4.0
diff --git a/tasks/robocasa-pnp-microwave-to-counter-i00-privileged/task.toml b/tasks/robocasa-pnp-microwave-to-counter-i00-privileged/task.toml
index 33ee638039..be67d38410 100644
--- a/tasks/robocasa-pnp-microwave-to-counter-i00-privileged/task.toml
+++ b/tasks/robocasa-pnp-microwave-to-counter-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa/PnPMicrowaveToCounter"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robocasa-pnp-microwave-to-counter-i00-standard/task.toml b/tasks/robocasa-pnp-microwave-to-counter-i00-standard/task.toml
index 34b0494a8a..f3c6171ca2 100644
--- a/tasks/robocasa-pnp-microwave-to-counter-i00-standard/task.toml
+++ b/tasks/robocasa-pnp-microwave-to-counter-i00-standard/task.toml
@@ -24,6 +24,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa/PnPMicrowaveToCounter"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
expert_time_estimate_hours = 4.0
diff --git a/tasks/robocasa-pnp-sink-to-counter-i00-privileged/task.toml b/tasks/robocasa-pnp-sink-to-counter-i00-privileged/task.toml
index a4086aa369..a165919621 100644
--- a/tasks/robocasa-pnp-sink-to-counter-i00-privileged/task.toml
+++ b/tasks/robocasa-pnp-sink-to-counter-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa/PnPSinkToCounter"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robocasa-pnp-sink-to-counter-i00-standard/task.toml b/tasks/robocasa-pnp-sink-to-counter-i00-standard/task.toml
index 469bfb6bb0..81635c6d99 100644
--- a/tasks/robocasa-pnp-sink-to-counter-i00-standard/task.toml
+++ b/tasks/robocasa-pnp-sink-to-counter-i00-standard/task.toml
@@ -24,6 +24,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa/PnPSinkToCounter"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
expert_time_estimate_hours = 4.0
diff --git a/tasks/robocasa-pnp-stove-to-counter-i00-privileged/task.toml b/tasks/robocasa-pnp-stove-to-counter-i00-privileged/task.toml
index 5aae7a94db..fe39c9be61 100644
--- a/tasks/robocasa-pnp-stove-to-counter-i00-privileged/task.toml
+++ b/tasks/robocasa-pnp-stove-to-counter-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa/PnPStoveToCounter"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robocasa-pnp-stove-to-counter-i00-standard/task.toml b/tasks/robocasa-pnp-stove-to-counter-i00-standard/task.toml
index 1761679470..ab85a5791e 100644
--- a/tasks/robocasa-pnp-stove-to-counter-i00-standard/task.toml
+++ b/tasks/robocasa-pnp-stove-to-counter-i00-standard/task.toml
@@ -24,6 +24,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa/PnPStoveToCounter"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
expert_time_estimate_hours = 4.0
diff --git a/tasks/robocasa-pre-soak-pan-i00-privileged/README.md b/tasks/robocasa-pre-soak-pan-i00-privileged/README.md
index a427f1a643..cef16f03c9 100644
--- a/tasks/robocasa-pre-soak-pan-i00-privileged/README.md
+++ b/tasks/robocasa-pre-soak-pan-i00-privileged/README.md
@@ -43,3 +43,6 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
| Instance | the initial state of official demonstration episode 0 (`demo_1`, v0.1/multi_stage/washing_dishes/PreSoakPan/2024-05-10); MuJoCo state, model arrays, task state and RNG frozen in `instance.npz` (SHA-256 `d4f029590af2…`, checked on load) |
| Deliverable | `(T, 12)` native controller commands, 1 ≤ T ≤ 3000, 20 Hz |
| Success | the environment's own check (`_check_success`) on the state after the last action |
+
+### Change Log
+- 2026-10-14 (board 0113, from board 0086): the success text now says that the faucet handle has two hinges at one pivot: the water on / off hinge (range 0 to 0.52 rad) and a temperature hinge (about ±0.79 rad) that never turns the water on or off, and that "off" means turned back toward 0 on that hinge, not lowered (board 0086: runs moved the temperature hinge or pressed the handle down). Text only: the scene, the frozen instance, the reference solution, the environment's success check and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/robocasa-pre-soak-pan-i00-privileged/instruction.md b/tasks/robocasa-pre-soak-pan-i00-privileged/instruction.md
index 4a2449e164..7c035ea473 100644
--- a/tasks/robocasa-pre-soak-pan-i00-privileged/instruction.md
+++ b/tasks/robocasa-pre-soak-pan-i00-privileged/instruction.md
@@ -18,7 +18,7 @@ achieved. `s.check()` returns the check's latest result.
What `PreSoakPan`'s check requires (RoboCasa @ 756598a), all on that final state:
- The pan and the sponge are each entirely inside the sink basin (all eight corners of its bounding box within the interior, with a 5 cm tolerance).
-- The water is on: the faucet handle's angle (mod 360°) is between 0.40 rad (about 23°) and π from its off position.
+- The water is on: the faucet handle's angle on its on / off hinge (mod 360°) is between 0.40 rad (about 23°) and π from its off position (the handle turns on two hinges at one pivot: the on / off hinge, whose range is 0 to 0.52 rad (so 0.40 rad is near the end of its travel), and a temperature hinge (about ±0.79 rad) that moves the handle in another direction and never turns the water on or off; which way each hinge turns depends on the sink model).
- The gripper's grip site (between the fingertips) is more than 25 cm from the pan's centre: let go and move the hand away.
A "centre" is a body's origin; "touches" means a MuJoCo contact.
diff --git a/tasks/robocasa-pre-soak-pan-i00-privileged/task.toml b/tasks/robocasa-pre-soak-pan-i00-privileged/task.toml
index dbc2fe4d28..b1c3735ee3 100644
--- a/tasks/robocasa-pre-soak-pan-i00-privileged/task.toml
+++ b/tasks/robocasa-pre-soak-pan-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa/PreSoakPan"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robocasa-pre-soak-pan-i00-standard/README.md b/tasks/robocasa-pre-soak-pan-i00-standard/README.md
index a979c70f03..128d49e4ad 100644
--- a/tasks/robocasa-pre-soak-pan-i00-standard/README.md
+++ b/tasks/robocasa-pre-soak-pan-i00-standard/README.md
@@ -48,3 +48,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
- 2026-10-10 (review round 3, board 0067): the sim service's request handling (`environment/sim/server_casa.py`) and `tests/grade.py`. A negative or non-numeric `Content-Length` is answered 400 without reading the body (the service read until the client hung up, without a bound, which could exhaust its memory); a camera named more than once in `cameras` is rendered and returned once (N copies of one name made it build N x N images within the frame budget); a port that is still held when the service restarts is waited for, not counted as a crash; a JSON `true` or `false` where `dt` or `seconds` takes a number is a 400 (`true` ran as 1 second). The grader: a `final.json` whose evidence fields (`control_steps`, `calls`, `elapsed_s`, `final_reward`) are not numbers no longer stops it before it writes a reward (they are reported as 0 and decide nothing). An episode's physics and the pass rule are unchanged.
- 2026-10-13 (owner, boards 0085 and 0087): the sim service's clock (`environment/sim/server_casa.py`, the docstring of `environment/rcb_client.py`). `meta["time_s"]` and `meta["time_left_s"]` count from the agent's first request to the service (any endpoint), and a restarted service keeps that moment; before, they counted from the first start of the service's container (`time_left_s`) and from when its scene was ready (`time_s`). Where the service is ready before the agent's phase begins (the harness is installed in between: a median of 79 s over the recorded standard trials), that clock ran ahead of the agent's and `time_left_s` reached 0 while the agent still had that time. `/spec` says what the clock counts from (`budget.clock`), and the instruction's `## Time limit` tells the agent to plan by `time_left_s`. The service never ended an episode on this clock: the scene, the frozen instance, the pass rule and the grader are unchanged, so the task keeps its name and earlier results stand.
- 2026-10-13 (board 0093; found in board 0086): the camera calibration (`environment/sim/server_casa.py`). `cam2world_gl` held the pose of MuJoCo's left-eye camera (`scene.camera[0]`), 3.4 cm (half the model's 0.068 m interpupillary distance) to the camera's left of the viewpoint the mono RGB and depth images are rendered from; it now is that viewpoint (`data.cam_xpos` / `data.cam_xmat`). Checked against MuJoCo's own renders of marker spheres at known world points, every served camera, before and after the robot moved: the reprojection error was 10.5-10.7 px median (up to 28 px for near points, always to the same side), now 0.15 px median and at most 0.4 px. The intrinsics were right (every camera here is given by its `fovy`, none by a sensor size) and are unchanged, as are the images, the scene, the frozen instance, the pass rule and the grader, so the task keeps its name. **Standard-mode results from before this date used the offset calibration**: a point located from the images and the reported calibration came out about 3.4 cm to the side (agents missed a microwave button by 3.5 cm), so they are not strictly comparable with later ones.
+- 2026-10-14 (board 0113, from board 0086): the success text now says that the faucet handle has two hinges at one pivot: the water on / off hinge (range 0 to 0.52 rad) and a temperature hinge (about ±0.79 rad) that never turns the water on or off, and that "off" means turned back toward 0 on that hinge, not lowered (board 0086: runs moved the temperature hinge or pressed the handle down). Text only: the scene, the frozen instance, the reference solution, the environment's success check and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/robocasa-pre-soak-pan-i00-standard/instruction.md b/tasks/robocasa-pre-soak-pan-i00-standard/instruction.md
index fd117f07ed..01f4bcb502 100644
--- a/tasks/robocasa-pre-soak-pan-i00-standard/instruction.md
+++ b/tasks/robocasa-pre-soak-pan-i00-standard/instruction.md
@@ -25,7 +25,7 @@ want graded and stop submitting.
What `PreSoakPan`'s check requires (RoboCasa @ 756598a), all on that final state:
- The pan and the sponge are each entirely inside the sink basin (all eight corners of its bounding box within the interior, with a 5 cm tolerance).
-- The water is on: the faucet handle's angle (mod 360°) is between 0.40 rad (about 23°) and π from its off position.
+- The water is on: the faucet handle's angle on its on / off hinge (mod 360°) is between 0.40 rad (about 23°) and π from its off position (the handle turns on two hinges at one pivot: the on / off hinge, whose range is 0 to 0.52 rad (so 0.40 rad is near the end of its travel), and a temperature hinge (about ±0.79 rad) that moves the handle in another direction and never turns the water on or off; which way each hinge turns depends on the sink model).
- The gripper's grip site (between the fingertips) is more than 25 cm from the pan's centre: let go and move the hand away.
A "centre" is a body's origin; "touches" means a MuJoCo contact.
diff --git a/tasks/robocasa-pre-soak-pan-i00-standard/task.toml b/tasks/robocasa-pre-soak-pan-i00-standard/task.toml
index b40cf0e51a..02b3c46186 100644
--- a/tasks/robocasa-pre-soak-pan-i00-standard/task.toml
+++ b/tasks/robocasa-pre-soak-pan-i00-standard/task.toml
@@ -24,6 +24,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa/PreSoakPan"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
expert_time_estimate_hours = 4.0
diff --git a/tasks/robocasa-prepare-coffee-i00-privileged/task.toml b/tasks/robocasa-prepare-coffee-i00-privileged/task.toml
index eeb37b6b19..463991e3ae 100644
--- a/tasks/robocasa-prepare-coffee-i00-privileged/task.toml
+++ b/tasks/robocasa-prepare-coffee-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa/PrepareCoffee"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robocasa-prepare-coffee-i00-standard/task.toml b/tasks/robocasa-prepare-coffee-i00-standard/task.toml
index e2d365ca9a..a68b18fad0 100644
--- a/tasks/robocasa-prepare-coffee-i00-standard/task.toml
+++ b/tasks/robocasa-prepare-coffee-i00-standard/task.toml
@@ -24,6 +24,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa/PrepareCoffee"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
expert_time_estimate_hours = 4.0
diff --git a/tasks/robocasa-restock-pantry-i00-privileged/task.toml b/tasks/robocasa-restock-pantry-i00-privileged/task.toml
index 81da63af27..a03cba3a0f 100644
--- a/tasks/robocasa-restock-pantry-i00-privileged/task.toml
+++ b/tasks/robocasa-restock-pantry-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa/RestockPantry"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robocasa-restock-pantry-i00-standard/task.toml b/tasks/robocasa-restock-pantry-i00-standard/task.toml
index 84dd3dbc00..e1f76a7321 100644
--- a/tasks/robocasa-restock-pantry-i00-standard/task.toml
+++ b/tasks/robocasa-restock-pantry-i00-standard/task.toml
@@ -24,6 +24,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa/RestockPantry"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
expert_time_estimate_hours = 4.0
diff --git a/tasks/robocasa-turn-off-microwave-i00-privileged/task.toml b/tasks/robocasa-turn-off-microwave-i00-privileged/task.toml
index f830bb0466..61ded6296d 100644
--- a/tasks/robocasa-turn-off-microwave-i00-privileged/task.toml
+++ b/tasks/robocasa-turn-off-microwave-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa/TurnOffMicrowave"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robocasa-turn-off-microwave-i00-standard/task.toml b/tasks/robocasa-turn-off-microwave-i00-standard/task.toml
index d9125086cf..74b2340c52 100644
--- a/tasks/robocasa-turn-off-microwave-i00-standard/task.toml
+++ b/tasks/robocasa-turn-off-microwave-i00-standard/task.toml
@@ -24,6 +24,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa/TurnOffMicrowave"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
expert_time_estimate_hours = 4.0
diff --git a/tasks/robocasa-turn-off-sink-faucet-i00-privileged/README.md b/tasks/robocasa-turn-off-sink-faucet-i00-privileged/README.md
index 571ecbf69a..7a0646c0dd 100644
--- a/tasks/robocasa-turn-off-sink-faucet-i00-privileged/README.md
+++ b/tasks/robocasa-turn-off-sink-faucet-i00-privileged/README.md
@@ -43,3 +43,6 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
| Instance | the initial state of official demonstration episode 0 (`demo_0`, v0.1/single_stage/kitchen_sink/TurnOffSinkFaucet/2024-04-25); MuJoCo state, model arrays, task state and RNG frozen in `instance.npz` (SHA-256 `24e4e530d062…`, checked on load) |
| Deliverable | `(T, 12)` native controller commands, 1 ≤ T ≤ 3000, 20 Hz |
| Success | the environment's own check (`_check_success`) on the state after the last action |
+
+### Change Log
+- 2026-10-14 (board 0113, from board 0086): the success text now says that the faucet handle has two hinges at one pivot: the water on / off hinge (range 0 to 0.52 rad) and a temperature hinge (about ±0.79 rad) that never turns the water on or off, and that "off" means turned back toward 0 on that hinge, not lowered (board 0086: runs moved the temperature hinge or pressed the handle down). Text only: the scene, the frozen instance, the reference solution, the environment's success check and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/robocasa-turn-off-sink-faucet-i00-privileged/instruction.md b/tasks/robocasa-turn-off-sink-faucet-i00-privileged/instruction.md
index fa23a45d64..62a2ebfc7f 100644
--- a/tasks/robocasa-turn-off-sink-faucet-i00-privileged/instruction.md
+++ b/tasks/robocasa-turn-off-sink-faucet-i00-privileged/instruction.md
@@ -17,7 +17,7 @@ the last action**. A success reached earlier and then lost does not count: end t
achieved. `s.check()` returns the check's latest result.
What `TurnOffSinkFaucet`'s check requires (RoboCasa @ 756598a), all on that final state:
-- The water is off: the faucet handle's angle (mod 360°) is outside 0.40 rad (about 23°) to π from its off position, i.e. lowered back to below 0.40 rad.
+- The water is off: the faucet handle's angle on its on / off hinge (mod 360°) is outside 0.40 rad (about 23°) to π from its off position, i.e. turned back toward 0 to below 0.40 rad (it starts at about 0.47 rad; the handle turns on two hinges at one pivot: the on / off hinge, whose range is 0 to 0.52 rad (so 0.40 rad is near the end of its travel), and a temperature hinge (about ±0.79 rad) that moves the handle in another direction and never turns the water on or off; which way each hinge turns depends on the sink model; what counts is that hinge's angle, not how high the handle is).
## Tooling (`/opt/tools/casa_tool.py`, importable as `casa_tool`)
```python
diff --git a/tasks/robocasa-turn-off-sink-faucet-i00-privileged/task.toml b/tasks/robocasa-turn-off-sink-faucet-i00-privileged/task.toml
index 3fd964c1fb..9d72491ecd 100644
--- a/tasks/robocasa-turn-off-sink-faucet-i00-privileged/task.toml
+++ b/tasks/robocasa-turn-off-sink-faucet-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa/TurnOffSinkFaucet"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robocasa-turn-off-sink-faucet-i00-standard/README.md b/tasks/robocasa-turn-off-sink-faucet-i00-standard/README.md
index d280c91e47..a0aa1c4bad 100644
--- a/tasks/robocasa-turn-off-sink-faucet-i00-standard/README.md
+++ b/tasks/robocasa-turn-off-sink-faucet-i00-standard/README.md
@@ -48,3 +48,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
- 2026-10-10 (review round 3, board 0067): the sim service's request handling (`environment/sim/server_casa.py`) and `tests/grade.py`. A negative or non-numeric `Content-Length` is answered 400 without reading the body (the service read until the client hung up, without a bound, which could exhaust its memory); a camera named more than once in `cameras` is rendered and returned once (N copies of one name made it build N x N images within the frame budget); a port that is still held when the service restarts is waited for, not counted as a crash; a JSON `true` or `false` where `dt` or `seconds` takes a number is a 400 (`true` ran as 1 second). The grader: a `final.json` whose evidence fields (`control_steps`, `calls`, `elapsed_s`, `final_reward`) are not numbers no longer stops it before it writes a reward (they are reported as 0 and decide nothing). An episode's physics and the pass rule are unchanged.
- 2026-10-13 (owner, boards 0085 and 0087): the sim service's clock (`environment/sim/server_casa.py`, the docstring of `environment/rcb_client.py`). `meta["time_s"]` and `meta["time_left_s"]` count from the agent's first request to the service (any endpoint), and a restarted service keeps that moment; before, they counted from the first start of the service's container (`time_left_s`) and from when its scene was ready (`time_s`). Where the service is ready before the agent's phase begins (the harness is installed in between: a median of 79 s over the recorded standard trials), that clock ran ahead of the agent's and `time_left_s` reached 0 while the agent still had that time. `/spec` says what the clock counts from (`budget.clock`), and the instruction's `## Time limit` tells the agent to plan by `time_left_s`. The service never ended an episode on this clock: the scene, the frozen instance, the pass rule and the grader are unchanged, so the task keeps its name and earlier results stand.
- 2026-10-13 (board 0093; found in board 0086): the camera calibration (`environment/sim/server_casa.py`). `cam2world_gl` held the pose of MuJoCo's left-eye camera (`scene.camera[0]`), 3.4 cm (half the model's 0.068 m interpupillary distance) to the camera's left of the viewpoint the mono RGB and depth images are rendered from; it now is that viewpoint (`data.cam_xpos` / `data.cam_xmat`). Checked against MuJoCo's own renders of marker spheres at known world points, every served camera, before and after the robot moved: the reprojection error was 10.5-10.7 px median (up to 28 px for near points, always to the same side), now 0.15 px median and at most 0.4 px. The intrinsics were right (every camera here is given by its `fovy`, none by a sensor size) and are unchanged, as are the images, the scene, the frozen instance, the pass rule and the grader, so the task keeps its name. **Standard-mode results from before this date used the offset calibration**: a point located from the images and the reported calibration came out about 3.4 cm to the side (agents missed a microwave button by 3.5 cm), so they are not strictly comparable with later ones.
+- 2026-10-14 (board 0113, from board 0086): the success text now says that the faucet handle has two hinges at one pivot: the water on / off hinge (range 0 to 0.52 rad) and a temperature hinge (about ±0.79 rad) that never turns the water on or off, and that "off" means turned back toward 0 on that hinge, not lowered (board 0086: runs moved the temperature hinge or pressed the handle down). Text only: the scene, the frozen instance, the reference solution, the environment's success check and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/robocasa-turn-off-sink-faucet-i00-standard/instruction.md b/tasks/robocasa-turn-off-sink-faucet-i00-standard/instruction.md
index 6fb07e50c5..f8cf5f8478 100644
--- a/tasks/robocasa-turn-off-sink-faucet-i00-standard/instruction.md
+++ b/tasks/robocasa-turn-off-sink-faucet-i00-standard/instruction.md
@@ -24,7 +24,7 @@ in the live episode and in the replay. A success reached earlier and then lost d
want graded and stop submitting.
What `TurnOffSinkFaucet`'s check requires (RoboCasa @ 756598a), all on that final state:
-- The water is off: the faucet handle's angle (mod 360°) is outside 0.40 rad (about 23°) to π from its off position, i.e. lowered back to below 0.40 rad.
+- The water is off: the faucet handle's angle on its on / off hinge (mod 360°) is outside 0.40 rad (about 23°) to π from its off position, i.e. turned back toward 0 to below 0.40 rad (it starts at about 0.47 rad; the handle turns on two hinges at one pivot: the on / off hinge, whose range is 0 to 0.52 rad (so 0.40 rad is near the end of its travel), and a temperature hinge (about ±0.79 rad) that moves the handle in another direction and never turns the water on or off; which way each hinge turns depends on the sink model; what counts is that hinge's angle, not how high the handle is).
## Client (`/opt/tools/rcb_client.py`, protocol `eai-standard/2.1`)
```python
diff --git a/tasks/robocasa-turn-off-sink-faucet-i00-standard/task.toml b/tasks/robocasa-turn-off-sink-faucet-i00-standard/task.toml
index 7fdc86f0e0..1458404709 100644
--- a/tasks/robocasa-turn-off-sink-faucet-i00-standard/task.toml
+++ b/tasks/robocasa-turn-off-sink-faucet-i00-standard/task.toml
@@ -24,6 +24,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa/TurnOffSinkFaucet"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
expert_time_estimate_hours = 4.0
diff --git a/tasks/robocasa-turn-off-stove-i00-privileged/task.toml b/tasks/robocasa-turn-off-stove-i00-privileged/task.toml
index 60a317c76f..826ce9d6e4 100644
--- a/tasks/robocasa-turn-off-stove-i00-privileged/task.toml
+++ b/tasks/robocasa-turn-off-stove-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa/TurnOffStove"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robocasa-turn-off-stove-i00-standard/task.toml b/tasks/robocasa-turn-off-stove-i00-standard/task.toml
index 378534466a..2cf4a6be23 100644
--- a/tasks/robocasa-turn-off-stove-i00-standard/task.toml
+++ b/tasks/robocasa-turn-off-stove-i00-standard/task.toml
@@ -24,6 +24,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa/TurnOffStove"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
expert_time_estimate_hours = 4.0
diff --git a/tasks/robocasa-turn-on-microwave-i00-privileged/task.toml b/tasks/robocasa-turn-on-microwave-i00-privileged/task.toml
index 43fa444bf8..ff246e6091 100644
--- a/tasks/robocasa-turn-on-microwave-i00-privileged/task.toml
+++ b/tasks/robocasa-turn-on-microwave-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa/TurnOnMicrowave"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robocasa-turn-on-microwave-i00-standard/task.toml b/tasks/robocasa-turn-on-microwave-i00-standard/task.toml
index 8f9f65f634..f60f5ff574 100644
--- a/tasks/robocasa-turn-on-microwave-i00-standard/task.toml
+++ b/tasks/robocasa-turn-on-microwave-i00-standard/task.toml
@@ -24,6 +24,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa/TurnOnMicrowave"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
expert_time_estimate_hours = 4.0
diff --git a/tasks/robocasa-turn-on-sink-faucet-i00-privileged/README.md b/tasks/robocasa-turn-on-sink-faucet-i00-privileged/README.md
index d705af757f..d256b4f5a2 100644
--- a/tasks/robocasa-turn-on-sink-faucet-i00-privileged/README.md
+++ b/tasks/robocasa-turn-on-sink-faucet-i00-privileged/README.md
@@ -43,3 +43,6 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
| Instance | the initial state of official demonstration episode 0 (`demo_1`, v0.1/single_stage/kitchen_sink/TurnOnSinkFaucet/2024-04-25); MuJoCo state, model arrays, task state and RNG frozen in `instance.npz` (SHA-256 `36053cc46ba6…`, checked on load) |
| Deliverable | `(T, 12)` native controller commands, 1 ≤ T ≤ 3000, 20 Hz |
| Success | the environment's own check (`_check_success`) on the state after the last action |
+
+### Change Log
+- 2026-10-14 (board 0113, from board 0086): the success text now says that the faucet handle has two hinges at one pivot: the water on / off hinge (range 0 to 0.52 rad) and a temperature hinge (about ±0.79 rad) that never turns the water on or off, and that "off" means turned back toward 0 on that hinge, not lowered (board 0086: runs moved the temperature hinge or pressed the handle down). Text only: the scene, the frozen instance, the reference solution, the environment's success check and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/robocasa-turn-on-sink-faucet-i00-privileged/instruction.md b/tasks/robocasa-turn-on-sink-faucet-i00-privileged/instruction.md
index 4ca8c7f872..54ddcf1e60 100644
--- a/tasks/robocasa-turn-on-sink-faucet-i00-privileged/instruction.md
+++ b/tasks/robocasa-turn-on-sink-faucet-i00-privileged/instruction.md
@@ -17,7 +17,7 @@ the last action**. A success reached earlier and then lost does not count: end t
achieved. `s.check()` returns the check's latest result.
What `TurnOnSinkFaucet`'s check requires (RoboCasa @ 756598a), all on that final state:
-- The water is on: the faucet handle's angle (mod 360°) is between 0.40 rad (about 23°) and π from its off position.
+- The water is on: the faucet handle's angle on its on / off hinge (mod 360°) is between 0.40 rad (about 23°) and π from its off position (the handle turns on two hinges at one pivot: the on / off hinge, whose range is 0 to 0.52 rad (so 0.40 rad is near the end of its travel), and a temperature hinge (about ±0.79 rad) that moves the handle in another direction and never turns the water on or off; which way each hinge turns depends on the sink model).
## Tooling (`/opt/tools/casa_tool.py`, importable as `casa_tool`)
```python
diff --git a/tasks/robocasa-turn-on-sink-faucet-i00-privileged/task.toml b/tasks/robocasa-turn-on-sink-faucet-i00-privileged/task.toml
index 3ace80de75..961474c5a9 100644
--- a/tasks/robocasa-turn-on-sink-faucet-i00-privileged/task.toml
+++ b/tasks/robocasa-turn-on-sink-faucet-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa/TurnOnSinkFaucet"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robocasa-turn-on-sink-faucet-i00-standard/README.md b/tasks/robocasa-turn-on-sink-faucet-i00-standard/README.md
index a620628de6..8c9617ed92 100644
--- a/tasks/robocasa-turn-on-sink-faucet-i00-standard/README.md
+++ b/tasks/robocasa-turn-on-sink-faucet-i00-standard/README.md
@@ -48,3 +48,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
- 2026-10-10 (review round 3, board 0067): the sim service's request handling (`environment/sim/server_casa.py`) and `tests/grade.py`. A negative or non-numeric `Content-Length` is answered 400 without reading the body (the service read until the client hung up, without a bound, which could exhaust its memory); a camera named more than once in `cameras` is rendered and returned once (N copies of one name made it build N x N images within the frame budget); a port that is still held when the service restarts is waited for, not counted as a crash; a JSON `true` or `false` where `dt` or `seconds` takes a number is a 400 (`true` ran as 1 second). The grader: a `final.json` whose evidence fields (`control_steps`, `calls`, `elapsed_s`, `final_reward`) are not numbers no longer stops it before it writes a reward (they are reported as 0 and decide nothing). An episode's physics and the pass rule are unchanged.
- 2026-10-13 (owner, boards 0085 and 0087): the sim service's clock (`environment/sim/server_casa.py`, the docstring of `environment/rcb_client.py`). `meta["time_s"]` and `meta["time_left_s"]` count from the agent's first request to the service (any endpoint), and a restarted service keeps that moment; before, they counted from the first start of the service's container (`time_left_s`) and from when its scene was ready (`time_s`). Where the service is ready before the agent's phase begins (the harness is installed in between: a median of 79 s over the recorded standard trials), that clock ran ahead of the agent's and `time_left_s` reached 0 while the agent still had that time. `/spec` says what the clock counts from (`budget.clock`), and the instruction's `## Time limit` tells the agent to plan by `time_left_s`. The service never ended an episode on this clock: the scene, the frozen instance, the pass rule and the grader are unchanged, so the task keeps its name and earlier results stand.
- 2026-10-13 (board 0093; found in board 0086): the camera calibration (`environment/sim/server_casa.py`). `cam2world_gl` held the pose of MuJoCo's left-eye camera (`scene.camera[0]`), 3.4 cm (half the model's 0.068 m interpupillary distance) to the camera's left of the viewpoint the mono RGB and depth images are rendered from; it now is that viewpoint (`data.cam_xpos` / `data.cam_xmat`). Checked against MuJoCo's own renders of marker spheres at known world points, every served camera, before and after the robot moved: the reprojection error was 10.5-10.7 px median (up to 28 px for near points, always to the same side), now 0.15 px median and at most 0.4 px. The intrinsics were right (every camera here is given by its `fovy`, none by a sensor size) and are unchanged, as are the images, the scene, the frozen instance, the pass rule and the grader, so the task keeps its name. **Standard-mode results from before this date used the offset calibration**: a point located from the images and the reported calibration came out about 3.4 cm to the side (agents missed a microwave button by 3.5 cm), so they are not strictly comparable with later ones.
+- 2026-10-14 (board 0113, from board 0086): the success text now says that the faucet handle has two hinges at one pivot: the water on / off hinge (range 0 to 0.52 rad) and a temperature hinge (about ±0.79 rad) that never turns the water on or off, and that "off" means turned back toward 0 on that hinge, not lowered (board 0086: runs moved the temperature hinge or pressed the handle down). Text only: the scene, the frozen instance, the reference solution, the environment's success check and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/robocasa-turn-on-sink-faucet-i00-standard/instruction.md b/tasks/robocasa-turn-on-sink-faucet-i00-standard/instruction.md
index 5130f538a5..b2ba5a90c3 100644
--- a/tasks/robocasa-turn-on-sink-faucet-i00-standard/instruction.md
+++ b/tasks/robocasa-turn-on-sink-faucet-i00-standard/instruction.md
@@ -24,7 +24,7 @@ in the live episode and in the replay. A success reached earlier and then lost d
want graded and stop submitting.
What `TurnOnSinkFaucet`'s check requires (RoboCasa @ 756598a), all on that final state:
-- The water is on: the faucet handle's angle (mod 360°) is between 0.40 rad (about 23°) and π from its off position.
+- The water is on: the faucet handle's angle on its on / off hinge (mod 360°) is between 0.40 rad (about 23°) and π from its off position (the handle turns on two hinges at one pivot: the on / off hinge, whose range is 0 to 0.52 rad (so 0.40 rad is near the end of its travel), and a temperature hinge (about ±0.79 rad) that moves the handle in another direction and never turns the water on or off; which way each hinge turns depends on the sink model).
## Client (`/opt/tools/rcb_client.py`, protocol `eai-standard/2.1`)
```python
diff --git a/tasks/robocasa-turn-on-sink-faucet-i00-standard/task.toml b/tasks/robocasa-turn-on-sink-faucet-i00-standard/task.toml
index e0ddc75374..e9293dcb31 100644
--- a/tasks/robocasa-turn-on-sink-faucet-i00-standard/task.toml
+++ b/tasks/robocasa-turn-on-sink-faucet-i00-standard/task.toml
@@ -24,6 +24,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa/TurnOnSinkFaucet"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
expert_time_estimate_hours = 4.0
diff --git a/tasks/robocasa-turn-on-stove-i00-privileged/task.toml b/tasks/robocasa-turn-on-stove-i00-privileged/task.toml
index aff4448402..45f435ea1f 100644
--- a/tasks/robocasa-turn-on-stove-i00-privileged/task.toml
+++ b/tasks/robocasa-turn-on-stove-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa/TurnOnStove"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robocasa-turn-on-stove-i00-standard/task.toml b/tasks/robocasa-turn-on-stove-i00-standard/task.toml
index 2f8cff6015..1665cec79f 100644
--- a/tasks/robocasa-turn-on-stove-i00-standard/task.toml
+++ b/tasks/robocasa-turn-on-stove-i00-standard/task.toml
@@ -24,6 +24,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa/TurnOnStove"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
expert_time_estimate_hours = 4.0
diff --git a/tasks/robocasa-turn-sink-spout-i00-privileged/task.toml b/tasks/robocasa-turn-sink-spout-i00-privileged/task.toml
index 2be74187c1..5db19b2191 100644
--- a/tasks/robocasa-turn-sink-spout-i00-privileged/task.toml
+++ b/tasks/robocasa-turn-sink-spout-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa/TurnSinkSpout"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robocasa-turn-sink-spout-i00-standard/task.toml b/tasks/robocasa-turn-sink-spout-i00-standard/task.toml
index 50818f070d..3025c823b6 100644
--- a/tasks/robocasa-turn-sink-spout-i00-standard/task.toml
+++ b/tasks/robocasa-turn-sink-spout-i00-standard/task.toml
@@ -24,6 +24,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa/TurnSinkSpout"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
expert_time_estimate_hours = 4.0
diff --git a/tasks/robocasa365-add-lemon-to-fish-i00-privileged/task.toml b/tasks/robocasa365-add-lemon-to-fish-i00-privileged/task.toml
index 0eeae430fa..d7bfaef842 100644
--- a/tasks/robocasa365-add-lemon-to-fish-i00-privileged/task.toml
+++ b/tasks/robocasa365-add-lemon-to-fish-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa365/AddLemonToFish"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robocasa365-add-lemon-to-fish-i00-standard/task.toml b/tasks/robocasa365-add-lemon-to-fish-i00-standard/task.toml
index 893e495814..93caf45e7c 100644
--- a/tasks/robocasa365-add-lemon-to-fish-i00-standard/task.toml
+++ b/tasks/robocasa365-add-lemon-to-fish-i00-standard/task.toml
@@ -24,6 +24,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa365/AddLemonToFish"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
expert_time_estimate_hours = 4.0
diff --git a/tasks/robocasa365-assemble-cooking-array-i00-privileged/task.toml b/tasks/robocasa365-assemble-cooking-array-i00-privileged/task.toml
index 372b06a902..f65d2ebf5f 100644
--- a/tasks/robocasa365-assemble-cooking-array-i00-privileged/task.toml
+++ b/tasks/robocasa365-assemble-cooking-array-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa365/AssembleCookingArray"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robocasa365-assemble-cooking-array-i00-standard/task.toml b/tasks/robocasa365-assemble-cooking-array-i00-standard/task.toml
index c6a816cde2..f06cf2bc0d 100644
--- a/tasks/robocasa365-assemble-cooking-array-i00-standard/task.toml
+++ b/tasks/robocasa365-assemble-cooking-array-i00-standard/task.toml
@@ -24,6 +24,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa365/AssembleCookingArray"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
expert_time_estimate_hours = 4.0
diff --git a/tasks/robocasa365-beverage-sorting-i00-privileged/task.toml b/tasks/robocasa365-beverage-sorting-i00-privileged/task.toml
index 0f1dcd9866..ffc909416e 100644
--- a/tasks/robocasa365-beverage-sorting-i00-privileged/task.toml
+++ b/tasks/robocasa365-beverage-sorting-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa365/BeverageSorting"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robocasa365-beverage-sorting-i00-standard/task.toml b/tasks/robocasa365-beverage-sorting-i00-standard/task.toml
index 98e314c7cd..dc9d5a15fd 100644
--- a/tasks/robocasa365-beverage-sorting-i00-standard/task.toml
+++ b/tasks/robocasa365-beverage-sorting-i00-standard/task.toml
@@ -24,6 +24,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa365/BeverageSorting"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
expert_time_estimate_hours = 4.0
diff --git a/tasks/robocasa365-bread-and-cheese-i00-privileged/task.toml b/tasks/robocasa365-bread-and-cheese-i00-privileged/task.toml
index a93e946751..f6e245b481 100644
--- a/tasks/robocasa365-bread-and-cheese-i00-privileged/task.toml
+++ b/tasks/robocasa365-bread-and-cheese-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa365/BreadAndCheese"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robocasa365-bread-and-cheese-i00-standard/task.toml b/tasks/robocasa365-bread-and-cheese-i00-standard/task.toml
index 5722a05b1e..59b83e451f 100644
--- a/tasks/robocasa365-bread-and-cheese-i00-standard/task.toml
+++ b/tasks/robocasa365-bread-and-cheese-i00-standard/task.toml
@@ -24,6 +24,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa365/BreadAndCheese"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
expert_time_estimate_hours = 4.0
diff --git a/tasks/robocasa365-candle-cleanup-i00-privileged/task.toml b/tasks/robocasa365-candle-cleanup-i00-privileged/task.toml
index 059d88f71f..9c2732b663 100644
--- a/tasks/robocasa365-candle-cleanup-i00-privileged/task.toml
+++ b/tasks/robocasa365-candle-cleanup-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa365/CandleCleanup"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robocasa365-candle-cleanup-i00-standard/task.toml b/tasks/robocasa365-candle-cleanup-i00-standard/task.toml
index ec03c97668..dbb9200ace 100644
--- a/tasks/robocasa365-candle-cleanup-i00-standard/task.toml
+++ b/tasks/robocasa365-candle-cleanup-i00-standard/task.toml
@@ -24,6 +24,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa365/CandleCleanup"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
expert_time_estimate_hours = 4.0
diff --git a/tasks/robocasa365-cereal-and-bowl-i00-privileged/task.toml b/tasks/robocasa365-cereal-and-bowl-i00-privileged/task.toml
index 0288bfffb7..c600e9a61f 100644
--- a/tasks/robocasa365-cereal-and-bowl-i00-privileged/task.toml
+++ b/tasks/robocasa365-cereal-and-bowl-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa365/CerealAndBowl"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robocasa365-cereal-and-bowl-i00-standard/task.toml b/tasks/robocasa365-cereal-and-bowl-i00-standard/task.toml
index 94182bfdc5..c851a8a50d 100644
--- a/tasks/robocasa365-cereal-and-bowl-i00-standard/task.toml
+++ b/tasks/robocasa365-cereal-and-bowl-i00-standard/task.toml
@@ -24,6 +24,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa365/CerealAndBowl"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
expert_time_estimate_hours = 4.0
diff --git a/tasks/robocasa365-choose-measuring-cup-i00-privileged/task.toml b/tasks/robocasa365-choose-measuring-cup-i00-privileged/task.toml
index b6f542284f..b531ebf926 100644
--- a/tasks/robocasa365-choose-measuring-cup-i00-privileged/task.toml
+++ b/tasks/robocasa365-choose-measuring-cup-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa365/ChooseMeasuringCup"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robocasa365-choose-measuring-cup-i00-standard/task.toml b/tasks/robocasa365-choose-measuring-cup-i00-standard/task.toml
index c0f3b665f1..fc2b8e984a 100644
--- a/tasks/robocasa365-choose-measuring-cup-i00-standard/task.toml
+++ b/tasks/robocasa365-choose-measuring-cup-i00-standard/task.toml
@@ -24,6 +24,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa365/ChooseMeasuringCup"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
expert_time_estimate_hours = 4.0
diff --git a/tasks/robocasa365-clear-food-waste-i00-privileged/task.toml b/tasks/robocasa365-clear-food-waste-i00-privileged/task.toml
index f6f2ede6dd..1b9ccff174 100644
--- a/tasks/robocasa365-clear-food-waste-i00-privileged/task.toml
+++ b/tasks/robocasa365-clear-food-waste-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa365/ClearFoodWaste"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robocasa365-clear-food-waste-i00-standard/task.toml b/tasks/robocasa365-clear-food-waste-i00-standard/task.toml
index a035999804..9d46c2ab74 100644
--- a/tasks/robocasa365-clear-food-waste-i00-standard/task.toml
+++ b/tasks/robocasa365-clear-food-waste-i00-standard/task.toml
@@ -24,6 +24,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa365/ClearFoodWaste"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
expert_time_estimate_hours = 4.0
diff --git a/tasks/robocasa365-clear-sink-i00-privileged/task.toml b/tasks/robocasa365-clear-sink-i00-privileged/task.toml
index 81c33bd9ad..9add65137a 100644
--- a/tasks/robocasa365-clear-sink-i00-privileged/task.toml
+++ b/tasks/robocasa365-clear-sink-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa365/ClearSink"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robocasa365-clear-sink-i00-standard/task.toml b/tasks/robocasa365-clear-sink-i00-standard/task.toml
index 1f9cd76549..aa6ecff630 100644
--- a/tasks/robocasa365-clear-sink-i00-standard/task.toml
+++ b/tasks/robocasa365-clear-sink-i00-standard/task.toml
@@ -24,6 +24,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa365/ClearSink"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
expert_time_estimate_hours = 4.0
diff --git a/tasks/robocasa365-close-drawer-i00-privileged/task.toml b/tasks/robocasa365-close-drawer-i00-privileged/task.toml
index e3d1bb5c74..329d387ee3 100644
--- a/tasks/robocasa365-close-drawer-i00-privileged/task.toml
+++ b/tasks/robocasa365-close-drawer-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa365/CloseDrawer"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robocasa365-close-drawer-i00-standard/task.toml b/tasks/robocasa365-close-drawer-i00-standard/task.toml
index df6c8e9f9c..18053bfb5d 100644
--- a/tasks/robocasa365-close-drawer-i00-standard/task.toml
+++ b/tasks/robocasa365-close-drawer-i00-standard/task.toml
@@ -24,6 +24,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa365/CloseDrawer"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
expert_time_estimate_hours = 4.0
diff --git a/tasks/robocasa365-close-oven-i00-privileged/task.toml b/tasks/robocasa365-close-oven-i00-privileged/task.toml
index 34f2159738..67168de364 100644
--- a/tasks/robocasa365-close-oven-i00-privileged/task.toml
+++ b/tasks/robocasa365-close-oven-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa365/CloseOven"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robocasa365-close-oven-i00-standard/task.toml b/tasks/robocasa365-close-oven-i00-standard/task.toml
index 688395f58a..a1a5c7a5e3 100644
--- a/tasks/robocasa365-close-oven-i00-standard/task.toml
+++ b/tasks/robocasa365-close-oven-i00-standard/task.toml
@@ -24,6 +24,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa365/CloseOven"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
expert_time_estimate_hours = 4.0
diff --git a/tasks/robocasa365-close-stand-mixer-head-i00-privileged/task.toml b/tasks/robocasa365-close-stand-mixer-head-i00-privileged/task.toml
index bb463e068f..3f71aac3a7 100644
--- a/tasks/robocasa365-close-stand-mixer-head-i00-privileged/task.toml
+++ b/tasks/robocasa365-close-stand-mixer-head-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa365/CloseStandMixerHead"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robocasa365-close-stand-mixer-head-i00-standard/task.toml b/tasks/robocasa365-close-stand-mixer-head-i00-standard/task.toml
index 6b8c8912a5..025994966c 100644
--- a/tasks/robocasa365-close-stand-mixer-head-i00-standard/task.toml
+++ b/tasks/robocasa365-close-stand-mixer-head-i00-standard/task.toml
@@ -24,6 +24,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa365/CloseStandMixerHead"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
expert_time_estimate_hours = 4.0
diff --git a/tasks/robocasa365-cookie-dough-prep-i00-privileged/task.toml b/tasks/robocasa365-cookie-dough-prep-i00-privileged/task.toml
index 47a312e26c..8a5b8fbb31 100644
--- a/tasks/robocasa365-cookie-dough-prep-i00-privileged/task.toml
+++ b/tasks/robocasa365-cookie-dough-prep-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa365/CookieDoughPrep"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robocasa365-cookie-dough-prep-i00-standard/task.toml b/tasks/robocasa365-cookie-dough-prep-i00-standard/task.toml
index 3a6bb1fae5..3ecbc44291 100644
--- a/tasks/robocasa365-cookie-dough-prep-i00-standard/task.toml
+++ b/tasks/robocasa365-cookie-dough-prep-i00-standard/task.toml
@@ -24,6 +24,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa365/CookieDoughPrep"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
expert_time_estimate_hours = 4.0
diff --git a/tasks/robocasa365-deliver-brewed-coffee-i00-privileged/task.toml b/tasks/robocasa365-deliver-brewed-coffee-i00-privileged/task.toml
index 378f4ae409..54d7f97cfd 100644
--- a/tasks/robocasa365-deliver-brewed-coffee-i00-privileged/task.toml
+++ b/tasks/robocasa365-deliver-brewed-coffee-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa365/DeliverBrewedCoffee"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robocasa365-deliver-brewed-coffee-i00-standard/task.toml b/tasks/robocasa365-deliver-brewed-coffee-i00-standard/task.toml
index 28b066ef3a..90f88fcf45 100644
--- a/tasks/robocasa365-deliver-brewed-coffee-i00-standard/task.toml
+++ b/tasks/robocasa365-deliver-brewed-coffee-i00-standard/task.toml
@@ -24,6 +24,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa365/DeliverBrewedCoffee"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
expert_time_estimate_hours = 4.0
diff --git a/tasks/robocasa365-dessert-upgrade-i00-privileged/task.toml b/tasks/robocasa365-dessert-upgrade-i00-privileged/task.toml
index e4105bb107..94fe4b8062 100644
--- a/tasks/robocasa365-dessert-upgrade-i00-privileged/task.toml
+++ b/tasks/robocasa365-dessert-upgrade-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa365/DessertUpgrade"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robocasa365-dessert-upgrade-i00-standard/task.toml b/tasks/robocasa365-dessert-upgrade-i00-standard/task.toml
index 1d6af7173e..cc101546ce 100644
--- a/tasks/robocasa365-dessert-upgrade-i00-standard/task.toml
+++ b/tasks/robocasa365-dessert-upgrade-i00-standard/task.toml
@@ -24,6 +24,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa365/DessertUpgrade"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
expert_time_estimate_hours = 4.0
diff --git a/tasks/robocasa365-display-meat-variety-i00-privileged/task.toml b/tasks/robocasa365-display-meat-variety-i00-privileged/task.toml
index fb84bfd5fa..3415addefa 100644
--- a/tasks/robocasa365-display-meat-variety-i00-privileged/task.toml
+++ b/tasks/robocasa365-display-meat-variety-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa365/DisplayMeatVariety"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robocasa365-display-meat-variety-i00-standard/task.toml b/tasks/robocasa365-display-meat-variety-i00-standard/task.toml
index f17342ccff..6efc468979 100644
--- a/tasks/robocasa365-display-meat-variety-i00-standard/task.toml
+++ b/tasks/robocasa365-display-meat-variety-i00-standard/task.toml
@@ -24,6 +24,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa365/DisplayMeatVariety"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
expert_time_estimate_hours = 4.0
diff --git a/tasks/robocasa365-divide-basins-i00-privileged/task.toml b/tasks/robocasa365-divide-basins-i00-privileged/task.toml
index 9917a6c0e3..6f76607e3f 100644
--- a/tasks/robocasa365-divide-basins-i00-privileged/task.toml
+++ b/tasks/robocasa365-divide-basins-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa365/DivideBasins"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robocasa365-divide-basins-i00-standard/task.toml b/tasks/robocasa365-divide-basins-i00-standard/task.toml
index 5ed7f141db..3e85a877a4 100644
--- a/tasks/robocasa365-divide-basins-i00-standard/task.toml
+++ b/tasks/robocasa365-divide-basins-i00-standard/task.toml
@@ -24,6 +24,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa365/DivideBasins"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
expert_time_estimate_hours = 4.0
diff --git a/tasks/robocasa365-drain-veggies-i00-privileged/README.md b/tasks/robocasa365-drain-veggies-i00-privileged/README.md
index 8faa36bfd5..d13aeba29f 100644
--- a/tasks/robocasa365-drain-veggies-i00-privileged/README.md
+++ b/tasks/robocasa365-drain-veggies-i00-privileged/README.md
@@ -42,4 +42,7 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
| Source | [RoboCasa365](https://github.com/robocasa/robocasa) @ 4f8a298 (robocasa 1.0.1; MIT), robosuite @ 5ce6643 (1.5.2), MuJoCo 3.3.1; [environment source](https://github.com/robocasa/robocasa/blob/4f8a2980def75a55dff96b990745b83540425f09/robocasa/environments/kitchen/composite/washing_fruits_and_vegetables/drain_veggies.py#L4) |
| Instance | the initial state of official demonstration episode 0 (`demo_0`); MuJoCo state, model arrays, task state and RNG frozen in `instance.npz` (SHA-256 `59dae1f7b031…`, checked on load) |
| Deliverable | `(T, 13)` native controller commands, 1 ≤ T ≤ 3000, 20 Hz |
-| Success | the environment's own check (`_check_success`) on the state after the last action: the lettuce has been washed: inside the sink with the water on for more than 10 consecutive checks (the check runs twice per control step: at least 6 control steps; once washed it stays washed); at the end the lettuce is in the pot and the water is off. |
+| Success | the environment's own check (`_check_success`) on the state after the last action: the lettuce has been washed: inside the sink with the water on for more than 10 consecutive checks (the check runs twice per control step: at least 6 control steps; once washed it stays washed); at the end the lettuce is in the pot and the water is off (the water is on while the faucet handle's on / off hinge is between 0.40 rad and π rad). The handle turns on two hinges at one pivot: the on / off hinge, whose range is 0 to 0.5 rad (so the water is on only over the last fifth of its travel), turns about the sink's front-to-back axis, so the handle swings sideways; the temperature hinge (about ±0.79 rad) tilts the handle toward or away from you and never turns the water on or off. |
+
+### Change Log
+- 2026-10-14 (board 0113, from board 0086): the success text now says that the faucet handle has two hinges at one pivot: the water on / off hinge (range 0 to 0.5 rad, about the sink's front-to-back axis, so the handle swings sideways) and a temperature hinge (about ±0.79 rad) that tilts it toward or away from you and never turns the water on; before, it named only "the handle joint" (board 0086: every standard run of `turn-on-sink-faucet` and `thaw-in-sink` moved the temperature hinge). Text only: the scene, the frozen instance, the reference solution, the environment's success check and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/robocasa365-drain-veggies-i00-privileged/instruction.md b/tasks/robocasa365-drain-veggies-i00-privileged/instruction.md
index bc927d9b2f..ef591ab01a 100644
--- a/tasks/robocasa365-drain-veggies-i00-privileged/instruction.md
+++ b/tasks/robocasa365-drain-veggies-i00-privileged/instruction.md
@@ -16,7 +16,7 @@ CPU is bit-exact) and applies the environment's own success check for `DrainVegg
the last action**. A success reached earlier and then lost does not count: end the trajectory while the task is
achieved. `s.check()` returns the check's latest result.
-The check for `DrainVeggies` requires that the lettuce has been washed: inside the sink with the water on for more than 10 consecutive checks (the check runs twice per control step: at least 6 control steps; once washed it stays washed); at the end the lettuce is in the pot and the water is off.
+The check for `DrainVeggies` requires that the lettuce has been washed: inside the sink with the water on for more than 10 consecutive checks (the check runs twice per control step: at least 6 control steps; once washed it stays washed); at the end the lettuce is in the pot and the water is off (the water is on while the faucet handle's on / off hinge is between 0.40 rad and π rad). The handle turns on two hinges at one pivot: the on / off hinge, whose range is 0 to 0.5 rad (so the water is on only over the last fifth of its travel), turns about the sink's front-to-back axis, so the handle swings sideways; the temperature hinge (about ±0.79 rad) tilts the handle toward or away from you and never turns the water on or off.
How RoboCasa365's checks compute these terms (`robocasa/utils/object_utils.py`):
- *the gripper ends more than d m from X*: the distance from the robot's end-effector grip site to X's origin exceeds
diff --git a/tasks/robocasa365-drain-veggies-i00-privileged/task.toml b/tasks/robocasa365-drain-veggies-i00-privileged/task.toml
index 237c017ccd..af59a26aa2 100644
--- a/tasks/robocasa365-drain-veggies-i00-privileged/task.toml
+++ b/tasks/robocasa365-drain-veggies-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa365/DrainVeggies"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robocasa365-drain-veggies-i00-standard/README.md b/tasks/robocasa365-drain-veggies-i00-standard/README.md
index 1a5f162aba..1047c3f8df 100644
--- a/tasks/robocasa365-drain-veggies-i00-standard/README.md
+++ b/tasks/robocasa365-drain-veggies-i00-standard/README.md
@@ -48,3 +48,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
- 2026-10-10 (review round 3, board 0067): the sim service's request handling (`environment/sim/server_casa.py`) and `tests/grade.py`. A negative or non-numeric `Content-Length` is answered 400 without reading the body (the service read until the client hung up, without a bound, which could exhaust its memory); a camera named more than once in `cameras` is rendered and returned once (N copies of one name made it build N x N images within the frame budget); a port that is still held when the service restarts is waited for, not counted as a crash; a JSON `true` or `false` where `dt` or `seconds` takes a number is a 400 (`true` ran as 1 second). The grader: a `final.json` whose evidence fields (`control_steps`, `calls`, `elapsed_s`, `final_reward`) are not numbers no longer stops it before it writes a reward (they are reported as 0 and decide nothing). An episode's physics and the pass rule are unchanged.
- 2026-10-13 (owner, boards 0085 and 0087): the sim service's clock (`environment/sim/server_casa.py`, the docstring of `environment/rcb_client.py`). `meta["time_s"]` and `meta["time_left_s"]` count from the agent's first request to the service (any endpoint), and a restarted service keeps that moment; before, they counted from the first start of the service's container (`time_left_s`) and from when its scene was ready (`time_s`). Where the service is ready before the agent's phase begins (the harness is installed in between: a median of 79 s over the recorded standard trials), that clock ran ahead of the agent's and `time_left_s` reached 0 while the agent still had that time. `/spec` says what the clock counts from (`budget.clock`), and the instruction's `## Time limit` tells the agent to plan by `time_left_s`. The service never ended an episode on this clock: the scene, the frozen instance, the pass rule and the grader are unchanged, so the task keeps its name and earlier results stand.
- 2026-10-13 (board 0093; found in board 0086): the camera calibration (`environment/sim/server_casa.py`). `cam2world_gl` held the pose of MuJoCo's left-eye camera (`scene.camera[0]`), 3.4 cm (half the model's 0.068 m interpupillary distance) to the camera's left of the viewpoint the mono RGB and depth images are rendered from; it now is that viewpoint (`data.cam_xpos` / `data.cam_xmat`). Checked against MuJoCo's own renders of marker spheres at known world points, every served camera, before and after the robot moved: the reprojection error was 10.5-10.7 px median (up to 28 px for near points, always to the same side), now 0.15 px median and at most 0.4 px. The intrinsics were right (every camera here is given by its `fovy`, none by a sensor size) and are unchanged, as are the images, the scene, the frozen instance, the pass rule and the grader, so the task keeps its name. **Standard-mode results from before this date used the offset calibration**: a point located from the images and the reported calibration came out about 3.4 cm to the side (agents missed a microwave button by 3.5 cm), so they are not strictly comparable with later ones.
+- 2026-10-14 (board 0113, from board 0086): the success text now says that the faucet handle has two hinges at one pivot: the water on / off hinge (range 0 to 0.5 rad, about the sink's front-to-back axis, so the handle swings sideways) and a temperature hinge (about ±0.79 rad) that tilts it toward or away from you and never turns the water on; before, it named only "the handle joint" (board 0086: every standard run of `turn-on-sink-faucet` and `thaw-in-sink` moved the temperature hinge). Text only: the scene, the frozen instance, the reference solution, the environment's success check and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/robocasa365-drain-veggies-i00-standard/instruction.md b/tasks/robocasa365-drain-veggies-i00-standard/instruction.md
index 24208567ad..efa56eb780 100644
--- a/tasks/robocasa365-drain-veggies-i00-standard/instruction.md
+++ b/tasks/robocasa365-drain-veggies-i00-standard/instruction.md
@@ -23,7 +23,7 @@ the environment's own success check for `DrainVeggies` to the state **after the
in the live episode and in the replay. A success reached earlier and then lost does not count: end in the state you
want graded and stop submitting.
-The check for `DrainVeggies` requires that the lettuce has been washed: inside the sink with the water on for more than 10 consecutive checks (the check runs twice per control step: at least 6 control steps; once washed it stays washed); at the end the lettuce is in the pot and the water is off.
+The check for `DrainVeggies` requires that the lettuce has been washed: inside the sink with the water on for more than 10 consecutive checks (the check runs twice per control step: at least 6 control steps; once washed it stays washed); at the end the lettuce is in the pot and the water is off (the water is on while the faucet handle's on / off hinge is between 0.40 rad and π rad). The handle turns on two hinges at one pivot: the on / off hinge, whose range is 0 to 0.5 rad (so the water is on only over the last fifth of its travel), turns about the sink's front-to-back axis, so the handle swings sideways; the temperature hinge (about ±0.79 rad) tilts the handle toward or away from you and never turns the water on or off.
How RoboCasa365's checks compute these terms (`robocasa/utils/object_utils.py`):
- *the gripper ends more than d m from X*: the distance from the robot's end-effector grip site to X's origin exceeds
diff --git a/tasks/robocasa365-drain-veggies-i00-standard/task.toml b/tasks/robocasa365-drain-veggies-i00-standard/task.toml
index def4a25e79..0caea3e574 100644
--- a/tasks/robocasa365-drain-veggies-i00-standard/task.toml
+++ b/tasks/robocasa365-drain-veggies-i00-standard/task.toml
@@ -24,6 +24,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa365/DrainVeggies"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
expert_time_estimate_hours = 4.0
diff --git a/tasks/robocasa365-drinkware-consolidation-i00-privileged/task.toml b/tasks/robocasa365-drinkware-consolidation-i00-privileged/task.toml
index 1738a73113..1c7bba8c2f 100644
--- a/tasks/robocasa365-drinkware-consolidation-i00-privileged/task.toml
+++ b/tasks/robocasa365-drinkware-consolidation-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa365/DrinkwareConsolidation"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robocasa365-drinkware-consolidation-i00-standard/task.toml b/tasks/robocasa365-drinkware-consolidation-i00-standard/task.toml
index befa59b646..275cade91d 100644
--- a/tasks/robocasa365-drinkware-consolidation-i00-standard/task.toml
+++ b/tasks/robocasa365-drinkware-consolidation-i00-standard/task.toml
@@ -24,6 +24,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa365/DrinkwareConsolidation"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
expert_time_estimate_hours = 4.0
diff --git a/tasks/robocasa365-make-ice-lemonade-i00-privileged/task.toml b/tasks/robocasa365-make-ice-lemonade-i00-privileged/task.toml
index 44ab9cb006..9b75f6387b 100644
--- a/tasks/robocasa365-make-ice-lemonade-i00-privileged/task.toml
+++ b/tasks/robocasa365-make-ice-lemonade-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa365/MakeIceLemonade"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robocasa365-make-ice-lemonade-i00-standard/task.toml b/tasks/robocasa365-make-ice-lemonade-i00-standard/task.toml
index e225ef5a3f..bdd2d35ae9 100644
--- a/tasks/robocasa365-make-ice-lemonade-i00-standard/task.toml
+++ b/tasks/robocasa365-make-ice-lemonade-i00-standard/task.toml
@@ -24,6 +24,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa365/MakeIceLemonade"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
expert_time_estimate_hours = 4.0
diff --git a/tasks/robocasa365-make-loaded-potato-i00-privileged/task.toml b/tasks/robocasa365-make-loaded-potato-i00-privileged/task.toml
index 88e3a9cf10..26e18b5613 100644
--- a/tasks/robocasa365-make-loaded-potato-i00-privileged/task.toml
+++ b/tasks/robocasa365-make-loaded-potato-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa365/MakeLoadedPotato"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robocasa365-make-loaded-potato-i00-standard/task.toml b/tasks/robocasa365-make-loaded-potato-i00-standard/task.toml
index 041edd13b3..ac434194d5 100644
--- a/tasks/robocasa365-make-loaded-potato-i00-standard/task.toml
+++ b/tasks/robocasa365-make-loaded-potato-i00-standard/task.toml
@@ -24,6 +24,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa365/MakeLoadedPotato"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
expert_time_estimate_hours = 4.0
diff --git a/tasks/robocasa365-meal-prep-staging-i00-privileged/task.toml b/tasks/robocasa365-meal-prep-staging-i00-privileged/task.toml
index 31f5553e73..f7b923ab39 100644
--- a/tasks/robocasa365-meal-prep-staging-i00-privileged/task.toml
+++ b/tasks/robocasa365-meal-prep-staging-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa365/MealPrepStaging"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robocasa365-meal-prep-staging-i00-standard/task.toml b/tasks/robocasa365-meal-prep-staging-i00-standard/task.toml
index a8115226d7..526cbdf08f 100644
--- a/tasks/robocasa365-meal-prep-staging-i00-standard/task.toml
+++ b/tasks/robocasa365-meal-prep-staging-i00-standard/task.toml
@@ -24,6 +24,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa365/MealPrepStaging"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
expert_time_estimate_hours = 4.0
diff --git a/tasks/robocasa365-microwave-defrost-meat-i00-privileged/task.toml b/tasks/robocasa365-microwave-defrost-meat-i00-privileged/task.toml
index e7affd7b72..2b743a7666 100644
--- a/tasks/robocasa365-microwave-defrost-meat-i00-privileged/task.toml
+++ b/tasks/robocasa365-microwave-defrost-meat-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa365/MicrowaveDefrostMeat"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robocasa365-microwave-defrost-meat-i00-standard/task.toml b/tasks/robocasa365-microwave-defrost-meat-i00-standard/task.toml
index d4fdc622f1..b587033965 100644
--- a/tasks/robocasa365-microwave-defrost-meat-i00-standard/task.toml
+++ b/tasks/robocasa365-microwave-defrost-meat-i00-standard/task.toml
@@ -24,6 +24,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa365/MicrowaveDefrostMeat"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
expert_time_estimate_hours = 4.0
diff --git a/tasks/robocasa365-open-dishwasher-i00-privileged/task.toml b/tasks/robocasa365-open-dishwasher-i00-privileged/task.toml
index 8e02bf8dc3..b2b39e7b41 100644
--- a/tasks/robocasa365-open-dishwasher-i00-privileged/task.toml
+++ b/tasks/robocasa365-open-dishwasher-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa365/OpenDishwasher"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robocasa365-open-dishwasher-i00-standard/task.toml b/tasks/robocasa365-open-dishwasher-i00-standard/task.toml
index 4b3abd756b..b6b14053ac 100644
--- a/tasks/robocasa365-open-dishwasher-i00-standard/task.toml
+++ b/tasks/robocasa365-open-dishwasher-i00-standard/task.toml
@@ -24,6 +24,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa365/OpenDishwasher"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
expert_time_estimate_hours = 4.0
diff --git a/tasks/robocasa365-organize-coffee-condiments-i00-privileged/task.toml b/tasks/robocasa365-organize-coffee-condiments-i00-privileged/task.toml
index 78015fa8a2..fc11282143 100644
--- a/tasks/robocasa365-organize-coffee-condiments-i00-privileged/task.toml
+++ b/tasks/robocasa365-organize-coffee-condiments-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa365/OrganizeCoffeeCondiments"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robocasa365-organize-coffee-condiments-i00-standard/task.toml b/tasks/robocasa365-organize-coffee-condiments-i00-standard/task.toml
index 2f447d5ae9..f1e8e19b2a 100644
--- a/tasks/robocasa365-organize-coffee-condiments-i00-standard/task.toml
+++ b/tasks/robocasa365-organize-coffee-condiments-i00-standard/task.toml
@@ -24,6 +24,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa365/OrganizeCoffeeCondiments"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
expert_time_estimate_hours = 4.0
diff --git a/tasks/robocasa365-oven-broil-fish-i00-privileged/task.toml b/tasks/robocasa365-oven-broil-fish-i00-privileged/task.toml
index 1cb48c7cdb..d5dd4db252 100644
--- a/tasks/robocasa365-oven-broil-fish-i00-privileged/task.toml
+++ b/tasks/robocasa365-oven-broil-fish-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa365/OvenBroilFish"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robocasa365-oven-broil-fish-i00-standard/task.toml b/tasks/robocasa365-oven-broil-fish-i00-standard/task.toml
index 3492468eca..be32ccf75b 100644
--- a/tasks/robocasa365-oven-broil-fish-i00-standard/task.toml
+++ b/tasks/robocasa365-oven-broil-fish-i00-standard/task.toml
@@ -24,6 +24,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa365/OvenBroilFish"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
expert_time_estimate_hours = 4.0
diff --git a/tasks/robocasa365-pack-dessert-i00-privileged/task.toml b/tasks/robocasa365-pack-dessert-i00-privileged/task.toml
index a1a26ebf9c..7f77dc704c 100644
--- a/tasks/robocasa365-pack-dessert-i00-privileged/task.toml
+++ b/tasks/robocasa365-pack-dessert-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa365/PackDessert"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robocasa365-pack-dessert-i00-standard/task.toml b/tasks/robocasa365-pack-dessert-i00-standard/task.toml
index 5e4da3a859..1aacdb4418 100644
--- a/tasks/robocasa365-pack-dessert-i00-standard/task.toml
+++ b/tasks/robocasa365-pack-dessert-i00-standard/task.toml
@@ -24,6 +24,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa365/PackDessert"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
expert_time_estimate_hours = 4.0
diff --git a/tasks/robocasa365-pick-place-cabinet-to-counter-i00-privileged/task.toml b/tasks/robocasa365-pick-place-cabinet-to-counter-i00-privileged/task.toml
index e301d83b9d..8656f97cf9 100644
--- a/tasks/robocasa365-pick-place-cabinet-to-counter-i00-privileged/task.toml
+++ b/tasks/robocasa365-pick-place-cabinet-to-counter-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa365/PickPlaceCabinetToCounter"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robocasa365-pick-place-cabinet-to-counter-i00-standard/task.toml b/tasks/robocasa365-pick-place-cabinet-to-counter-i00-standard/task.toml
index 39da1dff78..a1089e7db9 100644
--- a/tasks/robocasa365-pick-place-cabinet-to-counter-i00-standard/task.toml
+++ b/tasks/robocasa365-pick-place-cabinet-to-counter-i00-standard/task.toml
@@ -24,6 +24,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa365/PickPlaceCabinetToCounter"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
expert_time_estimate_hours = 4.0
diff --git a/tasks/robocasa365-pick-place-sink-to-counter-i00-privileged/task.toml b/tasks/robocasa365-pick-place-sink-to-counter-i00-privileged/task.toml
index b7f8d7b01f..bed404853c 100644
--- a/tasks/robocasa365-pick-place-sink-to-counter-i00-privileged/task.toml
+++ b/tasks/robocasa365-pick-place-sink-to-counter-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa365/PickPlaceSinkToCounter"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robocasa365-pick-place-sink-to-counter-i00-standard/task.toml b/tasks/robocasa365-pick-place-sink-to-counter-i00-standard/task.toml
index 1d5f798720..413714ef9c 100644
--- a/tasks/robocasa365-pick-place-sink-to-counter-i00-standard/task.toml
+++ b/tasks/robocasa365-pick-place-sink-to-counter-i00-standard/task.toml
@@ -24,6 +24,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa365/PickPlaceSinkToCounter"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
expert_time_estimate_hours = 4.0
diff --git a/tasks/robocasa365-plate-steak-meal-i00-privileged/task.toml b/tasks/robocasa365-plate-steak-meal-i00-privileged/task.toml
index 70877904bd..12f87ce77a 100644
--- a/tasks/robocasa365-plate-steak-meal-i00-privileged/task.toml
+++ b/tasks/robocasa365-plate-steak-meal-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa365/PlateSteakMeal"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robocasa365-plate-steak-meal-i00-standard/task.toml b/tasks/robocasa365-plate-steak-meal-i00-standard/task.toml
index e80df35a1d..28e18654e6 100644
--- a/tasks/robocasa365-plate-steak-meal-i00-standard/task.toml
+++ b/tasks/robocasa365-plate-steak-meal-i00-standard/task.toml
@@ -24,6 +24,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa365/PlateSteakMeal"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
expert_time_estimate_hours = 4.0
diff --git a/tasks/robocasa365-portion-in-tupperware-i00-privileged/task.toml b/tasks/robocasa365-portion-in-tupperware-i00-privileged/task.toml
index 09a42a25d7..042c261047 100644
--- a/tasks/robocasa365-portion-in-tupperware-i00-privileged/task.toml
+++ b/tasks/robocasa365-portion-in-tupperware-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa365/PortionInTupperware"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robocasa365-portion-in-tupperware-i00-standard/task.toml b/tasks/robocasa365-portion-in-tupperware-i00-standard/task.toml
index 592d382886..6729ce0506 100644
--- a/tasks/robocasa365-portion-in-tupperware-i00-standard/task.toml
+++ b/tasks/robocasa365-portion-in-tupperware-i00-standard/task.toml
@@ -24,6 +24,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa365/PortionInTupperware"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
expert_time_estimate_hours = 4.0
diff --git a/tasks/robocasa365-pre-rinse-station-i00-privileged/task.toml b/tasks/robocasa365-pre-rinse-station-i00-privileged/task.toml
index bb0aab74c8..88832b0ba8 100644
--- a/tasks/robocasa365-pre-rinse-station-i00-privileged/task.toml
+++ b/tasks/robocasa365-pre-rinse-station-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa365/PreRinseStation"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robocasa365-pre-rinse-station-i00-standard/task.toml b/tasks/robocasa365-pre-rinse-station-i00-standard/task.toml
index 077d350078..d47702ff27 100644
--- a/tasks/robocasa365-pre-rinse-station-i00-standard/task.toml
+++ b/tasks/robocasa365-pre-rinse-station-i00-standard/task.toml
@@ -24,6 +24,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa365/PreRinseStation"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
expert_time_estimate_hours = 4.0
diff --git a/tasks/robocasa365-prep-for-sanitizing-i00-privileged/task.toml b/tasks/robocasa365-prep-for-sanitizing-i00-privileged/task.toml
index 609d492faa..fbae3a7c2a 100644
--- a/tasks/robocasa365-prep-for-sanitizing-i00-privileged/task.toml
+++ b/tasks/robocasa365-prep-for-sanitizing-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa365/PrepForSanitizing"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robocasa365-prep-for-sanitizing-i00-standard/task.toml b/tasks/robocasa365-prep-for-sanitizing-i00-standard/task.toml
index 9d1bee99f3..f6d29e5893 100644
--- a/tasks/robocasa365-prep-for-sanitizing-i00-standard/task.toml
+++ b/tasks/robocasa365-prep-for-sanitizing-i00-standard/task.toml
@@ -24,6 +24,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa365/PrepForSanitizing"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
expert_time_estimate_hours = 4.0
diff --git a/tasks/robocasa365-prep-marinating-meat-i00-privileged/task.toml b/tasks/robocasa365-prep-marinating-meat-i00-privileged/task.toml
index cfa1f77804..30aaf97fff 100644
--- a/tasks/robocasa365-prep-marinating-meat-i00-privileged/task.toml
+++ b/tasks/robocasa365-prep-marinating-meat-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa365/PrepMarinatingMeat"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robocasa365-prep-marinating-meat-i00-standard/task.toml b/tasks/robocasa365-prep-marinating-meat-i00-standard/task.toml
index dba3c9cad5..abaf8b1580 100644
--- a/tasks/robocasa365-prep-marinating-meat-i00-standard/task.toml
+++ b/tasks/robocasa365-prep-marinating-meat-i00-standard/task.toml
@@ -24,6 +24,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa365/PrepMarinatingMeat"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
expert_time_estimate_hours = 4.0
diff --git a/tasks/robocasa365-reheat-meal-i00-privileged/task.toml b/tasks/robocasa365-reheat-meal-i00-privileged/task.toml
index a7fc51dd38..3e42fafabb 100644
--- a/tasks/robocasa365-reheat-meal-i00-privileged/task.toml
+++ b/tasks/robocasa365-reheat-meal-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa365/ReheatMeal"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robocasa365-reheat-meal-i00-standard/task.toml b/tasks/robocasa365-reheat-meal-i00-standard/task.toml
index 2d0c4141c9..64098336e9 100644
--- a/tasks/robocasa365-reheat-meal-i00-standard/task.toml
+++ b/tasks/robocasa365-reheat-meal-i00-standard/task.toml
@@ -24,6 +24,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa365/ReheatMeal"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
expert_time_estimate_hours = 4.0
diff --git a/tasks/robocasa365-retrieve-meat-i00-privileged/task.toml b/tasks/robocasa365-retrieve-meat-i00-privileged/task.toml
index 281ed0ffa9..8afc26148d 100644
--- a/tasks/robocasa365-retrieve-meat-i00-privileged/task.toml
+++ b/tasks/robocasa365-retrieve-meat-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa365/RetrieveMeat"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robocasa365-retrieve-meat-i00-standard/task.toml b/tasks/robocasa365-retrieve-meat-i00-standard/task.toml
index d19fcc0eb9..5b55d66120 100644
--- a/tasks/robocasa365-retrieve-meat-i00-standard/task.toml
+++ b/tasks/robocasa365-retrieve-meat-i00-standard/task.toml
@@ -24,6 +24,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa365/RetrieveMeat"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
expert_time_estimate_hours = 4.0
diff --git a/tasks/robocasa365-sanitize-prep-cutting-board-i00-privileged/task.toml b/tasks/robocasa365-sanitize-prep-cutting-board-i00-privileged/task.toml
index db21ca275a..41daa1997e 100644
--- a/tasks/robocasa365-sanitize-prep-cutting-board-i00-privileged/task.toml
+++ b/tasks/robocasa365-sanitize-prep-cutting-board-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa365/SanitizePrepCuttingBoard"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robocasa365-sanitize-prep-cutting-board-i00-standard/task.toml b/tasks/robocasa365-sanitize-prep-cutting-board-i00-standard/task.toml
index 3d07dbdbb7..d0f8c3337d 100644
--- a/tasks/robocasa365-sanitize-prep-cutting-board-i00-standard/task.toml
+++ b/tasks/robocasa365-sanitize-prep-cutting-board-i00-standard/task.toml
@@ -24,6 +24,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa365/SanitizePrepCuttingBoard"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
expert_time_estimate_hours = 4.0
diff --git a/tasks/robocasa365-scrub-bowl-i00-privileged/task.toml b/tasks/robocasa365-scrub-bowl-i00-privileged/task.toml
index b473e1623e..3c8b0764bb 100644
--- a/tasks/robocasa365-scrub-bowl-i00-privileged/task.toml
+++ b/tasks/robocasa365-scrub-bowl-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa365/ScrubBowl"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robocasa365-scrub-bowl-i00-standard/task.toml b/tasks/robocasa365-scrub-bowl-i00-standard/task.toml
index 4bf1ac68be..1da20712fb 100644
--- a/tasks/robocasa365-scrub-bowl-i00-standard/task.toml
+++ b/tasks/robocasa365-scrub-bowl-i00-standard/task.toml
@@ -24,6 +24,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa365/ScrubBowl"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
expert_time_estimate_hours = 4.0
diff --git a/tasks/robocasa365-searing-meat-i00-privileged/task.toml b/tasks/robocasa365-searing-meat-i00-privileged/task.toml
index 88f357cd79..b270ab95eb 100644
--- a/tasks/robocasa365-searing-meat-i00-privileged/task.toml
+++ b/tasks/robocasa365-searing-meat-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa365/SearingMeat"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robocasa365-searing-meat-i00-standard/task.toml b/tasks/robocasa365-searing-meat-i00-standard/task.toml
index 2341c9bb5f..5bf6fa9d7c 100644
--- a/tasks/robocasa365-searing-meat-i00-standard/task.toml
+++ b/tasks/robocasa365-searing-meat-i00-standard/task.toml
@@ -24,6 +24,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa365/SearingMeat"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
expert_time_estimate_hours = 4.0
diff --git a/tasks/robocasa365-seasoning-spice-setup-i00-privileged/task.toml b/tasks/robocasa365-seasoning-spice-setup-i00-privileged/task.toml
index 1024918b80..0bc9e6ff3b 100644
--- a/tasks/robocasa365-seasoning-spice-setup-i00-privileged/task.toml
+++ b/tasks/robocasa365-seasoning-spice-setup-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa365/SeasoningSpiceSetup"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robocasa365-seasoning-spice-setup-i00-standard/task.toml b/tasks/robocasa365-seasoning-spice-setup-i00-standard/task.toml
index 437ea48667..ccae0f87a7 100644
--- a/tasks/robocasa365-seasoning-spice-setup-i00-standard/task.toml
+++ b/tasks/robocasa365-seasoning-spice-setup-i00-standard/task.toml
@@ -24,6 +24,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa365/SeasoningSpiceSetup"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
expert_time_estimate_hours = 4.0
diff --git a/tasks/robocasa365-set-up-cutting-station-i00-privileged/task.toml b/tasks/robocasa365-set-up-cutting-station-i00-privileged/task.toml
index 646b73d38a..51fd24f932 100644
--- a/tasks/robocasa365-set-up-cutting-station-i00-privileged/task.toml
+++ b/tasks/robocasa365-set-up-cutting-station-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa365/SetUpCuttingStation"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robocasa365-set-up-cutting-station-i00-standard/task.toml b/tasks/robocasa365-set-up-cutting-station-i00-standard/task.toml
index 262185a99b..8763742fbe 100644
--- a/tasks/robocasa365-set-up-cutting-station-i00-standard/task.toml
+++ b/tasks/robocasa365-set-up-cutting-station-i00-standard/task.toml
@@ -24,6 +24,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa365/SetUpCuttingStation"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
expert_time_estimate_hours = 4.0
diff --git a/tasks/robocasa365-stocking-breakfast-foods-i00-privileged/task.toml b/tasks/robocasa365-stocking-breakfast-foods-i00-privileged/task.toml
index 25aff62a42..bffe8b67cb 100644
--- a/tasks/robocasa365-stocking-breakfast-foods-i00-privileged/task.toml
+++ b/tasks/robocasa365-stocking-breakfast-foods-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa365/StockingBreakfastFoods"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robocasa365-stocking-breakfast-foods-i00-standard/task.toml b/tasks/robocasa365-stocking-breakfast-foods-i00-standard/task.toml
index 66cb5431bd..025884de96 100644
--- a/tasks/robocasa365-stocking-breakfast-foods-i00-standard/task.toml
+++ b/tasks/robocasa365-stocking-breakfast-foods-i00-standard/task.toml
@@ -24,6 +24,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa365/StockingBreakfastFoods"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
expert_time_estimate_hours = 4.0
diff --git a/tasks/robocasa365-sweeten-coffee-i00-privileged/task.toml b/tasks/robocasa365-sweeten-coffee-i00-privileged/task.toml
index d5eb7d03d3..0e204bc694 100644
--- a/tasks/robocasa365-sweeten-coffee-i00-privileged/task.toml
+++ b/tasks/robocasa365-sweeten-coffee-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa365/SweetenCoffee"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robocasa365-sweeten-coffee-i00-standard/task.toml b/tasks/robocasa365-sweeten-coffee-i00-standard/task.toml
index aa2d368905..a669602c9f 100644
--- a/tasks/robocasa365-sweeten-coffee-i00-standard/task.toml
+++ b/tasks/robocasa365-sweeten-coffee-i00-standard/task.toml
@@ -24,6 +24,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa365/SweetenCoffee"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
expert_time_estimate_hours = 4.0
diff --git a/tasks/robocasa365-thaw-in-sink-i00-privileged/README.md b/tasks/robocasa365-thaw-in-sink-i00-privileged/README.md
index 13ca6203c1..e41a9ee3bc 100644
--- a/tasks/robocasa365-thaw-in-sink-i00-privileged/README.md
+++ b/tasks/robocasa365-thaw-in-sink-i00-privileged/README.md
@@ -42,4 +42,7 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
| Source | [RoboCasa365](https://github.com/robocasa/robocasa) @ 4f8a298 (robocasa 1.0.1; MIT), robosuite @ 5ce6643 (1.5.2), MuJoCo 3.3.1; [environment source](https://github.com/robocasa/robocasa/blob/4f8a2980def75a55dff96b990745b83540425f09/robocasa/environments/kitchen/composite/defrosting_food/thaw_in_sink.py#L4) |
| Instance | the initial state of official demonstration episode 0 (`demo_0`); MuJoCo state, model arrays, task state and RNG frozen in `instance.npz` (SHA-256 `3c90b079e163…`, checked on load) |
| Deliverable | `(T, 13)` native controller commands, 1 ≤ T ≤ 3000, 20 Hz |
-| Success | the environment's own check (`_check_success`) on the state after the last action: the orange is inside the sink and the water is on (faucet handle joint between 0.40 rad and π rad). |
+| Success | the environment's own check (`_check_success`) on the state after the last action: the orange is inside the sink and the water is on (the faucet handle's on / off hinge between 0.40 rad and π rad). The handle turns on two hinges at one pivot: the on / off hinge, whose range is 0 to 0.5 rad (so the water is on only over the last fifth of its travel), turns about the sink's front-to-back axis, so the handle swings sideways; the temperature hinge (about ±0.79 rad) tilts the handle toward or away from you and never turns the water on or off. |
+
+### Change Log
+- 2026-10-14 (board 0113, from board 0086): the success text now says that the faucet handle has two hinges at one pivot: the water on / off hinge (range 0 to 0.5 rad, about the sink's front-to-back axis, so the handle swings sideways) and a temperature hinge (about ±0.79 rad) that tilts it toward or away from you and never turns the water on; before, it named only "the handle joint" (board 0086: every standard run of `turn-on-sink-faucet` and `thaw-in-sink` moved the temperature hinge). Text only: the scene, the frozen instance, the reference solution, the environment's success check and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/robocasa365-thaw-in-sink-i00-privileged/instruction.md b/tasks/robocasa365-thaw-in-sink-i00-privileged/instruction.md
index 0abbf6e70d..1ae0b8f95d 100644
--- a/tasks/robocasa365-thaw-in-sink-i00-privileged/instruction.md
+++ b/tasks/robocasa365-thaw-in-sink-i00-privileged/instruction.md
@@ -16,7 +16,7 @@ CPU is bit-exact) and applies the environment's own success check for `ThawInSin
the last action**. A success reached earlier and then lost does not count: end the trajectory while the task is
achieved. `s.check()` returns the check's latest result.
-The check for `ThawInSink` requires that the orange is inside the sink and the water is on (faucet handle joint between 0.40 rad and π rad).
+The check for `ThawInSink` requires that the orange is inside the sink and the water is on (the faucet handle's on / off hinge between 0.40 rad and π rad). The handle turns on two hinges at one pivot: the on / off hinge, whose range is 0 to 0.5 rad (so the water is on only over the last fifth of its travel), turns about the sink's front-to-back axis, so the handle swings sideways; the temperature hinge (about ±0.79 rad) tilts the handle toward or away from you and never turns the water on or off.
How RoboCasa365's checks compute these terms (`robocasa/utils/object_utils.py`):
- *the gripper ends more than d m from X*: the distance from the robot's end-effector grip site to X's origin exceeds
diff --git a/tasks/robocasa365-thaw-in-sink-i00-privileged/task.toml b/tasks/robocasa365-thaw-in-sink-i00-privileged/task.toml
index f91b493f21..3a4e96a6b8 100644
--- a/tasks/robocasa365-thaw-in-sink-i00-privileged/task.toml
+++ b/tasks/robocasa365-thaw-in-sink-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa365/ThawInSink"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robocasa365-thaw-in-sink-i00-standard/README.md b/tasks/robocasa365-thaw-in-sink-i00-standard/README.md
index 29bc07619e..28fec7de2c 100644
--- a/tasks/robocasa365-thaw-in-sink-i00-standard/README.md
+++ b/tasks/robocasa365-thaw-in-sink-i00-standard/README.md
@@ -48,3 +48,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
- 2026-10-10 (review round 3, board 0067): the sim service's request handling (`environment/sim/server_casa.py`) and `tests/grade.py`. A negative or non-numeric `Content-Length` is answered 400 without reading the body (the service read until the client hung up, without a bound, which could exhaust its memory); a camera named more than once in `cameras` is rendered and returned once (N copies of one name made it build N x N images within the frame budget); a port that is still held when the service restarts is waited for, not counted as a crash; a JSON `true` or `false` where `dt` or `seconds` takes a number is a 400 (`true` ran as 1 second). The grader: a `final.json` whose evidence fields (`control_steps`, `calls`, `elapsed_s`, `final_reward`) are not numbers no longer stops it before it writes a reward (they are reported as 0 and decide nothing). An episode's physics and the pass rule are unchanged.
- 2026-10-13 (owner, boards 0085 and 0087): the sim service's clock (`environment/sim/server_casa.py`, the docstring of `environment/rcb_client.py`). `meta["time_s"]` and `meta["time_left_s"]` count from the agent's first request to the service (any endpoint), and a restarted service keeps that moment; before, they counted from the first start of the service's container (`time_left_s`) and from when its scene was ready (`time_s`). Where the service is ready before the agent's phase begins (the harness is installed in between: a median of 79 s over the recorded standard trials), that clock ran ahead of the agent's and `time_left_s` reached 0 while the agent still had that time. `/spec` says what the clock counts from (`budget.clock`), and the instruction's `## Time limit` tells the agent to plan by `time_left_s`. The service never ended an episode on this clock: the scene, the frozen instance, the pass rule and the grader are unchanged, so the task keeps its name and earlier results stand.
- 2026-10-13 (board 0093; found in board 0086): the camera calibration (`environment/sim/server_casa.py`). `cam2world_gl` held the pose of MuJoCo's left-eye camera (`scene.camera[0]`), 3.4 cm (half the model's 0.068 m interpupillary distance) to the camera's left of the viewpoint the mono RGB and depth images are rendered from; it now is that viewpoint (`data.cam_xpos` / `data.cam_xmat`). Checked against MuJoCo's own renders of marker spheres at known world points, every served camera, before and after the robot moved: the reprojection error was 10.5-10.7 px median (up to 28 px for near points, always to the same side), now 0.15 px median and at most 0.4 px. The intrinsics were right (every camera here is given by its `fovy`, none by a sensor size) and are unchanged, as are the images, the scene, the frozen instance, the pass rule and the grader, so the task keeps its name. **Standard-mode results from before this date used the offset calibration**: a point located from the images and the reported calibration came out about 3.4 cm to the side (agents missed a microwave button by 3.5 cm), so they are not strictly comparable with later ones.
+- 2026-10-14 (board 0113, from board 0086): the success text now says that the faucet handle has two hinges at one pivot: the water on / off hinge (range 0 to 0.5 rad, about the sink's front-to-back axis, so the handle swings sideways) and a temperature hinge (about ±0.79 rad) that tilts it toward or away from you and never turns the water on; before, it named only "the handle joint" (board 0086: every standard run of `turn-on-sink-faucet` and `thaw-in-sink` moved the temperature hinge). Text only: the scene, the frozen instance, the reference solution, the environment's success check and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/robocasa365-thaw-in-sink-i00-standard/instruction.md b/tasks/robocasa365-thaw-in-sink-i00-standard/instruction.md
index 7e3b326480..c05c7b2a8e 100644
--- a/tasks/robocasa365-thaw-in-sink-i00-standard/instruction.md
+++ b/tasks/robocasa365-thaw-in-sink-i00-standard/instruction.md
@@ -23,7 +23,7 @@ the environment's own success check for `ThawInSink` to the state **after the la
in the live episode and in the replay. A success reached earlier and then lost does not count: end in the state you
want graded and stop submitting.
-The check for `ThawInSink` requires that the orange is inside the sink and the water is on (faucet handle joint between 0.40 rad and π rad).
+The check for `ThawInSink` requires that the orange is inside the sink and the water is on (the faucet handle's on / off hinge between 0.40 rad and π rad). The handle turns on two hinges at one pivot: the on / off hinge, whose range is 0 to 0.5 rad (so the water is on only over the last fifth of its travel), turns about the sink's front-to-back axis, so the handle swings sideways; the temperature hinge (about ±0.79 rad) tilts the handle toward or away from you and never turns the water on or off.
How RoboCasa365's checks compute these terms (`robocasa/utils/object_utils.py`):
- *the gripper ends more than d m from X*: the distance from the robot's end-effector grip site to X's origin exceeds
diff --git a/tasks/robocasa365-thaw-in-sink-i00-standard/task.toml b/tasks/robocasa365-thaw-in-sink-i00-standard/task.toml
index fb46282da9..0ea449c2a5 100644
--- a/tasks/robocasa365-thaw-in-sink-i00-standard/task.toml
+++ b/tasks/robocasa365-thaw-in-sink-i00-standard/task.toml
@@ -24,6 +24,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa365/ThawInSink"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
expert_time_estimate_hours = 4.0
diff --git a/tasks/robocasa365-tong-buffet-setup-i00-privileged/task.toml b/tasks/robocasa365-tong-buffet-setup-i00-privileged/task.toml
index ae6b5b18e3..3d111314b4 100644
--- a/tasks/robocasa365-tong-buffet-setup-i00-privileged/task.toml
+++ b/tasks/robocasa365-tong-buffet-setup-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa365/TongBuffetSetup"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robocasa365-tong-buffet-setup-i00-standard/task.toml b/tasks/robocasa365-tong-buffet-setup-i00-standard/task.toml
index f8beb36465..4676ce972e 100644
--- a/tasks/robocasa365-tong-buffet-setup-i00-standard/task.toml
+++ b/tasks/robocasa365-tong-buffet-setup-i00-standard/task.toml
@@ -24,6 +24,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa365/TongBuffetSetup"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
expert_time_estimate_hours = 4.0
diff --git a/tasks/robocasa365-turn-off-microwave-i00-privileged/task.toml b/tasks/robocasa365-turn-off-microwave-i00-privileged/task.toml
index 5d1723c4dd..6e944234f7 100644
--- a/tasks/robocasa365-turn-off-microwave-i00-privileged/task.toml
+++ b/tasks/robocasa365-turn-off-microwave-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa365/TurnOffMicrowave"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robocasa365-turn-off-microwave-i00-standard/task.toml b/tasks/robocasa365-turn-off-microwave-i00-standard/task.toml
index 598ef69b22..b04735006c 100644
--- a/tasks/robocasa365-turn-off-microwave-i00-standard/task.toml
+++ b/tasks/robocasa365-turn-off-microwave-i00-standard/task.toml
@@ -24,6 +24,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa365/TurnOffMicrowave"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
expert_time_estimate_hours = 4.0
diff --git a/tasks/robocasa365-turn-on-microwave-i00-privileged/task.toml b/tasks/robocasa365-turn-on-microwave-i00-privileged/task.toml
index 4fdf2592ea..615090e08e 100644
--- a/tasks/robocasa365-turn-on-microwave-i00-privileged/task.toml
+++ b/tasks/robocasa365-turn-on-microwave-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa365/TurnOnMicrowave"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robocasa365-turn-on-microwave-i00-standard/task.toml b/tasks/robocasa365-turn-on-microwave-i00-standard/task.toml
index fab8dc798b..8b3ab7dac7 100644
--- a/tasks/robocasa365-turn-on-microwave-i00-standard/task.toml
+++ b/tasks/robocasa365-turn-on-microwave-i00-standard/task.toml
@@ -24,6 +24,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa365/TurnOnMicrowave"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
expert_time_estimate_hours = 4.0
diff --git a/tasks/robocasa365-turn-on-sink-faucet-i00-privileged/README.md b/tasks/robocasa365-turn-on-sink-faucet-i00-privileged/README.md
index 2eb6f19d82..520fa7ea8a 100644
--- a/tasks/robocasa365-turn-on-sink-faucet-i00-privileged/README.md
+++ b/tasks/robocasa365-turn-on-sink-faucet-i00-privileged/README.md
@@ -42,4 +42,7 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
| Source | [RoboCasa365](https://github.com/robocasa/robocasa) @ 4f8a298 (robocasa 1.0.1; MIT), robosuite @ 5ce6643 (1.5.2), MuJoCo 3.3.1; [environment source](https://github.com/robocasa/robocasa/blob/4f8a2980def75a55dff96b990745b83540425f09/robocasa/environments/kitchen/atomic/kitchen_sink.py#L112) |
| Instance | the initial state of official demonstration episode 0 (`demo_0`); MuJoCo state, model arrays, task state and RNG frozen in `instance.npz` (SHA-256 `f9eae292b33b…`, checked on load) |
| Deliverable | `(T, 13)` native controller commands, 1 ≤ T ≤ 3000, 20 Hz |
-| Success | the environment's own check (`_check_success`) on the state after the last action: the faucet handle is in its water-on range (handle joint between 0.40 rad and π rad). |
+| Success | the environment's own check (`_check_success`) on the state after the last action: the faucet handle is in its water-on range (its on / off hinge between 0.40 rad and π rad). The handle turns on two hinges at one pivot: the on / off hinge, whose range is 0 to 0.5 rad (so the water is on only over the last fifth of its travel), turns about the sink's front-to-back axis, so the handle swings sideways; the temperature hinge (about ±0.79 rad) tilts the handle toward or away from you and never turns the water on or off. |
+
+### Change Log
+- 2026-10-14 (board 0113, from board 0086): the success text now says that the faucet handle has two hinges at one pivot: the water on / off hinge (range 0 to 0.5 rad, about the sink's front-to-back axis, so the handle swings sideways) and a temperature hinge (about ±0.79 rad) that tilts it toward or away from you and never turns the water on; before, it named only "the handle joint" (board 0086: every standard run of `turn-on-sink-faucet` and `thaw-in-sink` moved the temperature hinge). Text only: the scene, the frozen instance, the reference solution, the environment's success check and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/robocasa365-turn-on-sink-faucet-i00-privileged/instruction.md b/tasks/robocasa365-turn-on-sink-faucet-i00-privileged/instruction.md
index fa7862bb8b..80ea1cab6c 100644
--- a/tasks/robocasa365-turn-on-sink-faucet-i00-privileged/instruction.md
+++ b/tasks/robocasa365-turn-on-sink-faucet-i00-privileged/instruction.md
@@ -16,7 +16,7 @@ CPU is bit-exact) and applies the environment's own success check for `TurnOnSin
the last action**. A success reached earlier and then lost does not count: end the trajectory while the task is
achieved. `s.check()` returns the check's latest result.
-The check for `TurnOnSinkFaucet` requires that the faucet handle is in its water-on range (handle joint between 0.40 rad and π rad).
+The check for `TurnOnSinkFaucet` requires that the faucet handle is in its water-on range (its on / off hinge between 0.40 rad and π rad). The handle turns on two hinges at one pivot: the on / off hinge, whose range is 0 to 0.5 rad (so the water is on only over the last fifth of its travel), turns about the sink's front-to-back axis, so the handle swings sideways; the temperature hinge (about ±0.79 rad) tilts the handle toward or away from you and never turns the water on or off.
## Tooling (`/opt/tools/casa_tool.py`, importable as `casa_tool`)
```python
diff --git a/tasks/robocasa365-turn-on-sink-faucet-i00-privileged/task.toml b/tasks/robocasa365-turn-on-sink-faucet-i00-privileged/task.toml
index f12c457cf5..ebc9814255 100644
--- a/tasks/robocasa365-turn-on-sink-faucet-i00-privileged/task.toml
+++ b/tasks/robocasa365-turn-on-sink-faucet-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa365/TurnOnSinkFaucet"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robocasa365-turn-on-sink-faucet-i00-standard/README.md b/tasks/robocasa365-turn-on-sink-faucet-i00-standard/README.md
index 32510441dd..3dbffa9fe1 100644
--- a/tasks/robocasa365-turn-on-sink-faucet-i00-standard/README.md
+++ b/tasks/robocasa365-turn-on-sink-faucet-i00-standard/README.md
@@ -48,3 +48,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
- 2026-10-10 (review round 3, board 0067): the sim service's request handling (`environment/sim/server_casa.py`) and `tests/grade.py`. A negative or non-numeric `Content-Length` is answered 400 without reading the body (the service read until the client hung up, without a bound, which could exhaust its memory); a camera named more than once in `cameras` is rendered and returned once (N copies of one name made it build N x N images within the frame budget); a port that is still held when the service restarts is waited for, not counted as a crash; a JSON `true` or `false` where `dt` or `seconds` takes a number is a 400 (`true` ran as 1 second). The grader: a `final.json` whose evidence fields (`control_steps`, `calls`, `elapsed_s`, `final_reward`) are not numbers no longer stops it before it writes a reward (they are reported as 0 and decide nothing). An episode's physics and the pass rule are unchanged.
- 2026-10-13 (owner, boards 0085 and 0087): the sim service's clock (`environment/sim/server_casa.py`, the docstring of `environment/rcb_client.py`). `meta["time_s"]` and `meta["time_left_s"]` count from the agent's first request to the service (any endpoint), and a restarted service keeps that moment; before, they counted from the first start of the service's container (`time_left_s`) and from when its scene was ready (`time_s`). Where the service is ready before the agent's phase begins (the harness is installed in between: a median of 79 s over the recorded standard trials), that clock ran ahead of the agent's and `time_left_s` reached 0 while the agent still had that time. `/spec` says what the clock counts from (`budget.clock`), and the instruction's `## Time limit` tells the agent to plan by `time_left_s`. The service never ended an episode on this clock: the scene, the frozen instance, the pass rule and the grader are unchanged, so the task keeps its name and earlier results stand.
- 2026-10-13 (board 0093; found in board 0086): the camera calibration (`environment/sim/server_casa.py`). `cam2world_gl` held the pose of MuJoCo's left-eye camera (`scene.camera[0]`), 3.4 cm (half the model's 0.068 m interpupillary distance) to the camera's left of the viewpoint the mono RGB and depth images are rendered from; it now is that viewpoint (`data.cam_xpos` / `data.cam_xmat`). Checked against MuJoCo's own renders of marker spheres at known world points, every served camera, before and after the robot moved: the reprojection error was 10.5-10.7 px median (up to 28 px for near points, always to the same side), now 0.15 px median and at most 0.4 px. The intrinsics were right (every camera here is given by its `fovy`, none by a sensor size) and are unchanged, as are the images, the scene, the frozen instance, the pass rule and the grader, so the task keeps its name. **Standard-mode results from before this date used the offset calibration**: a point located from the images and the reported calibration came out about 3.4 cm to the side (agents missed a microwave button by 3.5 cm), so they are not strictly comparable with later ones.
+- 2026-10-14 (board 0113, from board 0086): the success text now says that the faucet handle has two hinges at one pivot: the water on / off hinge (range 0 to 0.5 rad, about the sink's front-to-back axis, so the handle swings sideways) and a temperature hinge (about ±0.79 rad) that tilts it toward or away from you and never turns the water on; before, it named only "the handle joint" (board 0086: every standard run of `turn-on-sink-faucet` and `thaw-in-sink` moved the temperature hinge). Text only: the scene, the frozen instance, the reference solution, the environment's success check and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/robocasa365-turn-on-sink-faucet-i00-standard/instruction.md b/tasks/robocasa365-turn-on-sink-faucet-i00-standard/instruction.md
index cfd3240bd3..ca52366afc 100644
--- a/tasks/robocasa365-turn-on-sink-faucet-i00-standard/instruction.md
+++ b/tasks/robocasa365-turn-on-sink-faucet-i00-standard/instruction.md
@@ -23,7 +23,7 @@ the environment's own success check for `TurnOnSinkFaucet` to the state **after
in the live episode and in the replay. A success reached earlier and then lost does not count: end in the state you
want graded and stop submitting.
-The check for `TurnOnSinkFaucet` requires that the faucet handle is in its water-on range (handle joint between 0.40 rad and π rad).
+The check for `TurnOnSinkFaucet` requires that the faucet handle is in its water-on range (its on / off hinge between 0.40 rad and π rad). The handle turns on two hinges at one pivot: the on / off hinge, whose range is 0 to 0.5 rad (so the water is on only over the last fifth of its travel), turns about the sink's front-to-back axis, so the handle swings sideways; the temperature hinge (about ±0.79 rad) tilts the handle toward or away from you and never turns the water on or off.
## Client (`/opt/tools/rcb_client.py`, protocol `eai-standard/2.1`)
```python
diff --git a/tasks/robocasa365-turn-on-sink-faucet-i00-standard/task.toml b/tasks/robocasa365-turn-on-sink-faucet-i00-standard/task.toml
index 1105235369..9a5e1c589f 100644
--- a/tasks/robocasa365-turn-on-sink-faucet-i00-standard/task.toml
+++ b/tasks/robocasa365-turn-on-sink-faucet-i00-standard/task.toml
@@ -24,6 +24,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa365/TurnOnSinkFaucet"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
expert_time_estimate_hours = 4.0
diff --git a/tasks/robocasa365-turn-on-stove-i00-privileged/task.toml b/tasks/robocasa365-turn-on-stove-i00-privileged/task.toml
index 985fe1b2d0..ca0a406f86 100644
--- a/tasks/robocasa365-turn-on-stove-i00-privileged/task.toml
+++ b/tasks/robocasa365-turn-on-stove-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa365/TurnOnStove"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robocasa365-turn-on-stove-i00-standard/task.toml b/tasks/robocasa365-turn-on-stove-i00-standard/task.toml
index 63dc7d7071..ceb22a299c 100644
--- a/tasks/robocasa365-turn-on-stove-i00-standard/task.toml
+++ b/tasks/robocasa365-turn-on-stove-i00-standard/task.toml
@@ -24,6 +24,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa365/TurnOnStove"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
expert_time_estimate_hours = 4.0
diff --git a/tasks/robocasa365-turn-on-toaster-oven-i00-privileged/task.toml b/tasks/robocasa365-turn-on-toaster-oven-i00-privileged/task.toml
index 5091c8556a..f3bf25b681 100644
--- a/tasks/robocasa365-turn-on-toaster-oven-i00-privileged/task.toml
+++ b/tasks/robocasa365-turn-on-toaster-oven-i00-privileged/task.toml
@@ -15,6 +15,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa365/TurnOnToasterOven"
mode = "privileged"
+dynamics = "dynamic" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robocasa365-turn-on-toaster-oven-i00-standard/task.toml b/tasks/robocasa365-turn-on-toaster-oven-i00-standard/task.toml
index 8f3eba8c71..4c7a57012d 100644
--- a/tasks/robocasa365-turn-on-toaster-oven-i00-standard/task.toml
+++ b/tasks/robocasa365-turn-on-toaster-oven-i00-standard/task.toml
@@ -24,6 +24,7 @@ simulator = "robosuite"
robot = "franka-panda-omron"
family = "robocasa365/TurnOnToasterOven"
mode = "standard"
+dynamics = "dynamic" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
expert_time_estimate_hours = 4.0
diff --git a/tasks/robodojo-align-blocks-i00-privileged/task.toml b/tasks/robodojo-align-blocks-i00-privileged/task.toml
index 656b34aea0..2a0f13cad9 100644
--- a/tasks/robodojo-align-blocks-i00-privileged/task.toml
+++ b/tasks/robodojo-align-blocks-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "robodojo" # RoboDojo @ cb2a37d on its Isaac Lab 2.3 fork
robot = "arx-x5-bimanual"
family = "align_blocks"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "no positive example yet; no reference solution (RoboDojo releases no expert, and its training demonstrations carry no initial state); reported separately until one exists; human review in PR #142"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robodojo-align-blocks-i00-standard/task.toml b/tasks/robodojo-align-blocks-i00-standard/task.toml
index 7309b12531..3231186bb8 100644
--- a/tasks/robodojo-align-blocks-i00-standard/task.toml
+++ b/tasks/robodojo-align-blocks-i00-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "robodojo"
robot = "arx-x5-bimanual"
family = "align_blocks"
mode = "standard" # robot as a service: RGB, depth + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "no positive example yet; no reference solution (RoboDojo releases no expert, and its training demonstrations carry no initial state); reported separately until one exists; human review in PR #142"
diff --git a/tasks/robodojo-arrange-largest-number-i00-privileged/task.toml b/tasks/robodojo-arrange-largest-number-i00-privileged/task.toml
index 8eb1bda9c4..b843526ae5 100644
--- a/tasks/robodojo-arrange-largest-number-i00-privileged/task.toml
+++ b/tasks/robodojo-arrange-largest-number-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "robodojo" # RoboDojo @ cb2a37d on its Isaac Lab 2.3 fork
robot = "arx-x5-bimanual"
family = "arrange_largest_number"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "no positive example yet; no reference solution (RoboDojo releases no expert, and its training demonstrations carry no initial state); reported separately until one exists; human review in PR #142"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robodojo-arrange-largest-number-i00-standard/task.toml b/tasks/robodojo-arrange-largest-number-i00-standard/task.toml
index 0d57761592..6d09a57fb9 100644
--- a/tasks/robodojo-arrange-largest-number-i00-standard/task.toml
+++ b/tasks/robodojo-arrange-largest-number-i00-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "robodojo"
robot = "arx-x5-bimanual"
family = "arrange_largest_number"
mode = "standard" # robot as a service: RGB, depth + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "no positive example yet; no reference solution (RoboDojo releases no expert, and its training demonstrations carry no initial state); reported separately until one exists; human review in PR #142"
diff --git a/tasks/robodojo-arrange-largest-number-random-i00-privileged/task.toml b/tasks/robodojo-arrange-largest-number-random-i00-privileged/task.toml
index 25e7311b96..1ed4f4475f 100644
--- a/tasks/robodojo-arrange-largest-number-random-i00-privileged/task.toml
+++ b/tasks/robodojo-arrange-largest-number-random-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "robodojo" # RoboDojo @ cb2a37d on its Isaac Lab 2.3 fork
robot = "arx-x5-bimanual"
family = "arrange_largest_number_random"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "no positive example yet; no reference solution (RoboDojo releases no expert, and its training demonstrations carry no initial state); reported separately until one exists; human review in PR #142"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robodojo-arrange-largest-number-random-i00-standard/task.toml b/tasks/robodojo-arrange-largest-number-random-i00-standard/task.toml
index fb81d23061..1f3218c4d7 100644
--- a/tasks/robodojo-arrange-largest-number-random-i00-standard/task.toml
+++ b/tasks/robodojo-arrange-largest-number-random-i00-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "robodojo"
robot = "arx-x5-bimanual"
family = "arrange_largest_number_random"
mode = "standard" # robot as a service: RGB, depth + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "no positive example yet; no reference solution (RoboDojo releases no expert, and its training demonstrations carry no initial state); reported separately until one exists; human review in PR #142"
diff --git a/tasks/robodojo-build-tower-i00-privileged/task.toml b/tasks/robodojo-build-tower-i00-privileged/task.toml
index 91f492a6cb..ac9358e933 100644
--- a/tasks/robodojo-build-tower-i00-privileged/task.toml
+++ b/tasks/robodojo-build-tower-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "robodojo" # RoboDojo @ cb2a37d on its Isaac Lab 2.3 fork
robot = "arx-x5-bimanual"
family = "build_tower"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "no positive example yet; no reference solution (RoboDojo releases no expert, and its training demonstrations carry no initial state); reported separately until one exists; human review in PR #142"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robodojo-build-tower-i00-standard/task.toml b/tasks/robodojo-build-tower-i00-standard/task.toml
index 12a5f2c96e..c2be9265d4 100644
--- a/tasks/robodojo-build-tower-i00-standard/task.toml
+++ b/tasks/robodojo-build-tower-i00-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "robodojo"
robot = "arx-x5-bimanual"
family = "build_tower"
mode = "standard" # robot as a service: RGB, depth + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "no positive example yet; no reference solution (RoboDojo releases no expert, and its training demonstrations carry no initial state); reported separately until one exists; human review in PR #142"
diff --git a/tasks/robodojo-classify-objects-by-language-i00-privileged/task.toml b/tasks/robodojo-classify-objects-by-language-i00-privileged/task.toml
index a55177ca98..5772c249d5 100644
--- a/tasks/robodojo-classify-objects-by-language-i00-privileged/task.toml
+++ b/tasks/robodojo-classify-objects-by-language-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "robodojo" # RoboDojo @ cb2a37d on its Isaac Lab 2.3 fork
robot = "arx-x5-bimanual"
family = "classify_objects_by_language"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "no positive example yet; no reference solution (RoboDojo releases no expert, and its training demonstrations carry no initial state); reported separately until one exists; human review in PR #142"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robodojo-classify-objects-by-language-i00-standard/task.toml b/tasks/robodojo-classify-objects-by-language-i00-standard/task.toml
index 915be21d26..5b1bf0872c 100644
--- a/tasks/robodojo-classify-objects-by-language-i00-standard/task.toml
+++ b/tasks/robodojo-classify-objects-by-language-i00-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "robodojo"
robot = "arx-x5-bimanual"
family = "classify_objects_by_language"
mode = "standard" # robot as a service: RGB, depth + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "no positive example yet; no reference solution (RoboDojo releases no expert, and its training demonstrations carry no initial state); reported separately until one exists; human review in PR #142"
diff --git a/tasks/robodojo-classify-objects-i00-privileged/task.toml b/tasks/robodojo-classify-objects-i00-privileged/task.toml
index cc9f988097..656a4bcdbd 100644
--- a/tasks/robodojo-classify-objects-i00-privileged/task.toml
+++ b/tasks/robodojo-classify-objects-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "robodojo" # RoboDojo @ cb2a37d on its Isaac Lab 2.3 fork
robot = "arx-x5-bimanual"
family = "classify_objects"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "no positive example yet (a model trial failed (Codex + GPT-6 Luna, effort xhigh, run `codex-gpt6_luna-xhigh`, 2026-10-10)); no reference solution (RoboDojo releases no expert, and its training demonstrations carry no initial state); reported separately until one exists; human review in PR #142"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robodojo-classify-objects-i00-standard/task.toml b/tasks/robodojo-classify-objects-i00-standard/task.toml
index 796b7bc5e0..3ca606a8ae 100644
--- a/tasks/robodojo-classify-objects-i00-standard/task.toml
+++ b/tasks/robodojo-classify-objects-i00-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "robodojo"
robot = "arx-x5-bimanual"
family = "classify_objects"
mode = "standard" # robot as a service: RGB, depth + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "no positive example yet; no reference solution (RoboDojo releases no expert, and its training demonstrations carry no initial state); reported separately until one exists; human review in PR #142"
diff --git a/tasks/robodojo-cover-blocks-i00-privileged/task.toml b/tasks/robodojo-cover-blocks-i00-privileged/task.toml
index 733cac595b..6539e82d48 100644
--- a/tasks/robodojo-cover-blocks-i00-privileged/task.toml
+++ b/tasks/robodojo-cover-blocks-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "robodojo" # RoboDojo @ cb2a37d on its Isaac Lab 2.3 fork
robot = "arx-x5-bimanual"
family = "cover_blocks"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "no positive example yet; no reference solution (RoboDojo releases no expert, and its training demonstrations carry no initial state); reported separately until one exists; human review in PR #142"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robodojo-cover-blocks-i00-standard/task.toml b/tasks/robodojo-cover-blocks-i00-standard/task.toml
index e0683c4bd9..baab1a1695 100644
--- a/tasks/robodojo-cover-blocks-i00-standard/task.toml
+++ b/tasks/robodojo-cover-blocks-i00-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "robodojo"
robot = "arx-x5-bimanual"
family = "cover_blocks"
mode = "standard" # robot as a service: RGB, depth + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "no positive example yet; no reference solution (RoboDojo releases no expert, and its training demonstrations carry no initial state); reported separately until one exists; human review in PR #142"
diff --git a/tasks/robodojo-deposit-coin-i00-privileged/task.toml b/tasks/robodojo-deposit-coin-i00-privileged/task.toml
index 6cfd77a52d..a879c8db00 100644
--- a/tasks/robodojo-deposit-coin-i00-privileged/task.toml
+++ b/tasks/robodojo-deposit-coin-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "robodojo" # RoboDojo @ cb2a37d on its Isaac Lab 2.3 fork
robot = "arx-x5-bimanual"
family = "deposit_coin"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "positive example: a model trial solved this task (Codex + GPT-6 Luna, effort xhigh, run `codex-gpt6_luna-xhigh`, 2026-10-10; 491 actions, graded by this verifier); no reference solution (RoboDojo releases no expert); human review in PR #142"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robodojo-deposit-coin-i00-standard/task.toml b/tasks/robodojo-deposit-coin-i00-standard/task.toml
index 15e4cbbb9f..c7c1166494 100644
--- a/tasks/robodojo-deposit-coin-i00-standard/task.toml
+++ b/tasks/robodojo-deposit-coin-i00-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "robodojo"
robot = "arx-x5-bimanual"
family = "deposit_coin"
mode = "standard" # robot as a service: RGB, depth + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "no positive example yet (the privileged twin has a positive example (Codex + GPT-6 Luna, effort xhigh, run `codex-gpt6_luna-xhigh`, 2026-10-10)); no reference solution (RoboDojo releases no expert, and its training demonstrations carry no initial state); reported separately until one exists; human review in PR #142"
diff --git a/tasks/robodojo-fasten-screws-i00-privileged/task.toml b/tasks/robodojo-fasten-screws-i00-privileged/task.toml
index bd97501441..9a7e2a536f 100644
--- a/tasks/robodojo-fasten-screws-i00-privileged/task.toml
+++ b/tasks/robodojo-fasten-screws-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "robodojo" # RoboDojo @ cb2a37d on its Isaac Lab 2.3 fork
robot = "arx-x5-bimanual"
family = "fasten_screws"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "no positive example yet; no reference solution (RoboDojo releases no expert, and its training demonstrations carry no initial state); reported separately until one exists; human review in PR #142"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robodojo-fasten-screws-i00-standard/task.toml b/tasks/robodojo-fasten-screws-i00-standard/task.toml
index 1798c03762..c956aa758e 100644
--- a/tasks/robodojo-fasten-screws-i00-standard/task.toml
+++ b/tasks/robodojo-fasten-screws-i00-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "robodojo"
robot = "arx-x5-bimanual"
family = "fasten_screws"
mode = "standard" # robot as a service: RGB, depth + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "no positive example yet; no reference solution (RoboDojo releases no expert, and its training demonstrations carry no initial state); reported separately until one exists; human review in PR #142"
diff --git a/tasks/robodojo-fill-egg-holder-i00-privileged/task.toml b/tasks/robodojo-fill-egg-holder-i00-privileged/task.toml
index 7d0245d92c..c75345599d 100644
--- a/tasks/robodojo-fill-egg-holder-i00-privileged/task.toml
+++ b/tasks/robodojo-fill-egg-holder-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "robodojo" # RoboDojo @ cb2a37d on its Isaac Lab 2.3 fork
robot = "arx-x5-bimanual"
family = "fill_egg_holder"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "no positive example yet; no reference solution (RoboDojo releases no expert, and its training demonstrations carry no initial state); reported separately until one exists; human review in PR #142"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robodojo-fill-egg-holder-i00-standard/task.toml b/tasks/robodojo-fill-egg-holder-i00-standard/task.toml
index 64036904ba..37d82fbb33 100644
--- a/tasks/robodojo-fill-egg-holder-i00-standard/task.toml
+++ b/tasks/robodojo-fill-egg-holder-i00-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "robodojo"
robot = "arx-x5-bimanual"
family = "fill_egg_holder"
mode = "standard" # robot as a service: RGB, depth + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "no positive example yet; no reference solution (RoboDojo releases no expert, and its training demonstrations carry no initial state); reported separately until one exists; human review in PR #142"
diff --git a/tasks/robodojo-fill-pen-holder-i00-privileged/task.toml b/tasks/robodojo-fill-pen-holder-i00-privileged/task.toml
index a1f05b1239..50a456ddc1 100644
--- a/tasks/robodojo-fill-pen-holder-i00-privileged/task.toml
+++ b/tasks/robodojo-fill-pen-holder-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "robodojo" # RoboDojo @ cb2a37d on its Isaac Lab 2.3 fork
robot = "arx-x5-bimanual"
family = "fill_pen_holder"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "no positive example yet; no reference solution (RoboDojo releases no expert, and its training demonstrations carry no initial state); reported separately until one exists; human review in PR #142"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robodojo-fill-pen-holder-i00-standard/task.toml b/tasks/robodojo-fill-pen-holder-i00-standard/task.toml
index 3232614ac1..20bf28a423 100644
--- a/tasks/robodojo-fill-pen-holder-i00-standard/task.toml
+++ b/tasks/robodojo-fill-pen-holder-i00-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "robodojo"
robot = "arx-x5-bimanual"
family = "fill_pen_holder"
mode = "standard" # robot as a service: RGB, depth + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "no positive example yet; no reference solution (RoboDojo releases no expert, and its training demonstrations carry no initial state); reported separately until one exists; human review in PR #142"
diff --git a/tasks/robodojo-fold-clothes-i00-standard/task.toml b/tasks/robodojo-fold-clothes-i00-standard/task.toml
index 770571319e..5d9de70c72 100644
--- a/tasks/robodojo-fold-clothes-i00-standard/task.toml
+++ b/tasks/robodojo-fold-clothes-i00-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "robodojo"
robot = "arx-x5-bimanual"
family = "fold_clothes"
mode = "standard" # robot as a service: RGB, depth + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "no positive example yet; no reference solution (RoboDojo releases no expert, and its training demonstrations carry no initial state); reported separately until one exists; human review in PR #142"
diff --git a/tasks/robodojo-fold-clothes-random-i00-standard/task.toml b/tasks/robodojo-fold-clothes-random-i00-standard/task.toml
index 3113c8e6e4..72e0bba0c5 100644
--- a/tasks/robodojo-fold-clothes-random-i00-standard/task.toml
+++ b/tasks/robodojo-fold-clothes-random-i00-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "robodojo"
robot = "arx-x5-bimanual"
family = "fold_clothes_random"
mode = "standard" # robot as a service: RGB, depth + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "no positive example yet; no reference solution (RoboDojo releases no expert, and its training demonstrations carry no initial state); reported separately until one exists; human review in PR #142"
diff --git a/tasks/robodojo-general-pickup-i00-privileged/task.toml b/tasks/robodojo-general-pickup-i00-privileged/task.toml
index e7f89fce3c..a02d611de9 100644
--- a/tasks/robodojo-general-pickup-i00-privileged/task.toml
+++ b/tasks/robodojo-general-pickup-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "robodojo" # RoboDojo @ cb2a37d on its Isaac Lab 2.3 fork
robot = "arx-x5-bimanual"
family = "general_pickup"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "positive example: a model trial solved this task (Codex + GPT-6 Luna, effort xhigh, run `codex-gpt6_luna-xhigh`, 2026-10-10; 67 actions, graded by this verifier); no reference solution (RoboDojo releases no expert); human review in PR #142"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robodojo-general-pickup-i00-standard/task.toml b/tasks/robodojo-general-pickup-i00-standard/task.toml
index 8fdb74ed32..3a8380ba7e 100644
--- a/tasks/robodojo-general-pickup-i00-standard/task.toml
+++ b/tasks/robodojo-general-pickup-i00-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "robodojo"
robot = "arx-x5-bimanual"
family = "general_pickup"
mode = "standard" # robot as a service: RGB, depth + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "positive example: a model trial solved this task (Codex + GPT-6 Luna, effort xhigh, run `codex-gpt6_luna-xhigh`, 2026-10-10; 1131 actions, graded by this verifier); no reference solution (RoboDojo releases no expert); human review in PR #142"
diff --git a/tasks/robodojo-hang-mugs-i00-privileged/README.md b/tasks/robodojo-hang-mugs-i00-privileged/README.md
index 673254eeb0..3f1edd1ab2 100644
--- a/tasks/robodojo-hang-mugs-i00-privileged/README.md
+++ b/tasks/robodojo-hang-mugs-i00-privileged/README.md
@@ -58,3 +58,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
### Change Log
- 2026-10-11: agent time limit 5400 s -> 3600 s (`[agent] timeout_sec`; owner decision, docs/benchmarks/robodojo.md), and `instruction.md` states the cutoff at the top. The scene, the frozen layout, RoboDojo's success check and the verifier are unchanged, so the task keeps its name; trials recorded before this date ran with 5400 s and are kept apart (no official result was published with it).
- 2026-10-12: graded by ONE fresh-process replay (two before; owner decision, docs/benchmarks/robodojo.md): reward 1 iff that replay ends in RoboDojo's success. Before, both replays had to succeed and end in the same state hash; with one replay nothing is compared, so `reward.json` reports `deterministic` = 1 with `determinism_checked` = 0 and `replays` = 1, and a replay whose process did not rebuild the frozen layout to its recorded hash gives no verdict (`instance_mismatch`, `infra_error`). `[verifier] timeout_sec` 15000 s -> 7800 s (one replay of the 15,000 rows a trajectory may have, 7200 s, + 600 s). The scene, the frozen layout, RoboDojo's success check and what the agent hands in are unchanged, so the task keeps its name; `instruction.md` says one replay. No ranked result exists; a trial graded before this date is re-read from its saved `reward.json` (`all_success`, `state_identical`), not rerun: in the first 38 graded trials of the run that began on 2026-10-11 the two replays always ended in the same state, and no verdict changes.
+- 2026-10-14 (board 0113, from board 0086): the instruction now says that the two orientation conditions compare the mug's FIRST 'handle' functional point with the rack's FIRST 'handle_support_point' (RoboDojo's `is_axis_aligned` takes element [0] of each list), not with the hook the mug hangs on; the position condition is against any hook. Text only: the scene, the frozen layout, RoboDojo's success check and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/robodojo-hang-mugs-i00-privileged/instruction.md b/tasks/robodojo-hang-mugs-i00-privileged/instruction.md
index 3dc533facf..a818c98026 100644
--- a/tasks/robodojo-hang-mugs-i00-privileged/instruction.md
+++ b/tasks/robodojo-hang-mugs-i00-privileged/instruction.md
@@ -23,8 +23,8 @@ robot's start pose never change between resets. You may reset and retry as many
- For each mug m in {mug0, mug1, mug2} (all three, simultaneously):
- - lifted: m's current origin z is more than 0.045 m above its origin z at episode start (is_lift, z_threshold=0.045; current height, not 'ever lifted').
- - m's 'handle' active functional point has world z > 0.9 m (any handle point; is_A_functional_point_higher_than_z, z=0.9).
-- - m's 'handle' active functional-point frame z-axis is within 60 deg of the z-axis of the cup_holder's 'handle_support_point' passive functional-point frame.
-- - m's handle frame x-axis is within 60 deg of the cup_holder handle_support_point frame's +x OR -x axis (either sign accepted).
+- - m's 'handle' active functional-point frame z-axis is within 60 deg of the z-axis of the cup_holder's 'handle_support_point' passive functional-point frame. Both frames are the FIRST functional point of that tag (RoboDojo's is_axis_aligned takes element [0] of each list): the mug's first 'handle' point and the rack's first 'handle_support_point', whichever hook the mug actually hangs on.
+- - m's handle frame x-axis is within 60 deg of the cup_holder handle_support_point frame's +x OR -x axis (either sign accepted; again the first handle point and the rack's first support point).
- - m's 'handle' active functional point is within 0.045 m (3-D Euclidean) of a cup_holder 'handle_support_point' passive functional point (is_A_functional_point_close_to_B_functional_point, threshold=0.045, no orientation-alignment option).
- all_robot_back_to_origin(): every agent-controlled (target) arm's end-effector must be within 0.15 m on EACH of x, y, z (per-axis) and within 20 degrees of its pose at the start of the episode, at the same action boundary as all other conditions. Grippers are NOT required to be open by the success test (only by partial-credit scoring).
diff --git a/tasks/robodojo-hang-mugs-i00-privileged/task.toml b/tasks/robodojo-hang-mugs-i00-privileged/task.toml
index 3f4c697629..3db0d64bd3 100644
--- a/tasks/robodojo-hang-mugs-i00-privileged/task.toml
+++ b/tasks/robodojo-hang-mugs-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "robodojo" # RoboDojo @ cb2a37d on its Isaac Lab 2.3 fork
robot = "arx-x5-bimanual"
family = "hang_mugs"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "no positive example yet; no reference solution (RoboDojo releases no expert, and its training demonstrations carry no initial state); reported separately until one exists; human review in PR #142"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robodojo-hang-mugs-i00-standard/README.md b/tasks/robodojo-hang-mugs-i00-standard/README.md
index cfa71b3fc4..d57e0af1ab 100644
--- a/tasks/robodojo-hang-mugs-i00-standard/README.md
+++ b/tasks/robodojo-hang-mugs-i00-standard/README.md
@@ -56,3 +56,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
- 2026-10-11: agent time limit 5400 s -> 3600 s (`[agent] timeout_sec`, and the service's `RCB_WALL_CLOCK_S`, which only reports it: `spec["budget"]`, `meta["time_left_s"]`; owner decision, docs/benchmarks/robodojo.md), and `instruction.md` states the cutoff at the top. The scene, the frozen layout, RoboDojo's success check, the collect hook and the verifier are unchanged, so the task keeps its name; trials recorded before this date ran with 5400 s and are kept apart (no official result was published with it).
- 2026-10-12: graded with ONE fresh-process replay of the recorded episode (two before; owner decision, docs/benchmarks/robodojo.md): `success` = `live_success` and `replay_success`, one successful replay is enough. Before, two replays had to succeed and end in the same state hash. `deterministic` (= `state_identical` = `live_matches_replay`) now says whether the replay ended in the live episode's state; it is reported and no longer decides. The collect hook's limit 18000 s -> 10800 s (`[[verifier.collect]] timeout_sec` = `finalize.sh`'s `HOOK_BUDGET_S`); inside it the replay has max(1800 s, 1200 s + actions / 2.5) and is run once more when it gives no result in that time. The scene, the frozen layout, RoboDojo's success check and the service the agent talks to are unchanged, so the task keeps its name. No ranked result exists; a trial graded before this date is re-read from its saved `reward.json` (`live_success`, `replay_success`, `state_identical`), not rerun: in the first 38 graded trials of the run that began on 2026-10-11 no verdict changes.
- 2026-10-13 (owner, boards 0085 and 0087): the sim service's clock (`environment/sim/rcb_limited.py`, `environment/sim/server.py`, the docstring of `environment/rcb_client.py`). `meta["time_s"]` and `meta["time_left_s"]` count from the agent's first request to the service (any endpoint), and a restarted service keeps that moment; before, they counted from the service's own start of the episode (when its scene was ready). Where the service is ready before the agent's phase begins (the harness is installed in between: a median of 79 s over the recorded standard trials), that clock ran ahead of the agent's and `time_left_s` reached 0 while the agent still had that time. `/spec` says what the clock counts from (`budget.clock`), and the instruction's `## Time limit` tells the agent to plan by `time_left_s`. The service never ended an episode on this clock: the scene, the frozen instance, the pass rule and the grader are unchanged, so the task keeps its name and earlier results stand.
+- 2026-10-14 (board 0113, from board 0086): the instruction now says that the two orientation conditions compare the mug's FIRST 'handle' functional point with the rack's FIRST 'handle_support_point' (RoboDojo's `is_axis_aligned` takes element [0] of each list), not with the hook the mug hangs on; the position condition is against any hook. Text only: the scene, the frozen layout, RoboDojo's success check and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/robodojo-hang-mugs-i00-standard/instruction.md b/tasks/robodojo-hang-mugs-i00-standard/instruction.md
index d4fac1f6b4..9fa7d7b55a 100644
--- a/tasks/robodojo-hang-mugs-i00-standard/instruction.md
+++ b/tasks/robodojo-hang-mugs-i00-standard/instruction.md
@@ -16,8 +16,8 @@ RoboDojo's own instruction for this scene: *"Hang all the mugs on the mug rack."
- For each mug m in {mug0, mug1, mug2} (all three, simultaneously):
- - lifted: m's current origin z is more than 0.045 m above its origin z at episode start (is_lift, z_threshold=0.045; current height, not 'ever lifted').
- - m's 'handle' active functional point has world z > 0.9 m (any handle point; is_A_functional_point_higher_than_z, z=0.9).
-- - m's 'handle' active functional-point frame z-axis is within 60 deg of the z-axis of the cup_holder's 'handle_support_point' passive functional-point frame.
-- - m's handle frame x-axis is within 60 deg of the cup_holder handle_support_point frame's +x OR -x axis (either sign accepted).
+- - m's 'handle' active functional-point frame z-axis is within 60 deg of the z-axis of the cup_holder's 'handle_support_point' passive functional-point frame. Both frames are the FIRST functional point of that tag (RoboDojo's is_axis_aligned takes element [0] of each list): the mug's first 'handle' point and the rack's first 'handle_support_point', whichever hook the mug actually hangs on.
+- - m's handle frame x-axis is within 60 deg of the cup_holder handle_support_point frame's +x OR -x axis (either sign accepted; again the first handle point and the rack's first support point).
- - m's 'handle' active functional point is within 0.045 m (3-D Euclidean) of a cup_holder 'handle_support_point' passive functional point (is_A_functional_point_close_to_B_functional_point, threshold=0.045, no orientation-alignment option).
- all_robot_back_to_origin(): every agent-controlled (target) arm's end-effector must be within 0.15 m on EACH of x, y, z (per-axis) and within 20 degrees of its pose at the start of the episode, at the same action boundary as all other conditions. Grippers are NOT required to be open by the success test (only by partial-credit scoring).
diff --git a/tasks/robodojo-hang-mugs-i00-standard/task.toml b/tasks/robodojo-hang-mugs-i00-standard/task.toml
index 92f93b472c..ca7643eef3 100644
--- a/tasks/robodojo-hang-mugs-i00-standard/task.toml
+++ b/tasks/robodojo-hang-mugs-i00-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "robodojo"
robot = "arx-x5-bimanual"
family = "hang_mugs"
mode = "standard" # robot as a service: RGB, depth + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "no positive example yet; no reference solution (RoboDojo releases no expert, and its training demonstrations carry no initial state); reported separately until one exists; human review in PR #142"
diff --git a/tasks/robodojo-hang-mugs-random-i00-privileged/README.md b/tasks/robodojo-hang-mugs-random-i00-privileged/README.md
index 276928a392..1b9dd6f0d9 100644
--- a/tasks/robodojo-hang-mugs-random-i00-privileged/README.md
+++ b/tasks/robodojo-hang-mugs-random-i00-privileged/README.md
@@ -58,3 +58,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
### Change Log
- 2026-10-11: agent time limit 5400 s -> 3600 s (`[agent] timeout_sec`; owner decision, docs/benchmarks/robodojo.md), and `instruction.md` states the cutoff at the top. The scene, the frozen layout, RoboDojo's success check and the verifier are unchanged, so the task keeps its name; trials recorded before this date ran with 5400 s and are kept apart (no official result was published with it).
- 2026-10-12: graded by ONE fresh-process replay (two before; owner decision, docs/benchmarks/robodojo.md): reward 1 iff that replay ends in RoboDojo's success. Before, both replays had to succeed and end in the same state hash; with one replay nothing is compared, so `reward.json` reports `deterministic` = 1 with `determinism_checked` = 0 and `replays` = 1, and a replay whose process did not rebuild the frozen layout to its recorded hash gives no verdict (`instance_mismatch`, `infra_error`). `[verifier] timeout_sec` 15000 s -> 7800 s (one replay of the 15,000 rows a trajectory may have, 7200 s, + 600 s). The scene, the frozen layout, RoboDojo's success check and what the agent hands in are unchanged, so the task keeps its name; `instruction.md` says one replay. No ranked result exists; a trial graded before this date is re-read from its saved `reward.json` (`all_success`, `state_identical`), not rerun: in the first 38 graded trials of the run that began on 2026-10-11 the two replays always ended in the same state, and no verdict changes.
+- 2026-10-14 (board 0113, from board 0086): the instruction now says that the two orientation conditions compare the mug's FIRST 'handle' functional point with the rack's FIRST 'handle_support_point' (RoboDojo's `is_axis_aligned` takes element [0] of each list), not with the hook the mug hangs on; the position condition is against any hook. Text only: the scene, the frozen layout, RoboDojo's success check and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/robodojo-hang-mugs-random-i00-privileged/instruction.md b/tasks/robodojo-hang-mugs-random-i00-privileged/instruction.md
index 716551ec44..063cb4111c 100644
--- a/tasks/robodojo-hang-mugs-random-i00-privileged/instruction.md
+++ b/tasks/robodojo-hang-mugs-random-i00-privileged/instruction.md
@@ -23,8 +23,8 @@ robot's start pose never change between resets. You may reset and retry as many
## Success (RoboDojo's own check for this task)
- For each of mug0, mug1, mug2: root z is more than 0.045 m above the mug's initial root z (is_lift z_threshold=0.045, strict), i.e. the mug has been lifted at least 4.5 cm from where it started.
- For each mug: at least one of its active functional points tagged 'handle' has world z > 0.9 m (is_A_functional_point_higher_than_z, absolute z in the simulation world frame; the exact reference frame/table height is not determinable statically, so this is an absolute height).
-- For each mug: the handle point's local +z axis is within 60 degrees (strict <60) of the rack's passive functional point 'handle_support_point' local +z axis (is_axis_aligned with functional_point_A='handle' (active), functional_point_B='handle_support_point' (passive)).
-- For each mug: the handle point's local +x axis is within 60 degrees of either the support point's local +x or its local -x axis (parallel or anti-parallel).
+- For each mug: the handle point's local +z axis is within 60 degrees (strict <60) of the rack's passive functional point 'handle_support_point' local +z axis (is_axis_aligned with functional_point_A='handle' (active), functional_point_B='handle_support_point' (passive)). Both frames are the FIRST point of that tag (element [0] of each list): the mug's first 'handle' point and the rack's first 'handle_support_point', whichever hook the mug actually hangs on.
+- For each mug: the handle point's local +x axis is within 60 degrees of either the support point's local +x or its local -x axis (parallel or anti-parallel; again the first handle point and the rack's first support point).
- For each mug: the handle point (active) is within 0.045 m (3D Euclidean distance, strict <) of at least one 'handle_support_point' (passive) of cup_holder (is_A_functional_point_close_to_B_functional_point threshold=0.045, no orientation check on that one). Each mug is tested against any support point of the rack independently; the code does not require different mugs to use different hooks.
- All target robots back at start end-effector pose: |dx|,|dy|,|dz| each <= 0.15 m and rotation <= 20 degrees (all_robot_back_to_origin defaults).
diff --git a/tasks/robodojo-hang-mugs-random-i00-privileged/task.toml b/tasks/robodojo-hang-mugs-random-i00-privileged/task.toml
index cb1082cb15..6b88c8d920 100644
--- a/tasks/robodojo-hang-mugs-random-i00-privileged/task.toml
+++ b/tasks/robodojo-hang-mugs-random-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "robodojo" # RoboDojo @ cb2a37d on its Isaac Lab 2.3 fork
robot = "arx-x5-bimanual"
family = "hang_mugs_random"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "no positive example yet; no reference solution (RoboDojo releases no expert, and its training demonstrations carry no initial state); reported separately until one exists; human review in PR #142"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robodojo-hang-mugs-random-i00-standard/README.md b/tasks/robodojo-hang-mugs-random-i00-standard/README.md
index fbc80129d4..99bf94356a 100644
--- a/tasks/robodojo-hang-mugs-random-i00-standard/README.md
+++ b/tasks/robodojo-hang-mugs-random-i00-standard/README.md
@@ -56,3 +56,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
- 2026-10-11: agent time limit 5400 s -> 3600 s (`[agent] timeout_sec`, and the service's `RCB_WALL_CLOCK_S`, which only reports it: `spec["budget"]`, `meta["time_left_s"]`; owner decision, docs/benchmarks/robodojo.md), and `instruction.md` states the cutoff at the top. The scene, the frozen layout, RoboDojo's success check, the collect hook and the verifier are unchanged, so the task keeps its name; trials recorded before this date ran with 5400 s and are kept apart (no official result was published with it).
- 2026-10-12: graded with ONE fresh-process replay of the recorded episode (two before; owner decision, docs/benchmarks/robodojo.md): `success` = `live_success` and `replay_success`, one successful replay is enough. Before, two replays had to succeed and end in the same state hash. `deterministic` (= `state_identical` = `live_matches_replay`) now says whether the replay ended in the live episode's state; it is reported and no longer decides. The collect hook's limit 18000 s -> 10800 s (`[[verifier.collect]] timeout_sec` = `finalize.sh`'s `HOOK_BUDGET_S`); inside it the replay has max(1800 s, 1200 s + actions / 2.5) and is run once more when it gives no result in that time. The scene, the frozen layout, RoboDojo's success check and the service the agent talks to are unchanged, so the task keeps its name. No ranked result exists; a trial graded before this date is re-read from its saved `reward.json` (`live_success`, `replay_success`, `state_identical`), not rerun: in the first 38 graded trials of the run that began on 2026-10-11 no verdict changes.
- 2026-10-13 (owner, boards 0085 and 0087): the sim service's clock (`environment/sim/rcb_limited.py`, `environment/sim/server.py`, the docstring of `environment/rcb_client.py`). `meta["time_s"]` and `meta["time_left_s"]` count from the agent's first request to the service (any endpoint), and a restarted service keeps that moment; before, they counted from the service's own start of the episode (when its scene was ready). Where the service is ready before the agent's phase begins (the harness is installed in between: a median of 79 s over the recorded standard trials), that clock ran ahead of the agent's and `time_left_s` reached 0 while the agent still had that time. `/spec` says what the clock counts from (`budget.clock`), and the instruction's `## Time limit` tells the agent to plan by `time_left_s`. The service never ended an episode on this clock: the scene, the frozen instance, the pass rule and the grader are unchanged, so the task keeps its name and earlier results stand.
+- 2026-10-14 (board 0113, from board 0086): the instruction now says that the two orientation conditions compare the mug's FIRST 'handle' functional point with the rack's FIRST 'handle_support_point' (RoboDojo's `is_axis_aligned` takes element [0] of each list), not with the hook the mug hangs on; the position condition is against any hook. Text only: the scene, the frozen layout, RoboDojo's success check and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/robodojo-hang-mugs-random-i00-standard/instruction.md b/tasks/robodojo-hang-mugs-random-i00-standard/instruction.md
index 198da36b93..99196fce8e 100644
--- a/tasks/robodojo-hang-mugs-random-i00-standard/instruction.md
+++ b/tasks/robodojo-hang-mugs-random-i00-standard/instruction.md
@@ -15,8 +15,8 @@ RoboDojo's own instruction for this scene: *"Hang all the mugs on the mug rack."
## Success (RoboDojo's own check for this task)
- For each of mug0, mug1, mug2: root z is more than 0.045 m above the mug's initial root z (is_lift z_threshold=0.045, strict), i.e. the mug has been lifted at least 4.5 cm from where it started.
- For each mug: at least one of its active functional points tagged 'handle' has world z > 0.9 m (is_A_functional_point_higher_than_z, absolute z in the simulation world frame; the exact reference frame/table height is not determinable statically, so this is an absolute height).
-- For each mug: the handle point's local +z axis is within 60 degrees (strict <60) of the rack's passive functional point 'handle_support_point' local +z axis (is_axis_aligned with functional_point_A='handle' (active), functional_point_B='handle_support_point' (passive)).
-- For each mug: the handle point's local +x axis is within 60 degrees of either the support point's local +x or its local -x axis (parallel or anti-parallel).
+- For each mug: the handle point's local +z axis is within 60 degrees (strict <60) of the rack's passive functional point 'handle_support_point' local +z axis (is_axis_aligned with functional_point_A='handle' (active), functional_point_B='handle_support_point' (passive)). Both frames are the FIRST point of that tag (element [0] of each list): the mug's first 'handle' point and the rack's first 'handle_support_point', whichever hook the mug actually hangs on.
+- For each mug: the handle point's local +x axis is within 60 degrees of either the support point's local +x or its local -x axis (parallel or anti-parallel; again the first handle point and the rack's first support point).
- For each mug: the handle point (active) is within 0.045 m (3D Euclidean distance, strict <) of at least one 'handle_support_point' (passive) of cup_holder (is_A_functional_point_close_to_B_functional_point threshold=0.045, no orientation check on that one). Each mug is tested against any support point of the rack independently; the code does not require different mugs to use different hooks.
- All target robots back at start end-effector pose: |dx|,|dy|,|dz| each <= 0.15 m and rotation <= 20 degrees (all_robot_back_to_origin defaults).
diff --git a/tasks/robodojo-hang-mugs-random-i00-standard/task.toml b/tasks/robodojo-hang-mugs-random-i00-standard/task.toml
index 1f5718ea74..8e1c7adb81 100644
--- a/tasks/robodojo-hang-mugs-random-i00-standard/task.toml
+++ b/tasks/robodojo-hang-mugs-random-i00-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "robodojo"
robot = "arx-x5-bimanual"
family = "hang_mugs_random"
mode = "standard" # robot as a service: RGB, depth + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "no positive example yet; no reference solution (RoboDojo releases no expert, and its training demonstrations carry no initial state); reported separately until one exists; human review in PR #142"
diff --git a/tasks/robodojo-imitate-sorting-sequence-i00-privileged/task.toml b/tasks/robodojo-imitate-sorting-sequence-i00-privileged/task.toml
index 74ca368eba..4b8c84defb 100644
--- a/tasks/robodojo-imitate-sorting-sequence-i00-privileged/task.toml
+++ b/tasks/robodojo-imitate-sorting-sequence-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "robodojo" # RoboDojo @ cb2a37d on its Isaac Lab 2.3 fork
robot = "arx-x5-bimanual"
family = "imitate_sorting_sequence"
mode = "privileged"
+dynamics = "dynamic" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "no positive example yet; no reference solution (RoboDojo releases no expert, and its training demonstrations carry no initial state); reported separately until one exists; human review in PR #142"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robodojo-imitate-sorting-sequence-i00-standard/task.toml b/tasks/robodojo-imitate-sorting-sequence-i00-standard/task.toml
index eda6079d70..17aa325fbb 100644
--- a/tasks/robodojo-imitate-sorting-sequence-i00-standard/task.toml
+++ b/tasks/robodojo-imitate-sorting-sequence-i00-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "robodojo"
robot = "arx-x5-bimanual"
family = "imitate_sorting_sequence"
mode = "standard" # robot as a service: RGB, depth + proprioception only; no object poses, no success predicate, no reset
+dynamics = "dynamic" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "no positive example yet; no reference solution (RoboDojo releases no expert, and its training demonstrations carry no initial state); reported separately until one exists; human review in PR #142"
diff --git a/tasks/robodojo-insert-key-i00-privileged/task.toml b/tasks/robodojo-insert-key-i00-privileged/task.toml
index b4e0938f54..41ac4d99c5 100644
--- a/tasks/robodojo-insert-key-i00-privileged/task.toml
+++ b/tasks/robodojo-insert-key-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "robodojo" # RoboDojo @ cb2a37d on its Isaac Lab 2.3 fork
robot = "arx-x5-bimanual"
family = "insert_key"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "no positive example yet; no reference solution (RoboDojo releases no expert, and its training demonstrations carry no initial state); reported separately until one exists; human review in PR #142"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robodojo-insert-key-i00-standard/task.toml b/tasks/robodojo-insert-key-i00-standard/task.toml
index f8ce03768e..7a280cec18 100644
--- a/tasks/robodojo-insert-key-i00-standard/task.toml
+++ b/tasks/robodojo-insert-key-i00-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "robodojo"
robot = "arx-x5-bimanual"
family = "insert_key"
mode = "standard" # robot as a service: RGB, depth + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "no positive example yet; no reference solution (RoboDojo releases no expert, and its training demonstrations carry no initial state); reported separately until one exists; human review in PR #142"
diff --git a/tasks/robodojo-insert-tubes-i00-privileged/task.toml b/tasks/robodojo-insert-tubes-i00-privileged/task.toml
index 8bfd9f50fa..ee496c0d33 100644
--- a/tasks/robodojo-insert-tubes-i00-privileged/task.toml
+++ b/tasks/robodojo-insert-tubes-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "robodojo" # RoboDojo @ cb2a37d on its Isaac Lab 2.3 fork
robot = "arx-x5-bimanual"
family = "insert_tubes"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "no positive example yet; no reference solution (RoboDojo releases no expert, and its training demonstrations carry no initial state); reported separately until one exists; human review in PR #142"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robodojo-insert-tubes-i00-standard/task.toml b/tasks/robodojo-insert-tubes-i00-standard/task.toml
index 6da68f0ca0..8f0a394f41 100644
--- a/tasks/robodojo-insert-tubes-i00-standard/task.toml
+++ b/tasks/robodojo-insert-tubes-i00-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "robodojo"
robot = "arx-x5-bimanual"
family = "insert_tubes"
mode = "standard" # robot as a service: RGB, depth + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "no positive example yet; no reference solution (RoboDojo releases no expert, and its training demonstrations carry no initial state); reported separately until one exists; human review in PR #142"
diff --git a/tasks/robodojo-make-kong-i00-privileged/task.toml b/tasks/robodojo-make-kong-i00-privileged/task.toml
index d101b5ab9d..7475cdcb90 100644
--- a/tasks/robodojo-make-kong-i00-privileged/task.toml
+++ b/tasks/robodojo-make-kong-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "robodojo" # RoboDojo @ cb2a37d on its Isaac Lab 2.3 fork
robot = "arx-x5-bimanual"
family = "make_kong"
mode = "privileged"
+dynamics = "dynamic" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "no positive example yet; no reference solution (RoboDojo releases no expert, and its training demonstrations carry no initial state); reported separately until one exists; human review in PR #142"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robodojo-make-kong-i00-standard/task.toml b/tasks/robodojo-make-kong-i00-standard/task.toml
index bb232599ee..f3f6b93548 100644
--- a/tasks/robodojo-make-kong-i00-standard/task.toml
+++ b/tasks/robodojo-make-kong-i00-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "robodojo"
robot = "arx-x5-bimanual"
family = "make_kong"
mode = "standard" # robot as a service: RGB, depth + proprioception only; no object poses, no success predicate, no reset
+dynamics = "dynamic" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "no positive example yet; no reference solution (RoboDojo releases no expert, and its training demonstrations carry no initial state); reported separately until one exists; human review in PR #142"
diff --git a/tasks/robodojo-make-toast-i00-privileged/task.toml b/tasks/robodojo-make-toast-i00-privileged/task.toml
index f5fe945d9a..cd2a922b0a 100644
--- a/tasks/robodojo-make-toast-i00-privileged/task.toml
+++ b/tasks/robodojo-make-toast-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "robodojo" # RoboDojo @ cb2a37d on its Isaac Lab 2.3 fork
robot = "arx-x5-bimanual"
family = "make_toast"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "no positive example yet; no reference solution (RoboDojo releases no expert, and its training demonstrations carry no initial state); reported separately until one exists; human review in PR #142"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robodojo-make-toast-i00-standard/task.toml b/tasks/robodojo-make-toast-i00-standard/task.toml
index 7f6bc1975d..4de8ed8faa 100644
--- a/tasks/robodojo-make-toast-i00-standard/task.toml
+++ b/tasks/robodojo-make-toast-i00-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "robodojo"
robot = "arx-x5-bimanual"
family = "make_toast"
mode = "standard" # robot as a service: RGB, depth + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "no positive example yet; no reference solution (RoboDojo releases no expert, and its training demonstrations carry no initial state); reported separately until one exists; human review in PR #142"
diff --git a/tasks/robodojo-make-toast-random-i00-standard/task.toml b/tasks/robodojo-make-toast-random-i00-standard/task.toml
index 4933b58197..bcb394ed22 100644
--- a/tasks/robodojo-make-toast-random-i00-standard/task.toml
+++ b/tasks/robodojo-make-toast-random-i00-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "robodojo"
robot = "arx-x5-bimanual"
family = "make_toast_random"
mode = "standard" # robot as a service: RGB, depth + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "no positive example yet; no reference solution (RoboDojo releases no expert, and its training demonstrations carry no initial state); reported separately until one exists; human review in PR #142"
diff --git a/tasks/robodojo-match-and-pick-from-conveyor-i00-privileged/task.toml b/tasks/robodojo-match-and-pick-from-conveyor-i00-privileged/task.toml
index 6aebc59d11..2d5a95836a 100644
--- a/tasks/robodojo-match-and-pick-from-conveyor-i00-privileged/task.toml
+++ b/tasks/robodojo-match-and-pick-from-conveyor-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "robodojo" # RoboDojo @ cb2a37d on its Isaac Lab 2.3 fork
robot = "arx-x5-bimanual"
family = "match_and_pick_from_conveyor"
mode = "privileged"
+dynamics = "dynamic" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "positive example: a model trial solved this task (Codex + GPT-6 Luna, effort xhigh, run `codex-gpt6_luna-xhigh`, 2026-10-10; 307 actions, graded by this verifier); no reference solution (RoboDojo releases no expert); human review in PR #142"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robodojo-match-and-pick-from-conveyor-i00-standard/task.toml b/tasks/robodojo-match-and-pick-from-conveyor-i00-standard/task.toml
index bcc183c5b2..96f176141d 100644
--- a/tasks/robodojo-match-and-pick-from-conveyor-i00-standard/task.toml
+++ b/tasks/robodojo-match-and-pick-from-conveyor-i00-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "robodojo"
robot = "arx-x5-bimanual"
family = "match_and_pick_from_conveyor"
mode = "standard" # robot as a service: RGB, depth + proprioception only; no object poses, no success predicate, no reset
+dynamics = "dynamic" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "no positive example yet (the privileged twin has a positive example (Codex + GPT-6 Luna, effort xhigh, run `codex-gpt6_luna-xhigh`, 2026-10-10)); no reference solution (RoboDojo releases no expert, and its training demonstrations carry no initial state); reported separately until one exists; human review in PR #142"
diff --git a/tasks/robodojo-organize-table-i00-privileged/task.toml b/tasks/robodojo-organize-table-i00-privileged/task.toml
index 0a600f6780..155d89c5f5 100644
--- a/tasks/robodojo-organize-table-i00-privileged/task.toml
+++ b/tasks/robodojo-organize-table-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "robodojo" # RoboDojo @ cb2a37d on its Isaac Lab 2.3 fork
robot = "arx-x5-bimanual"
family = "organize_table"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "no positive example yet; no reference solution (RoboDojo releases no expert, and its training demonstrations carry no initial state); reported separately until one exists; human review in PR #142"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robodojo-organize-table-i00-standard/task.toml b/tasks/robodojo-organize-table-i00-standard/task.toml
index 440c930c85..f36689450f 100644
--- a/tasks/robodojo-organize-table-i00-standard/task.toml
+++ b/tasks/robodojo-organize-table-i00-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "robodojo"
robot = "arx-x5-bimanual"
family = "organize_table"
mode = "standard" # robot as a service: RGB, depth + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "no positive example yet; no reference solution (RoboDojo releases no expert, and its training demonstrations carry no initial state); reported separately until one exists; human review in PR #142"
diff --git a/tasks/robodojo-pack-objects-into-box-i00-privileged/task.toml b/tasks/robodojo-pack-objects-into-box-i00-privileged/task.toml
index cc4277b94c..da7e6c0d1b 100644
--- a/tasks/robodojo-pack-objects-into-box-i00-privileged/task.toml
+++ b/tasks/robodojo-pack-objects-into-box-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "robodojo" # RoboDojo @ cb2a37d on its Isaac Lab 2.3 fork
robot = "arx-x5-bimanual"
family = "pack_objects_into_box"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "no positive example yet; no reference solution (RoboDojo releases no expert, and its training demonstrations carry no initial state); reported separately until one exists; human review in PR #142"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robodojo-pack-objects-into-box-i00-standard/task.toml b/tasks/robodojo-pack-objects-into-box-i00-standard/task.toml
index 3b70c86a10..2dbd699921 100644
--- a/tasks/robodojo-pack-objects-into-box-i00-standard/task.toml
+++ b/tasks/robodojo-pack-objects-into-box-i00-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "robodojo"
robot = "arx-x5-bimanual"
family = "pack_objects_into_box"
mode = "standard" # robot as a service: RGB, depth + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "no positive example yet; no reference solution (RoboDojo releases no expert, and its training demonstrations carry no initial state); reported separately until one exists; human review in PR #142"
diff --git a/tasks/robodojo-pack-objects-into-box-random-i00-privileged/task.toml b/tasks/robodojo-pack-objects-into-box-random-i00-privileged/task.toml
index 2ddf684289..be222e97ef 100644
--- a/tasks/robodojo-pack-objects-into-box-random-i00-privileged/task.toml
+++ b/tasks/robodojo-pack-objects-into-box-random-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "robodojo" # RoboDojo @ cb2a37d on its Isaac Lab 2.3 fork
robot = "arx-x5-bimanual"
family = "pack_objects_into_box_random"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "no positive example yet; no reference solution (RoboDojo releases no expert, and its training demonstrations carry no initial state); reported separately until one exists; human review in PR #142"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robodojo-pack-objects-into-box-random-i00-standard/task.toml b/tasks/robodojo-pack-objects-into-box-random-i00-standard/task.toml
index 98f363e639..b9deca617a 100644
--- a/tasks/robodojo-pack-objects-into-box-random-i00-standard/task.toml
+++ b/tasks/robodojo-pack-objects-into-box-random-i00-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "robodojo"
robot = "arx-x5-bimanual"
family = "pack_objects_into_box_random"
mode = "standard" # robot as a service: RGB, depth + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "no positive example yet; no reference solution (RoboDojo releases no expert, and its training demonstrations carry no initial state); reported separately until one exists; human review in PR #142"
diff --git a/tasks/robodojo-pick-from-conveyor-by-image-i00-privileged/task.toml b/tasks/robodojo-pick-from-conveyor-by-image-i00-privileged/task.toml
index 4e55ad7d17..e5ed1fdf8f 100644
--- a/tasks/robodojo-pick-from-conveyor-by-image-i00-privileged/task.toml
+++ b/tasks/robodojo-pick-from-conveyor-by-image-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "robodojo" # RoboDojo @ cb2a37d on its Isaac Lab 2.3 fork
robot = "arx-x5-bimanual"
family = "pick_from_conveyor_by_image"
mode = "privileged"
+dynamics = "dynamic" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "no positive example yet; no reference solution (RoboDojo releases no expert, and its training demonstrations carry no initial state); reported separately until one exists; human review in PR #142"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robodojo-pick-from-conveyor-by-image-i00-standard/task.toml b/tasks/robodojo-pick-from-conveyor-by-image-i00-standard/task.toml
index 2144170eb8..d30fdf1f87 100644
--- a/tasks/robodojo-pick-from-conveyor-by-image-i00-standard/task.toml
+++ b/tasks/robodojo-pick-from-conveyor-by-image-i00-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "robodojo"
robot = "arx-x5-bimanual"
family = "pick_from_conveyor_by_image"
mode = "standard" # robot as a service: RGB, depth + proprioception only; no object poses, no success predicate, no reset
+dynamics = "dynamic" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "no positive example yet; no reference solution (RoboDojo releases no expert, and its training demonstrations carry no initial state); reported separately until one exists; human review in PR #142"
diff --git a/tasks/robodojo-play-stacking-toy-i00-privileged/task.toml b/tasks/robodojo-play-stacking-toy-i00-privileged/task.toml
index fe268f1dac..9b1a63480f 100644
--- a/tasks/robodojo-play-stacking-toy-i00-privileged/task.toml
+++ b/tasks/robodojo-play-stacking-toy-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "robodojo" # RoboDojo @ cb2a37d on its Isaac Lab 2.3 fork
robot = "arx-x5-bimanual"
family = "play_stacking_toy"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "no positive example yet; no reference solution (RoboDojo releases no expert, and its training demonstrations carry no initial state); reported separately until one exists; human review in PR #142"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robodojo-play-stacking-toy-i00-standard/task.toml b/tasks/robodojo-play-stacking-toy-i00-standard/task.toml
index dec3442816..c8d7330ee2 100644
--- a/tasks/robodojo-play-stacking-toy-i00-standard/task.toml
+++ b/tasks/robodojo-play-stacking-toy-i00-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "robodojo"
robot = "arx-x5-bimanual"
family = "play_stacking_toy"
mode = "standard" # robot as a service: RGB, depth + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "no positive example yet; no reference solution (RoboDojo releases no expert, and its training demonstrations carry no initial state); reported separately until one exists; human review in PR #142"
diff --git a/tasks/robodojo-play-tic-tac-toe-i00-privileged/task.toml b/tasks/robodojo-play-tic-tac-toe-i00-privileged/task.toml
index 895d282bcd..0e28f82814 100644
--- a/tasks/robodojo-play-tic-tac-toe-i00-privileged/task.toml
+++ b/tasks/robodojo-play-tic-tac-toe-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "robodojo" # RoboDojo @ cb2a37d on its Isaac Lab 2.3 fork
robot = "arx-x5-bimanual"
family = "play_tic_tac_toe"
mode = "privileged"
+dynamics = "dynamic" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "no positive example yet; no reference solution (RoboDojo releases no expert, and its training demonstrations carry no initial state); reported separately until one exists; human review in PR #142"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robodojo-play-tic-tac-toe-i00-standard/task.toml b/tasks/robodojo-play-tic-tac-toe-i00-standard/task.toml
index c0b2c8b1cb..c6af7da1dc 100644
--- a/tasks/robodojo-play-tic-tac-toe-i00-standard/task.toml
+++ b/tasks/robodojo-play-tic-tac-toe-i00-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "robodojo"
robot = "arx-x5-bimanual"
family = "play_tic_tac_toe"
mode = "standard" # robot as a service: RGB, depth + proprioception only; no object poses, no success predicate, no reset
+dynamics = "dynamic" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "no positive example yet; no reference solution (RoboDojo releases no expert, and its training demonstrations carry no initial state); reported separately until one exists; human review in PR #142"
diff --git a/tasks/robodojo-play-xylophone-i00-privileged/task.toml b/tasks/robodojo-play-xylophone-i00-privileged/task.toml
index 9b8839c05c..0083dab840 100644
--- a/tasks/robodojo-play-xylophone-i00-privileged/task.toml
+++ b/tasks/robodojo-play-xylophone-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "robodojo" # RoboDojo @ cb2a37d on its Isaac Lab 2.3 fork
robot = "arx-x5-bimanual"
family = "play_Xylophone"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "no positive example yet; no reference solution (RoboDojo releases no expert, and its training demonstrations carry no initial state); reported separately until one exists; human review in PR #142"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robodojo-play-xylophone-i00-standard/task.toml b/tasks/robodojo-play-xylophone-i00-standard/task.toml
index 2f8c1450e1..a9cdb18459 100644
--- a/tasks/robodojo-play-xylophone-i00-standard/task.toml
+++ b/tasks/robodojo-play-xylophone-i00-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "robodojo"
robot = "arx-x5-bimanual"
family = "play_Xylophone"
mode = "standard" # robot as a service: RGB, depth + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "no positive example yet; no reference solution (RoboDojo releases no expert, and its training demonstrations carry no initial state); reported separately until one exists; human review in PR #142"
diff --git a/tasks/robodojo-plug-in-charger-i00-privileged/task.toml b/tasks/robodojo-plug-in-charger-i00-privileged/task.toml
index 79f312f733..c58dc461c8 100644
--- a/tasks/robodojo-plug-in-charger-i00-privileged/task.toml
+++ b/tasks/robodojo-plug-in-charger-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "robodojo" # RoboDojo @ cb2a37d on its Isaac Lab 2.3 fork
robot = "arx-x5-bimanual"
family = "plug_in_charger"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "no positive example yet; no reference solution (RoboDojo releases no expert, and its training demonstrations carry no initial state); reported separately until one exists; human review in PR #142"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robodojo-plug-in-charger-i00-standard/task.toml b/tasks/robodojo-plug-in-charger-i00-standard/task.toml
index 8f82782495..a2931b7ece 100644
--- a/tasks/robodojo-plug-in-charger-i00-standard/task.toml
+++ b/tasks/robodojo-plug-in-charger-i00-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "robodojo"
robot = "arx-x5-bimanual"
family = "plug_in_charger"
mode = "standard" # robot as a service: RGB, depth + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "no positive example yet; no reference solution (RoboDojo releases no expert, and its training demonstrations carry no initial state); reported separately until one exists; human review in PR #142"
diff --git a/tasks/robodojo-pour-balls-into-vase-i00-privileged/task.toml b/tasks/robodojo-pour-balls-into-vase-i00-privileged/task.toml
index ed93f9c7d1..3b2a511ee4 100644
--- a/tasks/robodojo-pour-balls-into-vase-i00-privileged/task.toml
+++ b/tasks/robodojo-pour-balls-into-vase-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "robodojo" # RoboDojo @ cb2a37d on its Isaac Lab 2.3 fork
robot = "arx-x5-bimanual"
family = "pour_balls_into_vase"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "no positive example yet; no reference solution (RoboDojo releases no expert, and its training demonstrations carry no initial state); reported separately until one exists; human review in PR #142"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robodojo-pour-balls-into-vase-i00-standard/task.toml b/tasks/robodojo-pour-balls-into-vase-i00-standard/task.toml
index 4e2824485a..d35bb6be0b 100644
--- a/tasks/robodojo-pour-balls-into-vase-i00-standard/task.toml
+++ b/tasks/robodojo-pour-balls-into-vase-i00-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "robodojo"
robot = "arx-x5-bimanual"
family = "pour_balls_into_vase"
mode = "standard" # robot as a service: RGB, depth + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "no positive example yet; no reference solution (RoboDojo releases no expert, and its training demonstrations carry no initial state); reported separately until one exists; human review in PR #142"
diff --git a/tasks/robodojo-pour-by-language-i00-standard/task.toml b/tasks/robodojo-pour-by-language-i00-standard/task.toml
index 6a68388739..e8b4d3f05c 100644
--- a/tasks/robodojo-pour-by-language-i00-standard/task.toml
+++ b/tasks/robodojo-pour-by-language-i00-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "robodojo"
robot = "arx-x5-bimanual"
family = "pour_by_language"
mode = "standard" # robot as a service: RGB, depth + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "no positive example yet; no reference solution (RoboDojo releases no expert, and its training demonstrations carry no initial state); reported separately until one exists; human review in PR #142"
diff --git a/tasks/robodojo-pour-liquid-into-cup-i00-standard/task.toml b/tasks/robodojo-pour-liquid-into-cup-i00-standard/task.toml
index 2073d2521f..c658893a2f 100644
--- a/tasks/robodojo-pour-liquid-into-cup-i00-standard/task.toml
+++ b/tasks/robodojo-pour-liquid-into-cup-i00-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "robodojo"
robot = "arx-x5-bimanual"
family = "pour_liquid_into_cup"
mode = "standard" # robot as a service: RGB, depth + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "no positive example yet; no reference solution (RoboDojo releases no expert, and its training demonstrations carry no initial state); reported separately until one exists; human review in PR #142"
diff --git a/tasks/robodojo-pour-liquid-into-cup-random-i00-standard/task.toml b/tasks/robodojo-pour-liquid-into-cup-random-i00-standard/task.toml
index bc6e83ea87..ef7c93e991 100644
--- a/tasks/robodojo-pour-liquid-into-cup-random-i00-standard/task.toml
+++ b/tasks/robodojo-pour-liquid-into-cup-random-i00-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "robodojo"
robot = "arx-x5-bimanual"
family = "pour_liquid_into_cup_random"
mode = "standard" # robot as a service: RGB, depth + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "no positive example yet; no reference solution (RoboDojo releases no expert, and its training demonstrations carry no initial state); reported separately until one exists; human review in PR #142"
diff --git a/tasks/robodojo-press-by-number-i00-privileged/task.toml b/tasks/robodojo-press-by-number-i00-privileged/task.toml
index abed41d316..b134f87b00 100644
--- a/tasks/robodojo-press-by-number-i00-privileged/task.toml
+++ b/tasks/robodojo-press-by-number-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "robodojo" # RoboDojo @ cb2a37d on its Isaac Lab 2.3 fork
robot = "arx-x5-bimanual"
family = "press_by_number"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "no positive example yet; no reference solution (RoboDojo releases no expert, and its training demonstrations carry no initial state); reported separately until one exists; human review in PR #142"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robodojo-press-by-number-i00-standard/task.toml b/tasks/robodojo-press-by-number-i00-standard/task.toml
index e8e3564f06..464c40c3df 100644
--- a/tasks/robodojo-press-by-number-i00-standard/task.toml
+++ b/tasks/robodojo-press-by-number-i00-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "robodojo"
robot = "arx-x5-bimanual"
family = "press_by_number"
mode = "standard" # robot as a service: RGB, depth + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "no positive example yet; no reference solution (RoboDojo releases no expert, and its training demonstrations carry no initial state); reported separately until one exists; human review in PR #142"
diff --git a/tasks/robodojo-push-t-i00-privileged/task.toml b/tasks/robodojo-push-t-i00-privileged/task.toml
index 71553d4f8a..2b72b1ed25 100644
--- a/tasks/robodojo-push-t-i00-privileged/task.toml
+++ b/tasks/robodojo-push-t-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "robodojo" # RoboDojo @ cb2a37d on its Isaac Lab 2.3 fork
robot = "arx-x5-bimanual"
family = "push_T"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "no positive example yet; no reference solution (RoboDojo releases no expert, and its training demonstrations carry no initial state); reported separately until one exists; human review in PR #142"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robodojo-push-t-i00-standard/task.toml b/tasks/robodojo-push-t-i00-standard/task.toml
index 5a4d1f204b..8929183c9b 100644
--- a/tasks/robodojo-push-t-i00-standard/task.toml
+++ b/tasks/robodojo-push-t-i00-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "robodojo"
robot = "arx-x5-bimanual"
family = "push_T"
mode = "standard" # robot as a service: RGB, depth + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "no positive example yet; no reference solution (RoboDojo releases no expert, and its training demonstrations carry no initial state); reported separately until one exists; human review in PR #142"
diff --git a/tasks/robodojo-push-t-random-i00-privileged/task.toml b/tasks/robodojo-push-t-random-i00-privileged/task.toml
index 70e81ee22f..f1d9bf6c75 100644
--- a/tasks/robodojo-push-t-random-i00-privileged/task.toml
+++ b/tasks/robodojo-push-t-random-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "robodojo" # RoboDojo @ cb2a37d on its Isaac Lab 2.3 fork
robot = "arx-x5-bimanual"
family = "push_T_random"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "no positive example yet; no reference solution (RoboDojo releases no expert, and its training demonstrations carry no initial state); reported separately until one exists; human review in PR #142"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robodojo-push-t-random-i00-standard/task.toml b/tasks/robodojo-push-t-random-i00-standard/task.toml
index 3970f6ccbe..ae685c6cd0 100644
--- a/tasks/robodojo-push-t-random-i00-standard/task.toml
+++ b/tasks/robodojo-push-t-random-i00-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "robodojo"
robot = "arx-x5-bimanual"
family = "push_T_random"
mode = "standard" # robot as a service: RGB, depth + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "no positive example yet; no reference solution (RoboDojo releases no expert, and its training demonstrations carry no initial state); reported separately until one exists; human review in PR #142"
diff --git a/tasks/robodojo-put-bottles-into-dustbin-i00-privileged/task.toml b/tasks/robodojo-put-bottles-into-dustbin-i00-privileged/task.toml
index 9b4c43951e..5895de26e6 100644
--- a/tasks/robodojo-put-bottles-into-dustbin-i00-privileged/task.toml
+++ b/tasks/robodojo-put-bottles-into-dustbin-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "robodojo" # RoboDojo @ cb2a37d on its Isaac Lab 2.3 fork
robot = "arx-x5-bimanual"
family = "put_bottles_into_dustbin"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "no positive example yet; no reference solution (RoboDojo releases no expert, and its training demonstrations carry no initial state); reported separately until one exists; human review in PR #142"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robodojo-put-bottles-into-dustbin-i00-standard/task.toml b/tasks/robodojo-put-bottles-into-dustbin-i00-standard/task.toml
index 0110197bca..d64a4fa5c5 100644
--- a/tasks/robodojo-put-bottles-into-dustbin-i00-standard/task.toml
+++ b/tasks/robodojo-put-bottles-into-dustbin-i00-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "robodojo"
robot = "arx-x5-bimanual"
family = "put_bottles_into_dustbin"
mode = "standard" # robot as a service: RGB, depth + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "no positive example yet; no reference solution (RoboDojo releases no expert, and its training demonstrations carry no initial state); reported separately until one exists; human review in PR #142"
diff --git a/tasks/robodojo-solve-equation-i00-privileged/task.toml b/tasks/robodojo-solve-equation-i00-privileged/task.toml
index a84fd0c098..8f68542d8b 100644
--- a/tasks/robodojo-solve-equation-i00-privileged/task.toml
+++ b/tasks/robodojo-solve-equation-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "robodojo" # RoboDojo @ cb2a37d on its Isaac Lab 2.3 fork
robot = "arx-x5-bimanual"
family = "solve_equation"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "no positive example yet; no reference solution (RoboDojo releases no expert, and its training demonstrations carry no initial state); reported separately until one exists; human review in PR #142"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robodojo-solve-equation-i00-standard/task.toml b/tasks/robodojo-solve-equation-i00-standard/task.toml
index 80df9e7fcc..ccd40fae40 100644
--- a/tasks/robodojo-solve-equation-i00-standard/task.toml
+++ b/tasks/robodojo-solve-equation-i00-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "robodojo"
robot = "arx-x5-bimanual"
family = "solve_equation"
mode = "standard" # robot as a service: RGB, depth + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "no positive example yet; no reference solution (RoboDojo releases no expert, and its training demonstrations carry no initial state); reported separately until one exists; human review in PR #142"
diff --git a/tasks/robodojo-sort-nesting-dolls-by-size-i00-privileged/README.md b/tasks/robodojo-sort-nesting-dolls-by-size-i00-privileged/README.md
index 6b3e7f64e5..4ef6637d96 100644
--- a/tasks/robodojo-sort-nesting-dolls-by-size-i00-privileged/README.md
+++ b/tasks/robodojo-sort-nesting-dolls-by-size-i00-privileged/README.md
@@ -12,7 +12,7 @@
| Base image | `ghcr.io/mll-lab-nu/eai-robodojo:0.1.0` |
## Task
-Arrange the five nesting dolls upright in a straight row, ordered by size, from smallest to largest (left to right as the instruction says), and return the arms to their start pose. (RoboDojo's `sort_nesting_dolls_by_size`). RoboDojo capability dimension: Generalization. Instance i00 is RoboDojo's own evaluation layout
+Arrange the five nesting dolls upright in a straight row, ordered by size from the smallest at the smallest world x to the largest at the largest world x, and return the arms to their start pose. (RoboDojo's `sort_nesting_dolls_by_size`). RoboDojo capability dimension: Generalization. Instance i00 is RoboDojo's own evaluation layout
`arx_x5/0/sort_nesting_dolls_by_size_0.json` (sha256 `a01c86062db05ce7`), frozen in `instance.json`. The deliverable is one replayable trajectory of
RoboDojo's 14-D joint action at 25 Hz (`/app/output/trajectory.npz`), graded by RoboDojo's own reward check for the task,
with no step cap (RoboDojo's own evaluation gives a policy 1050 actions; PROTOCOL D1).
@@ -58,3 +58,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
### Change Log
- 2026-10-11: agent time limit 5400 s -> 3600 s (`[agent] timeout_sec`; owner decision, docs/benchmarks/robodojo.md), and `instruction.md` states the cutoff at the top. The scene, the frozen layout, RoboDojo's success check and the verifier are unchanged, so the task keeps its name; trials recorded before this date ran with 5400 s and are kept apart (no official result was published with it).
- 2026-10-12: graded by ONE fresh-process replay (two before; owner decision, docs/benchmarks/robodojo.md): reward 1 iff that replay ends in RoboDojo's success. Before, both replays had to succeed and end in the same state hash; with one replay nothing is compared, so `reward.json` reports `deterministic` = 1 with `determinism_checked` = 0 and `replays` = 1, and a replay whose process did not rebuild the frozen layout to its recorded hash gives no verdict (`instance_mismatch`, `infra_error`). `[verifier] timeout_sec` 15000 s -> 7800 s (one replay of the 15,000 rows a trajectory may have, 7200 s, + 600 s). The scene, the frozen layout, RoboDojo's success check and what the agent hands in are unchanged, so the task keeps its name; `instruction.md` says one replay. No ranked result exists; a trial graded before this date is re-read from its saved `reward.json` (`all_success`, `state_identical`), not rerun: in the first 38 graded trials of the run that began on 2026-10-11 the two replays always ended in the same state, and no verdict changes.
+- 2026-10-14 (board 0113, from board 0086): the instruction's description of RoboDojo's check (`instruction.md`, "What RoboDojo checks") had the size order the wrong way round. It said the LARGEST doll must have the smallest world x, following the comments in RoboDojo's config; the check actually wants the SMALLEST doll at the smallest world x and x growing with size (index mod 5 = 0 is the largest doll: the dolls' measured bounding boxes in the frozen layout shrink as index mod 5 grows), which is what RoboDojo's own sentence "from left to right, from smallest to largest" says. Text only: the scene, the frozen layout, RoboDojo's success check and the verifier are unchanged, so the task keeps its name. Trials that followed the old text and arranged the row as it said failed for that reason (board 0086: models that then mirrored the row passed).
diff --git a/tasks/robodojo-sort-nesting-dolls-by-size-i00-privileged/instruction.md b/tasks/robodojo-sort-nesting-dolls-by-size-i00-privileged/instruction.md
index b23eb9a163..c4ba2e82ed 100644
--- a/tasks/robodojo-sort-nesting-dolls-by-size-i00-privileged/instruction.md
+++ b/tasks/robodojo-sort-nesting-dolls-by-size-i00-privileged/instruction.md
@@ -7,7 +7,7 @@ setup on Isaac Sim 5.1 / Isaac Lab, PhysX) via Python. RoboDojo capability dimen
trajectory as soon as you have one and overwrite it as you improve it. Every Python process that opens a `Session` first
spends ~3-5 min launching Isaac Sim and building the scene, and that time counts.
-**Task:** Arrange the five nesting dolls upright in a straight row, ordered by size, from smallest to largest (left to right as the instruction says), and return the arms to their start pose.
+**Task:** Arrange the five nesting dolls upright in a straight row, ordered by size from the smallest at the smallest world x to the largest at the largest world x, and return the arms to their start pose.
RoboDojo's own instruction for this scene: *"Arrange the five nesting dolls in a row from left to right, from smallest to largest."*
@@ -19,9 +19,9 @@ robot's start pose never change between resets. You may reset and retry as many
**one replayable trajectory**, not a general policy.
## Success (RoboDojo's own check for this task)
-- Define the order of the dolls by their model index mod 5 (0 = smallest, 4 = largest; derived from the yml comments in the config: index 5 smallest half-extent 0.0152 m, index 9 largest 0.0335 m). The code lists the dolls sorted by (index mod 5) DESCENDING: L0 = largest, L1, L2, L3, L4 = smallest.
+- Define the order of the dolls by their model index mod 5, as the code does: 0 = the LARGEST doll, 4 = the SMALLEST (measured on the dolls of the frozen layout: their bounding boxes shrink as index mod 5 grows; the comments in RoboDojo's config, index 5 smallest and index 9 largest, have it the wrong way round). The code lists the dolls sorted by (index mod 5) DESCENDING: L0 = the smallest doll, L1, L2, L3, L4 = the largest.
- Row straightness: the maximum minus the minimum world Y of the five dolls' root positions is <= 0.035 m.
-- Order along X: for each consecutive pair (L_k, L_{k+1}), k=0..3: x(L_{k+1}) - x(L_k) > 0.03 m (strict), i.e. the largest doll has the smallest world x, and x increases by more than 3 cm with each smaller doll; the smallest doll has the largest world x. (The instruction says 'left to right, smallest to largest' while the code compares world/table x with the largest at the smallest x; presumably 'left' is as seen from the front camera, but this is not verifiable statically - use the numeric x rule.)
+- Order along X: for each consecutive pair (L_k, L_{k+1}), k=0..3: x(L_{k+1}) - x(L_k) > 0.03 m (strict), i.e. the SMALLEST doll has the smallest world x, and x increases by more than 3 cm with each larger doll; the largest doll has the largest world x. (This matches the instruction's 'left to right, from smallest to largest' with left = smaller world x.)
- Upright: each of doll0..doll4 has its local +Z axis within 15 deg (default threshold, strict) of world +Z (standing, not lying).
- Both target arms are back at origin: all target arms (left_arm and right_arm; the support arm is not counted) have their end-effector pose within 0.15 m on EACH axis (|dx|,|dy|,|dz| all <= 0.15 m) and within 20 degrees (quaternion angle) of the pose they had at the start of the episode (recorded after the scene settled).
- All conditions hold simultaneously after one action.
@@ -36,7 +36,7 @@ reads that limit and it does not apply here: where the description above mention
**At the end.** Both arms within 0.15 m per axis and 20 deg of their start pose in the same action as the dolls are correctly sorted, upright, and in a row.
-**Notes from RoboDojo's code.** The required X gap between consecutive dolls is only > 0.03 m between root positions (roots, not edges), while the row's Y spread must be <= 3.5 cm for all five, so the dolls need to be tightly aligned in y but only loosely spaced in x. The yml comments suggest the intended target positions are x = -0.24 (largest, index mod 5 = 4), -0.12, 0.0, 0.12, 0.24 (smallest) in the table frame, but any positions satisfying the stated rules succeed. The mapping index->size is inferred from yml comments, not verified from assets. Dolls must stay upright (within 15 deg) - a doll left lying fails.
+**Notes from RoboDojo's code.** The required X gap between consecutive dolls is only > 0.03 m between root positions (roots, not edges), while the row's Y spread must be <= 3.5 cm for all five, so the dolls need to be tightly aligned in y but only loosely spaced in x. The yml comments place the targets at table-frame x = -0.24 (index mod 5 = 4), -0.12, 0.0, 0.12, 0.24 (index mod 5 = 0) and call index mod 5 = 4 the largest; the dolls' measured sizes say the opposite (index mod 5 = 4 is the smallest), so the smallest doll goes at the smallest x. Any positions satisfying the stated rules succeed. Dolls must stay upright (within 15 deg) - a doll left lying fails.
## Deliverable
Write the final trajectory to `/app/output/trajectory.npz` with `Session.save_trajectory(...)` (below). It is replayed
diff --git a/tasks/robodojo-sort-nesting-dolls-by-size-i00-privileged/task.toml b/tasks/robodojo-sort-nesting-dolls-by-size-i00-privileged/task.toml
index ff66112add..72b5274a72 100644
--- a/tasks/robodojo-sort-nesting-dolls-by-size-i00-privileged/task.toml
+++ b/tasks/robodojo-sort-nesting-dolls-by-size-i00-privileged/task.toml
@@ -4,7 +4,7 @@ artifacts = [{ source = "/var/rcb-monitor/", service = "rcb_monitor" }, "/app/ou
[task]
name = "robot-coding-bench/robodojo-sort-nesting-dolls-by-size-i00-privileged"
-description = "RoboDojo sort_nesting_dolls_by_size (Generalization; two ARX X5 arms with parallel grippers, Isaac Sim 5.1 / Isaac Lab): produce one replayable trajectory of RoboDojo's 14-D joint actions. Task: Arrange the five nesting dolls upright in a straight row, ordered by size, from smallest to largest (left to right as the instruction says), and return the arms to their start pose. Frozen scene instance: RoboDojo's own evaluation layout arx_x5/0/sort_nesting_dolls_by_size_0.json; graded by RoboDojo's own reward check."
+description = "RoboDojo sort_nesting_dolls_by_size (Generalization; two ARX X5 arms with parallel grippers, Isaac Sim 5.1 / Isaac Lab): produce one replayable trajectory of RoboDojo's 14-D joint actions. Task: Arrange the five nesting dolls upright in a straight row, ordered by size from the smallest at the smallest world x to the largest at the largest world x, and return the arms to their start pose. Frozen scene instance: RoboDojo's own evaluation layout arx_x5/0/sort_nesting_dolls_by_size_0.json; graded by RoboDojo's own reward check."
authors = [{ name = "Changhe Chen" }]
keywords = ["robotics", "robodojo", "isaac-lab", "isaac-sim", "bimanual", "arx-x5", "code-as-policy", "trajectory", "privileged"]
@@ -16,6 +16,7 @@ simulator = "robodojo" # RoboDojo @ cb2a37d on its Isaac Lab 2.3 fork
robot = "arx-x5-bimanual"
family = "sort_nesting_dolls_by_size"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "no positive example yet; no reference solution (RoboDojo releases no expert, and its training demonstrations carry no initial state); reported separately until one exists; human review in PR #142"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robodojo-sort-nesting-dolls-by-size-i00-standard/README.md b/tasks/robodojo-sort-nesting-dolls-by-size-i00-standard/README.md
index a3fc811db9..a6eafb3581 100644
--- a/tasks/robodojo-sort-nesting-dolls-by-size-i00-standard/README.md
+++ b/tasks/robodojo-sort-nesting-dolls-by-size-i00-standard/README.md
@@ -12,7 +12,7 @@
| Base image | `ghcr.io/mll-lab-nu/eai-robodojo:0.1.0` |
## Task
-Standard-mode twin of `robodojo-sort-nesting-dolls-by-size-i00-privileged` (same frozen layout, RoboDojo's `arx_x5/0/sort_nesting_dolls_by_size_0.json`). Arrange the five nesting dolls upright in a straight row, ordered by size, from smallest to largest (left to right as the instruction says), and return the arms to their start pose. (RoboDojo's `sort_nesting_dolls_by_size`). The agent drives the
+Standard-mode twin of `robodojo-sort-nesting-dolls-by-size-i00-privileged` (same frozen layout, RoboDojo's `arx_x5/0/sort_nesting_dolls_by_size_0.json`). Arrange the five nesting dolls upright in a straight row, ordered by size from the smallest at the smallest world x to the largest at the largest world x, and return the arms to their start pose. (RoboDojo's `sort_nesting_dolls_by_size`). The agent drives the
robot only through the eai-standard/2.2 client (the simulator runs in the `sim` sidecar), which owns the scene and records
the episode as a (T, 14) trajectory of RoboDojo's own joint action at 25 Hz.
@@ -56,3 +56,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
- 2026-10-11: agent time limit 5400 s -> 3600 s (`[agent] timeout_sec`, and the service's `RCB_WALL_CLOCK_S`, which only reports it: `spec["budget"]`, `meta["time_left_s"]`; owner decision, docs/benchmarks/robodojo.md), and `instruction.md` states the cutoff at the top. The scene, the frozen layout, RoboDojo's success check, the collect hook and the verifier are unchanged, so the task keeps its name; trials recorded before this date ran with 5400 s and are kept apart (no official result was published with it).
- 2026-10-12: graded with ONE fresh-process replay of the recorded episode (two before; owner decision, docs/benchmarks/robodojo.md): `success` = `live_success` and `replay_success`, one successful replay is enough. Before, two replays had to succeed and end in the same state hash. `deterministic` (= `state_identical` = `live_matches_replay`) now says whether the replay ended in the live episode's state; it is reported and no longer decides. The collect hook's limit 18000 s -> 10800 s (`[[verifier.collect]] timeout_sec` = `finalize.sh`'s `HOOK_BUDGET_S`); inside it the replay has max(1800 s, 1200 s + actions / 2.5) and is run once more when it gives no result in that time. The scene, the frozen layout, RoboDojo's success check and the service the agent talks to are unchanged, so the task keeps its name. No ranked result exists; a trial graded before this date is re-read from its saved `reward.json` (`live_success`, `replay_success`, `state_identical`), not rerun: in the first 38 graded trials of the run that began on 2026-10-11 no verdict changes.
- 2026-10-13 (owner, boards 0085 and 0087): the sim service's clock (`environment/sim/rcb_limited.py`, `environment/sim/server.py`, the docstring of `environment/rcb_client.py`). `meta["time_s"]` and `meta["time_left_s"]` count from the agent's first request to the service (any endpoint), and a restarted service keeps that moment; before, they counted from the service's own start of the episode (when its scene was ready). Where the service is ready before the agent's phase begins (the harness is installed in between: a median of 79 s over the recorded standard trials), that clock ran ahead of the agent's and `time_left_s` reached 0 while the agent still had that time. `/spec` says what the clock counts from (`budget.clock`), and the instruction's `## Time limit` tells the agent to plan by `time_left_s`. The service never ended an episode on this clock: the scene, the frozen instance, the pass rule and the grader are unchanged, so the task keeps its name and earlier results stand.
+- 2026-10-14 (board 0113, from board 0086): the instruction's description of RoboDojo's check (`instruction.md`, "What RoboDojo checks") had the size order the wrong way round. It said the LARGEST doll must have the smallest world x, following the comments in RoboDojo's config; the check actually wants the SMALLEST doll at the smallest world x and x growing with size (index mod 5 = 0 is the largest doll: the dolls' measured bounding boxes in the frozen layout shrink as index mod 5 grows), which is what RoboDojo's own sentence "from left to right, from smallest to largest" says. Text only: the scene, the frozen layout, RoboDojo's success check and the verifier are unchanged, so the task keeps its name. Trials that followed the old text and arranged the row as it said failed for that reason (board 0086: models that then mirrored the row passed).
diff --git a/tasks/robodojo-sort-nesting-dolls-by-size-i00-standard/instruction.md b/tasks/robodojo-sort-nesting-dolls-by-size-i00-standard/instruction.md
index a31f2a35d2..3220c456e8 100644
--- a/tasks/robodojo-sort-nesting-dolls-by-size-i00-standard/instruction.md
+++ b/tasks/robodojo-sort-nesting-dolls-by-size-i00-standard/instruction.md
@@ -8,14 +8,14 @@ Nothing is handed in, so the task must be completed before the time limit (`## T
ends by itself when RoboDojo's check passes: `res.meta["done"]`). The time you wait for the service counts: its first
start in a trial takes ~4-15 min (Client, below).
-**Task:** Arrange the five nesting dolls upright in a straight row, ordered by size, from smallest to largest (left to right as the instruction says), and return the arms to their start pose.
+**Task:** Arrange the five nesting dolls upright in a straight row, ordered by size from the smallest at the smallest world x to the largest at the largest world x, and return the arms to their start pose.
RoboDojo's own instruction for this scene: *"Arrange the five nesting dolls in a row from left to right, from smallest to largest."*
## Success (RoboDojo's own check for this task)
-- Define the order of the dolls by their model index mod 5 (0 = smallest, 4 = largest; derived from the yml comments in the config: index 5 smallest half-extent 0.0152 m, index 9 largest 0.0335 m). The code lists the dolls sorted by (index mod 5) DESCENDING: L0 = largest, L1, L2, L3, L4 = smallest.
+- Define the order of the dolls by their model index mod 5, as the code does: 0 = the LARGEST doll, 4 = the SMALLEST (measured on the dolls of the frozen layout: their bounding boxes shrink as index mod 5 grows; the comments in RoboDojo's config, index 5 smallest and index 9 largest, have it the wrong way round). The code lists the dolls sorted by (index mod 5) DESCENDING: L0 = the smallest doll, L1, L2, L3, L4 = the largest.
- Row straightness: the maximum minus the minimum world Y of the five dolls' root positions is <= 0.035 m.
-- Order along X: for each consecutive pair (L_k, L_{k+1}), k=0..3: x(L_{k+1}) - x(L_k) > 0.03 m (strict), i.e. the largest doll has the smallest world x, and x increases by more than 3 cm with each smaller doll; the smallest doll has the largest world x. (The instruction says 'left to right, smallest to largest' while the code compares world/table x with the largest at the smallest x; presumably 'left' is as seen from the front camera, but this is not verifiable statically - use the numeric x rule.)
+- Order along X: for each consecutive pair (L_k, L_{k+1}), k=0..3: x(L_{k+1}) - x(L_k) > 0.03 m (strict), i.e. the SMALLEST doll has the smallest world x, and x increases by more than 3 cm with each larger doll; the largest doll has the largest world x. (This matches the instruction's 'left to right, from smallest to largest' with left = smaller world x.)
- Upright: each of doll0..doll4 has its local +Z axis within 15 deg (default threshold, strict) of world +Z (standing, not lying).
- Both target arms are back at origin: all target arms (left_arm and right_arm; the support arm is not counted) have their end-effector pose within 0.15 m on EACH axis (|dx|,|dy|,|dz| all <= 0.15 m) and within 20 degrees (quaternion angle) of the pose they had at the start of the episode (recorded after the scene settled).
- All conditions hold simultaneously after one action.
@@ -31,7 +31,7 @@ it does not apply here: where the description above mentions step_lim, read the
**At the end.** Both arms within 0.15 m per axis and 20 deg of their start pose in the same action as the dolls are correctly sorted, upright, and in a row.
-**Notes from RoboDojo's code.** The required X gap between consecutive dolls is only > 0.03 m between root positions (roots, not edges), while the row's Y spread must be <= 3.5 cm for all five, so the dolls need to be tightly aligned in y but only loosely spaced in x. The yml comments suggest the intended target positions are x = -0.24 (largest, index mod 5 = 4), -0.12, 0.0, 0.12, 0.24 (smallest) in the table frame, but any positions satisfying the stated rules succeed. The mapping index->size is inferred from yml comments, not verified from assets. Dolls must stay upright (within 15 deg) - a doll left lying fails.
+**Notes from RoboDojo's code.** The required X gap between consecutive dolls is only > 0.03 m between root positions (roots, not edges), while the row's Y spread must be <= 3.5 cm for all five, so the dolls need to be tightly aligned in y but only loosely spaced in x. The yml comments place the targets at table-frame x = -0.24 (index mod 5 = 4), -0.12, 0.0, 0.12, 0.24 (index mod 5 = 0) and call index mod 5 = 4 the largest; the dolls' measured sizes say the opposite (index mod 5 = 4 is the smallest), so the smallest doll goes at the smallest x. Any positions satisfying the stated rules succeed. Dolls must stay upright (within 15 deg) - a doll left lying fails.
## Rules (real-robot rules)
- **No reset.** The episode starts when the service comes up and ends as above, when your time is up, or on a simulator
diff --git a/tasks/robodojo-sort-nesting-dolls-by-size-i00-standard/task.toml b/tasks/robodojo-sort-nesting-dolls-by-size-i00-standard/task.toml
index 6986597894..b6d00ff208 100644
--- a/tasks/robodojo-sort-nesting-dolls-by-size-i00-standard/task.toml
+++ b/tasks/robodojo-sort-nesting-dolls-by-size-i00-standard/task.toml
@@ -15,7 +15,7 @@ artifacts = [{ source = "/var/rcb-monitor/", service = "rcb_monitor" },
[task]
name = "robot-coding-bench/robodojo-sort-nesting-dolls-by-size-i00-standard"
-description = "STANDARD MODE (robot as a service, eai-standard/2.2): RoboDojo sort_nesting_dolls_by_size (Generalization) on the same frozen layout as the privileged task (arx_x5/0/sort_nesting_dolls_by_size_0.json). Task: Arrange the five nesting dolls upright in a straight row, ordered by size, from smallest to largest (left to right as the instruction says), and return the arms to their start pose. The agent only gets a client: get_obs (RGB, depth + proprioception of both arms), submit (joint_position or ee_pose chunks). No reset, no privileged state; the sim service records the episode as a trajectory and grades it."
+description = "STANDARD MODE (robot as a service, eai-standard/2.2): RoboDojo sort_nesting_dolls_by_size (Generalization) on the same frozen layout as the privileged task (arx_x5/0/sort_nesting_dolls_by_size_0.json). Task: Arrange the five nesting dolls upright in a straight row, ordered by size from the smallest at the smallest world x to the largest at the largest world x, and return the arms to their start pose. The agent only gets a client: get_obs (RGB, depth + proprioception of both arms), submit (joint_position or ee_pose chunks). No reset, no privileged state; the sim service records the episode as a trajectory and grades it."
authors = [{ name = "Changhe Chen" }]
keywords = ["robotics", "robodojo", "isaac-lab", "isaac-sim", "bimanual", "arx-x5", "standard-mode", "trajectory"]
@@ -27,6 +27,7 @@ simulator = "robodojo"
robot = "arx-x5-bimanual"
family = "sort_nesting_dolls_by_size"
mode = "standard" # robot as a service: RGB, depth + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "no positive example yet; no reference solution (RoboDojo releases no expert, and its training demonstrations carry no initial state); reported separately until one exists; human review in PR #142"
diff --git a/tasks/robodojo-sort-nesting-dolls-by-size-random-i00-privileged/README.md b/tasks/robodojo-sort-nesting-dolls-by-size-random-i00-privileged/README.md
index 21c767a95f..60d888bf4d 100644
--- a/tasks/robodojo-sort-nesting-dolls-by-size-random-i00-privileged/README.md
+++ b/tasks/robodojo-sort-nesting-dolls-by-size-random-i00-privileged/README.md
@@ -12,7 +12,7 @@
| Base image | `ghcr.io/mll-lab-nu/eai-robodojo:0.1.0` |
## Task
-Arrange the five nesting dolls upright in a straight row, ordered by size, from smallest to largest (left to right as the instruction says), and return the arms to their start pose. (RoboDojo's `sort_nesting_dolls_by_size_random`). RoboDojo capability dimension: Generalization (randomised variant: other models, clutter). Instance i00 is RoboDojo's own evaluation layout
+Arrange the five nesting dolls upright in a straight row, ordered by size from the smallest at the smallest world x to the largest at the largest world x, and return the arms to their start pose. (RoboDojo's `sort_nesting_dolls_by_size_random`). RoboDojo capability dimension: Generalization (randomised variant: other models, clutter). Instance i00 is RoboDojo's own evaluation layout
`arx_x5/0/sort_nesting_dolls_by_size_random_0.json` (sha256 `f45438f654fd1ce7`), frozen in `instance.json`. The deliverable is one replayable trajectory of
RoboDojo's 14-D joint action at 25 Hz (`/app/output/trajectory.npz`), graded by RoboDojo's own reward check for the task,
with no step cap (RoboDojo's own evaluation gives a policy 1050 actions; PROTOCOL D1).
@@ -58,3 +58,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
### Change Log
- 2026-10-11: agent time limit 5400 s -> 3600 s (`[agent] timeout_sec`; owner decision, docs/benchmarks/robodojo.md), and `instruction.md` states the cutoff at the top. The scene, the frozen layout, RoboDojo's success check and the verifier are unchanged, so the task keeps its name; trials recorded before this date ran with 5400 s and are kept apart (no official result was published with it).
- 2026-10-12: graded by ONE fresh-process replay (two before; owner decision, docs/benchmarks/robodojo.md): reward 1 iff that replay ends in RoboDojo's success. Before, both replays had to succeed and end in the same state hash; with one replay nothing is compared, so `reward.json` reports `deterministic` = 1 with `determinism_checked` = 0 and `replays` = 1, and a replay whose process did not rebuild the frozen layout to its recorded hash gives no verdict (`instance_mismatch`, `infra_error`). `[verifier] timeout_sec` 15000 s -> 7800 s (one replay of the 15,000 rows a trajectory may have, 7200 s, + 600 s). The scene, the frozen layout, RoboDojo's success check and what the agent hands in are unchanged, so the task keeps its name; `instruction.md` says one replay. No ranked result exists; a trial graded before this date is re-read from its saved `reward.json` (`all_success`, `state_identical`), not rerun: in the first 38 graded trials of the run that began on 2026-10-11 the two replays always ended in the same state, and no verdict changes.
+- 2026-10-14 (board 0113, from board 0086): the instruction's description of RoboDojo's check (`instruction.md`, "What RoboDojo checks") had the size order the wrong way round. It said the LARGEST doll must have the smallest world x, following the comments in RoboDojo's config; the check actually wants the SMALLEST doll at the smallest world x and x growing with size (index mod 5 = 0 is the largest doll: the dolls' measured bounding boxes in the frozen layout shrink as index mod 5 grows), which is what RoboDojo's own sentence "from left to right, from smallest to largest" says. Text only: the scene, the frozen layout, RoboDojo's success check and the verifier are unchanged, so the task keeps its name. Trials that followed the old text and arranged the row as it said failed for that reason (board 0086: models that then mirrored the row passed).
diff --git a/tasks/robodojo-sort-nesting-dolls-by-size-random-i00-privileged/instruction.md b/tasks/robodojo-sort-nesting-dolls-by-size-random-i00-privileged/instruction.md
index 8b8d489c92..9f9a398eb8 100644
--- a/tasks/robodojo-sort-nesting-dolls-by-size-random-i00-privileged/instruction.md
+++ b/tasks/robodojo-sort-nesting-dolls-by-size-random-i00-privileged/instruction.md
@@ -7,7 +7,7 @@ setup on Isaac Sim 5.1 / Isaac Lab, PhysX) via Python. RoboDojo capability dimen
trajectory as soon as you have one and overwrite it as you improve it. Every Python process that opens a `Session` first
spends ~3-5 min launching Isaac Sim and building the scene, and that time counts.
-**Task:** Arrange the five nesting dolls upright in a straight row, ordered by size, from smallest to largest (left to right as the instruction says), and return the arms to their start pose.
+**Task:** Arrange the five nesting dolls upright in a straight row, ordered by size from the smallest at the smallest world x to the largest at the largest world x, and return the arms to their start pose.
RoboDojo's own instruction for this scene: *"Arrange the five nesting dolls in a row from left to right, from smallest to largest."*
@@ -20,9 +20,9 @@ robot's start pose never change between resets. You may reset and retry as many
**one replayable trajectory**, not a general policy.
## Success (RoboDojo's own check for this task)
-- Define the order of the dolls by their model index mod 5 (0 = smallest, 4 = largest; derived from the yml comments in the config: index 5 smallest half-extent 0.0152 m, index 9 largest 0.0335 m). The code lists the dolls sorted by (index mod 5) DESCENDING: L0 = largest, L1, L2, L3, L4 = smallest.
+- Define the order of the dolls by their model index mod 5, as the code does: 0 = the LARGEST doll, 4 = the SMALLEST (measured on the dolls of the frozen layout: their bounding boxes shrink as index mod 5 grows; the comments in RoboDojo's config, index 5 smallest and index 9 largest, have it the wrong way round). The code lists the dolls sorted by (index mod 5) DESCENDING: L0 = the smallest doll, L1, L2, L3, L4 = the largest.
- Row straightness: the maximum minus the minimum world Y of the five dolls' root positions is <= 0.035 m.
-- Order along X: for each consecutive pair (L_k, L_{k+1}), k=0..3: x(L_{k+1}) - x(L_k) > 0.03 m (strict), i.e. the largest doll has the smallest world x, and x increases by more than 3 cm with each smaller doll; the smallest doll has the largest world x. (The instruction says 'left to right, smallest to largest' while the code compares world/table x with the largest at the smallest x; presumably 'left' is as seen from the front camera, but this is not verifiable statically - use the numeric x rule.)
+- Order along X: for each consecutive pair (L_k, L_{k+1}), k=0..3: x(L_{k+1}) - x(L_k) > 0.03 m (strict), i.e. the SMALLEST doll has the smallest world x, and x increases by more than 3 cm with each larger doll; the largest doll has the largest world x. (This matches the instruction's 'left to right, from smallest to largest' with left = smaller world x.)
- Upright: each of doll0..doll4 has its local +Z axis within 15 deg (default threshold, strict) of world +Z (standing, not lying).
- Both target arms are back at origin: all target arms (left_arm and right_arm; the support arm is not counted) have their end-effector pose within 0.15 m on EACH axis (|dx|,|dy|,|dz| all <= 0.15 m) and within 20 degrees (quaternion angle) of the pose they had at the start of the episode (recorded after the scene settled).
- All conditions hold simultaneously after one action.
@@ -37,7 +37,7 @@ reads that limit and it does not apply here: where the description above mention
**At the end.** Both arms within 0.15 m per axis and 20 deg of their start pose in the same action as the dolls are correctly sorted, upright, and in a row.
-**Notes from RoboDojo's code.** The required X gap between consecutive dolls is only > 0.03 m between root positions (roots, not edges), while the row's Y spread must be <= 3.5 cm for all five, so the dolls need to be tightly aligned in y but only loosely spaced in x. The yml comments suggest the intended target positions are x = -0.24 (largest, index mod 5 = 4), -0.12, 0.0, 0.12, 0.24 (smallest) in the table frame, but any positions satisfying the stated rules succeed. The mapping index->size is inferred from yml comments, not verified from assets. Dolls must stay upright (within 15 deg) - a doll left lying fails. The random variant has different texture models and clutter distractors; reward identical.
+**Notes from RoboDojo's code.** The required X gap between consecutive dolls is only > 0.03 m between root positions (roots, not edges), while the row's Y spread must be <= 3.5 cm for all five, so the dolls need to be tightly aligned in y but only loosely spaced in x. The yml comments place the targets at table-frame x = -0.24 (index mod 5 = 4), -0.12, 0.0, 0.12, 0.24 (index mod 5 = 0) and call index mod 5 = 4 the largest; the dolls' measured sizes say the opposite (index mod 5 = 4 is the smallest), so the smallest doll goes at the smallest x. Any positions satisfying the stated rules succeed. Dolls must stay upright (within 15 deg) - a doll left lying fails. The random variant has different texture models and clutter distractors; reward identical.
## Deliverable
Write the final trajectory to `/app/output/trajectory.npz` with `Session.save_trajectory(...)` (below). It is replayed
diff --git a/tasks/robodojo-sort-nesting-dolls-by-size-random-i00-privileged/task.toml b/tasks/robodojo-sort-nesting-dolls-by-size-random-i00-privileged/task.toml
index 90f980a555..51c57fda75 100644
--- a/tasks/robodojo-sort-nesting-dolls-by-size-random-i00-privileged/task.toml
+++ b/tasks/robodojo-sort-nesting-dolls-by-size-random-i00-privileged/task.toml
@@ -4,7 +4,7 @@ artifacts = [{ source = "/var/rcb-monitor/", service = "rcb_monitor" }, "/app/ou
[task]
name = "robot-coding-bench/robodojo-sort-nesting-dolls-by-size-random-i00-privileged"
-description = "RoboDojo sort_nesting_dolls_by_size_random (Generalization (randomised variant: other models, clutter); two ARX X5 arms with parallel grippers, Isaac Sim 5.1 / Isaac Lab): produce one replayable trajectory of RoboDojo's 14-D joint actions. Task: Arrange the five nesting dolls upright in a straight row, ordered by size, from smallest to largest (left to right as the instruction says), and return the arms to their start pose. Frozen scene instance: RoboDojo's own evaluation layout arx_x5/0/sort_nesting_dolls_by_size_random_0.json; graded by RoboDojo's own reward check."
+description = "RoboDojo sort_nesting_dolls_by_size_random (Generalization (randomised variant: other models, clutter); two ARX X5 arms with parallel grippers, Isaac Sim 5.1 / Isaac Lab): produce one replayable trajectory of RoboDojo's 14-D joint actions. Task: Arrange the five nesting dolls upright in a straight row, ordered by size from the smallest at the smallest world x to the largest at the largest world x, and return the arms to their start pose. Frozen scene instance: RoboDojo's own evaluation layout arx_x5/0/sort_nesting_dolls_by_size_random_0.json; graded by RoboDojo's own reward check."
authors = [{ name = "Changhe Chen" }]
keywords = ["robotics", "robodojo", "isaac-lab", "isaac-sim", "bimanual", "arx-x5", "code-as-policy", "trajectory", "privileged"]
@@ -16,6 +16,7 @@ simulator = "robodojo" # RoboDojo @ cb2a37d on its Isaac Lab 2.3 fork
robot = "arx-x5-bimanual"
family = "sort_nesting_dolls_by_size_random"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "no positive example yet; no reference solution (RoboDojo releases no expert, and its training demonstrations carry no initial state); reported separately until one exists; human review in PR #142"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robodojo-sort-nesting-dolls-by-size-random-i00-standard/README.md b/tasks/robodojo-sort-nesting-dolls-by-size-random-i00-standard/README.md
index b41206b2d8..716a8ce70c 100644
--- a/tasks/robodojo-sort-nesting-dolls-by-size-random-i00-standard/README.md
+++ b/tasks/robodojo-sort-nesting-dolls-by-size-random-i00-standard/README.md
@@ -12,7 +12,7 @@
| Base image | `ghcr.io/mll-lab-nu/eai-robodojo:0.1.0` |
## Task
-Standard-mode twin of `robodojo-sort-nesting-dolls-by-size-random-i00-privileged` (same frozen layout, RoboDojo's `arx_x5/0/sort_nesting_dolls_by_size_random_0.json`). Arrange the five nesting dolls upright in a straight row, ordered by size, from smallest to largest (left to right as the instruction says), and return the arms to their start pose. (RoboDojo's `sort_nesting_dolls_by_size_random`). The agent drives the
+Standard-mode twin of `robodojo-sort-nesting-dolls-by-size-random-i00-privileged` (same frozen layout, RoboDojo's `arx_x5/0/sort_nesting_dolls_by_size_random_0.json`). Arrange the five nesting dolls upright in a straight row, ordered by size from the smallest at the smallest world x to the largest at the largest world x, and return the arms to their start pose. (RoboDojo's `sort_nesting_dolls_by_size_random`). The agent drives the
robot only through the eai-standard/2.2 client (the simulator runs in the `sim` sidecar), which owns the scene and records
the episode as a (T, 14) trajectory of RoboDojo's own joint action at 25 Hz.
@@ -56,3 +56,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
- 2026-10-11: agent time limit 5400 s -> 3600 s (`[agent] timeout_sec`, and the service's `RCB_WALL_CLOCK_S`, which only reports it: `spec["budget"]`, `meta["time_left_s"]`; owner decision, docs/benchmarks/robodojo.md), and `instruction.md` states the cutoff at the top. The scene, the frozen layout, RoboDojo's success check, the collect hook and the verifier are unchanged, so the task keeps its name; trials recorded before this date ran with 5400 s and are kept apart (no official result was published with it).
- 2026-10-12: graded with ONE fresh-process replay of the recorded episode (two before; owner decision, docs/benchmarks/robodojo.md): `success` = `live_success` and `replay_success`, one successful replay is enough. Before, two replays had to succeed and end in the same state hash. `deterministic` (= `state_identical` = `live_matches_replay`) now says whether the replay ended in the live episode's state; it is reported and no longer decides. The collect hook's limit 18000 s -> 10800 s (`[[verifier.collect]] timeout_sec` = `finalize.sh`'s `HOOK_BUDGET_S`); inside it the replay has max(1800 s, 1200 s + actions / 2.5) and is run once more when it gives no result in that time. The scene, the frozen layout, RoboDojo's success check and the service the agent talks to are unchanged, so the task keeps its name. No ranked result exists; a trial graded before this date is re-read from its saved `reward.json` (`live_success`, `replay_success`, `state_identical`), not rerun: in the first 38 graded trials of the run that began on 2026-10-11 no verdict changes.
- 2026-10-13 (owner, boards 0085 and 0087): the sim service's clock (`environment/sim/rcb_limited.py`, `environment/sim/server.py`, the docstring of `environment/rcb_client.py`). `meta["time_s"]` and `meta["time_left_s"]` count from the agent's first request to the service (any endpoint), and a restarted service keeps that moment; before, they counted from the service's own start of the episode (when its scene was ready). Where the service is ready before the agent's phase begins (the harness is installed in between: a median of 79 s over the recorded standard trials), that clock ran ahead of the agent's and `time_left_s` reached 0 while the agent still had that time. `/spec` says what the clock counts from (`budget.clock`), and the instruction's `## Time limit` tells the agent to plan by `time_left_s`. The service never ended an episode on this clock: the scene, the frozen instance, the pass rule and the grader are unchanged, so the task keeps its name and earlier results stand.
+- 2026-10-14 (board 0113, from board 0086): the instruction's description of RoboDojo's check (`instruction.md`, "What RoboDojo checks") had the size order the wrong way round. It said the LARGEST doll must have the smallest world x, following the comments in RoboDojo's config; the check actually wants the SMALLEST doll at the smallest world x and x growing with size (index mod 5 = 0 is the largest doll: the dolls' measured bounding boxes in the frozen layout shrink as index mod 5 grows), which is what RoboDojo's own sentence "from left to right, from smallest to largest" says. Text only: the scene, the frozen layout, RoboDojo's success check and the verifier are unchanged, so the task keeps its name. Trials that followed the old text and arranged the row as it said failed for that reason (board 0086: models that then mirrored the row passed).
diff --git a/tasks/robodojo-sort-nesting-dolls-by-size-random-i00-standard/instruction.md b/tasks/robodojo-sort-nesting-dolls-by-size-random-i00-standard/instruction.md
index dbc33aaed2..cfebc92422 100644
--- a/tasks/robodojo-sort-nesting-dolls-by-size-random-i00-standard/instruction.md
+++ b/tasks/robodojo-sort-nesting-dolls-by-size-random-i00-standard/instruction.md
@@ -8,14 +8,14 @@ Nothing is handed in, so the task must be completed before the time limit (`## T
ends by itself when RoboDojo's check passes: `res.meta["done"]`). The time you wait for the service counts: its first
start in a trial takes ~4-15 min (Client, below).
-**Task:** Arrange the five nesting dolls upright in a straight row, ordered by size, from smallest to largest (left to right as the instruction says), and return the arms to their start pose.
+**Task:** Arrange the five nesting dolls upright in a straight row, ordered by size from the smallest at the smallest world x to the largest at the largest world x, and return the arms to their start pose.
RoboDojo's own instruction for this scene: *"Arrange the five nesting dolls in a row from left to right, from smallest to largest."*
## Success (RoboDojo's own check for this task)
-- Define the order of the dolls by their model index mod 5 (0 = smallest, 4 = largest; derived from the yml comments in the config: index 5 smallest half-extent 0.0152 m, index 9 largest 0.0335 m). The code lists the dolls sorted by (index mod 5) DESCENDING: L0 = largest, L1, L2, L3, L4 = smallest.
+- Define the order of the dolls by their model index mod 5, as the code does: 0 = the LARGEST doll, 4 = the SMALLEST (measured on the dolls of the frozen layout: their bounding boxes shrink as index mod 5 grows; the comments in RoboDojo's config, index 5 smallest and index 9 largest, have it the wrong way round). The code lists the dolls sorted by (index mod 5) DESCENDING: L0 = the smallest doll, L1, L2, L3, L4 = the largest.
- Row straightness: the maximum minus the minimum world Y of the five dolls' root positions is <= 0.035 m.
-- Order along X: for each consecutive pair (L_k, L_{k+1}), k=0..3: x(L_{k+1}) - x(L_k) > 0.03 m (strict), i.e. the largest doll has the smallest world x, and x increases by more than 3 cm with each smaller doll; the smallest doll has the largest world x. (The instruction says 'left to right, smallest to largest' while the code compares world/table x with the largest at the smallest x; presumably 'left' is as seen from the front camera, but this is not verifiable statically - use the numeric x rule.)
+- Order along X: for each consecutive pair (L_k, L_{k+1}), k=0..3: x(L_{k+1}) - x(L_k) > 0.03 m (strict), i.e. the SMALLEST doll has the smallest world x, and x increases by more than 3 cm with each larger doll; the largest doll has the largest world x. (This matches the instruction's 'left to right, from smallest to largest' with left = smaller world x.)
- Upright: each of doll0..doll4 has its local +Z axis within 15 deg (default threshold, strict) of world +Z (standing, not lying).
- Both target arms are back at origin: all target arms (left_arm and right_arm; the support arm is not counted) have their end-effector pose within 0.15 m on EACH axis (|dx|,|dy|,|dz| all <= 0.15 m) and within 20 degrees (quaternion angle) of the pose they had at the start of the episode (recorded after the scene settled).
- All conditions hold simultaneously after one action.
@@ -31,7 +31,7 @@ it does not apply here: where the description above mentions step_lim, read the
**At the end.** Both arms within 0.15 m per axis and 20 deg of their start pose in the same action as the dolls are correctly sorted, upright, and in a row.
-**Notes from RoboDojo's code.** The required X gap between consecutive dolls is only > 0.03 m between root positions (roots, not edges), while the row's Y spread must be <= 3.5 cm for all five, so the dolls need to be tightly aligned in y but only loosely spaced in x. The yml comments suggest the intended target positions are x = -0.24 (largest, index mod 5 = 4), -0.12, 0.0, 0.12, 0.24 (smallest) in the table frame, but any positions satisfying the stated rules succeed. The mapping index->size is inferred from yml comments, not verified from assets. Dolls must stay upright (within 15 deg) - a doll left lying fails. The random variant has different texture models and clutter distractors; reward identical.
+**Notes from RoboDojo's code.** The required X gap between consecutive dolls is only > 0.03 m between root positions (roots, not edges), while the row's Y spread must be <= 3.5 cm for all five, so the dolls need to be tightly aligned in y but only loosely spaced in x. The yml comments place the targets at table-frame x = -0.24 (index mod 5 = 4), -0.12, 0.0, 0.12, 0.24 (index mod 5 = 0) and call index mod 5 = 4 the largest; the dolls' measured sizes say the opposite (index mod 5 = 4 is the smallest), so the smallest doll goes at the smallest x. Any positions satisfying the stated rules succeed. Dolls must stay upright (within 15 deg) - a doll left lying fails. The random variant has different texture models and clutter distractors; reward identical.
## Rules (real-robot rules)
- **No reset.** The episode starts when the service comes up and ends as above, when your time is up, or on a simulator
diff --git a/tasks/robodojo-sort-nesting-dolls-by-size-random-i00-standard/task.toml b/tasks/robodojo-sort-nesting-dolls-by-size-random-i00-standard/task.toml
index 38cf3fce80..64a9830b38 100644
--- a/tasks/robodojo-sort-nesting-dolls-by-size-random-i00-standard/task.toml
+++ b/tasks/robodojo-sort-nesting-dolls-by-size-random-i00-standard/task.toml
@@ -15,7 +15,7 @@ artifacts = [{ source = "/var/rcb-monitor/", service = "rcb_monitor" },
[task]
name = "robot-coding-bench/robodojo-sort-nesting-dolls-by-size-random-i00-standard"
-description = "STANDARD MODE (robot as a service, eai-standard/2.2): RoboDojo sort_nesting_dolls_by_size_random (Generalization (randomised variant: other models, clutter)) on the same frozen layout as the privileged task (arx_x5/0/sort_nesting_dolls_by_size_random_0.json). Task: Arrange the five nesting dolls upright in a straight row, ordered by size, from smallest to largest (left to right as the instruction says), and return the arms to their start pose. The agent only gets a client: get_obs (RGB, depth + proprioception of both arms), submit (joint_position or ee_pose chunks). No reset, no privileged state; the sim service records the episode as a trajectory and grades it."
+description = "STANDARD MODE (robot as a service, eai-standard/2.2): RoboDojo sort_nesting_dolls_by_size_random (Generalization (randomised variant: other models, clutter)) on the same frozen layout as the privileged task (arx_x5/0/sort_nesting_dolls_by_size_random_0.json). Task: Arrange the five nesting dolls upright in a straight row, ordered by size from the smallest at the smallest world x to the largest at the largest world x, and return the arms to their start pose. The agent only gets a client: get_obs (RGB, depth + proprioception of both arms), submit (joint_position or ee_pose chunks). No reset, no privileged state; the sim service records the episode as a trajectory and grades it."
authors = [{ name = "Changhe Chen" }]
keywords = ["robotics", "robodojo", "isaac-lab", "isaac-sim", "bimanual", "arx-x5", "standard-mode", "trajectory"]
@@ -27,6 +27,7 @@ simulator = "robodojo"
robot = "arx-x5-bimanual"
family = "sort_nesting_dolls_by_size_random"
mode = "standard" # robot as a service: RGB, depth + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "no positive example yet; no reference solution (RoboDojo releases no expert, and its training demonstrations carry no initial state); reported separately until one exists; human review in PR #142"
diff --git a/tasks/robodojo-stack-blocks-by-language-i00-privileged/task.toml b/tasks/robodojo-stack-blocks-by-language-i00-privileged/task.toml
index 8f3b84cc0f..ca23a55cd4 100644
--- a/tasks/robodojo-stack-blocks-by-language-i00-privileged/task.toml
+++ b/tasks/robodojo-stack-blocks-by-language-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "robodojo" # RoboDojo @ cb2a37d on its Isaac Lab 2.3 fork
robot = "arx-x5-bimanual"
family = "stack_blocks_by_language"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "positive example: a model trial solved this task (Codex + GPT-6 Luna, effort xhigh, run `codex-gpt6_luna-xhigh`, 2026-10-10; 1616 actions, graded by this verifier); no reference solution (RoboDojo releases no expert); human review in PR #142"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robodojo-stack-blocks-by-language-i00-standard/task.toml b/tasks/robodojo-stack-blocks-by-language-i00-standard/task.toml
index 0881a84368..917026b63e 100644
--- a/tasks/robodojo-stack-blocks-by-language-i00-standard/task.toml
+++ b/tasks/robodojo-stack-blocks-by-language-i00-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "robodojo"
robot = "arx-x5-bimanual"
family = "stack_blocks_by_language"
mode = "standard" # robot as a service: RGB, depth + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "no positive example yet (the privileged twin has a positive example (Codex + GPT-6 Luna, effort xhigh, run `codex-gpt6_luna-xhigh`, 2026-10-10)); no reference solution (RoboDojo releases no expert, and its training demonstrations carry no initial state); reported separately until one exists; human review in PR #142"
diff --git a/tasks/robodojo-stack-blocks-i00-privileged/task.toml b/tasks/robodojo-stack-blocks-i00-privileged/task.toml
index 36d496ac5c..19ecfb0a39 100644
--- a/tasks/robodojo-stack-blocks-i00-privileged/task.toml
+++ b/tasks/robodojo-stack-blocks-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "robodojo" # RoboDojo @ cb2a37d on its Isaac Lab 2.3 fork
robot = "arx-x5-bimanual"
family = "stack_blocks"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "no positive example yet; no reference solution (RoboDojo releases no expert, and its training demonstrations carry no initial state); reported separately until one exists; human review in PR #142"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robodojo-stack-blocks-i00-standard/task.toml b/tasks/robodojo-stack-blocks-i00-standard/task.toml
index cea4d59f51..a10de36e14 100644
--- a/tasks/robodojo-stack-blocks-i00-standard/task.toml
+++ b/tasks/robodojo-stack-blocks-i00-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "robodojo"
robot = "arx-x5-bimanual"
family = "stack_blocks"
mode = "standard" # robot as a service: RGB, depth + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "no positive example yet; no reference solution (RoboDojo releases no expert, and its training demonstrations carry no initial state); reported separately until one exists; human review in PR #142"
diff --git a/tasks/robodojo-stack-blocks-random-i00-privileged/task.toml b/tasks/robodojo-stack-blocks-random-i00-privileged/task.toml
index ad2bef8b84..73a3e0f554 100644
--- a/tasks/robodojo-stack-blocks-random-i00-privileged/task.toml
+++ b/tasks/robodojo-stack-blocks-random-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "robodojo" # RoboDojo @ cb2a37d on its Isaac Lab 2.3 fork
robot = "arx-x5-bimanual"
family = "stack_blocks_random"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "no positive example yet; no reference solution (RoboDojo releases no expert, and its training demonstrations carry no initial state); reported separately until one exists; human review in PR #142"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robodojo-stack-blocks-random-i00-standard/task.toml b/tasks/robodojo-stack-blocks-random-i00-standard/task.toml
index 253f44824d..a54ecaa0ac 100644
--- a/tasks/robodojo-stack-blocks-random-i00-standard/task.toml
+++ b/tasks/robodojo-stack-blocks-random-i00-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "robodojo"
robot = "arx-x5-bimanual"
family = "stack_blocks_random"
mode = "standard" # robot as a service: RGB, depth + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "no positive example yet; no reference solution (RoboDojo releases no expert, and its training demonstrations carry no initial state); reported separately until one exists; human review in PR #142"
diff --git a/tasks/robodojo-stack-bowls-i00-privileged/task.toml b/tasks/robodojo-stack-bowls-i00-privileged/task.toml
index 94acceb240..f17865402e 100644
--- a/tasks/robodojo-stack-bowls-i00-privileged/task.toml
+++ b/tasks/robodojo-stack-bowls-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "robodojo" # RoboDojo @ cb2a37d on its Isaac Lab 2.3 fork
robot = "arx-x5-bimanual"
family = "stack_bowls"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "no positive example yet (a model trial failed (Codex + GPT-6 Luna, effort xhigh, run `codex-gpt6_luna-xhigh`, 2026-10-10)); no reference solution (RoboDojo releases no expert, and its training demonstrations carry no initial state); reported separately until one exists; human review in PR #142"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robodojo-stack-bowls-i00-standard/task.toml b/tasks/robodojo-stack-bowls-i00-standard/task.toml
index b6e7afc655..132d871f54 100644
--- a/tasks/robodojo-stack-bowls-i00-standard/task.toml
+++ b/tasks/robodojo-stack-bowls-i00-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "robodojo"
robot = "arx-x5-bimanual"
family = "stack_bowls"
mode = "standard" # robot as a service: RGB, depth + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "no positive example yet; no reference solution (RoboDojo releases no expert, and its training demonstrations carry no initial state); reported separately until one exists; human review in PR #142"
diff --git a/tasks/robodojo-stack-bowls-random-i00-privileged/task.toml b/tasks/robodojo-stack-bowls-random-i00-privileged/task.toml
index 625299a11e..732217ffba 100644
--- a/tasks/robodojo-stack-bowls-random-i00-privileged/task.toml
+++ b/tasks/robodojo-stack-bowls-random-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "robodojo" # RoboDojo @ cb2a37d on its Isaac Lab 2.3 fork
robot = "arx-x5-bimanual"
family = "stack_bowls_random"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "no positive example yet; no reference solution (RoboDojo releases no expert, and its training demonstrations carry no initial state); reported separately until one exists; human review in PR #142"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robodojo-stack-bowls-random-i00-standard/task.toml b/tasks/robodojo-stack-bowls-random-i00-standard/task.toml
index 5f70baa733..9a9b428f09 100644
--- a/tasks/robodojo-stack-bowls-random-i00-standard/task.toml
+++ b/tasks/robodojo-stack-bowls-random-i00-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "robodojo"
robot = "arx-x5-bimanual"
family = "stack_bowls_random"
mode = "standard" # robot as a service: RGB, depth + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "no positive example yet; no reference solution (RoboDojo releases no expert, and its training demonstrations carry no initial state); reported separately until one exists; human review in PR #142"
diff --git a/tasks/robodojo-store-laptop-and-headphones-i00-privileged/task.toml b/tasks/robodojo-store-laptop-and-headphones-i00-privileged/task.toml
index 32c78af13e..45d92557bd 100644
--- a/tasks/robodojo-store-laptop-and-headphones-i00-privileged/task.toml
+++ b/tasks/robodojo-store-laptop-and-headphones-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "robodojo" # RoboDojo @ cb2a37d on its Isaac Lab 2.3 fork
robot = "arx-x5-bimanual"
family = "store_laptop_and_headphones"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "no positive example yet; no reference solution (RoboDojo releases no expert, and its training demonstrations carry no initial state); reported separately until one exists; human review in PR #142"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robodojo-store-laptop-and-headphones-i00-standard/task.toml b/tasks/robodojo-store-laptop-and-headphones-i00-standard/task.toml
index 8c3c5ac8a4..0cda3f3940 100644
--- a/tasks/robodojo-store-laptop-and-headphones-i00-standard/task.toml
+++ b/tasks/robodojo-store-laptop-and-headphones-i00-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "robodojo"
robot = "arx-x5-bimanual"
family = "store_laptop_and_headphones"
mode = "standard" # robot as a service: RGB, depth + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "no positive example yet; no reference solution (RoboDojo releases no expert, and its training demonstrations carry no initial state); reported separately until one exists; human review in PR #142"
diff --git a/tasks/robodojo-store-laptop-and-headphones-random-i00-privileged/task.toml b/tasks/robodojo-store-laptop-and-headphones-random-i00-privileged/task.toml
index de2d095b78..f20cc41749 100644
--- a/tasks/robodojo-store-laptop-and-headphones-random-i00-privileged/task.toml
+++ b/tasks/robodojo-store-laptop-and-headphones-random-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "robodojo" # RoboDojo @ cb2a37d on its Isaac Lab 2.3 fork
robot = "arx-x5-bimanual"
family = "store_laptop_and_headphones_random"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "no positive example yet; no reference solution (RoboDojo releases no expert, and its training demonstrations carry no initial state); reported separately until one exists; human review in PR #142"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robodojo-store-laptop-and-headphones-random-i00-standard/task.toml b/tasks/robodojo-store-laptop-and-headphones-random-i00-standard/task.toml
index 0e0b52658b..526a0aa551 100644
--- a/tasks/robodojo-store-laptop-and-headphones-random-i00-standard/task.toml
+++ b/tasks/robodojo-store-laptop-and-headphones-random-i00-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "robodojo"
robot = "arx-x5-bimanual"
family = "store_laptop_and_headphones_random"
mode = "standard" # robot as a service: RGB, depth + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "no positive example yet; no reference solution (RoboDojo releases no expert, and its training demonstrations carry no initial state); reported separately until one exists; human review in PR #142"
diff --git a/tasks/robodojo-store-tools-in-toolbox-i00-privileged/task.toml b/tasks/robodojo-store-tools-in-toolbox-i00-privileged/task.toml
index 8799e49177..eb1fd27156 100644
--- a/tasks/robodojo-store-tools-in-toolbox-i00-privileged/task.toml
+++ b/tasks/robodojo-store-tools-in-toolbox-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "robodojo" # RoboDojo @ cb2a37d on its Isaac Lab 2.3 fork
robot = "arx-x5-bimanual"
family = "store_tools_in_toolbox"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "no positive example yet; no reference solution (RoboDojo releases no expert, and its training demonstrations carry no initial state); reported separately until one exists; human review in PR #142"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robodojo-store-tools-in-toolbox-i00-standard/task.toml b/tasks/robodojo-store-tools-in-toolbox-i00-standard/task.toml
index a4db6470f5..c1714ada1c 100644
--- a/tasks/robodojo-store-tools-in-toolbox-i00-standard/task.toml
+++ b/tasks/robodojo-store-tools-in-toolbox-i00-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "robodojo"
robot = "arx-x5-bimanual"
family = "store_tools_in_toolbox"
mode = "standard" # robot as a service: RGB, depth + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "no positive example yet; no reference solution (RoboDojo releases no expert, and its training demonstrations carry no initial state); reported separately until one exists; human review in PR #142"
diff --git a/tasks/robodojo-swap-blocks-i00-privileged/README.md b/tasks/robodojo-swap-blocks-i00-privileged/README.md
index aa31f68c56..6bd2ff7fb6 100644
--- a/tasks/robodojo-swap-blocks-i00-privileged/README.md
+++ b/tasks/robodojo-swap-blocks-i00-privileged/README.md
@@ -58,3 +58,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
### Change Log
- 2026-10-11: agent time limit 5400 s -> 3600 s (`[agent] timeout_sec`; owner decision, docs/benchmarks/robodojo.md), and `instruction.md` states the cutoff at the top. The scene, the frozen layout, RoboDojo's success check and the verifier are unchanged, so the task keeps its name; trials recorded before this date ran with 5400 s and are kept apart (no official result was published with it).
- 2026-10-12: graded by ONE fresh-process replay (two before; owner decision, docs/benchmarks/robodojo.md): reward 1 iff that replay ends in RoboDojo's success. Before, both replays had to succeed and end in the same state hash; with one replay nothing is compared, so `reward.json` reports `deterministic` = 1 with `determinism_checked` = 0 and `replays` = 1, and a replay whose process did not rebuild the frozen layout to its recorded hash gives no verdict (`instance_mismatch`, `infra_error`). `[verifier] timeout_sec` 15000 s -> 7800 s (one replay of the 15,000 rows a trajectory may have, 7200 s, + 600 s). The scene, the frozen layout, RoboDojo's success check and what the agent hands in are unchanged, so the task keeps its name; `instruction.md` says one replay. No ranked result exists; a trial graded before this date is re-read from its saved `reward.json` (`all_success`, `state_identical`), not rerun: in the first 38 graded trials of the run that began on 2026-10-11 the two replays always ended in the same state, and no verdict changes.
+- 2026-10-14 (board 0113, from board 0086): the instruction said the spring button returns to its rest position by itself after release. In this scene it comes back only to about 0.79 of its travel ratio (measured in two trials on this frozen layout), below the 0.9 that the 'released' stages need and the 0.95 that re-arms RoboDojo's press counter; the instruction now says so, and that the cap has to be lifted back up after each press. Text only: the scene, the frozen layout, RoboDojo's success check and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/robodojo-swap-blocks-i00-privileged/instruction.md b/tasks/robodojo-swap-blocks-i00-privileged/instruction.md
index a97a80137a..031a6c2569 100644
--- a/tasks/robodojo-swap-blocks-i00-privileged/instruction.md
+++ b/tasks/robodojo-swap-blocks-i00-privileged/instruction.md
@@ -43,11 +43,11 @@ simulated time). RoboDojo's own evaluation gives a policy 700 actions (28 s) for
reads that limit and it does not apply here: where the description above mentions step_lim, read the guard.
`s.status()` reports the end.
-**What changes by itself.** The spring button returns to its rest position by itself after being released (spring). Otherwise none.
+**What changes by itself.** The spring button pushes back up by itself after being released, but in this scene it comes to rest at about 0.79 of its travel ratio (measured in two trials on this frozen layout), below the 0.9 that a 'released' stage needs and the 0.95 that re-arms the press counter: after each press the button cap has to be pulled back up (above 0.95) before the next stage or press can count. Otherwise none.
**At the end.** Final stage requires all target arms back within 0.15 m per axis and 20 degrees of their start end-effector pose, simultaneously with button0 ratio > 0.9 (released). Nothing about gripper opening.
-**Notes from RoboDojo's code.** Many predicates use OR inside the nested brackets (a bracket list inside a stage is evaluated with OR semantics in check_once; only flat items of a stage are ANDed), so individual stages are looser than the intent (e.g. 'either block lifted'), but the ordered sequence plus the final stage 10 (both blocks swapped) and the exact-3-presses query effectively force the 3-move swap: move one block to the empty mat, press, move the other block into the first block's old mat, press, move the first block into the second block's old mat, press. Which block goes first is free (either order passes since the checks use OR). The button must be pressed fully (<0.5 of travel) and released (>0.9 for stage checks, >0.95 for a new press to be counted by the query) after every move, and extra presses (e.g. a double tap that re-arms above 0.95) are fatal. Pressing the button before the first lift/move counts toward the 3-press total. 'Lifted' is relative to the z recorded at the start or at the last re-record (stages 2 and 6). Block release is not checked directly, but final positions must be within 0.03 m (root to root) of the mat centres. Success is evaluated on object root poses/points, not on gripper contact. Object geometry sizes/asset dimensions are not in the source and cannot be determined statically.
+**Notes from RoboDojo's code.** Many predicates use OR inside the nested brackets (a bracket list inside a stage is evaluated with OR semantics in check_once; only flat items of a stage are ANDed), so individual stages are looser than the intent (e.g. 'either block lifted'), but the ordered sequence plus the final stage 10 (both blocks swapped) and the exact-3-presses query effectively force the 3-move swap: move one block to the empty mat, press, move the other block into the first block's old mat, press, move the first block into the second block's old mat, press. Which block goes first is free (either order passes since the checks use OR). The button must be pressed fully (<0.5 of travel) and released (>0.9 for stage checks, >0.95 for a new press to be counted by the query) after every move; its spring alone returns it only to about 0.79, so lift its cap back up after each press, and extra presses (e.g. a double tap that re-arms above 0.95) are fatal. Pressing the button before the first lift/move counts toward the 3-press total. 'Lifted' is relative to the z recorded at the start or at the last re-record (stages 2 and 6). Block release is not checked directly, but final positions must be within 0.03 m (root to root) of the mat centres. Success is evaluated on object root poses/points, not on gripper contact. Object geometry sizes/asset dimensions are not in the source and cannot be determined statically.
## Deliverable
Write the final trajectory to `/app/output/trajectory.npz` with `Session.save_trajectory(...)` (below). It is replayed
diff --git a/tasks/robodojo-swap-blocks-i00-privileged/task.toml b/tasks/robodojo-swap-blocks-i00-privileged/task.toml
index baf97a4e82..64c195809a 100644
--- a/tasks/robodojo-swap-blocks-i00-privileged/task.toml
+++ b/tasks/robodojo-swap-blocks-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "robodojo" # RoboDojo @ cb2a37d on its Isaac Lab 2.3 fork
robot = "arx-x5-bimanual"
family = "swap_blocks"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "no positive example yet; no reference solution (RoboDojo releases no expert, and its training demonstrations carry no initial state); reported separately until one exists; human review in PR #142"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robodojo-swap-blocks-i00-standard/README.md b/tasks/robodojo-swap-blocks-i00-standard/README.md
index ac86655d4f..00ff405ebc 100644
--- a/tasks/robodojo-swap-blocks-i00-standard/README.md
+++ b/tasks/robodojo-swap-blocks-i00-standard/README.md
@@ -56,3 +56,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
- 2026-10-11: agent time limit 5400 s -> 3600 s (`[agent] timeout_sec`, and the service's `RCB_WALL_CLOCK_S`, which only reports it: `spec["budget"]`, `meta["time_left_s"]`; owner decision, docs/benchmarks/robodojo.md), and `instruction.md` states the cutoff at the top. The scene, the frozen layout, RoboDojo's success check, the collect hook and the verifier are unchanged, so the task keeps its name; trials recorded before this date ran with 5400 s and are kept apart (no official result was published with it).
- 2026-10-12: graded with ONE fresh-process replay of the recorded episode (two before; owner decision, docs/benchmarks/robodojo.md): `success` = `live_success` and `replay_success`, one successful replay is enough. Before, two replays had to succeed and end in the same state hash. `deterministic` (= `state_identical` = `live_matches_replay`) now says whether the replay ended in the live episode's state; it is reported and no longer decides. The collect hook's limit 18000 s -> 10800 s (`[[verifier.collect]] timeout_sec` = `finalize.sh`'s `HOOK_BUDGET_S`); inside it the replay has max(1800 s, 1200 s + actions / 2.5) and is run once more when it gives no result in that time. The scene, the frozen layout, RoboDojo's success check and the service the agent talks to are unchanged, so the task keeps its name. No ranked result exists; a trial graded before this date is re-read from its saved `reward.json` (`live_success`, `replay_success`, `state_identical`), not rerun: in the first 38 graded trials of the run that began on 2026-10-11 no verdict changes.
- 2026-10-13 (owner, boards 0085 and 0087): the sim service's clock (`environment/sim/rcb_limited.py`, `environment/sim/server.py`, the docstring of `environment/rcb_client.py`). `meta["time_s"]` and `meta["time_left_s"]` count from the agent's first request to the service (any endpoint), and a restarted service keeps that moment; before, they counted from the service's own start of the episode (when its scene was ready). Where the service is ready before the agent's phase begins (the harness is installed in between: a median of 79 s over the recorded standard trials), that clock ran ahead of the agent's and `time_left_s` reached 0 while the agent still had that time. `/spec` says what the clock counts from (`budget.clock`), and the instruction's `## Time limit` tells the agent to plan by `time_left_s`. The service never ended an episode on this clock: the scene, the frozen instance, the pass rule and the grader are unchanged, so the task keeps its name and earlier results stand.
+- 2026-10-14 (board 0113, from board 0086): the instruction said the spring button returns to its rest position by itself after release. In this scene it comes back only to about 0.79 of its travel ratio (measured in two trials on this frozen layout), below the 0.9 that the 'released' stages need and the 0.95 that re-arms RoboDojo's press counter; the instruction now says so, and that the cap has to be lifted back up after each press. Text only: the scene, the frozen layout, RoboDojo's success check and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/robodojo-swap-blocks-i00-standard/instruction.md b/tasks/robodojo-swap-blocks-i00-standard/instruction.md
index 3b61e6182a..a190c2811f 100644
--- a/tasks/robodojo-swap-blocks-i00-standard/instruction.md
+++ b/tasks/robodojo-swap-blocks-i00-standard/instruction.md
@@ -36,11 +36,11 @@ RoboDojo's own evaluation gives a policy 700 actions (28 s) for this task; no ch
it does not apply here: where the description above mentions step_lim, read the guard. `res.meta["done"]` /
`res.meta["reason"]` report the end.
-**What changes by itself.** The spring button returns to its rest position by itself after being released (spring). Otherwise none.
+**What changes by itself.** The spring button pushes back up by itself after being released, but in this scene it comes to rest at about 0.79 of its travel ratio (measured in two trials on this frozen layout), below the 0.9 that a 'released' stage needs and the 0.95 that re-arms the press counter: after each press the button cap has to be pulled back up (above 0.95) before the next stage or press can count. Otherwise none.
**At the end.** Final stage requires all target arms back within 0.15 m per axis and 20 degrees of their start end-effector pose, simultaneously with button0 ratio > 0.9 (released). Nothing about gripper opening.
-**Notes from RoboDojo's code.** Many predicates use OR inside the nested brackets (a bracket list inside a stage is evaluated with OR semantics in check_once; only flat items of a stage are ANDed), so individual stages are looser than the intent (e.g. 'either block lifted'), but the ordered sequence plus the final stage 10 (both blocks swapped) and the exact-3-presses query effectively force the 3-move swap: move one block to the empty mat, press, move the other block into the first block's old mat, press, move the first block into the second block's old mat, press. Which block goes first is free (either order passes since the checks use OR). The button must be pressed fully (<0.5 of travel) and released (>0.9 for stage checks, >0.95 for a new press to be counted by the query) after every move, and extra presses (e.g. a double tap that re-arms above 0.95) are fatal. Pressing the button before the first lift/move counts toward the 3-press total. 'Lifted' is relative to the z recorded at the start or at the last re-record (stages 2 and 6). Block release is not checked directly, but final positions must be within 0.03 m (root to root) of the mat centres. Success is evaluated on object root poses/points, not on gripper contact. Object geometry sizes/asset dimensions are not in the source and cannot be determined statically.
+**Notes from RoboDojo's code.** Many predicates use OR inside the nested brackets (a bracket list inside a stage is evaluated with OR semantics in check_once; only flat items of a stage are ANDed), so individual stages are looser than the intent (e.g. 'either block lifted'), but the ordered sequence plus the final stage 10 (both blocks swapped) and the exact-3-presses query effectively force the 3-move swap: move one block to the empty mat, press, move the other block into the first block's old mat, press, move the first block into the second block's old mat, press. Which block goes first is free (either order passes since the checks use OR). The button must be pressed fully (<0.5 of travel) and released (>0.9 for stage checks, >0.95 for a new press to be counted by the query) after every move; its spring alone returns it only to about 0.79, so lift its cap back up after each press, and extra presses (e.g. a double tap that re-arms above 0.95) are fatal. Pressing the button before the first lift/move counts toward the 3-press total. 'Lifted' is relative to the z recorded at the start or at the last re-record (stages 2 and 6). Block release is not checked directly, but final positions must be within 0.03 m (root to root) of the mat centres. Success is evaluated on object root poses/points, not on gripper contact. Object geometry sizes/asset dimensions are not in the source and cannot be determined statically.
## Rules (real-robot rules)
- **No reset.** The episode starts when the service comes up and ends as above, when your time is up, or on a simulator
diff --git a/tasks/robodojo-swap-blocks-i00-standard/task.toml b/tasks/robodojo-swap-blocks-i00-standard/task.toml
index 8308977b54..8e432794c3 100644
--- a/tasks/robodojo-swap-blocks-i00-standard/task.toml
+++ b/tasks/robodojo-swap-blocks-i00-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "robodojo"
robot = "arx-x5-bimanual"
family = "swap_blocks"
mode = "standard" # robot as a service: RGB, depth + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "no positive example yet; no reference solution (RoboDojo releases no expert, and its training demonstrations carry no initial state); reported separately until one exists; human review in PR #142"
diff --git a/tasks/robodojo-swap-t-i00-privileged/task.toml b/tasks/robodojo-swap-t-i00-privileged/task.toml
index 456c8f24be..c6ae2f2cfb 100644
--- a/tasks/robodojo-swap-t-i00-privileged/task.toml
+++ b/tasks/robodojo-swap-t-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "robodojo" # RoboDojo @ cb2a37d on its Isaac Lab 2.3 fork
robot = "arx-x5-bimanual"
family = "swap_T"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "no positive example yet; no reference solution (RoboDojo releases no expert, and its training demonstrations carry no initial state); reported separately until one exists; human review in PR #142"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robodojo-swap-t-i00-standard/task.toml b/tasks/robodojo-swap-t-i00-standard/task.toml
index 7e77321c42..6ff4d2e9c4 100644
--- a/tasks/robodojo-swap-t-i00-standard/task.toml
+++ b/tasks/robodojo-swap-t-i00-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "robodojo"
robot = "arx-x5-bimanual"
family = "swap_T"
mode = "standard" # robot as a service: RGB, depth + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "no positive example yet; no reference solution (RoboDojo releases no expert, and its training demonstrations carry no initial state); reported separately until one exists; human review in PR #142"
diff --git a/tasks/robodojo-sweep-blocks-i00-privileged/task.toml b/tasks/robodojo-sweep-blocks-i00-privileged/task.toml
index b22d46247d..24065d71a1 100644
--- a/tasks/robodojo-sweep-blocks-i00-privileged/task.toml
+++ b/tasks/robodojo-sweep-blocks-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "robodojo" # RoboDojo @ cb2a37d on its Isaac Lab 2.3 fork
robot = "arx-x5-bimanual"
family = "sweep_blocks"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "no positive example yet; no reference solution (RoboDojo releases no expert, and its training demonstrations carry no initial state); reported separately until one exists; human review in PR #142"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robodojo-sweep-blocks-i00-standard/task.toml b/tasks/robodojo-sweep-blocks-i00-standard/task.toml
index 81b246862b..5cf9bd00de 100644
--- a/tasks/robodojo-sweep-blocks-i00-standard/task.toml
+++ b/tasks/robodojo-sweep-blocks-i00-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "robodojo"
robot = "arx-x5-bimanual"
family = "sweep_blocks"
mode = "standard" # robot as a service: RGB, depth + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "no positive example yet; no reference solution (RoboDojo releases no expert, and its training demonstrations carry no initial state); reported separately until one exists; human review in PR #142"
diff --git a/tasks/robodojo-sweep-blocks-random-i00-privileged/task.toml b/tasks/robodojo-sweep-blocks-random-i00-privileged/task.toml
index 4d0dab4f97..daf1f4e4dd 100644
--- a/tasks/robodojo-sweep-blocks-random-i00-privileged/task.toml
+++ b/tasks/robodojo-sweep-blocks-random-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "robodojo" # RoboDojo @ cb2a37d on its Isaac Lab 2.3 fork
robot = "arx-x5-bimanual"
family = "sweep_blocks_random"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "no positive example yet; no reference solution (RoboDojo releases no expert, and its training demonstrations carry no initial state); reported separately until one exists; human review in PR #142"
expert_time_estimate_hours = 2.0
diff --git a/tasks/robodojo-sweep-blocks-random-i00-standard/task.toml b/tasks/robodojo-sweep-blocks-random-i00-standard/task.toml
index 4cee2dfcb9..dfbe6333b0 100644
--- a/tasks/robodojo-sweep-blocks-random-i00-standard/task.toml
+++ b/tasks/robodojo-sweep-blocks-random-i00-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "robodojo"
robot = "arx-x5-bimanual"
family = "sweep_blocks_random"
mode = "standard" # robot as a service: RGB, depth + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "no positive example yet; no reference solution (RoboDojo releases no expert, and its training demonstrations carry no initial state); reported separately until one exists; human review in PR #142"
diff --git a/tasks/robofunolympics-backflip-i00-privileged/task.toml b/tasks/robofunolympics-backflip-i00-privileged/task.toml
index a17be2cb91..ff2790af5c 100644
--- a/tasks/robofunolympics-backflip-i00-privileged/task.toml
+++ b/tasks/robofunolympics-backflip-i00-privileged/task.toml
@@ -15,6 +15,7 @@ robot = "unitree-g1"
family = "backflip"
category = "parkour-apparatus-balance"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
# scene = environment/instance.json (the family's parameters + its t = 0 state hash); identity = the dataset.toml digest
instance_seed = 0
expert_time_estimate_hours = 10.0
diff --git a/tasks/robofunolympics-backflip-i00-standard/task.toml b/tasks/robofunolympics-backflip-i00-standard/task.toml
index ff4580b082..31b2be7345 100644
--- a/tasks/robofunolympics-backflip-i00-standard/task.toml
+++ b/tasks/robofunolympics-backflip-i00-standard/task.toml
@@ -26,6 +26,7 @@ robot = "unitree-g1"
family = "backflip"
category = "parkour-apparatus-balance"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
instance_seed = 0
reset_budget = 50
diff --git a/tasks/robofunolympics-balance-beam-i00-privileged/task.toml b/tasks/robofunolympics-balance-beam-i00-privileged/task.toml
index f4cb453b5d..b39bce73f5 100644
--- a/tasks/robofunolympics-balance-beam-i00-privileged/task.toml
+++ b/tasks/robofunolympics-balance-beam-i00-privileged/task.toml
@@ -15,6 +15,7 @@ robot = "unitree-g1"
family = "balance_beam"
category = "parkour-apparatus-balance"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
# scene = environment/instance.json (the family's parameters + its t = 0 state hash); identity = the dataset.toml digest
instance_seed = 0
expert_time_estimate_hours = 8.0
diff --git a/tasks/robofunolympics-balance-beam-i00-standard/task.toml b/tasks/robofunolympics-balance-beam-i00-standard/task.toml
index 56ed71ae0c..d2e672015e 100644
--- a/tasks/robofunolympics-balance-beam-i00-standard/task.toml
+++ b/tasks/robofunolympics-balance-beam-i00-standard/task.toml
@@ -26,6 +26,7 @@ robot = "unitree-g1"
family = "balance_beam"
category = "parkour-apparatus-balance"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
instance_seed = 0
reset_budget = 50
diff --git a/tasks/robofunolympics-balance-board-i00-privileged/task.toml b/tasks/robofunolympics-balance-board-i00-privileged/task.toml
index da8715a55b..2783a7762f 100644
--- a/tasks/robofunolympics-balance-board-i00-privileged/task.toml
+++ b/tasks/robofunolympics-balance-board-i00-privileged/task.toml
@@ -15,6 +15,7 @@ robot = "unitree-g1"
family = "balance_board"
category = "parkour-apparatus-balance"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
# scene = environment/instance.json (the family's parameters + its t = 0 state hash); identity = the dataset.toml digest
instance_seed = 0
expert_time_estimate_hours = 6.0
diff --git a/tasks/robofunolympics-balance-board-i00-standard/task.toml b/tasks/robofunolympics-balance-board-i00-standard/task.toml
index e1418742ef..dd57a72956 100644
--- a/tasks/robofunolympics-balance-board-i00-standard/task.toml
+++ b/tasks/robofunolympics-balance-board-i00-standard/task.toml
@@ -26,6 +26,7 @@ robot = "unitree-g1"
family = "balance_board"
category = "parkour-apparatus-balance"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
instance_seed = 0
reset_budget = 50
diff --git a/tasks/robofunolympics-basketball-dribble-i00-privileged/task.toml b/tasks/robofunolympics-basketball-dribble-i00-privileged/task.toml
index 73b9e1cc8f..77bc48d1af 100644
--- a/tasks/robofunolympics-basketball-dribble-i00-privileged/task.toml
+++ b/tasks/robofunolympics-basketball-dribble-i00-privileged/task.toml
@@ -15,6 +15,7 @@ robot = "unitree-h1-2"
family = "basketball_dribble"
category = "basketball-and-rackets"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
# scene = environment/instance.json (the family's parameters + its t = 0 state hash); identity = the dataset.toml digest
instance_seed = 0
expert_time_estimate_hours = 8.0
diff --git a/tasks/robofunolympics-basketball-dribble-i00-standard/task.toml b/tasks/robofunolympics-basketball-dribble-i00-standard/task.toml
index 4cfe7f90b9..6b6f1faae1 100644
--- a/tasks/robofunolympics-basketball-dribble-i00-standard/task.toml
+++ b/tasks/robofunolympics-basketball-dribble-i00-standard/task.toml
@@ -26,6 +26,7 @@ robot = "unitree-h1-2"
family = "basketball_dribble"
category = "basketball-and-rackets"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
instance_seed = 0
reset_budget = 50
diff --git a/tasks/robofunolympics-bowling-i00-privileged/task.toml b/tasks/robofunolympics-bowling-i00-privileged/task.toml
index 93b2676f94..30faf14e45 100644
--- a/tasks/robofunolympics-bowling-i00-privileged/task.toml
+++ b/tasks/robofunolympics-bowling-i00-privileged/task.toml
@@ -15,6 +15,7 @@ robot = "unitree-h1-2"
family = "bowling"
category = "throwing-catching-striking"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
# scene = environment/instance.json (the family's parameters + its t = 0 state hash); identity = the dataset.toml digest
instance_seed = 0
expert_time_estimate_hours = 8.0
diff --git a/tasks/robofunolympics-bowling-i00-standard/task.toml b/tasks/robofunolympics-bowling-i00-standard/task.toml
index 45dc17d762..577c0b2702 100644
--- a/tasks/robofunolympics-bowling-i00-standard/task.toml
+++ b/tasks/robofunolympics-bowling-i00-standard/task.toml
@@ -26,6 +26,7 @@ robot = "unitree-h1-2"
family = "bowling"
category = "throwing-catching-striking"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
instance_seed = 0
reset_budget = 50
diff --git a/tasks/robofunolympics-box-jump-i00-privileged/task.toml b/tasks/robofunolympics-box-jump-i00-privileged/task.toml
index e0c2dec331..0a912301be 100644
--- a/tasks/robofunolympics-box-jump-i00-privileged/task.toml
+++ b/tasks/robofunolympics-box-jump-i00-privileged/task.toml
@@ -15,6 +15,7 @@ robot = "unitree-h1-2"
family = "box_jump"
category = "parkour-apparatus-balance"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
# scene = environment/instance.json (the family's parameters + its t = 0 state hash); identity = the dataset.toml digest
instance_seed = 0
expert_time_estimate_hours = 8.0
diff --git a/tasks/robofunolympics-box-jump-i00-standard/task.toml b/tasks/robofunolympics-box-jump-i00-standard/task.toml
index 15190d7ab1..8caacbeee8 100644
--- a/tasks/robofunolympics-box-jump-i00-standard/task.toml
+++ b/tasks/robofunolympics-box-jump-i00-standard/task.toml
@@ -26,6 +26,7 @@ robot = "unitree-h1-2"
family = "box_jump"
category = "parkour-apparatus-balance"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
instance_seed = 0
reset_budget = 50
diff --git a/tasks/robofunolympics-boxing-i00-privileged/task.toml b/tasks/robofunolympics-boxing-i00-privileged/task.toml
index 039a67194d..50fc9cc890 100644
--- a/tasks/robofunolympics-boxing-i00-privileged/task.toml
+++ b/tasks/robofunolympics-boxing-i00-privileged/task.toml
@@ -15,6 +15,7 @@ robot = "unitree-h1-2"
family = "boxing"
category = "throwing-catching-striking"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
# scene = environment/instance.json (the family's parameters + its t = 0 state hash); identity = the dataset.toml digest
instance_seed = 0
expert_time_estimate_hours = 4.0
diff --git a/tasks/robofunolympics-boxing-i00-standard/task.toml b/tasks/robofunolympics-boxing-i00-standard/task.toml
index 1789689b91..0ca495fa34 100644
--- a/tasks/robofunolympics-boxing-i00-standard/task.toml
+++ b/tasks/robofunolympics-boxing-i00-standard/task.toml
@@ -26,6 +26,7 @@ robot = "unitree-h1-2"
family = "boxing"
category = "throwing-catching-striking"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
instance_seed = 0
reset_budget = 50
diff --git a/tasks/robofunolympics-catch-and-shoot-i00-privileged/task.toml b/tasks/robofunolympics-catch-and-shoot-i00-privileged/task.toml
index df558a96dc..57a601054b 100644
--- a/tasks/robofunolympics-catch-and-shoot-i00-privileged/task.toml
+++ b/tasks/robofunolympics-catch-and-shoot-i00-privileged/task.toml
@@ -15,6 +15,7 @@ robot = "unitree-h1-2"
family = "catch_and_shoot"
category = "basketball-and-rackets"
mode = "privileged"
+dynamics = "dynamic" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
# scene = environment/instance.json (the family's parameters + its t = 0 state hash); identity = the dataset.toml digest
instance_seed = 0
expert_time_estimate_hours = 6.0
diff --git a/tasks/robofunolympics-catch-and-shoot-i00-standard/task.toml b/tasks/robofunolympics-catch-and-shoot-i00-standard/task.toml
index bbc6767d84..156c735a1b 100644
--- a/tasks/robofunolympics-catch-and-shoot-i00-standard/task.toml
+++ b/tasks/robofunolympics-catch-and-shoot-i00-standard/task.toml
@@ -26,6 +26,7 @@ robot = "unitree-h1-2"
family = "catch_and_shoot"
category = "basketball-and-rackets"
mode = "standard"
+dynamics = "dynamic" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
instance_seed = 0
reset_budget = 50
diff --git a/tasks/robofunolympics-catch-i00-privileged/task.toml b/tasks/robofunolympics-catch-i00-privileged/task.toml
index 3f81d4bc34..8198500d34 100644
--- a/tasks/robofunolympics-catch-i00-privileged/task.toml
+++ b/tasks/robofunolympics-catch-i00-privileged/task.toml
@@ -15,6 +15,7 @@ robot = "unitree-h1-2"
family = "catch"
category = "throwing-catching-striking"
mode = "privileged"
+dynamics = "dynamic" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
# scene = environment/instance.json (the family's parameters + its t = 0 state hash); identity = the dataset.toml digest
instance_seed = 0
expert_time_estimate_hours = 8.0
diff --git a/tasks/robofunolympics-catch-i00-standard/task.toml b/tasks/robofunolympics-catch-i00-standard/task.toml
index 880cded6d4..522a7ade4b 100644
--- a/tasks/robofunolympics-catch-i00-standard/task.toml
+++ b/tasks/robofunolympics-catch-i00-standard/task.toml
@@ -26,6 +26,7 @@ robot = "unitree-h1-2"
family = "catch"
category = "throwing-catching-striking"
mode = "standard"
+dynamics = "dynamic" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
instance_seed = 0
reset_budget = 50
diff --git a/tasks/robofunolympics-clean-and-jerk-i00-privileged/task.toml b/tasks/robofunolympics-clean-and-jerk-i00-privileged/task.toml
index d90942945a..3d05692118 100644
--- a/tasks/robofunolympics-clean-and-jerk-i00-privileged/task.toml
+++ b/tasks/robofunolympics-clean-and-jerk-i00-privileged/task.toml
@@ -15,6 +15,7 @@ robot = "unitree-h1-2"
family = "clean_and_jerk"
category = "parkour-apparatus-balance"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
# scene = environment/instance.json (the family's parameters + its t = 0 state hash); identity = the dataset.toml digest
instance_seed = 0
expert_time_estimate_hours = 8.0
diff --git a/tasks/robofunolympics-clean-and-jerk-i00-standard/task.toml b/tasks/robofunolympics-clean-and-jerk-i00-standard/task.toml
index 294a9fe1d9..61e7f1cb03 100644
--- a/tasks/robofunolympics-clean-and-jerk-i00-standard/task.toml
+++ b/tasks/robofunolympics-clean-and-jerk-i00-standard/task.toml
@@ -26,6 +26,7 @@ robot = "unitree-h1-2"
family = "clean_and_jerk"
category = "parkour-apparatus-balance"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
instance_seed = 0
reset_budget = 50
diff --git a/tasks/robofunolympics-crawl-i00-privileged/task.toml b/tasks/robofunolympics-crawl-i00-privileged/task.toml
index 1160174a31..b5c4797836 100644
--- a/tasks/robofunolympics-crawl-i00-privileged/task.toml
+++ b/tasks/robofunolympics-crawl-i00-privileged/task.toml
@@ -15,6 +15,7 @@ robot = "unitree-g1"
family = "crawl"
category = "parkour-apparatus-balance"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
# scene = environment/instance.json (the family's parameters + its t = 0 state hash); identity = the dataset.toml digest
instance_seed = 0
expert_time_estimate_hours = 8.0
diff --git a/tasks/robofunolympics-crawl-i00-standard/task.toml b/tasks/robofunolympics-crawl-i00-standard/task.toml
index b6b2f79aa8..b479b9fd70 100644
--- a/tasks/robofunolympics-crawl-i00-standard/task.toml
+++ b/tasks/robofunolympics-crawl-i00-standard/task.toml
@@ -26,6 +26,7 @@ robot = "unitree-g1"
family = "crawl"
category = "parkour-apparatus-balance"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
instance_seed = 0
reset_budget = 50
diff --git a/tasks/robofunolympics-discus-i00-privileged/task.toml b/tasks/robofunolympics-discus-i00-privileged/task.toml
index 855a84e807..33ae38c466 100644
--- a/tasks/robofunolympics-discus-i00-privileged/task.toml
+++ b/tasks/robofunolympics-discus-i00-privileged/task.toml
@@ -15,6 +15,7 @@ robot = "unitree-h1-2"
family = "discus"
category = "throwing-catching-striking"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
# scene = environment/instance.json (the family's parameters + its t = 0 state hash); identity = the dataset.toml digest
instance_seed = 0
expert_time_estimate_hours = 8.0
diff --git a/tasks/robofunolympics-discus-i00-standard/task.toml b/tasks/robofunolympics-discus-i00-standard/task.toml
index d239d93230..4bddd80144 100644
--- a/tasks/robofunolympics-discus-i00-standard/task.toml
+++ b/tasks/robofunolympics-discus-i00-standard/task.toml
@@ -26,6 +26,7 @@ robot = "unitree-h1-2"
family = "discus"
category = "throwing-catching-striking"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
instance_seed = 0
reset_budget = 50
diff --git a/tasks/robofunolympics-dribble-slalom-i00-privileged/task.toml b/tasks/robofunolympics-dribble-slalom-i00-privileged/task.toml
index 0019171b73..52fae78129 100644
--- a/tasks/robofunolympics-dribble-slalom-i00-privileged/task.toml
+++ b/tasks/robofunolympics-dribble-slalom-i00-privileged/task.toml
@@ -15,6 +15,7 @@ robot = "unitree-g1"
family = "dribble_slalom"
category = "football-and-volleyball"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
# scene = environment/instance.json (the family's parameters + its t = 0 state hash); identity = the dataset.toml digest
instance_seed = 0
expert_time_estimate_hours = 5.0
diff --git a/tasks/robofunolympics-dribble-slalom-i00-standard/task.toml b/tasks/robofunolympics-dribble-slalom-i00-standard/task.toml
index 2f314eeafe..46bc558fc9 100644
--- a/tasks/robofunolympics-dribble-slalom-i00-standard/task.toml
+++ b/tasks/robofunolympics-dribble-slalom-i00-standard/task.toml
@@ -26,6 +26,7 @@ robot = "unitree-g1"
family = "dribble_slalom"
category = "football-and-volleyball"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
instance_seed = 0
reset_budget = 50
diff --git a/tasks/robofunolympics-dunk-i00-privileged/task.toml b/tasks/robofunolympics-dunk-i00-privileged/task.toml
index 9f365662c2..7ff4d437a3 100644
--- a/tasks/robofunolympics-dunk-i00-privileged/task.toml
+++ b/tasks/robofunolympics-dunk-i00-privileged/task.toml
@@ -15,6 +15,7 @@ robot = "unitree-h1-2"
family = "dunk"
category = "basketball-and-rackets"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
# scene = environment/instance.json (the family's parameters + its t = 0 state hash); identity = the dataset.toml digest
instance_seed = 0
expert_time_estimate_hours = 10.0
diff --git a/tasks/robofunolympics-dunk-i00-standard/task.toml b/tasks/robofunolympics-dunk-i00-standard/task.toml
index f03e2fef00..1cd9dd0e25 100644
--- a/tasks/robofunolympics-dunk-i00-standard/task.toml
+++ b/tasks/robofunolympics-dunk-i00-standard/task.toml
@@ -26,6 +26,7 @@ robot = "unitree-h1-2"
family = "dunk"
category = "basketball-and-rackets"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
instance_seed = 0
reset_budget = 50
diff --git a/tasks/robofunolympics-free-throw-i00-privileged/task.toml b/tasks/robofunolympics-free-throw-i00-privileged/task.toml
index b43b98117d..fc4eb930b4 100644
--- a/tasks/robofunolympics-free-throw-i00-privileged/task.toml
+++ b/tasks/robofunolympics-free-throw-i00-privileged/task.toml
@@ -15,6 +15,7 @@ robot = "unitree-h1-2"
family = "free_throw"
category = "basketball-and-rackets"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
# scene = environment/instance.json (the family's parameters + its t = 0 state hash); identity = the dataset.toml digest
instance_seed = 0
expert_time_estimate_hours = 6.0
diff --git a/tasks/robofunolympics-free-throw-i00-standard/task.toml b/tasks/robofunolympics-free-throw-i00-standard/task.toml
index dfe71586ee..c0e139376f 100644
--- a/tasks/robofunolympics-free-throw-i00-standard/task.toml
+++ b/tasks/robofunolympics-free-throw-i00-standard/task.toml
@@ -26,6 +26,7 @@ robot = "unitree-h1-2"
family = "free_throw"
category = "basketball-and-rackets"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
instance_seed = 0
reset_budget = 50
diff --git a/tasks/robofunolympics-golf-putt-i00-privileged/task.toml b/tasks/robofunolympics-golf-putt-i00-privileged/task.toml
index 0c8997be8b..24f2369cbb 100644
--- a/tasks/robofunolympics-golf-putt-i00-privileged/task.toml
+++ b/tasks/robofunolympics-golf-putt-i00-privileged/task.toml
@@ -15,6 +15,7 @@ robot = "unitree-g1"
family = "golf_putt"
category = "basketball-and-rackets"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
# scene = environment/instance.json (the family's parameters + its t = 0 state hash); identity = the dataset.toml digest
instance_seed = 0
expert_time_estimate_hours = 3.0
diff --git a/tasks/robofunolympics-golf-putt-i00-standard/task.toml b/tasks/robofunolympics-golf-putt-i00-standard/task.toml
index 16cc519d33..42d039b32f 100644
--- a/tasks/robofunolympics-golf-putt-i00-standard/task.toml
+++ b/tasks/robofunolympics-golf-putt-i00-standard/task.toml
@@ -26,6 +26,7 @@ robot = "unitree-g1"
family = "golf_putt"
category = "basketball-and-rackets"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
instance_seed = 0
reset_budget = 50
diff --git a/tasks/robofunolympics-header-i00-privileged/task.toml b/tasks/robofunolympics-header-i00-privileged/task.toml
index 71caa70a5f..cdfb2b1ef8 100644
--- a/tasks/robofunolympics-header-i00-privileged/task.toml
+++ b/tasks/robofunolympics-header-i00-privileged/task.toml
@@ -15,6 +15,7 @@ robot = "unitree-h1-2"
family = "header"
category = "football-and-volleyball"
mode = "privileged"
+dynamics = "dynamic" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
# scene = environment/instance.json (the family's parameters + its t = 0 state hash); identity = the dataset.toml digest
instance_seed = 0
expert_time_estimate_hours = 8.0
diff --git a/tasks/robofunolympics-header-i00-standard/task.toml b/tasks/robofunolympics-header-i00-standard/task.toml
index 93562215ce..e9fc91be4f 100644
--- a/tasks/robofunolympics-header-i00-standard/task.toml
+++ b/tasks/robofunolympics-header-i00-standard/task.toml
@@ -26,6 +26,7 @@ robot = "unitree-h1-2"
family = "header"
category = "football-and-volleyball"
mode = "standard"
+dynamics = "dynamic" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
instance_seed = 0
reset_budget = 50
diff --git a/tasks/robofunolympics-high-bar-swing-i00-privileged/task.toml b/tasks/robofunolympics-high-bar-swing-i00-privileged/task.toml
index 88b55f4010..c7e0bbf3cf 100644
--- a/tasks/robofunolympics-high-bar-swing-i00-privileged/task.toml
+++ b/tasks/robofunolympics-high-bar-swing-i00-privileged/task.toml
@@ -15,6 +15,7 @@ robot = "unitree-g1"
family = "high_bar_swing"
category = "parkour-apparatus-balance"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
# scene = environment/instance.json (the family's parameters + its t = 0 state hash); identity = the dataset.toml digest
instance_seed = 0
expert_time_estimate_hours = 4.0
diff --git a/tasks/robofunolympics-high-bar-swing-i00-standard/task.toml b/tasks/robofunolympics-high-bar-swing-i00-standard/task.toml
index 9308239bb8..bd0ecb7ae9 100644
--- a/tasks/robofunolympics-high-bar-swing-i00-standard/task.toml
+++ b/tasks/robofunolympics-high-bar-swing-i00-standard/task.toml
@@ -26,6 +26,7 @@ robot = "unitree-g1"
family = "high_bar_swing"
category = "parkour-apparatus-balance"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
instance_seed = 0
reset_budget = 50
diff --git a/tasks/robofunolympics-high-jump-i00-privileged/task.toml b/tasks/robofunolympics-high-jump-i00-privileged/task.toml
index f407a7103d..434d160fb2 100644
--- a/tasks/robofunolympics-high-jump-i00-privileged/task.toml
+++ b/tasks/robofunolympics-high-jump-i00-privileged/task.toml
@@ -15,6 +15,7 @@ robot = "unitree-h1-2"
family = "high_jump"
category = "track-and-jumps"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
# scene = environment/instance.json (the family's parameters + its t = 0 state hash); identity = the dataset.toml digest
instance_seed = 0
expert_time_estimate_hours = 8.0
diff --git a/tasks/robofunolympics-high-jump-i00-standard/task.toml b/tasks/robofunolympics-high-jump-i00-standard/task.toml
index 40c0ca323f..219339a882 100644
--- a/tasks/robofunolympics-high-jump-i00-standard/task.toml
+++ b/tasks/robofunolympics-high-jump-i00-standard/task.toml
@@ -26,6 +26,7 @@ robot = "unitree-h1-2"
family = "high_jump"
category = "track-and-jumps"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
instance_seed = 0
reset_budget = 50
diff --git a/tasks/robofunolympics-hurdles-i00-privileged/task.toml b/tasks/robofunolympics-hurdles-i00-privileged/task.toml
index 82f3f80c17..d227bc6604 100644
--- a/tasks/robofunolympics-hurdles-i00-privileged/task.toml
+++ b/tasks/robofunolympics-hurdles-i00-privileged/task.toml
@@ -15,6 +15,7 @@ robot = "unitree-h1-2"
family = "hurdles"
category = "track-and-jumps"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
# scene = environment/instance.json (the family's parameters + its t = 0 state hash); identity = the dataset.toml digest
instance_seed = 0
expert_time_estimate_hours = 8.0
diff --git a/tasks/robofunolympics-hurdles-i00-standard/task.toml b/tasks/robofunolympics-hurdles-i00-standard/task.toml
index d5991fb9bd..5dafe0eda1 100644
--- a/tasks/robofunolympics-hurdles-i00-standard/task.toml
+++ b/tasks/robofunolympics-hurdles-i00-standard/task.toml
@@ -26,6 +26,7 @@ robot = "unitree-h1-2"
family = "hurdles"
category = "track-and-jumps"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
instance_seed = 0
reset_budget = 50
diff --git a/tasks/robofunolympics-javelin-i00-privileged/task.toml b/tasks/robofunolympics-javelin-i00-privileged/task.toml
index 9df558c750..9d396b4764 100644
--- a/tasks/robofunolympics-javelin-i00-privileged/task.toml
+++ b/tasks/robofunolympics-javelin-i00-privileged/task.toml
@@ -15,6 +15,7 @@ robot = "unitree-h1-2"
family = "javelin"
category = "throwing-catching-striking"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
# scene = environment/instance.json (the family's parameters + its t = 0 state hash); identity = the dataset.toml digest
instance_seed = 0
expert_time_estimate_hours = 8.0
diff --git a/tasks/robofunolympics-javelin-i00-standard/task.toml b/tasks/robofunolympics-javelin-i00-standard/task.toml
index 72f107ba39..fb59e9305a 100644
--- a/tasks/robofunolympics-javelin-i00-standard/task.toml
+++ b/tasks/robofunolympics-javelin-i00-standard/task.toml
@@ -26,6 +26,7 @@ robot = "unitree-h1-2"
family = "javelin"
category = "throwing-catching-striking"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
instance_seed = 0
reset_budget = 50
diff --git a/tasks/robofunolympics-juggle-three-i00-privileged/task.toml b/tasks/robofunolympics-juggle-three-i00-privileged/task.toml
index 8a70cf6c95..82ed6213af 100644
--- a/tasks/robofunolympics-juggle-three-i00-privileged/task.toml
+++ b/tasks/robofunolympics-juggle-three-i00-privileged/task.toml
@@ -15,6 +15,7 @@ robot = "unitree-g1"
family = "juggle_three"
category = "throwing-catching-striking"
mode = "privileged"
+dynamics = "dynamic" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
# scene = environment/instance.json (the family's parameters + its t = 0 state hash); identity = the dataset.toml digest
instance_seed = 0
expert_time_estimate_hours = 10.0
diff --git a/tasks/robofunolympics-juggle-three-i00-standard/task.toml b/tasks/robofunolympics-juggle-three-i00-standard/task.toml
index 2d8f8da4c8..3361eade50 100644
--- a/tasks/robofunolympics-juggle-three-i00-standard/task.toml
+++ b/tasks/robofunolympics-juggle-three-i00-standard/task.toml
@@ -26,6 +26,7 @@ robot = "unitree-g1"
family = "juggle_three"
category = "throwing-catching-striking"
mode = "standard"
+dynamics = "dynamic" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
instance_seed = 0
reset_budget = 50
diff --git a/tasks/robofunolympics-juggling-i00-privileged/task.toml b/tasks/robofunolympics-juggling-i00-privileged/task.toml
index ffbde71174..dbfb80816a 100644
--- a/tasks/robofunolympics-juggling-i00-privileged/task.toml
+++ b/tasks/robofunolympics-juggling-i00-privileged/task.toml
@@ -15,6 +15,7 @@ robot = "unitree-g1"
family = "juggling"
category = "football-and-volleyball"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
# scene = environment/instance.json (the family's parameters + its t = 0 state hash); identity = the dataset.toml digest
instance_seed = 0
expert_time_estimate_hours = 8.0
diff --git a/tasks/robofunolympics-juggling-i00-standard/task.toml b/tasks/robofunolympics-juggling-i00-standard/task.toml
index 6994e9d172..2640166068 100644
--- a/tasks/robofunolympics-juggling-i00-standard/task.toml
+++ b/tasks/robofunolympics-juggling-i00-standard/task.toml
@@ -26,6 +26,7 @@ robot = "unitree-g1"
family = "juggling"
category = "football-and-volleyball"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
instance_seed = 0
reset_budget = 50
diff --git a/tasks/robofunolympics-kick-target-i00-privileged/task.toml b/tasks/robofunolympics-kick-target-i00-privileged/task.toml
index 7f09ee4de3..8aeef423bc 100644
--- a/tasks/robofunolympics-kick-target-i00-privileged/task.toml
+++ b/tasks/robofunolympics-kick-target-i00-privileged/task.toml
@@ -15,6 +15,7 @@ robot = "unitree-h1-2"
family = "kick_target"
category = "throwing-catching-striking"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
# scene = environment/instance.json (the family's parameters + its t = 0 state hash); identity = the dataset.toml digest
instance_seed = 0
expert_time_estimate_hours = 8.0
diff --git a/tasks/robofunolympics-kick-target-i00-standard/task.toml b/tasks/robofunolympics-kick-target-i00-standard/task.toml
index 1e00e38396..6c279bb4ad 100644
--- a/tasks/robofunolympics-kick-target-i00-standard/task.toml
+++ b/tasks/robofunolympics-kick-target-i00-standard/task.toml
@@ -26,6 +26,7 @@ robot = "unitree-h1-2"
family = "kick_target"
category = "throwing-catching-striking"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
instance_seed = 0
reset_budget = 50
diff --git a/tasks/robofunolympics-layup-i00-privileged/task.toml b/tasks/robofunolympics-layup-i00-privileged/task.toml
index 28a6cd7c98..ff643558a0 100644
--- a/tasks/robofunolympics-layup-i00-privileged/task.toml
+++ b/tasks/robofunolympics-layup-i00-privileged/task.toml
@@ -15,6 +15,7 @@ robot = "unitree-h1-2"
family = "layup"
category = "basketball-and-rackets"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
# scene = environment/instance.json (the family's parameters + its t = 0 state hash); identity = the dataset.toml digest
instance_seed = 0
expert_time_estimate_hours = 8.0
diff --git a/tasks/robofunolympics-layup-i00-standard/task.toml b/tasks/robofunolympics-layup-i00-standard/task.toml
index 575597429f..0b75b34998 100644
--- a/tasks/robofunolympics-layup-i00-standard/task.toml
+++ b/tasks/robofunolympics-layup-i00-standard/task.toml
@@ -26,6 +26,7 @@ robot = "unitree-h1-2"
family = "layup"
category = "basketball-and-rackets"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
instance_seed = 0
reset_budget = 50
diff --git a/tasks/robofunolympics-long-jump-i00-privileged/task.toml b/tasks/robofunolympics-long-jump-i00-privileged/task.toml
index b2e9613640..f95c6f5adb 100644
--- a/tasks/robofunolympics-long-jump-i00-privileged/task.toml
+++ b/tasks/robofunolympics-long-jump-i00-privileged/task.toml
@@ -15,6 +15,7 @@ robot = "unitree-g1"
family = "long_jump"
category = "track-and-jumps"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
# scene = environment/instance.json (the family's parameters + its t = 0 state hash); identity = the dataset.toml digest
instance_seed = 0
expert_time_estimate_hours = 8.0
diff --git a/tasks/robofunolympics-long-jump-i00-standard/task.toml b/tasks/robofunolympics-long-jump-i00-standard/task.toml
index 30c232836d..d1faf731c8 100644
--- a/tasks/robofunolympics-long-jump-i00-standard/task.toml
+++ b/tasks/robofunolympics-long-jump-i00-standard/task.toml
@@ -26,6 +26,7 @@ robot = "unitree-g1"
family = "long_jump"
category = "track-and-jumps"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
instance_seed = 0
reset_budget = 50
diff --git a/tasks/robofunolympics-paddle-bounce-i00-privileged/task.toml b/tasks/robofunolympics-paddle-bounce-i00-privileged/task.toml
index 6921d50936..4a76631d9f 100644
--- a/tasks/robofunolympics-paddle-bounce-i00-privileged/task.toml
+++ b/tasks/robofunolympics-paddle-bounce-i00-privileged/task.toml
@@ -15,6 +15,7 @@ robot = "unitree-g1"
family = "paddle_bounce"
category = "basketball-and-rackets"
mode = "privileged"
+dynamics = "dynamic" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
# scene = environment/instance.json (the family's parameters + its t = 0 state hash); identity = the dataset.toml digest
instance_seed = 0
expert_time_estimate_hours = 4.0
diff --git a/tasks/robofunolympics-paddle-bounce-i00-standard/task.toml b/tasks/robofunolympics-paddle-bounce-i00-standard/task.toml
index 059b5edaaa..6aa4393567 100644
--- a/tasks/robofunolympics-paddle-bounce-i00-standard/task.toml
+++ b/tasks/robofunolympics-paddle-bounce-i00-standard/task.toml
@@ -26,6 +26,7 @@ robot = "unitree-g1"
family = "paddle_bounce"
category = "basketball-and-rackets"
mode = "standard"
+dynamics = "dynamic" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
instance_seed = 0
reset_budget = 50
diff --git a/tasks/robofunolympics-penalty-kick-i00-privileged/task.toml b/tasks/robofunolympics-penalty-kick-i00-privileged/task.toml
index 8d093eb6d2..1af0d88dd1 100644
--- a/tasks/robofunolympics-penalty-kick-i00-privileged/task.toml
+++ b/tasks/robofunolympics-penalty-kick-i00-privileged/task.toml
@@ -15,6 +15,7 @@ robot = "unitree-g1"
family = "penalty_kick"
category = "football-and-volleyball"
mode = "privileged"
+dynamics = "dynamic" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
# scene = environment/instance.json (the family's parameters + its t = 0 state hash); identity = the dataset.toml digest
instance_seed = 0
expert_time_estimate_hours = 8.0
diff --git a/tasks/robofunolympics-penalty-kick-i00-standard/task.toml b/tasks/robofunolympics-penalty-kick-i00-standard/task.toml
index adae2f75fb..a4514a9a29 100644
--- a/tasks/robofunolympics-penalty-kick-i00-standard/task.toml
+++ b/tasks/robofunolympics-penalty-kick-i00-standard/task.toml
@@ -26,6 +26,7 @@ robot = "unitree-g1"
family = "penalty_kick"
category = "football-and-volleyball"
mode = "standard"
+dynamics = "dynamic" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
instance_seed = 0
reset_budget = 50
diff --git a/tasks/robofunolympics-pitch-target-i00-privileged/task.toml b/tasks/robofunolympics-pitch-target-i00-privileged/task.toml
index 6aa7d095c9..3f4655d037 100644
--- a/tasks/robofunolympics-pitch-target-i00-privileged/task.toml
+++ b/tasks/robofunolympics-pitch-target-i00-privileged/task.toml
@@ -15,6 +15,7 @@ robot = "unitree-h1-2"
family = "pitch_target"
category = "throwing-catching-striking"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
# scene = environment/instance.json (the family's parameters + its t = 0 state hash); identity = the dataset.toml digest
instance_seed = 0
expert_time_estimate_hours = 8.0
diff --git a/tasks/robofunolympics-pitch-target-i00-standard/task.toml b/tasks/robofunolympics-pitch-target-i00-standard/task.toml
index f6ab1cc676..678e2aeee5 100644
--- a/tasks/robofunolympics-pitch-target-i00-standard/task.toml
+++ b/tasks/robofunolympics-pitch-target-i00-standard/task.toml
@@ -26,6 +26,7 @@ robot = "unitree-h1-2"
family = "pitch_target"
category = "throwing-catching-striking"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
instance_seed = 0
reset_budget = 50
diff --git a/tasks/robofunolympics-pole-vault-i00-privileged/task.toml b/tasks/robofunolympics-pole-vault-i00-privileged/task.toml
index 908169c612..eee9d1378c 100644
--- a/tasks/robofunolympics-pole-vault-i00-privileged/task.toml
+++ b/tasks/robofunolympics-pole-vault-i00-privileged/task.toml
@@ -15,6 +15,7 @@ robot = "unitree-h1-2"
family = "pole_vault"
category = "track-and-jumps"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
# scene = environment/instance.json (the family's parameters + its t = 0 state hash); identity = the dataset.toml digest
instance_seed = 0
expert_time_estimate_hours = 10.0
diff --git a/tasks/robofunolympics-pole-vault-i00-standard/task.toml b/tasks/robofunolympics-pole-vault-i00-standard/task.toml
index a3f9eb9e7a..d8feb5bc3d 100644
--- a/tasks/robofunolympics-pole-vault-i00-standard/task.toml
+++ b/tasks/robofunolympics-pole-vault-i00-standard/task.toml
@@ -26,6 +26,7 @@ robot = "unitree-h1-2"
family = "pole_vault"
category = "track-and-jumps"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
instance_seed = 0
reset_budget = 50
diff --git a/tasks/robofunolympics-shot-put-i00-privileged/task.toml b/tasks/robofunolympics-shot-put-i00-privileged/task.toml
index 24e4dc4d0a..e4198b5142 100644
--- a/tasks/robofunolympics-shot-put-i00-privileged/task.toml
+++ b/tasks/robofunolympics-shot-put-i00-privileged/task.toml
@@ -15,6 +15,7 @@ robot = "unitree-h1-2"
family = "shot_put"
category = "throwing-catching-striking"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
# scene = environment/instance.json (the family's parameters + its t = 0 state hash); identity = the dataset.toml digest
instance_seed = 0
expert_time_estimate_hours = 8.0
diff --git a/tasks/robofunolympics-shot-put-i00-standard/task.toml b/tasks/robofunolympics-shot-put-i00-standard/task.toml
index 8824476662..4bc9a58c88 100644
--- a/tasks/robofunolympics-shot-put-i00-standard/task.toml
+++ b/tasks/robofunolympics-shot-put-i00-standard/task.toml
@@ -26,6 +26,7 @@ robot = "unitree-h1-2"
family = "shot_put"
category = "throwing-catching-striking"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
instance_seed = 0
reset_budget = 50
diff --git a/tasks/robofunolympics-shuttle-run-i00-privileged/task.toml b/tasks/robofunolympics-shuttle-run-i00-privileged/task.toml
index 27ea020e4b..2a99019ebf 100644
--- a/tasks/robofunolympics-shuttle-run-i00-privileged/task.toml
+++ b/tasks/robofunolympics-shuttle-run-i00-privileged/task.toml
@@ -15,6 +15,7 @@ robot = "unitree-g1"
family = "shuttle_run"
category = "track-and-jumps"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
# scene = environment/instance.json (the family's parameters + its t = 0 state hash); identity = the dataset.toml digest
instance_seed = 0
expert_time_estimate_hours = 4.0
diff --git a/tasks/robofunolympics-shuttle-run-i00-standard/task.toml b/tasks/robofunolympics-shuttle-run-i00-standard/task.toml
index d310418dfb..c6ed71be0e 100644
--- a/tasks/robofunolympics-shuttle-run-i00-standard/task.toml
+++ b/tasks/robofunolympics-shuttle-run-i00-standard/task.toml
@@ -26,6 +26,7 @@ robot = "unitree-g1"
family = "shuttle_run"
category = "track-and-jumps"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
instance_seed = 0
reset_budget = 50
diff --git a/tasks/robofunolympics-slalom-i00-privileged/task.toml b/tasks/robofunolympics-slalom-i00-privileged/task.toml
index 571bdd8463..dbc92a4102 100644
--- a/tasks/robofunolympics-slalom-i00-privileged/task.toml
+++ b/tasks/robofunolympics-slalom-i00-privileged/task.toml
@@ -15,6 +15,7 @@ robot = "unitree-g1"
family = "slalom"
category = "track-and-jumps"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
# scene = environment/instance.json (the family's parameters + its t = 0 state hash); identity = the dataset.toml digest
instance_seed = 0
expert_time_estimate_hours = 5.0
diff --git a/tasks/robofunolympics-slalom-i00-standard/task.toml b/tasks/robofunolympics-slalom-i00-standard/task.toml
index 0b6210cad4..968712a7da 100644
--- a/tasks/robofunolympics-slalom-i00-standard/task.toml
+++ b/tasks/robofunolympics-slalom-i00-standard/task.toml
@@ -26,6 +26,7 @@ robot = "unitree-g1"
family = "slalom"
category = "track-and-jumps"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
instance_seed = 0
reset_budget = 50
diff --git a/tasks/robofunolympics-sprint-i00-privileged/task.toml b/tasks/robofunolympics-sprint-i00-privileged/task.toml
index 4c92497c05..4f89c454ca 100644
--- a/tasks/robofunolympics-sprint-i00-privileged/task.toml
+++ b/tasks/robofunolympics-sprint-i00-privileged/task.toml
@@ -15,6 +15,7 @@ robot = "unitree-g1"
family = "sprint"
category = "track-and-jumps"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
# scene = environment/instance.json (the family's parameters + its t = 0 state hash); identity = the dataset.toml digest
instance_seed = 0
expert_time_estimate_hours = 2.0
diff --git a/tasks/robofunolympics-sprint-i00-standard/task.toml b/tasks/robofunolympics-sprint-i00-standard/task.toml
index cad46ab922..332bb8737d 100644
--- a/tasks/robofunolympics-sprint-i00-standard/task.toml
+++ b/tasks/robofunolympics-sprint-i00-standard/task.toml
@@ -26,6 +26,7 @@ robot = "unitree-g1"
family = "sprint"
category = "track-and-jumps"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
instance_seed = 0
reset_budget = 50
diff --git a/tasks/robofunolympics-standing-long-jump-i00-privileged/task.toml b/tasks/robofunolympics-standing-long-jump-i00-privileged/task.toml
index 99b7de4b77..6413e77eab 100644
--- a/tasks/robofunolympics-standing-long-jump-i00-privileged/task.toml
+++ b/tasks/robofunolympics-standing-long-jump-i00-privileged/task.toml
@@ -15,6 +15,7 @@ robot = "unitree-g1"
family = "standing_long_jump"
category = "track-and-jumps"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
# scene = environment/instance.json (the family's parameters + its t = 0 state hash); identity = the dataset.toml digest
instance_seed = 0
expert_time_estimate_hours = 8.0
diff --git a/tasks/robofunolympics-standing-long-jump-i00-standard/task.toml b/tasks/robofunolympics-standing-long-jump-i00-standard/task.toml
index d912b8180d..cfc4e16b25 100644
--- a/tasks/robofunolympics-standing-long-jump-i00-standard/task.toml
+++ b/tasks/robofunolympics-standing-long-jump-i00-standard/task.toml
@@ -26,6 +26,7 @@ robot = "unitree-g1"
family = "standing_long_jump"
category = "track-and-jumps"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
instance_seed = 0
reset_budget = 50
diff --git a/tasks/robofunolympics-steeplechase-i00-privileged/task.toml b/tasks/robofunolympics-steeplechase-i00-privileged/task.toml
index 3ebfac9f56..005e3740bf 100644
--- a/tasks/robofunolympics-steeplechase-i00-privileged/task.toml
+++ b/tasks/robofunolympics-steeplechase-i00-privileged/task.toml
@@ -15,6 +15,7 @@ robot = "unitree-h1-2"
family = "steeplechase"
category = "parkour-apparatus-balance"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
# scene = environment/instance.json (the family's parameters + its t = 0 state hash); identity = the dataset.toml digest
instance_seed = 0
expert_time_estimate_hours = 8.0
diff --git a/tasks/robofunolympics-steeplechase-i00-standard/task.toml b/tasks/robofunolympics-steeplechase-i00-standard/task.toml
index ce8f2e3f3d..19fdc343ea 100644
--- a/tasks/robofunolympics-steeplechase-i00-standard/task.toml
+++ b/tasks/robofunolympics-steeplechase-i00-standard/task.toml
@@ -26,6 +26,7 @@ robot = "unitree-h1-2"
family = "steeplechase"
category = "parkour-apparatus-balance"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
instance_seed = 0
reset_budget = 50
diff --git a/tasks/robofunolympics-table-tennis-return-i00-privileged/task.toml b/tasks/robofunolympics-table-tennis-return-i00-privileged/task.toml
index 7e840d6972..abd777fb15 100644
--- a/tasks/robofunolympics-table-tennis-return-i00-privileged/task.toml
+++ b/tasks/robofunolympics-table-tennis-return-i00-privileged/task.toml
@@ -15,6 +15,7 @@ robot = "unitree-g1"
family = "table_tennis_return"
category = "basketball-and-rackets"
mode = "privileged"
+dynamics = "dynamic" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
# scene = environment/instance.json (the family's parameters + its t = 0 state hash); identity = the dataset.toml digest
instance_seed = 0
expert_time_estimate_hours = 8.0
diff --git a/tasks/robofunolympics-table-tennis-return-i00-standard/task.toml b/tasks/robofunolympics-table-tennis-return-i00-standard/task.toml
index 9b7a94dfb9..7dbc0dc1d5 100644
--- a/tasks/robofunolympics-table-tennis-return-i00-standard/task.toml
+++ b/tasks/robofunolympics-table-tennis-return-i00-standard/task.toml
@@ -26,6 +26,7 @@ robot = "unitree-g1"
family = "table_tennis_return"
category = "basketball-and-rackets"
mode = "standard"
+dynamics = "dynamic" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
instance_seed = 0
reset_budget = 50
diff --git a/tasks/robofunolympics-tennis-return-i00-privileged/task.toml b/tasks/robofunolympics-tennis-return-i00-privileged/task.toml
index 55e5c8ee0e..66a210bcdd 100644
--- a/tasks/robofunolympics-tennis-return-i00-privileged/task.toml
+++ b/tasks/robofunolympics-tennis-return-i00-privileged/task.toml
@@ -15,6 +15,7 @@ robot = "unitree-h1-2"
family = "tennis_return"
category = "basketball-and-rackets"
mode = "privileged"
+dynamics = "dynamic" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
# scene = environment/instance.json (the family's parameters + its t = 0 state hash); identity = the dataset.toml digest
instance_seed = 0
expert_time_estimate_hours = 8.0
diff --git a/tasks/robofunolympics-tennis-return-i00-standard/task.toml b/tasks/robofunolympics-tennis-return-i00-standard/task.toml
index 5227a86caa..74bc272d1b 100644
--- a/tasks/robofunolympics-tennis-return-i00-standard/task.toml
+++ b/tasks/robofunolympics-tennis-return-i00-standard/task.toml
@@ -26,6 +26,7 @@ robot = "unitree-h1-2"
family = "tennis_return"
category = "basketball-and-rackets"
mode = "standard"
+dynamics = "dynamic" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
instance_seed = 0
reset_budget = 50
diff --git a/tasks/robofunolympics-tennis-serve-i00-privileged/task.toml b/tasks/robofunolympics-tennis-serve-i00-privileged/task.toml
index 5a5273388c..6d504e58c1 100644
--- a/tasks/robofunolympics-tennis-serve-i00-privileged/task.toml
+++ b/tasks/robofunolympics-tennis-serve-i00-privileged/task.toml
@@ -15,6 +15,7 @@ robot = "unitree-h1-2"
family = "tennis_serve"
category = "basketball-and-rackets"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
# scene = environment/instance.json (the family's parameters + its t = 0 state hash); identity = the dataset.toml digest
instance_seed = 0
expert_time_estimate_hours = 8.0
diff --git a/tasks/robofunolympics-tennis-serve-i00-standard/task.toml b/tasks/robofunolympics-tennis-serve-i00-standard/task.toml
index 4f8cd987ec..a847efba44 100644
--- a/tasks/robofunolympics-tennis-serve-i00-standard/task.toml
+++ b/tasks/robofunolympics-tennis-serve-i00-standard/task.toml
@@ -26,6 +26,7 @@ robot = "unitree-h1-2"
family = "tennis_serve"
category = "basketball-and-rackets"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
instance_seed = 0
reset_budget = 50
diff --git a/tasks/robofunolympics-toes-to-bar-i00-privileged/task.toml b/tasks/robofunolympics-toes-to-bar-i00-privileged/task.toml
index 9aada2ef3e..0494030309 100644
--- a/tasks/robofunolympics-toes-to-bar-i00-privileged/task.toml
+++ b/tasks/robofunolympics-toes-to-bar-i00-privileged/task.toml
@@ -15,6 +15,7 @@ robot = "unitree-g1"
family = "toes_to_bar"
category = "parkour-apparatus-balance"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
# scene = environment/instance.json (the family's parameters + its t = 0 state hash); identity = the dataset.toml digest
instance_seed = 0
expert_time_estimate_hours = 3.0
diff --git a/tasks/robofunolympics-toes-to-bar-i00-standard/task.toml b/tasks/robofunolympics-toes-to-bar-i00-standard/task.toml
index 082a6d52e0..b4e7706f16 100644
--- a/tasks/robofunolympics-toes-to-bar-i00-standard/task.toml
+++ b/tasks/robofunolympics-toes-to-bar-i00-standard/task.toml
@@ -26,6 +26,7 @@ robot = "unitree-g1"
family = "toes_to_bar"
category = "parkour-apparatus-balance"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
instance_seed = 0
reset_budget = 50
diff --git a/tasks/robofunolympics-trap-ball-i00-privileged/task.toml b/tasks/robofunolympics-trap-ball-i00-privileged/task.toml
index 4a2b7c192e..316b67b6de 100644
--- a/tasks/robofunolympics-trap-ball-i00-privileged/task.toml
+++ b/tasks/robofunolympics-trap-ball-i00-privileged/task.toml
@@ -15,6 +15,7 @@ robot = "unitree-h1-2"
family = "trap_ball"
category = "football-and-volleyball"
mode = "privileged"
+dynamics = "dynamic" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
# scene = environment/instance.json (the family's parameters + its t = 0 state hash); identity = the dataset.toml digest
instance_seed = 0
expert_time_estimate_hours = 5.0
diff --git a/tasks/robofunolympics-trap-ball-i00-standard/task.toml b/tasks/robofunolympics-trap-ball-i00-standard/task.toml
index 2cbb87d3a5..032a312c88 100644
--- a/tasks/robofunolympics-trap-ball-i00-standard/task.toml
+++ b/tasks/robofunolympics-trap-ball-i00-standard/task.toml
@@ -26,6 +26,7 @@ robot = "unitree-h1-2"
family = "trap_ball"
category = "football-and-volleyball"
mode = "standard"
+dynamics = "dynamic" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
instance_seed = 0
reset_budget = 50
diff --git a/tasks/robofunolympics-triple-jump-i00-privileged/task.toml b/tasks/robofunolympics-triple-jump-i00-privileged/task.toml
index 7849b14ba2..69be8e2e86 100644
--- a/tasks/robofunolympics-triple-jump-i00-privileged/task.toml
+++ b/tasks/robofunolympics-triple-jump-i00-privileged/task.toml
@@ -15,6 +15,7 @@ robot = "unitree-g1"
family = "triple_jump"
category = "track-and-jumps"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
# scene = environment/instance.json (the family's parameters + its t = 0 state hash); identity = the dataset.toml digest
instance_seed = 0
expert_time_estimate_hours = 8.0
diff --git a/tasks/robofunolympics-triple-jump-i00-standard/task.toml b/tasks/robofunolympics-triple-jump-i00-standard/task.toml
index 861a1b97e8..3c45264382 100644
--- a/tasks/robofunolympics-triple-jump-i00-standard/task.toml
+++ b/tasks/robofunolympics-triple-jump-i00-standard/task.toml
@@ -26,6 +26,7 @@ robot = "unitree-g1"
family = "triple_jump"
category = "track-and-jumps"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
instance_seed = 0
reset_budget = 50
diff --git a/tasks/robofunolympics-vault-i00-privileged/task.toml b/tasks/robofunolympics-vault-i00-privileged/task.toml
index ef44b98401..db9581a861 100644
--- a/tasks/robofunolympics-vault-i00-privileged/task.toml
+++ b/tasks/robofunolympics-vault-i00-privileged/task.toml
@@ -15,6 +15,7 @@ robot = "unitree-h1-2"
family = "vault"
category = "parkour-apparatus-balance"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
# scene = environment/instance.json (the family's parameters + its t = 0 state hash); identity = the dataset.toml digest
instance_seed = 0
expert_time_estimate_hours = 8.0
diff --git a/tasks/robofunolympics-vault-i00-standard/task.toml b/tasks/robofunolympics-vault-i00-standard/task.toml
index ed596d0ef1..22839f0ec7 100644
--- a/tasks/robofunolympics-vault-i00-standard/task.toml
+++ b/tasks/robofunolympics-vault-i00-standard/task.toml
@@ -26,6 +26,7 @@ robot = "unitree-h1-2"
family = "vault"
category = "parkour-apparatus-balance"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
instance_seed = 0
reset_budget = 50
diff --git a/tasks/robofunolympics-vertical-jump-i00-privileged/task.toml b/tasks/robofunolympics-vertical-jump-i00-privileged/task.toml
index 83365b5bed..d74aaacfdf 100644
--- a/tasks/robofunolympics-vertical-jump-i00-privileged/task.toml
+++ b/tasks/robofunolympics-vertical-jump-i00-privileged/task.toml
@@ -15,6 +15,7 @@ robot = "unitree-g1"
family = "vertical_jump"
category = "track-and-jumps"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
# scene = environment/instance.json (the family's parameters + its t = 0 state hash); identity = the dataset.toml digest
instance_seed = 0
expert_time_estimate_hours = 6.0
diff --git a/tasks/robofunolympics-vertical-jump-i00-standard/task.toml b/tasks/robofunolympics-vertical-jump-i00-standard/task.toml
index 7bc93fc517..ba5cd8cdb3 100644
--- a/tasks/robofunolympics-vertical-jump-i00-standard/task.toml
+++ b/tasks/robofunolympics-vertical-jump-i00-standard/task.toml
@@ -26,6 +26,7 @@ robot = "unitree-g1"
family = "vertical_jump"
category = "track-and-jumps"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
instance_seed = 0
reset_budget = 50
diff --git a/tasks/robofunolympics-volley-shot-i00-privileged/task.toml b/tasks/robofunolympics-volley-shot-i00-privileged/task.toml
index 576dac4241..b9ca366641 100644
--- a/tasks/robofunolympics-volley-shot-i00-privileged/task.toml
+++ b/tasks/robofunolympics-volley-shot-i00-privileged/task.toml
@@ -15,6 +15,7 @@ robot = "unitree-h1-2"
family = "volley_shot"
category = "football-and-volleyball"
mode = "privileged"
+dynamics = "dynamic" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
# scene = environment/instance.json (the family's parameters + its t = 0 state hash); identity = the dataset.toml digest
instance_seed = 0
expert_time_estimate_hours = 8.0
diff --git a/tasks/robofunolympics-volley-shot-i00-standard/task.toml b/tasks/robofunolympics-volley-shot-i00-standard/task.toml
index 8074f00ccb..51e4533c5d 100644
--- a/tasks/robofunolympics-volley-shot-i00-standard/task.toml
+++ b/tasks/robofunolympics-volley-shot-i00-standard/task.toml
@@ -26,6 +26,7 @@ robot = "unitree-h1-2"
family = "volley_shot"
category = "football-and-volleyball"
mode = "standard"
+dynamics = "dynamic" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
instance_seed = 0
reset_budget = 50
diff --git a/tasks/robofunolympics-volleyball-bump-i00-privileged/task.toml b/tasks/robofunolympics-volleyball-bump-i00-privileged/task.toml
index ee6a141b1e..6992121224 100644
--- a/tasks/robofunolympics-volleyball-bump-i00-privileged/task.toml
+++ b/tasks/robofunolympics-volleyball-bump-i00-privileged/task.toml
@@ -15,6 +15,7 @@ robot = "unitree-h1-2"
family = "volleyball_bump"
category = "football-and-volleyball"
mode = "privileged"
+dynamics = "dynamic" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
# scene = environment/instance.json (the family's parameters + its t = 0 state hash); identity = the dataset.toml digest
instance_seed = 0
expert_time_estimate_hours = 6.0
diff --git a/tasks/robofunolympics-volleyball-bump-i00-standard/task.toml b/tasks/robofunolympics-volleyball-bump-i00-standard/task.toml
index b2139dc68e..1dce972f5d 100644
--- a/tasks/robofunolympics-volleyball-bump-i00-standard/task.toml
+++ b/tasks/robofunolympics-volleyball-bump-i00-standard/task.toml
@@ -26,6 +26,7 @@ robot = "unitree-h1-2"
family = "volleyball_bump"
category = "football-and-volleyball"
mode = "standard"
+dynamics = "dynamic" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
instance_seed = 0
reset_budget = 50
diff --git a/tasks/robofunolympics-volleyball-serve-i00-privileged/task.toml b/tasks/robofunolympics-volleyball-serve-i00-privileged/task.toml
index 7b551ad641..44f23619b4 100644
--- a/tasks/robofunolympics-volleyball-serve-i00-privileged/task.toml
+++ b/tasks/robofunolympics-volleyball-serve-i00-privileged/task.toml
@@ -15,6 +15,7 @@ robot = "unitree-h1-2"
family = "volleyball_serve"
category = "football-and-volleyball"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
# scene = environment/instance.json (the family's parameters + its t = 0 state hash); identity = the dataset.toml digest
instance_seed = 0
expert_time_estimate_hours = 6.0
diff --git a/tasks/robofunolympics-volleyball-serve-i00-standard/task.toml b/tasks/robofunolympics-volleyball-serve-i00-standard/task.toml
index bd7d7273f7..5ed44e98d2 100644
--- a/tasks/robofunolympics-volleyball-serve-i00-standard/task.toml
+++ b/tasks/robofunolympics-volleyball-serve-i00-standard/task.toml
@@ -26,6 +26,7 @@ robot = "unitree-h1-2"
family = "volleyball_serve"
category = "football-and-volleyball"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
instance_seed = 0
reset_budget = 50
diff --git a/tasks/robofunolympics-volleyball-set-i00-privileged/task.toml b/tasks/robofunolympics-volleyball-set-i00-privileged/task.toml
index f4f5cd3933..3b7de651f2 100644
--- a/tasks/robofunolympics-volleyball-set-i00-privileged/task.toml
+++ b/tasks/robofunolympics-volleyball-set-i00-privileged/task.toml
@@ -15,6 +15,7 @@ robot = "unitree-h1-2"
family = "volleyball_set"
category = "football-and-volleyball"
mode = "privileged"
+dynamics = "dynamic" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
# scene = environment/instance.json (the family's parameters + its t = 0 state hash); identity = the dataset.toml digest
instance_seed = 0
expert_time_estimate_hours = 8.0
diff --git a/tasks/robofunolympics-volleyball-set-i00-standard/task.toml b/tasks/robofunolympics-volleyball-set-i00-standard/task.toml
index dc28f84130..d6b3bdd493 100644
--- a/tasks/robofunolympics-volleyball-set-i00-standard/task.toml
+++ b/tasks/robofunolympics-volleyball-set-i00-standard/task.toml
@@ -26,6 +26,7 @@ robot = "unitree-h1-2"
family = "volleyball_set"
category = "football-and-volleyball"
mode = "standard"
+dynamics = "dynamic" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
instance_seed = 0
reset_budget = 50
diff --git a/tasks/robofunolympics-volleyball-spike-i00-privileged/task.toml b/tasks/robofunolympics-volleyball-spike-i00-privileged/task.toml
index 64e016bb18..0ea8262a4c 100644
--- a/tasks/robofunolympics-volleyball-spike-i00-privileged/task.toml
+++ b/tasks/robofunolympics-volleyball-spike-i00-privileged/task.toml
@@ -15,6 +15,7 @@ robot = "unitree-h1-2"
family = "volleyball_spike"
category = "football-and-volleyball"
mode = "privileged"
+dynamics = "dynamic" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
# scene = environment/instance.json (the family's parameters + its t = 0 state hash); identity = the dataset.toml digest
instance_seed = 0
expert_time_estimate_hours = 8.0
diff --git a/tasks/robofunolympics-volleyball-spike-i00-standard/task.toml b/tasks/robofunolympics-volleyball-spike-i00-standard/task.toml
index b4e80e5b8a..62b619ab63 100644
--- a/tasks/robofunolympics-volleyball-spike-i00-standard/task.toml
+++ b/tasks/robofunolympics-volleyball-spike-i00-standard/task.toml
@@ -26,6 +26,7 @@ robot = "unitree-h1-2"
family = "volleyball_spike"
category = "football-and-volleyball"
mode = "standard"
+dynamics = "dynamic" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
instance_seed = 0
reset_budget = 50
diff --git a/tasks/robofunolympics-wall-ball-i00-privileged/task.toml b/tasks/robofunolympics-wall-ball-i00-privileged/task.toml
index f69a3eaffe..abe73e9e5a 100644
--- a/tasks/robofunolympics-wall-ball-i00-privileged/task.toml
+++ b/tasks/robofunolympics-wall-ball-i00-privileged/task.toml
@@ -15,6 +15,7 @@ robot = "unitree-h1-2"
family = "wall_ball"
category = "throwing-catching-striking"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
# scene = environment/instance.json (the family's parameters + its t = 0 state hash); identity = the dataset.toml digest
instance_seed = 0
expert_time_estimate_hours = 8.0
diff --git a/tasks/robofunolympics-wall-ball-i00-standard/task.toml b/tasks/robofunolympics-wall-ball-i00-standard/task.toml
index 6674198449..fcb04e236b 100644
--- a/tasks/robofunolympics-wall-ball-i00-standard/task.toml
+++ b/tasks/robofunolympics-wall-ball-i00-standard/task.toml
@@ -26,6 +26,7 @@ robot = "unitree-h1-2"
family = "wall_ball"
category = "throwing-catching-striking"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
instance_seed = 0
reset_budget = 50
diff --git a/tasks/robolab-apple-and-yogurt-in-bowl-i00-privileged/task.toml b/tasks/robolab-apple-and-yogurt-in-bowl-i00-privileged/task.toml
index d485950d26..a9705de651 100644
--- a/tasks/robolab-apple-and-yogurt-in-bowl-i00-privileged/task.toml
+++ b/tasks/robolab-apple-and-yogurt-in-bowl-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "robolab" # NVIDIA RoboLab v0.3.1 on Isaac Lab 2.3.2 / I
robot = "franka-robotiq-2f85"
family = "apple_and_yogurt_in_bowl"
mode = "privileged" # privileged state, resets and primitives through the tooling (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution: RoboLab ships no expert or planner. Positive examples graded 1 by the separate verifier: robolab-apple-and-yogurt-in-bowl__pZKiavK (robolab-apple-and-yogurt-in-bowl-i00-privileged, run codex-gpt6_luna-xhigh, robolab-lunacx-0925-2038, ws; https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/apple_and_yogurt_in_bowl/unlimited/). Human review per family (PROTOCOL §2.7)."
instance_seed = 0 # scene = environment/instance.json (the full scene state after the seeded reset + its hash); identity = dataset.toml digest
diff --git a/tasks/robolab-apple-and-yogurt-in-bowl-i00-standard/README.md b/tasks/robolab-apple-and-yogurt-in-bowl-i00-standard/README.md
index 9e18e464d3..8c965fe3c7 100644
--- a/tasks/robolab-apple-and-yogurt-in-bowl-i00-standard/README.md
+++ b/tasks/robolab-apple-and-yogurt-in-bowl-i00-standard/README.md
@@ -72,3 +72,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
- 2026-10-10 (review round 3, board 0067): the shared service layer and the client only (`environment/sim/rcb_limited.py`, `environment/rcb_client.py`). A camera named more than once in `cameras` is rendered and returned once (N copies of one name made the service build N x N images within the frame budget, which could exhaust its memory). A service that has given up on the episode (it could not start it, or could not resume it after its crashes) answers every request with 409 and the reason, and `Robot()` / `wait_ready()` raise `EpisodeOver` at once; before, nothing answered and they waited out their timeout (up to an hour). The adapter, the collect hook, the grader, the scene and the pass rule are unchanged.
- 2026-10-10 (owner decision, board 0072): the agent's network egress is restricted to the model APIs (`[agent] network_mode = "allowlist"`, the hosts of `scripts/set_agent_egress.py`; since 2026-10-11 these include the package registries, PyPI and the Debian / Ubuntu archives, so that `pip install` and `apt-get install` work); it was open before (Harbor's default), and agents could download upstream sources, reference solutions and robot models from their own shell. The simulator image is `eai-robolab:0.1.2` (was 0.1.1): it holds the MDL materials and textures that RoboLab's asset library fetched from NVIDIA's Omniverse content bucket whenever a scene was built, so that the cameras show under the allowlist what they showed with open egress (on 0.1.1 without that download the table renders red). The scene, the frozen instance, the physics, the tooling and the pass rule are unchanged, so the task keeps its name; results from before this date were obtained with open egress.
- 2026-10-13 (owner, boards 0085 and 0087): the sim service's clock (`environment/sim/rcb_limited.py`, `environment/sim/server.py`, the docstring of `environment/rcb_client.py`). `meta["time_s"]` and `meta["time_left_s"]` count from the agent's first request to the service (any endpoint), and a restarted service keeps that moment; before, they counted from the service's own start of the episode (when its scene was ready). Where the service is ready before the agent's phase begins (the harness is installed in between: a median of 79 s over the recorded standard trials), that clock ran ahead of the agent's and `time_left_s` reached 0 while the agent still had that time. `/spec` says what the clock counts from (`budget.clock`), and the instruction's `## Time limit` tells the agent to plan by `time_left_s`. The service never ended an episode on this clock: the scene, the frozen instance, the pass rule and the grader are unchanged, so the task keeps its name and earlier results stand.
+- 2026-10-14 (board 0113, from board 0086): the success text says that the gripper must not be touching the objects at the moment of success ("at that moment"); the old wording "the gripper must touch none of them" was read by an agent as "never touch them" (board 0086). Text only: the scene, the frozen instance, RoboLab's success term and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/robolab-apple-and-yogurt-in-bowl-i00-standard/instruction.md b/tasks/robolab-apple-and-yogurt-in-bowl-i00-standard/instruction.md
index 4d6e12789a..6948525108 100644
--- a/tasks/robolab-apple-and-yogurt-in-bowl-i00-standard/instruction.md
+++ b/tasks/robolab-apple-and-yogurt-in-bowl-i00-standard/instruction.md
@@ -1,7 +1,7 @@
# RoboLab · apple_and_yogurt_in_bowl — STANDARD MODE (robot as a service, eai-standard/2.2)
You control a **Franka Panda arm with a Robotiq 2F-85 gripper** (NVIDIA RoboLab's DROID setup) **as if it were real
-hardware**: the simulator runs in a service you cannot open. You get one client and one episode. **Task:** Put the apple and yogurt in the bowl. Success: The red apple and the small yogurt cup must end inside the wooden bowl and the gripper must touch neither. "Inside" means the centre of the object's convex hull lies within the container's convex hull, extended without limit along the container's own up axis (the open side; straight up for a container standing normally).
+hardware**: the simulator runs in a service you cannot open. You get one client and one episode. **Task:** Put the apple and yogurt in the bowl. Success: The red apple and the small yogurt cup must end inside the wooden bowl and, at that moment, the gripper must touch neither (let go of both at the end). "Inside" means the centre of the object's convex hull lies within the container's convex hull, extended without limit along the container's own up axis (the open side; straight up for a container standing normally).
## Rules (real-robot rules)
- **No reset.** The episode starts when the service comes up and ends on success (at the first control step where the
diff --git a/tasks/robolab-apple-and-yogurt-in-bowl-i00-standard/task.toml b/tasks/robolab-apple-and-yogurt-in-bowl-i00-standard/task.toml
index fee50fe3ca..e9a8d0c4eb 100644
--- a/tasks/robolab-apple-and-yogurt-in-bowl-i00-standard/task.toml
+++ b/tasks/robolab-apple-and-yogurt-in-bowl-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "robolab"
robot = "franka-robotiq-2f85"
family = "apple_and_yogurt_in_bowl"
mode = "standard" # robot as a service: RGB, depth + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution: RoboLab ships no expert or planner. Positive examples graded 1 by the separate verifier: robolab-apple-and-yogurt-in-bowl__pZKiavK (robolab-apple-and-yogurt-in-bowl-i00-privileged, run codex-gpt6_luna-xhigh, robolab-lunacx-0925-2038, ws; https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/apple_and_yogurt_in_bowl/unlimited/). Human review per family (PROTOCOL §2.7)."
diff --git a/tasks/robolab-banana-in-bowl-i00-privileged/task.toml b/tasks/robolab-banana-in-bowl-i00-privileged/task.toml
index 2dfcfcc0e9..5aacf8e4c6 100644
--- a/tasks/robolab-banana-in-bowl-i00-privileged/task.toml
+++ b/tasks/robolab-banana-in-bowl-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "robolab" # NVIDIA RoboLab v0.3.1 on Isaac Lab 2.3.2 / I
robot = "franka-robotiq-2f85"
family = "banana_in_bowl"
mode = "privileged" # privileged state, resets and primitives through the tooling (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
oracle_evidence = "The oracle scored 1 (deterministic, instance_exact) in Harbor runs robolab-banana-in-bowl-i00__SQLoScr (laptop, job tpl11-0927-1823-oracle-robolab-banana-in-bowl-i00, at eaf2aff0), robolab-banana-in-bowl-i00__CLgmEWt (ws3, job tpl19-0927-0430-oracle-robolab-banana-in-bowl-i00, at fe045825), and robolab-banana-in-bowl-i00__xi9pT2B, robolab-banana-in-bowl-i00__WxusWvq (laptop, 2026-09-28, the 2.2 tooling, uncommitted trees), on a byte-identical tests/grade.py, frozen instance and oracle code (tests/test.sh differs only in the command that empties /logs/verifier before grading: rm -rf, the protocol v1.0 convention); the tooling (rcb_robolab.py) changed since only off the oracle and replay path: eaf2aff0 adds other families' registry entries; b3c7ca89 adds fail-safe 2.2 helpers and was re-checked through the same tooling path (the banana_in_bowl oracle trajectory ends in the same final state sha 61afa384566be3a5 with the old and the new CLI); cde31660 flips the depth default, which run-oracle and replay pass explicitly (depth=False); one docstring line now says standard mode."
instance_seed = 0 # scene = environment/instance.json (the full scene state after the seeded reset + its hash); identity = dataset.toml digest
diff --git a/tasks/robolab-banana-in-bowl-i00-standard/task.toml b/tasks/robolab-banana-in-bowl-i00-standard/task.toml
index 08c05f6271..be6e8d1be8 100644
--- a/tasks/robolab-banana-in-bowl-i00-standard/task.toml
+++ b/tasks/robolab-banana-in-bowl-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "robolab"
robot = "franka-robotiq-2f85"
family = "banana_in_bowl"
mode = "standard" # robot as a service: RGB, depth + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "partial"
oracle_evidence = "solution/oracle_replay.py scored 1 in 6/6 recorded runs on the pre-crash-recovery service with the same verdict formula (success = live, replay and deterministic), the latest robolab-banana-in-bowl-i00-limit__vABaxSX (ws3, job eq-ws3-oracle-banana-in-bowl, 2026-09-29, the dev/qineng 9eb18edf version); not re-run on this version (crash recovery, the collect hook's `end`, one-second chunks): static validation only, 2026-10-01. Human review per family (PROTOCOL §2.7)."
diff --git a/tasks/robolab-banana-then-rubiks-cube-i00-privileged/task.toml b/tasks/robolab-banana-then-rubiks-cube-i00-privileged/task.toml
index 9da014b74b..034d7b40eb 100644
--- a/tasks/robolab-banana-then-rubiks-cube-i00-privileged/task.toml
+++ b/tasks/robolab-banana-then-rubiks-cube-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "robolab" # NVIDIA RoboLab v0.3.1 on Isaac Lab 2.3.2 / I
robot = "franka-robotiq-2f85"
family = "banana_then_rubiks_cube"
mode = "privileged" # privileged state, resets and primitives through the tooling (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution: RoboLab ships no expert or planner. Positive examples graded 1 by the separate verifier: robolab-banana-then-rubiks-cube__QRpikDi (robolab-banana-then-rubiks-cube-i00-privileged, run codex-gpt6_luna-xhigh, robolab-lunacx-0925-2038, ws3; https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/banana_then_rubiks_cube/unlimited/); robolab-banana-then-rubiks-cube__HwFn4uZ (robolab-banana-then-rubiks-cube-i00-standard, run codex-gpt6_luna-xhigh, robolab-lunaxh-limited-0928, lab; https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/banana_then_rubiks_cube/limited/). Human review per family (PROTOCOL §2.7)."
instance_seed = 0 # scene = environment/instance.json (the full scene state after the seeded reset + its hash); identity = dataset.toml digest
diff --git a/tasks/robolab-banana-then-rubiks-cube-i00-standard/README.md b/tasks/robolab-banana-then-rubiks-cube-i00-standard/README.md
index fdb4592925..b057243b57 100644
--- a/tasks/robolab-banana-then-rubiks-cube-i00-standard/README.md
+++ b/tasks/robolab-banana-then-rubiks-cube-i00-standard/README.md
@@ -73,3 +73,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
- 2026-10-10 (review round 3, board 0067): the shared service layer and the client only (`environment/sim/rcb_limited.py`, `environment/rcb_client.py`). A camera named more than once in `cameras` is rendered and returned once (N copies of one name made the service build N x N images within the frame budget, which could exhaust its memory). A service that has given up on the episode (it could not start it, or could not resume it after its crashes) answers every request with 409 and the reason, and `Robot()` / `wait_ready()` raise `EpisodeOver` at once; before, nothing answered and they waited out their timeout (up to an hour). The adapter, the collect hook, the grader, the scene and the pass rule are unchanged.
- 2026-10-10 (owner decision, board 0072): the agent's network egress is restricted to the model APIs (`[agent] network_mode = "allowlist"`, the hosts of `scripts/set_agent_egress.py`; since 2026-10-11 these include the package registries, PyPI and the Debian / Ubuntu archives, so that `pip install` and `apt-get install` work); it was open before (Harbor's default), and agents could download upstream sources, reference solutions and robot models from their own shell. The simulator image is `eai-robolab:0.1.2` (was 0.1.1): it holds the MDL materials and textures that RoboLab's asset library fetched from NVIDIA's Omniverse content bucket whenever a scene was built, so that the cameras show under the allowlist what they showed with open egress (on 0.1.1 without that download the table renders red). The scene, the frozen instance, the physics, the tooling and the pass rule are unchanged, so the task keeps its name; results from before this date were obtained with open egress.
- 2026-10-13 (owner, boards 0085 and 0087): the sim service's clock (`environment/sim/rcb_limited.py`, `environment/sim/server.py`, the docstring of `environment/rcb_client.py`). `meta["time_s"]` and `meta["time_left_s"]` count from the agent's first request to the service (any endpoint), and a restarted service keeps that moment; before, they counted from the service's own start of the episode (when its scene was ready). Where the service is ready before the agent's phase begins (the harness is installed in between: a median of 79 s over the recorded standard trials), that clock ran ahead of the agent's and `time_left_s` reached 0 while the agent still had that time. `/spec` says what the clock counts from (`budget.clock`), and the instruction's `## Time limit` tells the agent to plan by `time_left_s`. The service never ended an episode on this clock: the scene, the frozen instance, the pass rule and the grader are unchanged, so the task keeps its name and earlier results stand.
+- 2026-10-14 (board 0113, from board 0086): the success text says that the gripper must not be touching the objects at the moment of success ("at that moment"); the old wording "the gripper must touch none of them" was read by an agent as "never touch them" (board 0086). Text only: the scene, the frozen instance, RoboLab's success term and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/robolab-banana-then-rubiks-cube-i00-standard/instruction.md b/tasks/robolab-banana-then-rubiks-cube-i00-standard/instruction.md
index abaf20a670..d49fa1abac 100644
--- a/tasks/robolab-banana-then-rubiks-cube-i00-standard/instruction.md
+++ b/tasks/robolab-banana-then-rubiks-cube-i00-standard/instruction.md
@@ -1,7 +1,7 @@
# RoboLab · banana_then_rubiks_cube — STANDARD MODE (robot as a service, eai-standard/2.2)
You control a **Franka Panda arm with a Robotiq 2F-85 gripper** (NVIDIA RoboLab's DROID setup) **as if it were real
-hardware**: the simulator runs in a service you cannot open. You get one client and one episode. **Task:** Put the banana then the cube in the bowl. Success: The yellow banana and then the Rubik's cube must end inside the bowl, both touching it, and the gripper must touch neither. Order counts: the service remembers the first moment each object is placed and a later removal does not erase it, so if the Rubik's cube is placed first the episode can no longer succeed. "Inside" means the centre of the object's convex hull lies within the container's convex hull, extended without limit along the container's own up axis (the open side; straight up for a container standing normally).
+hardware**: the simulator runs in a service you cannot open. You get one client and one episode. **Task:** Put the banana then the cube in the bowl. Success: The yellow banana and then the Rubik's cube must end inside the bowl, both touching it, and, at that moment, the gripper must touch neither (let go of both at the end). Order counts: the service remembers the first moment each object is placed and a later removal does not erase it, so if the Rubik's cube is placed first the episode can no longer succeed. "Inside" means the centre of the object's convex hull lies within the container's convex hull, extended without limit along the container's own up axis (the open side; straight up for a container standing normally).
## Rules (real-robot rules)
- **No reset.** The episode starts when the service comes up and ends on success (at the first control step where the
diff --git a/tasks/robolab-banana-then-rubiks-cube-i00-standard/task.toml b/tasks/robolab-banana-then-rubiks-cube-i00-standard/task.toml
index 16c1c2e8a2..1a3ae8bf6f 100644
--- a/tasks/robolab-banana-then-rubiks-cube-i00-standard/task.toml
+++ b/tasks/robolab-banana-then-rubiks-cube-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "robolab"
robot = "franka-robotiq-2f85"
family = "banana_then_rubiks_cube"
mode = "standard" # robot as a service: RGB, depth + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution: RoboLab ships no expert or planner. Positive examples graded 1 by the separate verifier: robolab-banana-then-rubiks-cube__QRpikDi (robolab-banana-then-rubiks-cube-i00-privileged, run codex-gpt6_luna-xhigh, robolab-lunacx-0925-2038, ws3; https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/banana_then_rubiks_cube/unlimited/); robolab-banana-then-rubiks-cube__HwFn4uZ (robolab-banana-then-rubiks-cube-i00-standard, run codex-gpt6_luna-xhigh, robolab-lunaxh-limited-0928, lab; https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/banana_then_rubiks_cube/limited/). Human review per family (PROTOCOL §2.7)."
diff --git a/tasks/robolab-bananas-in-bin-one-more-i00-privileged/task.toml b/tasks/robolab-bananas-in-bin-one-more-i00-privileged/task.toml
index db7c301873..adea073c80 100644
--- a/tasks/robolab-bananas-in-bin-one-more-i00-privileged/task.toml
+++ b/tasks/robolab-bananas-in-bin-one-more-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "robolab" # NVIDIA RoboLab v0.3.1 on Isaac Lab 2.3.2 / I
robot = "franka-robotiq-2f85"
family = "bananas_in_bin_one_more"
mode = "privileged" # privileged state, resets and primitives through the tooling (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution: RoboLab ships no expert or planner. Positive examples graded 1 by the separate verifier: robolab-bananas-in-bin-one-more__uypUNUt (robolab-bananas-in-bin-one-more-i00-privileged, run codex-gpt6_luna-xhigh, robolab-lunacx-0925-2038, ws; https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/bananas_in_bin_one_more/unlimited/). Human review per family (PROTOCOL §2.7)."
instance_seed = 0 # scene = environment/instance.json (the full scene state after the seeded reset + its hash); identity = dataset.toml digest
diff --git a/tasks/robolab-bananas-in-bin-one-more-i00-standard/README.md b/tasks/robolab-bananas-in-bin-one-more-i00-standard/README.md
index 1a27372e29..63c099d991 100644
--- a/tasks/robolab-bananas-in-bin-one-more-i00-standard/README.md
+++ b/tasks/robolab-bananas-in-bin-one-more-i00-standard/README.md
@@ -72,3 +72,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
- 2026-10-10 (review round 3, board 0067): the shared service layer and the client only (`environment/sim/rcb_limited.py`, `environment/rcb_client.py`). A camera named more than once in `cameras` is rendered and returned once (N copies of one name made the service build N x N images within the frame budget, which could exhaust its memory). A service that has given up on the episode (it could not start it, or could not resume it after its crashes) answers every request with 409 and the reason, and `Robot()` / `wait_ready()` raise `EpisodeOver` at once; before, nothing answered and they waited out their timeout (up to an hour). The adapter, the collect hook, the grader, the scene and the pass rule are unchanged.
- 2026-10-10 (owner decision, board 0072): the agent's network egress is restricted to the model APIs (`[agent] network_mode = "allowlist"`, the hosts of `scripts/set_agent_egress.py`; since 2026-10-11 these include the package registries, PyPI and the Debian / Ubuntu archives, so that `pip install` and `apt-get install` work); it was open before (Harbor's default), and agents could download upstream sources, reference solutions and robot models from their own shell. The simulator image is `eai-robolab:0.1.2` (was 0.1.1): it holds the MDL materials and textures that RoboLab's asset library fetched from NVIDIA's Omniverse content bucket whenever a scene was built, so that the cameras show under the allowlist what they showed with open egress (on 0.1.1 without that download the table renders red). The scene, the frozen instance, the physics, the tooling and the pass rule are unchanged, so the task keeps its name; results from before this date were obtained with open egress.
- 2026-10-13 (owner, boards 0085 and 0087): the sim service's clock (`environment/sim/rcb_limited.py`, `environment/sim/server.py`, the docstring of `environment/rcb_client.py`). `meta["time_s"]` and `meta["time_left_s"]` count from the agent's first request to the service (any endpoint), and a restarted service keeps that moment; before, they counted from the service's own start of the episode (when its scene was ready). Where the service is ready before the agent's phase begins (the harness is installed in between: a median of 79 s over the recorded standard trials), that clock ran ahead of the agent's and `time_left_s` reached 0 while the agent still had that time. `/spec` says what the clock counts from (`budget.clock`), and the instruction's `## Time limit` tells the agent to plan by `time_left_s`. The service never ended an episode on this clock: the scene, the frozen instance, the pass rule and the grader are unchanged, so the task keeps its name and earlier results stand.
+- 2026-10-14 (board 0113, from board 0086): the success text says that the gripper must not be touching the objects at the moment of success ("at that moment"); the old wording "the gripper must touch none of them" was read by an agent as "never touch them" (board 0086). Text only: the scene, the frozen instance, RoboLab's success term and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/robolab-bananas-in-bin-one-more-i00-standard/instruction.md b/tasks/robolab-bananas-in-bin-one-more-i00-standard/instruction.md
index 9f61170400..d4cda64f60 100644
--- a/tasks/robolab-bananas-in-bin-one-more-i00-standard/instruction.md
+++ b/tasks/robolab-bananas-in-bin-one-more-i00-standard/instruction.md
@@ -1,7 +1,7 @@
# RoboLab · bananas_in_bin_one_more — STANDARD MODE (robot as a service, eai-standard/2.2)
You control a **Franka Panda arm with a Robotiq 2F-85 gripper** (NVIDIA RoboLab's DROID setup) **as if it were real
-hardware**: the simulator runs in a service you cannot open. You get one client and one episode. **Task:** Put one (1) more bananas in the grey bin. Success: Exactly 1 of the three bananas that start outside the grey bin (no more, no fewer) must end inside the grey bin and the gripper must touch none of them. "Inside" means the centre of the object's convex hull lies within the container's convex hull, extended without limit along the container's own up axis (the open side; straight up for a container standing normally).
+hardware**: the simulator runs in a service you cannot open. You get one client and one episode. **Task:** Put one (1) more bananas in the grey bin. Success: Exactly 1 of the three bananas that start outside the grey bin (no more, no fewer) must end inside the grey bin and, at that moment, the gripper must touch none of them (moving them with the gripper is how the task is done; let go of them at the end). "Inside" means the centre of the object's convex hull lies within the container's convex hull, extended without limit along the container's own up axis (the open side; straight up for a container standing normally).
## Rules (real-robot rules)
- **No reset.** The episode starts when the service comes up and ends on success (at the first control step where the
diff --git a/tasks/robolab-bananas-in-bin-one-more-i00-standard/task.toml b/tasks/robolab-bananas-in-bin-one-more-i00-standard/task.toml
index 3a6e15709e..8489f6cf05 100644
--- a/tasks/robolab-bananas-in-bin-one-more-i00-standard/task.toml
+++ b/tasks/robolab-bananas-in-bin-one-more-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "robolab"
robot = "franka-robotiq-2f85"
family = "bananas_in_bin_one_more"
mode = "standard" # robot as a service: RGB, depth + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution: RoboLab ships no expert or planner. Positive examples graded 1 by the separate verifier: robolab-bananas-in-bin-one-more__uypUNUt (robolab-bananas-in-bin-one-more-i00-privileged, run codex-gpt6_luna-xhigh, robolab-lunacx-0925-2038, ws; https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/bananas_in_bin_one_more/unlimited/). Human review per family (PROTOCOL §2.7)."
diff --git a/tasks/robolab-bananas-in-bin-three-total-i00-privileged/task.toml b/tasks/robolab-bananas-in-bin-three-total-i00-privileged/task.toml
index c21faab78f..8812c89ddf 100644
--- a/tasks/robolab-bananas-in-bin-three-total-i00-privileged/task.toml
+++ b/tasks/robolab-bananas-in-bin-three-total-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "robolab" # NVIDIA RoboLab v0.3.1 on Isaac Lab 2.3.2 / I
robot = "franka-robotiq-2f85"
family = "bananas_in_bin_three_total"
mode = "privileged" # privileged state, resets and primitives through the tooling (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution: RoboLab ships no expert or planner. Positive examples graded 1 by the separate verifier: robolab-bananas-in-bin-three-tot__HWLcbcG (robolab-bananas-in-bin-three-total-i00-privileged, run codex-gpt6_luna-xhigh, robolab-lunacx-0925-2038, ws; https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/bananas_in_bin_three_total/unlimited/). Human review per family (PROTOCOL §2.7)."
instance_seed = 0 # scene = environment/instance.json (the full scene state after the seeded reset + its hash); identity = dataset.toml digest
diff --git a/tasks/robolab-bananas-in-bin-three-total-i00-standard/README.md b/tasks/robolab-bananas-in-bin-three-total-i00-standard/README.md
index fa3b474e84..3e5afb9508 100644
--- a/tasks/robolab-bananas-in-bin-three-total-i00-standard/README.md
+++ b/tasks/robolab-bananas-in-bin-three-total-i00-standard/README.md
@@ -72,3 +72,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
- 2026-10-10 (review round 3, board 0067): the shared service layer and the client only (`environment/sim/rcb_limited.py`, `environment/rcb_client.py`). A camera named more than once in `cameras` is rendered and returned once (N copies of one name made the service build N x N images within the frame budget, which could exhaust its memory). A service that has given up on the episode (it could not start it, or could not resume it after its crashes) answers every request with 409 and the reason, and `Robot()` / `wait_ready()` raise `EpisodeOver` at once; before, nothing answered and they waited out their timeout (up to an hour). The adapter, the collect hook, the grader, the scene and the pass rule are unchanged.
- 2026-10-10 (owner decision, board 0072): the agent's network egress is restricted to the model APIs (`[agent] network_mode = "allowlist"`, the hosts of `scripts/set_agent_egress.py`; since 2026-10-11 these include the package registries, PyPI and the Debian / Ubuntu archives, so that `pip install` and `apt-get install` work); it was open before (Harbor's default), and agents could download upstream sources, reference solutions and robot models from their own shell. The simulator image is `eai-robolab:0.1.2` (was 0.1.1): it holds the MDL materials and textures that RoboLab's asset library fetched from NVIDIA's Omniverse content bucket whenever a scene was built, so that the cameras show under the allowlist what they showed with open egress (on 0.1.1 without that download the table renders red). The scene, the frozen instance, the physics, the tooling and the pass rule are unchanged, so the task keeps its name; results from before this date were obtained with open egress.
- 2026-10-13 (owner, boards 0085 and 0087): the sim service's clock (`environment/sim/rcb_limited.py`, `environment/sim/server.py`, the docstring of `environment/rcb_client.py`). `meta["time_s"]` and `meta["time_left_s"]` count from the agent's first request to the service (any endpoint), and a restarted service keeps that moment; before, they counted from the service's own start of the episode (when its scene was ready). Where the service is ready before the agent's phase begins (the harness is installed in between: a median of 79 s over the recorded standard trials), that clock ran ahead of the agent's and `time_left_s` reached 0 while the agent still had that time. `/spec` says what the clock counts from (`budget.clock`), and the instruction's `## Time limit` tells the agent to plan by `time_left_s`. The service never ended an episode on this clock: the scene, the frozen instance, the pass rule and the grader are unchanged, so the task keeps its name and earlier results stand.
+- 2026-10-14 (board 0113, from board 0086): the success text says that the gripper must not be touching the objects at the moment of success ("at that moment"); the old wording "the gripper must touch none of them" was read by an agent as "never touch them" (board 0086). Text only: the scene, the frozen instance, RoboLab's success term and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/robolab-bananas-in-bin-three-total-i00-standard/instruction.md b/tasks/robolab-bananas-in-bin-three-total-i00-standard/instruction.md
index 0636001b5e..dd686c0bde 100644
--- a/tasks/robolab-bananas-in-bin-three-total-i00-standard/instruction.md
+++ b/tasks/robolab-bananas-in-bin-three-total-i00-standard/instruction.md
@@ -1,7 +1,7 @@
# RoboLab · bananas_in_bin_three_total — STANDARD MODE (robot as a service, eai-standard/2.2)
You control a **Franka Panda arm with a Robotiq 2F-85 gripper** (NVIDIA RoboLab's DROID setup) **as if it were real
-hardware**: the simulator runs in a service you cannot open. You get one client and one episode. **Task:** Make sure there are 3 (three) bananas in the grey bin. Success: Exactly 1 of the three bananas that start outside the grey bin (no more, no fewer) must end inside the grey bin and the gripper must touch none of them. "Inside" means the centre of the object's convex hull lies within the container's convex hull, extended without limit along the container's own up axis (the open side; straight up for a container standing normally).
+hardware**: the simulator runs in a service you cannot open. You get one client and one episode. **Task:** Make sure there are 3 (three) bananas in the grey bin. Success: Exactly 1 of the three bananas that start outside the grey bin (no more, no fewer) must end inside the grey bin and, at that moment, the gripper must touch none of them (moving them with the gripper is how the task is done; let go of them at the end). "Inside" means the centre of the object's convex hull lies within the container's convex hull, extended without limit along the container's own up axis (the open side; straight up for a container standing normally).
## Rules (real-robot rules)
- **No reset.** The episode starts when the service comes up and ends on success (at the first control step where the
diff --git a/tasks/robolab-bananas-in-bin-three-total-i00-standard/task.toml b/tasks/robolab-bananas-in-bin-three-total-i00-standard/task.toml
index e9d4ff0fcb..0be94866e8 100644
--- a/tasks/robolab-bananas-in-bin-three-total-i00-standard/task.toml
+++ b/tasks/robolab-bananas-in-bin-three-total-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "robolab"
robot = "franka-robotiq-2f85"
family = "bananas_in_bin_three_total"
mode = "standard" # robot as a service: RGB, depth + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution: RoboLab ships no expert or planner. Positive examples graded 1 by the separate verifier: robolab-bananas-in-bin-three-tot__HWLcbcG (robolab-bananas-in-bin-three-total-i00-privileged, run codex-gpt6_luna-xhigh, robolab-lunacx-0925-2038, ws; https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/bananas_in_bin_three_total/unlimited/). Human review per family (PROTOCOL §2.7)."
diff --git a/tasks/robolab-bananas-out-of-bin-i00-privileged/task.toml b/tasks/robolab-bananas-out-of-bin-i00-privileged/task.toml
index 72d650d990..029e8ca152 100644
--- a/tasks/robolab-bananas-out-of-bin-i00-privileged/task.toml
+++ b/tasks/robolab-bananas-out-of-bin-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "robolab" # NVIDIA RoboLab v0.3.1 on Isaac Lab 2.3.2 / I
robot = "franka-robotiq-2f85"
family = "bananas_out_of_bin"
mode = "privileged" # privileged state, resets and primitives through the tooling (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution: RoboLab ships no expert or planner. Positive examples graded 1 by the separate verifier: robolab-bananas-out-of-bin-i00__ZBYE469 (robolab-bananas-out-of-bin-i00-privileged, run codex-gpt6_luna-xhigh, robolab-lunacx-0925-2038, ws3; https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/bananas_out_of_bin/unlimited/); robolab-bananas-out-of-bin-i00-l__sYgutMj (robolab-bananas-out-of-bin-i00-standard, run codex-gpt6_luna-xhigh, robolab-lunaxh-limited-0928, lab; https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/bananas_out_of_bin/limited/). Human review per family (PROTOCOL §2.7)."
instance_seed = 0 # scene = environment/instance.json (the full scene state after the seeded reset + its hash); identity = dataset.toml digest
diff --git a/tasks/robolab-bananas-out-of-bin-i00-standard/README.md b/tasks/robolab-bananas-out-of-bin-i00-standard/README.md
index b026148d8e..a4883d5bad 100644
--- a/tasks/robolab-bananas-out-of-bin-i00-standard/README.md
+++ b/tasks/robolab-bananas-out-of-bin-i00-standard/README.md
@@ -73,3 +73,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
- 2026-10-10 (review round 3, board 0067): the shared service layer and the client only (`environment/sim/rcb_limited.py`, `environment/rcb_client.py`). A camera named more than once in `cameras` is rendered and returned once (N copies of one name made the service build N x N images within the frame budget, which could exhaust its memory). A service that has given up on the episode (it could not start it, or could not resume it after its crashes) answers every request with 409 and the reason, and `Robot()` / `wait_ready()` raise `EpisodeOver` at once; before, nothing answered and they waited out their timeout (up to an hour). The adapter, the collect hook, the grader, the scene and the pass rule are unchanged.
- 2026-10-10 (owner decision, board 0072): the agent's network egress is restricted to the model APIs (`[agent] network_mode = "allowlist"`, the hosts of `scripts/set_agent_egress.py`; since 2026-10-11 these include the package registries, PyPI and the Debian / Ubuntu archives, so that `pip install` and `apt-get install` work); it was open before (Harbor's default), and agents could download upstream sources, reference solutions and robot models from their own shell. The simulator image is `eai-robolab:0.1.2` (was 0.1.1): it holds the MDL materials and textures that RoboLab's asset library fetched from NVIDIA's Omniverse content bucket whenever a scene was built, so that the cameras show under the allowlist what they showed with open egress (on 0.1.1 without that download the table renders red). The scene, the frozen instance, the physics, the tooling and the pass rule are unchanged, so the task keeps its name; results from before this date were obtained with open egress.
- 2026-10-13 (owner, boards 0085 and 0087): the sim service's clock (`environment/sim/rcb_limited.py`, `environment/sim/server.py`, the docstring of `environment/rcb_client.py`). `meta["time_s"]` and `meta["time_left_s"]` count from the agent's first request to the service (any endpoint), and a restarted service keeps that moment; before, they counted from the service's own start of the episode (when its scene was ready). Where the service is ready before the agent's phase begins (the harness is installed in between: a median of 79 s over the recorded standard trials), that clock ran ahead of the agent's and `time_left_s` reached 0 while the agent still had that time. `/spec` says what the clock counts from (`budget.clock`), and the instruction's `## Time limit` tells the agent to plan by `time_left_s`. The service never ended an episode on this clock: the scene, the frozen instance, the pass rule and the grader are unchanged, so the task keeps its name and earlier results stand.
+- 2026-10-14 (board 0113, from board 0086): the success text says that the gripper must not be touching the objects at the moment of success ("at that moment"); the old wording "the gripper must touch none of them" was read by an agent as "never touch them" (board 0086). Text only: the scene, the frozen instance, RoboLab's success term and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/robolab-bananas-out-of-bin-i00-standard/instruction.md b/tasks/robolab-bananas-out-of-bin-i00-standard/instruction.md
index 9a6acb812d..88e3df5361 100644
--- a/tasks/robolab-bananas-out-of-bin-i00-standard/instruction.md
+++ b/tasks/robolab-bananas-out-of-bin-i00-standard/instruction.md
@@ -1,7 +1,7 @@
# RoboLab · bananas_out_of_bin — STANDARD MODE (robot as a service, eai-standard/2.2)
You control a **Franka Panda arm with a Robotiq 2F-85 gripper** (NVIDIA RoboLab's DROID setup) **as if it were real
-hardware**: the simulator runs in a service you cannot open. You get one client and one episode. **Task:** Take the bananas out. Success: Both bananas that start inside the grey bin must end outside the grey bin, and the gripper must touch neither. "Outside" means the centre of the object's convex hull is not within the container's convex hull extended along the container's own up axis.
+hardware**: the simulator runs in a service you cannot open. You get one client and one episode. **Task:** Take the bananas out. Success: Both bananas that start inside the grey bin must end outside the grey bin, and, at that moment, the gripper must touch neither (let go of both at the end). "Outside" means the centre of the object's convex hull is not within the container's convex hull extended along the container's own up axis.
## Rules (real-robot rules)
- **No reset.** The episode starts when the service comes up and ends on success (at the first control step where the
diff --git a/tasks/robolab-bananas-out-of-bin-i00-standard/task.toml b/tasks/robolab-bananas-out-of-bin-i00-standard/task.toml
index 20ec038738..ee8d0b77a9 100644
--- a/tasks/robolab-bananas-out-of-bin-i00-standard/task.toml
+++ b/tasks/robolab-bananas-out-of-bin-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "robolab"
robot = "franka-robotiq-2f85"
family = "bananas_out_of_bin"
mode = "standard" # robot as a service: RGB, depth + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution: RoboLab ships no expert or planner. Positive examples graded 1 by the separate verifier: robolab-bananas-out-of-bin-i00__ZBYE469 (robolab-bananas-out-of-bin-i00-privileged, run codex-gpt6_luna-xhigh, robolab-lunacx-0925-2038, ws3; https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/bananas_out_of_bin/unlimited/); robolab-bananas-out-of-bin-i00-l__sYgutMj (robolab-bananas-out-of-bin-i00-standard, run codex-gpt6_luna-xhigh, robolab-lunaxh-limited-0928, lab; https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/bananas_out_of_bin/limited/). Human review per family (PROTOCOL §2.7)."
diff --git a/tasks/robolab-bbq-sauce-in-bin-i00-privileged/task.toml b/tasks/robolab-bbq-sauce-in-bin-i00-privileged/task.toml
index fef6e85b57..ea7360bef1 100644
--- a/tasks/robolab-bbq-sauce-in-bin-i00-privileged/task.toml
+++ b/tasks/robolab-bbq-sauce-in-bin-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "robolab" # NVIDIA RoboLab v0.3.1 on Isaac Lab 2.3.2 / I
robot = "franka-robotiq-2f85"
family = "bbq_sauce_in_bin"
mode = "privileged" # privileged state, resets and primitives through the tooling (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution: RoboLab ships no expert or planner. Positive examples graded 1 by the separate verifier: robolab-bbq-sauce-in-bin-i00__kHjYXj6 (robolab-bbq-sauce-in-bin-i00-privileged, run codex-gpt6_luna-xhigh, robolab-lunacx-0925-2038, ws3; https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/bbq_sauce_in_bin/unlimited/). Human review per family (PROTOCOL §2.7)."
instance_seed = 0 # scene = environment/instance.json (the full scene state after the seeded reset + its hash); identity = dataset.toml digest
diff --git a/tasks/robolab-bbq-sauce-in-bin-i00-standard/README.md b/tasks/robolab-bbq-sauce-in-bin-i00-standard/README.md
index 7278417b65..2e0e23f3b3 100644
--- a/tasks/robolab-bbq-sauce-in-bin-i00-standard/README.md
+++ b/tasks/robolab-bbq-sauce-in-bin-i00-standard/README.md
@@ -72,3 +72,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
- 2026-10-10 (review round 3, board 0067): the shared service layer and the client only (`environment/sim/rcb_limited.py`, `environment/rcb_client.py`). A camera named more than once in `cameras` is rendered and returned once (N copies of one name made the service build N x N images within the frame budget, which could exhaust its memory). A service that has given up on the episode (it could not start it, or could not resume it after its crashes) answers every request with 409 and the reason, and `Robot()` / `wait_ready()` raise `EpisodeOver` at once; before, nothing answered and they waited out their timeout (up to an hour). The adapter, the collect hook, the grader, the scene and the pass rule are unchanged.
- 2026-10-10 (owner decision, board 0072): the agent's network egress is restricted to the model APIs (`[agent] network_mode = "allowlist"`, the hosts of `scripts/set_agent_egress.py`; since 2026-10-11 these include the package registries, PyPI and the Debian / Ubuntu archives, so that `pip install` and `apt-get install` work); it was open before (Harbor's default), and agents could download upstream sources, reference solutions and robot models from their own shell. The simulator image is `eai-robolab:0.1.2` (was 0.1.1): it holds the MDL materials and textures that RoboLab's asset library fetched from NVIDIA's Omniverse content bucket whenever a scene was built, so that the cameras show under the allowlist what they showed with open egress (on 0.1.1 without that download the table renders red). The scene, the frozen instance, the physics, the tooling and the pass rule are unchanged, so the task keeps its name; results from before this date were obtained with open egress.
- 2026-10-13 (owner, boards 0085 and 0087): the sim service's clock (`environment/sim/rcb_limited.py`, `environment/sim/server.py`, the docstring of `environment/rcb_client.py`). `meta["time_s"]` and `meta["time_left_s"]` count from the agent's first request to the service (any endpoint), and a restarted service keeps that moment; before, they counted from the service's own start of the episode (when its scene was ready). Where the service is ready before the agent's phase begins (the harness is installed in between: a median of 79 s over the recorded standard trials), that clock ran ahead of the agent's and `time_left_s` reached 0 while the agent still had that time. `/spec` says what the clock counts from (`budget.clock`), and the instruction's `## Time limit` tells the agent to plan by `time_left_s`. The service never ended an episode on this clock: the scene, the frozen instance, the pass rule and the grader are unchanged, so the task keeps its name and earlier results stand.
+- 2026-10-14 (board 0113, from board 0086): the success text says that the gripper must not be touching the objects at the moment of success ("at that moment"); the old wording "the gripper must touch none of them" was read by an agent as "never touch them" (board 0086). Text only: the scene, the frozen instance, RoboLab's success term and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/robolab-bbq-sauce-in-bin-i00-standard/instruction.md b/tasks/robolab-bbq-sauce-in-bin-i00-standard/instruction.md
index d40d97058c..e1e05495bb 100644
--- a/tasks/robolab-bbq-sauce-in-bin-i00-standard/instruction.md
+++ b/tasks/robolab-bbq-sauce-in-bin-i00-standard/instruction.md
@@ -1,7 +1,7 @@
# RoboLab · bbq_sauce_in_bin — STANDARD MODE (robot as a service, eai-standard/2.2)
You control a **Franka Panda arm with a Robotiq 2F-85 gripper** (NVIDIA RoboLab's DROID setup) **as if it were real
-hardware**: the simulator runs in a service you cannot open. You get one client and one episode. **Task:** Put the red BBQ sauce bottles in the grey bin. Success: Both red BBQ sauce bottles must end inside the grey bin and the gripper must touch neither. "Inside" means the centre of the object's convex hull lies within the container's convex hull, extended without limit along the container's own up axis (the open side; straight up for a container standing normally).
+hardware**: the simulator runs in a service you cannot open. You get one client and one episode. **Task:** Put the red BBQ sauce bottles in the grey bin. Success: Both red BBQ sauce bottles must end inside the grey bin and, at that moment, the gripper must touch neither (let go of both at the end). "Inside" means the centre of the object's convex hull lies within the container's convex hull, extended without limit along the container's own up axis (the open side; straight up for a container standing normally).
## Rules (real-robot rules)
- **No reset.** The episode starts when the service comes up and ends on success (at the first control step where the
diff --git a/tasks/robolab-bbq-sauce-in-bin-i00-standard/task.toml b/tasks/robolab-bbq-sauce-in-bin-i00-standard/task.toml
index 079900276a..c27b65ad39 100644
--- a/tasks/robolab-bbq-sauce-in-bin-i00-standard/task.toml
+++ b/tasks/robolab-bbq-sauce-in-bin-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "robolab"
robot = "franka-robotiq-2f85"
family = "bbq_sauce_in_bin"
mode = "standard" # robot as a service: RGB, depth + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution: RoboLab ships no expert or planner. Positive examples graded 1 by the separate verifier: robolab-bbq-sauce-in-bin-i00__kHjYXj6 (robolab-bbq-sauce-in-bin-i00-privileged, run codex-gpt6_luna-xhigh, robolab-lunacx-0925-2038, ws3; https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/bbq_sauce_in_bin/unlimited/). Human review per family (PROTOCOL §2.7)."
diff --git a/tasks/robolab-block-stacking-order-agnostic-i00-privileged/task.toml b/tasks/robolab-block-stacking-order-agnostic-i00-privileged/task.toml
index e0496a29a3..7e2cdcd68b 100644
--- a/tasks/robolab-block-stacking-order-agnostic-i00-privileged/task.toml
+++ b/tasks/robolab-block-stacking-order-agnostic-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "robolab" # NVIDIA RoboLab v0.3.1 on Isaac Lab 2.3.2 / I
robot = "franka-robotiq-2f85"
family = "block_stacking_order_agnostic"
mode = "privileged" # privileged state, resets and primitives through the tooling (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution: RoboLab ships no expert or planner. Positive examples graded 1 by the separate verifier: robolab-block-stacking-order-agn__kqJQMLf (robolab-block-stacking-order-agnostic-i00-privileged, run codex-gpt6_luna-xhigh, robolab-lunacx-0925-2038, ws3; https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/block_stacking_order_agnostic/unlimited/); robolab-block-stacking-order-agn__S6o3W3Z (robolab-block-stacking-order-agnostic-i00-standard, run codex-gpt6_luna-xhigh, robolab-lunaxh-limited-rerun-0928, lab; https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/block_stacking_order_agnostic/limited/). Human review per family (PROTOCOL §2.7)."
instance_seed = 0 # scene = environment/instance.json (the full scene state after the seeded reset + its hash); identity = dataset.toml digest
diff --git a/tasks/robolab-block-stacking-order-agnostic-i00-standard/task.toml b/tasks/robolab-block-stacking-order-agnostic-i00-standard/task.toml
index aa55d07a36..bd23d424a5 100644
--- a/tasks/robolab-block-stacking-order-agnostic-i00-standard/task.toml
+++ b/tasks/robolab-block-stacking-order-agnostic-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "robolab"
robot = "franka-robotiq-2f85"
family = "block_stacking_order_agnostic"
mode = "standard" # robot as a service: RGB, depth + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution: RoboLab ships no expert or planner. Positive examples graded 1 by the separate verifier: robolab-block-stacking-order-agn__kqJQMLf (robolab-block-stacking-order-agnostic-i00-privileged, run codex-gpt6_luna-xhigh, robolab-lunacx-0925-2038, ws3; https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/block_stacking_order_agnostic/unlimited/); robolab-block-stacking-order-agn__S6o3W3Z (robolab-block-stacking-order-agnostic-i00-standard, run codex-gpt6_luna-xhigh, robolab-lunaxh-limited-rerun-0928, lab; https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/block_stacking_order_agnostic/limited/). Human review per family (PROTOCOL §2.7)."
diff --git a/tasks/robolab-block-stacking-specified-order-i00-privileged/task.toml b/tasks/robolab-block-stacking-specified-order-i00-privileged/task.toml
index 62e6641e96..56b6e21a4b 100644
--- a/tasks/robolab-block-stacking-specified-order-i00-privileged/task.toml
+++ b/tasks/robolab-block-stacking-specified-order-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "robolab" # NVIDIA RoboLab v0.3.1 on Isaac Lab 2.3.2 / I
robot = "franka-robotiq-2f85"
family = "block_stacking_specified_order"
mode = "privileged" # privileged state, resets and primitives through the tooling (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution: RoboLab ships no expert or planner. Positive examples graded 1 by the separate verifier: robolab-block-stacking-specified__DGoPthS (robolab-block-stacking-specified-order-i00-privileged, run codex-gpt6_luna-xhigh, robolab-lunacx-0925-2038, ws3; https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/block_stacking_specified_order/unlimited/); robolab-block-stacking-specified__Mvd28Jf (robolab-block-stacking-specified-order-i00-standard, run codex-gpt6_luna-xhigh, robolab-lunaxh-limited-0928, lab; https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/block_stacking_specified_order/limited/). Human review per family (PROTOCOL §2.7)."
instance_seed = 0 # scene = environment/instance.json (the full scene state after the seeded reset + its hash); identity = dataset.toml digest
diff --git a/tasks/robolab-block-stacking-specified-order-i00-standard/task.toml b/tasks/robolab-block-stacking-specified-order-i00-standard/task.toml
index e666c5d099..2df6b55f7f 100644
--- a/tasks/robolab-block-stacking-specified-order-i00-standard/task.toml
+++ b/tasks/robolab-block-stacking-specified-order-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "robolab"
robot = "franka-robotiq-2f85"
family = "block_stacking_specified_order"
mode = "standard" # robot as a service: RGB, depth + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution: RoboLab ships no expert or planner. Positive examples graded 1 by the separate verifier: robolab-block-stacking-specified__DGoPthS (robolab-block-stacking-specified-order-i00-privileged, run codex-gpt6_luna-xhigh, robolab-lunacx-0925-2038, ws3; https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/block_stacking_specified_order/unlimited/); robolab-block-stacking-specified__Mvd28Jf (robolab-block-stacking-specified-order-i00-standard, run codex-gpt6_luna-xhigh, robolab-lunaxh-limited-0928, lab; https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/block_stacking_specified_order/limited/). Human review per family (PROTOCOL §2.7)."
diff --git a/tasks/robolab-blocks-in-bin-i00-privileged/task.toml b/tasks/robolab-blocks-in-bin-i00-privileged/task.toml
index d5fde65ac0..da2e062da8 100644
--- a/tasks/robolab-blocks-in-bin-i00-privileged/task.toml
+++ b/tasks/robolab-blocks-in-bin-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "robolab" # NVIDIA RoboLab v0.3.1 on Isaac Lab 2.3.2 / I
robot = "franka-robotiq-2f85"
family = "blocks_in_bin"
mode = "privileged" # privileged state, resets and primitives through the tooling (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution: RoboLab ships no expert or planner. Positive examples graded 1 by the separate verifier: robolab-blocks-in-bin-i00__QNeJQ9b (robolab-blocks-in-bin-i00-privileged, run codex-gpt6_luna-xhigh, robolab-lunacx-0925-2038, ws3; https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/blocks_in_bin/unlimited/); robolab-blocks-in-bin-i00-limite__iBaJYXh (robolab-blocks-in-bin-i00-standard, run codex-gpt6_luna-xhigh, robolab-lunaxh-limited-0928, lab; https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/blocks_in_bin/limited/). Human review per family (PROTOCOL §2.7)."
instance_seed = 0 # scene = environment/instance.json (the full scene state after the seeded reset + its hash); identity = dataset.toml digest
diff --git a/tasks/robolab-blocks-in-bin-i00-standard/README.md b/tasks/robolab-blocks-in-bin-i00-standard/README.md
index a23ed0fc1a..a154366836 100644
--- a/tasks/robolab-blocks-in-bin-i00-standard/README.md
+++ b/tasks/robolab-blocks-in-bin-i00-standard/README.md
@@ -73,3 +73,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
- 2026-10-10 (review round 3, board 0067): the shared service layer and the client only (`environment/sim/rcb_limited.py`, `environment/rcb_client.py`). A camera named more than once in `cameras` is rendered and returned once (N copies of one name made the service build N x N images within the frame budget, which could exhaust its memory). A service that has given up on the episode (it could not start it, or could not resume it after its crashes) answers every request with 409 and the reason, and `Robot()` / `wait_ready()` raise `EpisodeOver` at once; before, nothing answered and they waited out their timeout (up to an hour). The adapter, the collect hook, the grader, the scene and the pass rule are unchanged.
- 2026-10-10 (owner decision, board 0072): the agent's network egress is restricted to the model APIs (`[agent] network_mode = "allowlist"`, the hosts of `scripts/set_agent_egress.py`; since 2026-10-11 these include the package registries, PyPI and the Debian / Ubuntu archives, so that `pip install` and `apt-get install` work); it was open before (Harbor's default), and agents could download upstream sources, reference solutions and robot models from their own shell. The simulator image is `eai-robolab:0.1.2` (was 0.1.1): it holds the MDL materials and textures that RoboLab's asset library fetched from NVIDIA's Omniverse content bucket whenever a scene was built, so that the cameras show under the allowlist what they showed with open egress (on 0.1.1 without that download the table renders red). The scene, the frozen instance, the physics, the tooling and the pass rule are unchanged, so the task keeps its name; results from before this date were obtained with open egress.
- 2026-10-13 (owner, boards 0085 and 0087): the sim service's clock (`environment/sim/rcb_limited.py`, `environment/sim/server.py`, the docstring of `environment/rcb_client.py`). `meta["time_s"]` and `meta["time_left_s"]` count from the agent's first request to the service (any endpoint), and a restarted service keeps that moment; before, they counted from the service's own start of the episode (when its scene was ready). Where the service is ready before the agent's phase begins (the harness is installed in between: a median of 79 s over the recorded standard trials), that clock ran ahead of the agent's and `time_left_s` reached 0 while the agent still had that time. `/spec` says what the clock counts from (`budget.clock`), and the instruction's `## Time limit` tells the agent to plan by `time_left_s`. The service never ended an episode on this clock: the scene, the frozen instance, the pass rule and the grader are unchanged, so the task keeps its name and earlier results stand.
+- 2026-10-14 (board 0113, from board 0086): the success text says that the gripper must not be touching the objects at the moment of success ("at that moment"); the old wording "the gripper must touch none of them" was read by an agent as "never touch them" (board 0086). Text only: the scene, the frozen instance, RoboLab's success term and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/robolab-blocks-in-bin-i00-standard/instruction.md b/tasks/robolab-blocks-in-bin-i00-standard/instruction.md
index fd53e7fb06..ee6194f5c0 100644
--- a/tasks/robolab-blocks-in-bin-i00-standard/instruction.md
+++ b/tasks/robolab-blocks-in-bin-i00-standard/instruction.md
@@ -1,7 +1,7 @@
# RoboLab · blocks_in_bin — STANDARD MODE (robot as a service, eai-standard/2.2)
You control a **Franka Panda arm with a Robotiq 2F-85 gripper** (NVIDIA RoboLab's DROID setup) **as if it were real
-hardware**: the simulator runs in a service you cannot open. You get one client and one episode. **Task:** Sort all colored blocks into the bin. Success: The four coloured blocks (yellow, red, green, blue) must end inside the grey bin and the gripper must touch none of them. "Inside" means the centre of the object's convex hull lies within the container's convex hull, extended without limit along the container's own up axis (the open side; straight up for a container standing normally).
+hardware**: the simulator runs in a service you cannot open. You get one client and one episode. **Task:** Sort all colored blocks into the bin. Success: The four coloured blocks (yellow, red, green, blue) must end inside the grey bin and, at that moment, the gripper must touch none of them (moving them with the gripper is how the task is done; let go of them at the end). "Inside" means the centre of the object's convex hull lies within the container's convex hull, extended without limit along the container's own up axis (the open side; straight up for a container standing normally).
## Rules (real-robot rules)
- **No reset.** The episode starts when the service comes up and ends on success (at the first control step where the
diff --git a/tasks/robolab-blocks-in-bin-i00-standard/task.toml b/tasks/robolab-blocks-in-bin-i00-standard/task.toml
index 9689facad2..340d91cca1 100644
--- a/tasks/robolab-blocks-in-bin-i00-standard/task.toml
+++ b/tasks/robolab-blocks-in-bin-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "robolab"
robot = "franka-robotiq-2f85"
family = "blocks_in_bin"
mode = "standard" # robot as a service: RGB, depth + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution: RoboLab ships no expert or planner. Positive examples graded 1 by the separate verifier: robolab-blocks-in-bin-i00__QNeJQ9b (robolab-blocks-in-bin-i00-privileged, run codex-gpt6_luna-xhigh, robolab-lunacx-0925-2038, ws3; https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/blocks_in_bin/unlimited/); robolab-blocks-in-bin-i00-limite__iBaJYXh (robolab-blocks-in-bin-i00-standard, run codex-gpt6_luna-xhigh, robolab-lunaxh-limited-0928, lab; https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/blocks_in_bin/limited/). Human review per family (PROTOCOL §2.7)."
diff --git a/tasks/robolab-butter-above-raisin-i00-privileged/task.toml b/tasks/robolab-butter-above-raisin-i00-privileged/task.toml
index 4427613723..f11a4b0008 100644
--- a/tasks/robolab-butter-above-raisin-i00-privileged/task.toml
+++ b/tasks/robolab-butter-above-raisin-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "robolab" # NVIDIA RoboLab v0.3.1 on Isaac Lab 2.3.2 / I
robot = "franka-robotiq-2f85"
family = "butter_above_raisin"
mode = "privileged" # privileged state, resets and primitives through the tooling (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution: RoboLab ships no expert or planner. Positive examples graded 1 by the separate verifier: robolab-butter-above-raisin-i00__SHtnec6 (robolab-butter-above-raisin-i00-privileged, run codex-gpt6_luna-xhigh, robolab-lunacx-0925-2038, ws; https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/butter_above_raisin/unlimited/). Human review per family (PROTOCOL §2.7)."
instance_seed = 0 # scene = environment/instance.json (the full scene state after the seeded reset + its hash); identity = dataset.toml digest
diff --git a/tasks/robolab-butter-above-raisin-i00-standard/task.toml b/tasks/robolab-butter-above-raisin-i00-standard/task.toml
index 4b0ea3ca8f..c1974de0c7 100644
--- a/tasks/robolab-butter-above-raisin-i00-standard/task.toml
+++ b/tasks/robolab-butter-above-raisin-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "robolab"
robot = "franka-robotiq-2f85"
family = "butter_above_raisin"
mode = "standard" # robot as a service: RGB, depth + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution: RoboLab ships no expert or planner. Positive examples graded 1 by the separate verifier: robolab-butter-above-raisin-i00__SHtnec6 (robolab-butter-above-raisin-i00-privileged, run codex-gpt6_luna-xhigh, robolab-lunacx-0925-2038, ws; https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/butter_above_raisin/unlimited/). Human review per family (PROTOCOL §2.7)."
diff --git a/tasks/robolab-clutter-plastic-i00-privileged/task.toml b/tasks/robolab-clutter-plastic-i00-privileged/task.toml
index 0a63f36213..488617adc0 100644
--- a/tasks/robolab-clutter-plastic-i00-privileged/task.toml
+++ b/tasks/robolab-clutter-plastic-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "robolab" # NVIDIA RoboLab v0.3.1 on Isaac Lab 2.3.2 / I
robot = "franka-robotiq-2f85"
family = "clutter_plastic"
mode = "privileged" # privileged state, resets and primitives through the tooling (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution: RoboLab ships no expert or planner. Positive examples graded 1 by the separate verifier: robolab-clutter-plastic-i00__iFRwjNF (robolab-clutter-plastic-i00-privileged, run codex-gpt6_luna-xhigh, robolab-lunacx-0925-2038, ws3; https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/clutter_plastic/unlimited/). Human review per family (PROTOCOL §2.7)."
instance_seed = 0 # scene = environment/instance.json (the full scene state after the seeded reset + its hash); identity = dataset.toml digest
diff --git a/tasks/robolab-clutter-plastic-i00-standard/README.md b/tasks/robolab-clutter-plastic-i00-standard/README.md
index 6f700c2670..2d324451ee 100644
--- a/tasks/robolab-clutter-plastic-i00-standard/README.md
+++ b/tasks/robolab-clutter-plastic-i00-standard/README.md
@@ -72,3 +72,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
- 2026-10-10 (review round 3, board 0067): the shared service layer and the client only (`environment/sim/rcb_limited.py`, `environment/rcb_client.py`). A camera named more than once in `cameras` is rendered and returned once (N copies of one name made the service build N x N images within the frame budget, which could exhaust its memory). A service that has given up on the episode (it could not start it, or could not resume it after its crashes) answers every request with 409 and the reason, and `Robot()` / `wait_ready()` raise `EpisodeOver` at once; before, nothing answered and they waited out their timeout (up to an hour). The adapter, the collect hook, the grader, the scene and the pass rule are unchanged.
- 2026-10-10 (owner decision, board 0072): the agent's network egress is restricted to the model APIs (`[agent] network_mode = "allowlist"`, the hosts of `scripts/set_agent_egress.py`; since 2026-10-11 these include the package registries, PyPI and the Debian / Ubuntu archives, so that `pip install` and `apt-get install` work); it was open before (Harbor's default), and agents could download upstream sources, reference solutions and robot models from their own shell. The simulator image is `eai-robolab:0.1.2` (was 0.1.1): it holds the MDL materials and textures that RoboLab's asset library fetched from NVIDIA's Omniverse content bucket whenever a scene was built, so that the cameras show under the allowlist what they showed with open egress (on 0.1.1 without that download the table renders red). The scene, the frozen instance, the physics, the tooling and the pass rule are unchanged, so the task keeps its name; results from before this date were obtained with open egress.
- 2026-10-13 (owner, boards 0085 and 0087): the sim service's clock (`environment/sim/rcb_limited.py`, `environment/sim/server.py`, the docstring of `environment/rcb_client.py`). `meta["time_s"]` and `meta["time_left_s"]` count from the agent's first request to the service (any endpoint), and a restarted service keeps that moment; before, they counted from the service's own start of the episode (when its scene was ready). Where the service is ready before the agent's phase begins (the harness is installed in between: a median of 79 s over the recorded standard trials), that clock ran ahead of the agent's and `time_left_s` reached 0 while the agent still had that time. `/spec` says what the clock counts from (`budget.clock`), and the instruction's `## Time limit` tells the agent to plan by `time_left_s`. The service never ended an episode on this clock: the scene, the frozen instance, the pass rule and the grader are unchanged, so the task keeps its name and earlier results stand.
+- 2026-10-14 (board 0113, from board 0086): the success text says that the gripper must not be touching the objects at the moment of success ("at that moment"); the old wording "the gripper must touch none of them" was read by an agent as "never touch them" (board 0086). Text only: the scene, the frozen instance, RoboLab's success term and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/robolab-clutter-plastic-i00-standard/instruction.md b/tasks/robolab-clutter-plastic-i00-standard/instruction.md
index 4abfae793f..7f7b92a890 100644
--- a/tasks/robolab-clutter-plastic-i00-standard/instruction.md
+++ b/tasks/robolab-clutter-plastic-i00-standard/instruction.md
@@ -1,7 +1,7 @@
# RoboLab · clutter_plastic — STANDARD MODE (robot as a service, eai-standard/2.2)
You control a **Franka Panda arm with a Robotiq 2F-85 gripper** (NVIDIA RoboLab's DROID setup) **as if it were real
-hardware**: the simulator runs in a service you cannot open. You get one client and one episode. **Task:** Put all plastic bottles away in the bin. Success: The three plastic bottles (the small white plastic bottle, the milk jug and the utility jug) must end inside the blue bin and the gripper must touch none of them. "Inside" means the centre of the object's convex hull lies within the container's convex hull, extended without limit along the container's own up axis (the open side; straight up for a container standing normally).
+hardware**: the simulator runs in a service you cannot open. You get one client and one episode. **Task:** Put all plastic bottles away in the bin. Success: The three plastic bottles (the small white plastic bottle, the milk jug and the utility jug) must end inside the blue bin and, at that moment, the gripper must touch none of them (moving them with the gripper is how the task is done; let go of them at the end). "Inside" means the centre of the object's convex hull lies within the container's convex hull, extended without limit along the container's own up axis (the open side; straight up for a container standing normally).
## Rules (real-robot rules)
- **No reset.** The episode starts when the service comes up and ends on success (at the first control step where the
diff --git a/tasks/robolab-clutter-plastic-i00-standard/task.toml b/tasks/robolab-clutter-plastic-i00-standard/task.toml
index f15b09f5bc..d9e3d3d13d 100644
--- a/tasks/robolab-clutter-plastic-i00-standard/task.toml
+++ b/tasks/robolab-clutter-plastic-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "robolab"
robot = "franka-robotiq-2f85"
family = "clutter_plastic"
mode = "standard" # robot as a service: RGB, depth + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution: RoboLab ships no expert or planner. Positive examples graded 1 by the separate verifier: robolab-clutter-plastic-i00__iFRwjNF (robolab-clutter-plastic-i00-privileged, run codex-gpt6_luna-xhigh, robolab-lunacx-0925-2038, ws3; https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/clutter_plastic/unlimited/). Human review per family (PROTOCOL §2.7)."
diff --git a/tasks/robolab-clutter-pumpkin-i00-privileged/task.toml b/tasks/robolab-clutter-pumpkin-i00-privileged/task.toml
index f6cadce9d1..0396511cdf 100644
--- a/tasks/robolab-clutter-pumpkin-i00-privileged/task.toml
+++ b/tasks/robolab-clutter-pumpkin-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "robolab" # NVIDIA RoboLab v0.3.1 on Isaac Lab 2.3.2 / I
robot = "franka-robotiq-2f85"
family = "clutter_pumpkin"
mode = "privileged" # privileged state, resets and primitives through the tooling (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution: RoboLab ships no expert or planner. Positive examples graded 1 by the separate verifier: robolab-clutter-pumpkin-i00__wa3DTDa (robolab-clutter-pumpkin-i00-privileged, run codex-gpt6_luna-xhigh, robolab-lunacx-0925-2038, ws; https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/clutter_pumpkin/unlimited/). Human review per family (PROTOCOL §2.7)."
instance_seed = 0 # scene = environment/instance.json (the full scene state after the seeded reset + its hash); identity = dataset.toml digest
diff --git a/tasks/robolab-clutter-pumpkin-i00-standard/README.md b/tasks/robolab-clutter-pumpkin-i00-standard/README.md
index 6f9a4649c2..337908283a 100644
--- a/tasks/robolab-clutter-pumpkin-i00-standard/README.md
+++ b/tasks/robolab-clutter-pumpkin-i00-standard/README.md
@@ -72,3 +72,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
- 2026-10-10 (review round 3, board 0067): the shared service layer and the client only (`environment/sim/rcb_limited.py`, `environment/rcb_client.py`). A camera named more than once in `cameras` is rendered and returned once (N copies of one name made the service build N x N images within the frame budget, which could exhaust its memory). A service that has given up on the episode (it could not start it, or could not resume it after its crashes) answers every request with 409 and the reason, and `Robot()` / `wait_ready()` raise `EpisodeOver` at once; before, nothing answered and they waited out their timeout (up to an hour). The adapter, the collect hook, the grader, the scene and the pass rule are unchanged.
- 2026-10-10 (owner decision, board 0072): the agent's network egress is restricted to the model APIs (`[agent] network_mode = "allowlist"`, the hosts of `scripts/set_agent_egress.py`; since 2026-10-11 these include the package registries, PyPI and the Debian / Ubuntu archives, so that `pip install` and `apt-get install` work); it was open before (Harbor's default), and agents could download upstream sources, reference solutions and robot models from their own shell. The simulator image is `eai-robolab:0.1.2` (was 0.1.1): it holds the MDL materials and textures that RoboLab's asset library fetched from NVIDIA's Omniverse content bucket whenever a scene was built, so that the cameras show under the allowlist what they showed with open egress (on 0.1.1 without that download the table renders red). The scene, the frozen instance, the physics, the tooling and the pass rule are unchanged, so the task keeps its name; results from before this date were obtained with open egress.
- 2026-10-13 (owner, boards 0085 and 0087): the sim service's clock (`environment/sim/rcb_limited.py`, `environment/sim/server.py`, the docstring of `environment/rcb_client.py`). `meta["time_s"]` and `meta["time_left_s"]` count from the agent's first request to the service (any endpoint), and a restarted service keeps that moment; before, they counted from the service's own start of the episode (when its scene was ready). Where the service is ready before the agent's phase begins (the harness is installed in between: a median of 79 s over the recorded standard trials), that clock ran ahead of the agent's and `time_left_s` reached 0 while the agent still had that time. `/spec` says what the clock counts from (`budget.clock`), and the instruction's `## Time limit` tells the agent to plan by `time_left_s`. The service never ended an episode on this clock: the scene, the frozen instance, the pass rule and the grader are unchanged, so the task keeps its name and earlier results stand.
+- 2026-10-14 (board 0113, from board 0086): the success text says that the gripper must not be touching the objects at the moment of success ("at that moment"); the old wording "the gripper must touch none of them" was read by an agent as "never touch them" (board 0086). Text only: the scene, the frozen instance, RoboLab's success term and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/robolab-clutter-pumpkin-i00-standard/instruction.md b/tasks/robolab-clutter-pumpkin-i00-standard/instruction.md
index 5a1647abb1..2baf101453 100644
--- a/tasks/robolab-clutter-pumpkin-i00-standard/instruction.md
+++ b/tasks/robolab-clutter-pumpkin-i00-standard/instruction.md
@@ -1,7 +1,7 @@
# RoboLab · clutter_pumpkin — STANDARD MODE (robot as a service, eai-standard/2.2)
You control a **Franka Panda arm with a Robotiq 2F-85 gripper** (NVIDIA RoboLab's DROID setup) **as if it were real
-hardware**: the simulator runs in a service you cannot open. You get one client and one episode. **Task:** Put all the pumpkins away in the bin. Success: The large and the small pumpkin must end inside the blue bin and the gripper must touch neither. "Inside" means the centre of the object's convex hull lies within the container's convex hull, extended without limit along the container's own up axis (the open side; straight up for a container standing normally).
+hardware**: the simulator runs in a service you cannot open. You get one client and one episode. **Task:** Put all the pumpkins away in the bin. Success: The large and the small pumpkin must end inside the blue bin and, at that moment, the gripper must touch neither (let go of both at the end). "Inside" means the centre of the object's convex hull lies within the container's convex hull, extended without limit along the container's own up axis (the open side; straight up for a container standing normally).
## Rules (real-robot rules)
- **No reset.** The episode starts when the service comes up and ends on success (at the first control step where the
diff --git a/tasks/robolab-clutter-pumpkin-i00-standard/task.toml b/tasks/robolab-clutter-pumpkin-i00-standard/task.toml
index 87e9683742..3fbdae12d4 100644
--- a/tasks/robolab-clutter-pumpkin-i00-standard/task.toml
+++ b/tasks/robolab-clutter-pumpkin-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "robolab"
robot = "franka-robotiq-2f85"
family = "clutter_pumpkin"
mode = "standard" # robot as a service: RGB, depth + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution: RoboLab ships no expert or planner. Positive examples graded 1 by the separate verifier: robolab-clutter-pumpkin-i00__wa3DTDa (robolab-clutter-pumpkin-i00-privileged, run codex-gpt6_luna-xhigh, robolab-lunacx-0925-2038, ws; https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/clutter_pumpkin/unlimited/). Human review per family (PROTOCOL §2.7)."
diff --git a/tasks/robolab-condiments-in-bin-i00-privileged/task.toml b/tasks/robolab-condiments-in-bin-i00-privileged/task.toml
index 61ff5cd448..ca2dc993b3 100644
--- a/tasks/robolab-condiments-in-bin-i00-privileged/task.toml
+++ b/tasks/robolab-condiments-in-bin-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "robolab" # NVIDIA RoboLab v0.3.1 on Isaac Lab 2.3.2 / I
robot = "franka-robotiq-2f85"
family = "condiments_in_bin"
mode = "privileged" # privileged state, resets and primitives through the tooling (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution: RoboLab ships no expert or planner. Positive examples graded 1 by the separate verifier: robolab-condiments-in-bin-i00__XZbFmBx (robolab-condiments-in-bin-i00-privileged, run codex-gpt6_luna-xhigh, robolab-lunacx-0925-2038, ws; https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/condiments_in_bin/unlimited/). Human review per family (PROTOCOL §2.7)."
instance_seed = 0 # scene = environment/instance.json (the full scene state after the seeded reset + its hash); identity = dataset.toml digest
diff --git a/tasks/robolab-condiments-in-bin-i00-standard/README.md b/tasks/robolab-condiments-in-bin-i00-standard/README.md
index 7ebe33aad3..74073af466 100644
--- a/tasks/robolab-condiments-in-bin-i00-standard/README.md
+++ b/tasks/robolab-condiments-in-bin-i00-standard/README.md
@@ -72,3 +72,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
- 2026-10-10 (review round 3, board 0067): the shared service layer and the client only (`environment/sim/rcb_limited.py`, `environment/rcb_client.py`). A camera named more than once in `cameras` is rendered and returned once (N copies of one name made the service build N x N images within the frame budget, which could exhaust its memory). A service that has given up on the episode (it could not start it, or could not resume it after its crashes) answers every request with 409 and the reason, and `Robot()` / `wait_ready()` raise `EpisodeOver` at once; before, nothing answered and they waited out their timeout (up to an hour). The adapter, the collect hook, the grader, the scene and the pass rule are unchanged.
- 2026-10-10 (owner decision, board 0072): the agent's network egress is restricted to the model APIs (`[agent] network_mode = "allowlist"`, the hosts of `scripts/set_agent_egress.py`; since 2026-10-11 these include the package registries, PyPI and the Debian / Ubuntu archives, so that `pip install` and `apt-get install` work); it was open before (Harbor's default), and agents could download upstream sources, reference solutions and robot models from their own shell. The simulator image is `eai-robolab:0.1.2` (was 0.1.1): it holds the MDL materials and textures that RoboLab's asset library fetched from NVIDIA's Omniverse content bucket whenever a scene was built, so that the cameras show under the allowlist what they showed with open egress (on 0.1.1 without that download the table renders red). The scene, the frozen instance, the physics, the tooling and the pass rule are unchanged, so the task keeps its name; results from before this date were obtained with open egress.
- 2026-10-13 (owner, boards 0085 and 0087): the sim service's clock (`environment/sim/rcb_limited.py`, `environment/sim/server.py`, the docstring of `environment/rcb_client.py`). `meta["time_s"]` and `meta["time_left_s"]` count from the agent's first request to the service (any endpoint), and a restarted service keeps that moment; before, they counted from the service's own start of the episode (when its scene was ready). Where the service is ready before the agent's phase begins (the harness is installed in between: a median of 79 s over the recorded standard trials), that clock ran ahead of the agent's and `time_left_s` reached 0 while the agent still had that time. `/spec` says what the clock counts from (`budget.clock`), and the instruction's `## Time limit` tells the agent to plan by `time_left_s`. The service never ended an episode on this clock: the scene, the frozen instance, the pass rule and the grader are unchanged, so the task keeps its name and earlier results stand.
+- 2026-10-14 (board 0113, from board 0086): the success text says that the gripper must not be touching the objects at the moment of success ("at that moment"); the old wording "the gripper must touch none of them" was read by an agent as "never touch them" (board 0086). Text only: the scene, the frozen instance, RoboLab's success term and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/robolab-condiments-in-bin-i00-standard/instruction.md b/tasks/robolab-condiments-in-bin-i00-standard/instruction.md
index d6b2918cbe..0b13d40187 100644
--- a/tasks/robolab-condiments-in-bin-i00-standard/instruction.md
+++ b/tasks/robolab-condiments-in-bin-i00-standard/instruction.md
@@ -1,7 +1,7 @@
# RoboLab · condiments_in_bin — STANDARD MODE (robot as a service, eai-standard/2.2)
You control a **Franka Panda arm with a Robotiq 2F-85 gripper** (NVIDIA RoboLab's DROID setup) **as if it were real
-hardware**: the simulator runs in a service you cannot open. You get one client and one episode. **Task:** Sort the sauce condiments into the grey bin. Success: The four sauce bottles (the mustard, the ranch dressing and both red BBQ sauce bottles) must end inside the grey bin and the gripper must touch none of them. "Inside" means the centre of the object's convex hull lies within the container's convex hull, extended without limit along the container's own up axis (the open side; straight up for a container standing normally).
+hardware**: the simulator runs in a service you cannot open. You get one client and one episode. **Task:** Sort the sauce condiments into the grey bin. Success: The four sauce bottles (the mustard, the ranch dressing and both red BBQ sauce bottles) must end inside the grey bin and, at that moment, the gripper must touch none of them (moving them with the gripper is how the task is done; let go of them at the end). "Inside" means the centre of the object's convex hull lies within the container's convex hull, extended without limit along the container's own up axis (the open side; straight up for a container standing normally).
## Rules (real-robot rules)
- **No reset.** The episode starts when the service comes up and ends on success (at the first control step where the
diff --git a/tasks/robolab-condiments-in-bin-i00-standard/task.toml b/tasks/robolab-condiments-in-bin-i00-standard/task.toml
index 9fd89ea11e..ddc36def9d 100644
--- a/tasks/robolab-condiments-in-bin-i00-standard/task.toml
+++ b/tasks/robolab-condiments-in-bin-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "robolab"
robot = "franka-robotiq-2f85"
family = "condiments_in_bin"
mode = "standard" # robot as a service: RGB, depth + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution: RoboLab ships no expert or planner. Positive examples graded 1 by the separate verifier: robolab-condiments-in-bin-i00__XZbFmBx (robolab-condiments-in-bin-i00-privileged, run codex-gpt6_luna-xhigh, robolab-lunacx-0925-2038, ws; https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/condiments_in_bin/unlimited/). Human review per family (PROTOCOL §2.7)."
diff --git a/tasks/robolab-cooking-clear-plate-i00-privileged/task.toml b/tasks/robolab-cooking-clear-plate-i00-privileged/task.toml
index 8b4d51b34e..44acc2c514 100644
--- a/tasks/robolab-cooking-clear-plate-i00-privileged/task.toml
+++ b/tasks/robolab-cooking-clear-plate-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "robolab" # NVIDIA RoboLab v0.3.1 on Isaac Lab 2.3.2 / I
robot = "franka-robotiq-2f85"
family = "cooking_clear_plate"
mode = "privileged" # privileged state, resets and primitives through the tooling (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution: RoboLab ships no expert or planner. Positive examples graded 1 by the separate verifier: robolab-cooking-clear-plate-i00__M9JWo2C (robolab-cooking-clear-plate-i00-privileged, run codex-gpt6_luna-xhigh, robolab-lunacx-0925-2038, ws; https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/cooking_clear_plate/unlimited/); robolab-cooking-clear-plate-i00__pjCZgi2 (robolab-cooking-clear-plate-i00-standard, run codex-gpt6_luna-xhigh, robolab-lunacx-0925-2038, ws; not published: its eai-standard/2.1 batch was withdrawn for the agent's stale wrist-camera pose; grader unaffected). Human review per family (PROTOCOL §2.7)."
instance_seed = 0 # scene = environment/instance.json (the full scene state after the seeded reset + its hash); identity = dataset.toml digest
diff --git a/tasks/robolab-cooking-clear-plate-i00-standard/README.md b/tasks/robolab-cooking-clear-plate-i00-standard/README.md
index fa31f631b6..9c0b0b9994 100644
--- a/tasks/robolab-cooking-clear-plate-i00-standard/README.md
+++ b/tasks/robolab-cooking-clear-plate-i00-standard/README.md
@@ -73,3 +73,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
- 2026-10-10 (review round 3, board 0067): the shared service layer and the client only (`environment/sim/rcb_limited.py`, `environment/rcb_client.py`). A camera named more than once in `cameras` is rendered and returned once (N copies of one name made the service build N x N images within the frame budget, which could exhaust its memory). A service that has given up on the episode (it could not start it, or could not resume it after its crashes) answers every request with 409 and the reason, and `Robot()` / `wait_ready()` raise `EpisodeOver` at once; before, nothing answered and they waited out their timeout (up to an hour). The adapter, the collect hook, the grader, the scene and the pass rule are unchanged.
- 2026-10-10 (owner decision, board 0072): the agent's network egress is restricted to the model APIs (`[agent] network_mode = "allowlist"`, the hosts of `scripts/set_agent_egress.py`; since 2026-10-11 these include the package registries, PyPI and the Debian / Ubuntu archives, so that `pip install` and `apt-get install` work); it was open before (Harbor's default), and agents could download upstream sources, reference solutions and robot models from their own shell. The simulator image is `eai-robolab:0.1.2` (was 0.1.1): it holds the MDL materials and textures that RoboLab's asset library fetched from NVIDIA's Omniverse content bucket whenever a scene was built, so that the cameras show under the allowlist what they showed with open egress (on 0.1.1 without that download the table renders red). The scene, the frozen instance, the physics, the tooling and the pass rule are unchanged, so the task keeps its name; results from before this date were obtained with open egress.
- 2026-10-13 (owner, boards 0085 and 0087): the sim service's clock (`environment/sim/rcb_limited.py`, `environment/sim/server.py`, the docstring of `environment/rcb_client.py`). `meta["time_s"]` and `meta["time_left_s"]` count from the agent's first request to the service (any endpoint), and a restarted service keeps that moment; before, they counted from the service's own start of the episode (when its scene was ready). Where the service is ready before the agent's phase begins (the harness is installed in between: a median of 79 s over the recorded standard trials), that clock ran ahead of the agent's and `time_left_s` reached 0 while the agent still had that time. `/spec` says what the clock counts from (`budget.clock`), and the instruction's `## Time limit` tells the agent to plan by `time_left_s`. The service never ended an episode on this clock: the scene, the frozen instance, the pass rule and the grader are unchanged, so the task keeps its name and earlier results stand.
+- 2026-10-14 (board 0113, from board 0086): the success text says that the gripper must not be touching the objects at the moment of success ("at that moment"); the old wording "the gripper must touch none of them" was read by an agent as "never touch them" (board 0086). Text only: the scene, the frozen instance, RoboLab's success term and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/robolab-cooking-clear-plate-i00-standard/instruction.md b/tasks/robolab-cooking-clear-plate-i00-standard/instruction.md
index f47048619b..1db263ba3b 100644
--- a/tasks/robolab-cooking-clear-plate-i00-standard/instruction.md
+++ b/tasks/robolab-cooking-clear-plate-i00-standard/instruction.md
@@ -1,7 +1,7 @@
# RoboLab · cooking_clear_plate — STANDARD MODE (robot as a service, eai-standard/2.2)
You control a **Franka Panda arm with a Robotiq 2F-85 gripper** (NVIDIA RoboLab's DROID setup) **as if it were real
-hardware**: the simulator runs in a service you cannot open. You get one client and one episode. **Task:** Put the two measuring cups outside of the plate. Success: The orange measuring cup and the small blue measuring scoop, which start on the clay plate, must end outside the clay plate, and the gripper must touch neither. "Outside" means the centre of the object's convex hull is not within the container's convex hull extended along the container's own up axis.
+hardware**: the simulator runs in a service you cannot open. You get one client and one episode. **Task:** Put the two measuring cups outside of the plate. Success: The orange measuring cup and the small blue measuring scoop, which start on the clay plate, must end outside the clay plate, and, at that moment, the gripper must touch neither (let go of both at the end). "Outside" means the centre of the object's convex hull is not within the container's convex hull extended along the container's own up axis.
## Rules (real-robot rules)
- **No reset.** The episode starts when the service comes up and ends on success (at the first control step where the
diff --git a/tasks/robolab-cooking-clear-plate-i00-standard/task.toml b/tasks/robolab-cooking-clear-plate-i00-standard/task.toml
index aa4c92a85a..14cb641352 100644
--- a/tasks/robolab-cooking-clear-plate-i00-standard/task.toml
+++ b/tasks/robolab-cooking-clear-plate-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "robolab"
robot = "franka-robotiq-2f85"
family = "cooking_clear_plate"
mode = "standard" # robot as a service: RGB, depth + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution: RoboLab ships no expert or planner. Positive examples graded 1 by the separate verifier: robolab-cooking-clear-plate-i00__M9JWo2C (robolab-cooking-clear-plate-i00-privileged, run codex-gpt6_luna-xhigh, robolab-lunacx-0925-2038, ws; https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/cooking_clear_plate/unlimited/); robolab-cooking-clear-plate-i00__pjCZgi2 (robolab-cooking-clear-plate-i00-standard, run codex-gpt6_luna-xhigh, robolab-lunacx-0925-2038, ws; not published: its eai-standard/2.1 batch was withdrawn for the agent's stale wrist-camera pose; grader unaffected). Human review per family (PROTOCOL §2.7)."
diff --git a/tasks/robolab-cubes-and-blocks-in-bin-i00-privileged/task.toml b/tasks/robolab-cubes-and-blocks-in-bin-i00-privileged/task.toml
index 419f8fc507..9ff8fbe806 100644
--- a/tasks/robolab-cubes-and-blocks-in-bin-i00-privileged/task.toml
+++ b/tasks/robolab-cubes-and-blocks-in-bin-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "robolab" # NVIDIA RoboLab v0.3.1 on Isaac Lab 2.3.2 / I
robot = "franka-robotiq-2f85"
family = "cubes_and_blocks_in_bin"
mode = "privileged" # privileged state, resets and primitives through the tooling (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution: RoboLab ships no expert or planner. Positive examples graded 1 by the separate verifier: robolab-cubes-and-blocks-in-bin__MJxbnQb (robolab-cubes-and-blocks-in-bin-i00-privileged, run codex-gpt6_luna-xhigh, robolab-lunacx-0925-2038, ws; https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/cubes_and_blocks_in_bin/unlimited/); robolab-cubes-and-blocks-in-bin__YG25byB (robolab-cubes-and-blocks-in-bin-i00-standard, run codex-gpt6_luna-xhigh, robolab-lunaxh-limited-0928, lab; https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/cubes_and_blocks_in_bin/limited/). Human review per family (PROTOCOL §2.7)."
instance_seed = 0 # scene = environment/instance.json (the full scene state after the seeded reset + its hash); identity = dataset.toml digest
diff --git a/tasks/robolab-cubes-and-blocks-in-bin-i00-standard/README.md b/tasks/robolab-cubes-and-blocks-in-bin-i00-standard/README.md
index 43f2c513e9..b33af970f5 100644
--- a/tasks/robolab-cubes-and-blocks-in-bin-i00-standard/README.md
+++ b/tasks/robolab-cubes-and-blocks-in-bin-i00-standard/README.md
@@ -73,3 +73,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
- 2026-10-10 (review round 3, board 0067): the shared service layer and the client only (`environment/sim/rcb_limited.py`, `environment/rcb_client.py`). A camera named more than once in `cameras` is rendered and returned once (N copies of one name made the service build N x N images within the frame budget, which could exhaust its memory). A service that has given up on the episode (it could not start it, or could not resume it after its crashes) answers every request with 409 and the reason, and `Robot()` / `wait_ready()` raise `EpisodeOver` at once; before, nothing answered and they waited out their timeout (up to an hour). The adapter, the collect hook, the grader, the scene and the pass rule are unchanged.
- 2026-10-10 (owner decision, board 0072): the agent's network egress is restricted to the model APIs (`[agent] network_mode = "allowlist"`, the hosts of `scripts/set_agent_egress.py`; since 2026-10-11 these include the package registries, PyPI and the Debian / Ubuntu archives, so that `pip install` and `apt-get install` work); it was open before (Harbor's default), and agents could download upstream sources, reference solutions and robot models from their own shell. The simulator image is `eai-robolab:0.1.2` (was 0.1.1): it holds the MDL materials and textures that RoboLab's asset library fetched from NVIDIA's Omniverse content bucket whenever a scene was built, so that the cameras show under the allowlist what they showed with open egress (on 0.1.1 without that download the table renders red). The scene, the frozen instance, the physics, the tooling and the pass rule are unchanged, so the task keeps its name; results from before this date were obtained with open egress.
- 2026-10-13 (owner, boards 0085 and 0087): the sim service's clock (`environment/sim/rcb_limited.py`, `environment/sim/server.py`, the docstring of `environment/rcb_client.py`). `meta["time_s"]` and `meta["time_left_s"]` count from the agent's first request to the service (any endpoint), and a restarted service keeps that moment; before, they counted from the service's own start of the episode (when its scene was ready). Where the service is ready before the agent's phase begins (the harness is installed in between: a median of 79 s over the recorded standard trials), that clock ran ahead of the agent's and `time_left_s` reached 0 while the agent still had that time. `/spec` says what the clock counts from (`budget.clock`), and the instruction's `## Time limit` tells the agent to plan by `time_left_s`. The service never ended an episode on this clock: the scene, the frozen instance, the pass rule and the grader are unchanged, so the task keeps its name and earlier results stand.
+- 2026-10-14 (board 0113, from board 0086): the success text says that the gripper must not be touching the objects at the moment of success ("at that moment"); the old wording "the gripper must touch none of them" was read by an agent as "never touch them" (board 0086). Text only: the scene, the frozen instance, RoboLab's success term and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/robolab-cubes-and-blocks-in-bin-i00-standard/instruction.md b/tasks/robolab-cubes-and-blocks-in-bin-i00-standard/instruction.md
index 27424a0665..b03014f901 100644
--- a/tasks/robolab-cubes-and-blocks-in-bin-i00-standard/instruction.md
+++ b/tasks/robolab-cubes-and-blocks-in-bin-i00-standard/instruction.md
@@ -1,7 +1,7 @@
# RoboLab · cubes_and_blocks_in_bin — STANDARD MODE (robot as a service, eai-standard/2.2)
You control a **Franka Panda arm with a Robotiq 2F-85 gripper** (NVIDIA RoboLab's DROID setup) **as if it were real
-hardware**: the simulator runs in a service you cannot open. You get one client and one episode. **Task:** Put all the cubes and blocks in the bin. Success: The seven cubes (the three Rubik's cubes and the four coloured blocks (yellow, red, green, blue)) must end inside the grey bin and the gripper must touch none of them. "Inside" means the centre of the object's convex hull lies within the container's convex hull, extended without limit along the container's own up axis (the open side; straight up for a container standing normally).
+hardware**: the simulator runs in a service you cannot open. You get one client and one episode. **Task:** Put all the cubes and blocks in the bin. Success: The seven cubes (the three Rubik's cubes and the four coloured blocks (yellow, red, green, blue)) must end inside the grey bin and, at that moment, the gripper must touch none of them (moving them with the gripper is how the task is done; let go of them at the end). "Inside" means the centre of the object's convex hull lies within the container's convex hull, extended without limit along the container's own up axis (the open side; straight up for a container standing normally).
## Rules (real-robot rules)
- **No reset.** The episode starts when the service comes up and ends on success (at the first control step where the
diff --git a/tasks/robolab-cubes-and-blocks-in-bin-i00-standard/task.toml b/tasks/robolab-cubes-and-blocks-in-bin-i00-standard/task.toml
index 1dd77562ee..4f96294a00 100644
--- a/tasks/robolab-cubes-and-blocks-in-bin-i00-standard/task.toml
+++ b/tasks/robolab-cubes-and-blocks-in-bin-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "robolab"
robot = "franka-robotiq-2f85"
family = "cubes_and_blocks_in_bin"
mode = "standard" # robot as a service: RGB, depth + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution: RoboLab ships no expert or planner. Positive examples graded 1 by the separate verifier: robolab-cubes-and-blocks-in-bin__MJxbnQb (robolab-cubes-and-blocks-in-bin-i00-privileged, run codex-gpt6_luna-xhigh, robolab-lunacx-0925-2038, ws; https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/cubes_and_blocks_in_bin/unlimited/); robolab-cubes-and-blocks-in-bin__YG25byB (robolab-cubes-and-blocks-in-bin-i00-standard, run codex-gpt6_luna-xhigh, robolab-lunaxh-limited-0928, lab; https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/cubes_and_blocks_in_bin/limited/). Human review per family (PROTOCOL §2.7)."
diff --git a/tasks/robolab-food-packing2-cans-i00-privileged/task.toml b/tasks/robolab-food-packing2-cans-i00-privileged/task.toml
index 3a6fcb9526..561a7e0298 100644
--- a/tasks/robolab-food-packing2-cans-i00-privileged/task.toml
+++ b/tasks/robolab-food-packing2-cans-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "robolab" # NVIDIA RoboLab v0.3.1 on Isaac Lab 2.3.2 / I
robot = "franka-robotiq-2f85"
family = "food_packing2_cans"
mode = "privileged" # privileged state, resets and primitives through the tooling (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution: RoboLab ships no expert or planner. Positive examples graded 1 by the separate verifier: robolab-food-packing2-cans-i00__ne3U8Jh (robolab-food-packing2-cans-i00-privileged, run codex-gpt6_luna-xhigh, robolab-lunacx-0925-2038, ws; https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/food_packing2_cans/unlimited/). Human review per family (PROTOCOL §2.7)."
instance_seed = 0 # scene = environment/instance.json (the full scene state after the seeded reset + its hash); identity = dataset.toml digest
diff --git a/tasks/robolab-food-packing2-cans-i00-standard/README.md b/tasks/robolab-food-packing2-cans-i00-standard/README.md
index e07b0c6c98..4d59ac8bb1 100644
--- a/tasks/robolab-food-packing2-cans-i00-standard/README.md
+++ b/tasks/robolab-food-packing2-cans-i00-standard/README.md
@@ -72,3 +72,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
- 2026-10-10 (review round 3, board 0067): the shared service layer and the client only (`environment/sim/rcb_limited.py`, `environment/rcb_client.py`). A camera named more than once in `cameras` is rendered and returned once (N copies of one name made the service build N x N images within the frame budget, which could exhaust its memory). A service that has given up on the episode (it could not start it, or could not resume it after its crashes) answers every request with 409 and the reason, and `Robot()` / `wait_ready()` raise `EpisodeOver` at once; before, nothing answered and they waited out their timeout (up to an hour). The adapter, the collect hook, the grader, the scene and the pass rule are unchanged.
- 2026-10-10 (owner decision, board 0072): the agent's network egress is restricted to the model APIs (`[agent] network_mode = "allowlist"`, the hosts of `scripts/set_agent_egress.py`; since 2026-10-11 these include the package registries, PyPI and the Debian / Ubuntu archives, so that `pip install` and `apt-get install` work); it was open before (Harbor's default), and agents could download upstream sources, reference solutions and robot models from their own shell. The simulator image is `eai-robolab:0.1.2` (was 0.1.1): it holds the MDL materials and textures that RoboLab's asset library fetched from NVIDIA's Omniverse content bucket whenever a scene was built, so that the cameras show under the allowlist what they showed with open egress (on 0.1.1 without that download the table renders red). The scene, the frozen instance, the physics, the tooling and the pass rule are unchanged, so the task keeps its name; results from before this date were obtained with open egress.
- 2026-10-13 (owner, boards 0085 and 0087): the sim service's clock (`environment/sim/rcb_limited.py`, `environment/sim/server.py`, the docstring of `environment/rcb_client.py`). `meta["time_s"]` and `meta["time_left_s"]` count from the agent's first request to the service (any endpoint), and a restarted service keeps that moment; before, they counted from the service's own start of the episode (when its scene was ready). Where the service is ready before the agent's phase begins (the harness is installed in between: a median of 79 s over the recorded standard trials), that clock ran ahead of the agent's and `time_left_s` reached 0 while the agent still had that time. `/spec` says what the clock counts from (`budget.clock`), and the instruction's `## Time limit` tells the agent to plan by `time_left_s`. The service never ended an episode on this clock: the scene, the frozen instance, the pass rule and the grader are unchanged, so the task keeps its name and earlier results stand.
+- 2026-10-14 (board 0113, from board 0086): the success text says that the gripper must not be touching the objects at the moment of success ("at that moment"); the old wording "the gripper must touch none of them" was read by an agent as "never touch them" (board 0086). Text only: the scene, the frozen instance, RoboLab's success term and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/robolab-food-packing2-cans-i00-standard/instruction.md b/tasks/robolab-food-packing2-cans-i00-standard/instruction.md
index a1e8af9a3a..320d825d2e 100644
--- a/tasks/robolab-food-packing2-cans-i00-standard/instruction.md
+++ b/tasks/robolab-food-packing2-cans-i00-standard/instruction.md
@@ -1,7 +1,7 @@
# RoboLab · food_packing2_cans — STANDARD MODE (robot as a service, eai-standard/2.2)
You control a **Franka Panda arm with a Robotiq 2F-85 gripper** (NVIDIA RoboLab's DROID setup) **as if it were real
-hardware**: the simulator runs in a service you cannot open. You get one client and one episode. **Task:** Pack canned foods into the bin. Success: Both cans (the tomato soup can and the tuna can) must end inside the dark grey bin and the gripper must touch neither. "Inside" means the centre of the object's convex hull lies within the container's convex hull, extended without limit along the container's own up axis (the open side; straight up for a container standing normally).
+hardware**: the simulator runs in a service you cannot open. You get one client and one episode. **Task:** Pack canned foods into the bin. Success: Both cans (the tomato soup can and the tuna can) must end inside the dark grey bin and, at that moment, the gripper must touch neither (let go of both at the end). "Inside" means the centre of the object's convex hull lies within the container's convex hull, extended without limit along the container's own up axis (the open side; straight up for a container standing normally).
## Rules (real-robot rules)
- **No reset.** The episode starts when the service comes up and ends on success (at the first control step where the
diff --git a/tasks/robolab-food-packing2-cans-i00-standard/task.toml b/tasks/robolab-food-packing2-cans-i00-standard/task.toml
index 53def3b52e..d9ecba6250 100644
--- a/tasks/robolab-food-packing2-cans-i00-standard/task.toml
+++ b/tasks/robolab-food-packing2-cans-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "robolab"
robot = "franka-robotiq-2f85"
family = "food_packing2_cans"
mode = "standard" # robot as a service: RGB, depth + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution: RoboLab ships no expert or planner. Positive examples graded 1 by the separate verifier: robolab-food-packing2-cans-i00__ne3U8Jh (robolab-food-packing2-cans-i00-privileged, run codex-gpt6_luna-xhigh, robolab-lunacx-0925-2038, ws; https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/food_packing2_cans/unlimited/). Human review per family (PROTOCOL §2.7)."
diff --git a/tasks/robolab-food-packing3-cans-i00-privileged/task.toml b/tasks/robolab-food-packing3-cans-i00-privileged/task.toml
index bac92b2adc..d6287e67d8 100644
--- a/tasks/robolab-food-packing3-cans-i00-privileged/task.toml
+++ b/tasks/robolab-food-packing3-cans-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "robolab" # NVIDIA RoboLab v0.3.1 on Isaac Lab 2.3.2 / I
robot = "franka-robotiq-2f85"
family = "food_packing3_cans"
mode = "privileged" # privileged state, resets and primitives through the tooling (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution: RoboLab ships no expert or planner. Positive examples graded 1 by the separate verifier: robolab-food-packing3-cans-i00__LsaHDUU (robolab-food-packing3-cans-i00-privileged, run codex-gpt6_luna-xhigh, robolab-lunacx-0925-2038, ws; https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/food_packing3_cans/unlimited/). Human review per family (PROTOCOL §2.7)."
instance_seed = 0 # scene = environment/instance.json (the full scene state after the seeded reset + its hash); identity = dataset.toml digest
diff --git a/tasks/robolab-food-packing3-cans-i00-standard/README.md b/tasks/robolab-food-packing3-cans-i00-standard/README.md
index bdf7e3dd61..1ab7c2bdd4 100644
--- a/tasks/robolab-food-packing3-cans-i00-standard/README.md
+++ b/tasks/robolab-food-packing3-cans-i00-standard/README.md
@@ -72,3 +72,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
- 2026-10-10 (review round 3, board 0067): the shared service layer and the client only (`environment/sim/rcb_limited.py`, `environment/rcb_client.py`). A camera named more than once in `cameras` is rendered and returned once (N copies of one name made the service build N x N images within the frame budget, which could exhaust its memory). A service that has given up on the episode (it could not start it, or could not resume it after its crashes) answers every request with 409 and the reason, and `Robot()` / `wait_ready()` raise `EpisodeOver` at once; before, nothing answered and they waited out their timeout (up to an hour). The adapter, the collect hook, the grader, the scene and the pass rule are unchanged.
- 2026-10-10 (owner decision, board 0072): the agent's network egress is restricted to the model APIs (`[agent] network_mode = "allowlist"`, the hosts of `scripts/set_agent_egress.py`; since 2026-10-11 these include the package registries, PyPI and the Debian / Ubuntu archives, so that `pip install` and `apt-get install` work); it was open before (Harbor's default), and agents could download upstream sources, reference solutions and robot models from their own shell. The simulator image is `eai-robolab:0.1.2` (was 0.1.1): it holds the MDL materials and textures that RoboLab's asset library fetched from NVIDIA's Omniverse content bucket whenever a scene was built, so that the cameras show under the allowlist what they showed with open egress (on 0.1.1 without that download the table renders red). The scene, the frozen instance, the physics, the tooling and the pass rule are unchanged, so the task keeps its name; results from before this date were obtained with open egress.
- 2026-10-13 (owner, boards 0085 and 0087): the sim service's clock (`environment/sim/rcb_limited.py`, `environment/sim/server.py`, the docstring of `environment/rcb_client.py`). `meta["time_s"]` and `meta["time_left_s"]` count from the agent's first request to the service (any endpoint), and a restarted service keeps that moment; before, they counted from the service's own start of the episode (when its scene was ready). Where the service is ready before the agent's phase begins (the harness is installed in between: a median of 79 s over the recorded standard trials), that clock ran ahead of the agent's and `time_left_s` reached 0 while the agent still had that time. `/spec` says what the clock counts from (`budget.clock`), and the instruction's `## Time limit` tells the agent to plan by `time_left_s`. The service never ended an episode on this clock: the scene, the frozen instance, the pass rule and the grader are unchanged, so the task keeps its name and earlier results stand.
+- 2026-10-14 (board 0113, from board 0086): the success text says that the gripper must not be touching the objects at the moment of success ("at that moment"); the old wording "the gripper must touch none of them" was read by an agent as "never touch them" (board 0086). Text only: the scene, the frozen instance, RoboLab's success term and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/robolab-food-packing3-cans-i00-standard/instruction.md b/tasks/robolab-food-packing3-cans-i00-standard/instruction.md
index fce7be0709..aa9f579699 100644
--- a/tasks/robolab-food-packing3-cans-i00-standard/instruction.md
+++ b/tasks/robolab-food-packing3-cans-i00-standard/instruction.md
@@ -1,7 +1,7 @@
# RoboLab · food_packing3_cans — STANDARD MODE (robot as a service, eai-standard/2.2)
You control a **Franka Panda arm with a Robotiq 2F-85 gripper** (NVIDIA RoboLab's DROID setup) **as if it were real
-hardware**: the simulator runs in a service you cannot open. You get one client and one episode. **Task:** Pack canned foods into the bin. Success: The three cans (the Spam can, the tomato soup can and the tuna can) must end inside the dark grey bin and the gripper must touch none of them. "Inside" means the centre of the object's convex hull lies within the container's convex hull, extended without limit along the container's own up axis (the open side; straight up for a container standing normally).
+hardware**: the simulator runs in a service you cannot open. You get one client and one episode. **Task:** Pack canned foods into the bin. Success: The three cans (the Spam can, the tomato soup can and the tuna can) must end inside the dark grey bin and, at that moment, the gripper must touch none of them (moving them with the gripper is how the task is done; let go of them at the end). "Inside" means the centre of the object's convex hull lies within the container's convex hull, extended without limit along the container's own up axis (the open side; straight up for a container standing normally).
## Rules (real-robot rules)
- **No reset.** The episode starts when the service comes up and ends on success (at the first control step where the
diff --git a/tasks/robolab-food-packing3-cans-i00-standard/task.toml b/tasks/robolab-food-packing3-cans-i00-standard/task.toml
index 5be3652065..77f5bc34f4 100644
--- a/tasks/robolab-food-packing3-cans-i00-standard/task.toml
+++ b/tasks/robolab-food-packing3-cans-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "robolab"
robot = "franka-robotiq-2f85"
family = "food_packing3_cans"
mode = "standard" # robot as a service: RGB, depth + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution: RoboLab ships no expert or planner. Positive examples graded 1 by the separate verifier: robolab-food-packing3-cans-i00__LsaHDUU (robolab-food-packing3-cans-i00-privileged, run codex-gpt6_luna-xhigh, robolab-lunacx-0925-2038, ws; https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/food_packing3_cans/unlimited/). Human review per family (PROTOCOL §2.7)."
diff --git a/tasks/robolab-fruits-green-limes-on-plate-i00-privileged/task.toml b/tasks/robolab-fruits-green-limes-on-plate-i00-privileged/task.toml
index 8285ef897c..8a5225926f 100644
--- a/tasks/robolab-fruits-green-limes-on-plate-i00-privileged/task.toml
+++ b/tasks/robolab-fruits-green-limes-on-plate-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "robolab" # NVIDIA RoboLab v0.3.1 on Isaac Lab 2.3.2 / I
robot = "franka-robotiq-2f85"
family = "fruits_green_limes_on_plate"
mode = "privileged" # privileged state, resets and primitives through the tooling (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution: RoboLab ships no expert or planner. Positive examples graded 1 by the separate verifier: robolab-fruits-green-limes-on-pl__3NH9BN4 (robolab-fruits-green-limes-on-plate-i00-standard, run codex-gpt6_luna-xhigh, robolab-lunaxh-limited-0928, lab; https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/fruits_green_limes_on_plate/limited/). Human review per family (PROTOCOL §2.7)."
instance_seed = 0 # scene = environment/instance.json (the full scene state after the seeded reset + its hash); identity = dataset.toml digest
diff --git a/tasks/robolab-fruits-green-limes-on-plate-i00-standard/README.md b/tasks/robolab-fruits-green-limes-on-plate-i00-standard/README.md
index 9e50aa4e7a..2a1f094b41 100644
--- a/tasks/robolab-fruits-green-limes-on-plate-i00-standard/README.md
+++ b/tasks/robolab-fruits-green-limes-on-plate-i00-standard/README.md
@@ -72,3 +72,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
- 2026-10-10 (review round 3, board 0067): the shared service layer and the client only (`environment/sim/rcb_limited.py`, `environment/rcb_client.py`). A camera named more than once in `cameras` is rendered and returned once (N copies of one name made the service build N x N images within the frame budget, which could exhaust its memory). A service that has given up on the episode (it could not start it, or could not resume it after its crashes) answers every request with 409 and the reason, and `Robot()` / `wait_ready()` raise `EpisodeOver` at once; before, nothing answered and they waited out their timeout (up to an hour). The adapter, the collect hook, the grader, the scene and the pass rule are unchanged.
- 2026-10-10 (owner decision, board 0072): the agent's network egress is restricted to the model APIs (`[agent] network_mode = "allowlist"`, the hosts of `scripts/set_agent_egress.py`; since 2026-10-11 these include the package registries, PyPI and the Debian / Ubuntu archives, so that `pip install` and `apt-get install` work); it was open before (Harbor's default), and agents could download upstream sources, reference solutions and robot models from their own shell. The simulator image is `eai-robolab:0.1.2` (was 0.1.1): it holds the MDL materials and textures that RoboLab's asset library fetched from NVIDIA's Omniverse content bucket whenever a scene was built, so that the cameras show under the allowlist what they showed with open egress (on 0.1.1 without that download the table renders red). The scene, the frozen instance, the physics, the tooling and the pass rule are unchanged, so the task keeps its name; results from before this date were obtained with open egress.
- 2026-10-13 (owner, boards 0085 and 0087): the sim service's clock (`environment/sim/rcb_limited.py`, `environment/sim/server.py`, the docstring of `environment/rcb_client.py`). `meta["time_s"]` and `meta["time_left_s"]` count from the agent's first request to the service (any endpoint), and a restarted service keeps that moment; before, they counted from the service's own start of the episode (when its scene was ready). Where the service is ready before the agent's phase begins (the harness is installed in between: a median of 79 s over the recorded standard trials), that clock ran ahead of the agent's and `time_left_s` reached 0 while the agent still had that time. `/spec` says what the clock counts from (`budget.clock`), and the instruction's `## Time limit` tells the agent to plan by `time_left_s`. The service never ended an episode on this clock: the scene, the frozen instance, the pass rule and the grader are unchanged, so the task keeps its name and earlier results stand.
+- 2026-10-14 (board 0113, from board 0086): the success text says that the gripper must not be touching the objects at the moment of success ("at that moment"); the old wording "the gripper must touch none of them" was read by an agent as "never touch them" (board 0086). Text only: the scene, the frozen instance, RoboLab's success term and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/robolab-fruits-green-limes-on-plate-i00-standard/instruction.md b/tasks/robolab-fruits-green-limes-on-plate-i00-standard/instruction.md
index a388e3425a..ee6039d2ab 100644
--- a/tasks/robolab-fruits-green-limes-on-plate-i00-standard/instruction.md
+++ b/tasks/robolab-fruits-green-limes-on-plate-i00-standard/instruction.md
@@ -1,7 +1,7 @@
# RoboLab · fruits_green_limes_on_plate — STANDARD MODE (robot as a service, eai-standard/2.2)
You control a **Franka Panda arm with a Robotiq 2F-85 gripper** (NVIDIA RoboLab's DROID setup) **as if it were real
-hardware**: the simulator runs in a service you cannot open. You get one client and one episode. **Task:** Put all the green fruit on the plate. Success: Both green limes must end on the flat beige plate, and the gripper must touch neither. "On" means the surface supports the object (the contact force it exerts on the object points upwards, within 45 deg of vertical) and the object's centre lies over the surface's bounding-box footprint, +/- 1 cm.
+hardware**: the simulator runs in a service you cannot open. You get one client and one episode. **Task:** Put all the green fruit on the plate. Success: Both green limes must end on the flat beige plate, and, at that moment, the gripper must touch neither (let go of both at the end). "On" means the surface supports the object (the contact force it exerts on the object points upwards, within 45 deg of vertical) and the object's centre lies over the surface's bounding-box footprint, +/- 1 cm.
## Rules (real-robot rules)
- **No reset.** The episode starts when the service comes up and ends on success (at the first control step where the
diff --git a/tasks/robolab-fruits-green-limes-on-plate-i00-standard/task.toml b/tasks/robolab-fruits-green-limes-on-plate-i00-standard/task.toml
index 2a6ed6011e..bbb7af0d25 100644
--- a/tasks/robolab-fruits-green-limes-on-plate-i00-standard/task.toml
+++ b/tasks/robolab-fruits-green-limes-on-plate-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "robolab"
robot = "franka-robotiq-2f85"
family = "fruits_green_limes_on_plate"
mode = "standard" # robot as a service: RGB, depth + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution: RoboLab ships no expert or planner. Positive examples graded 1 by the separate verifier: robolab-fruits-green-limes-on-pl__3NH9BN4 (robolab-fruits-green-limes-on-plate-i00-standard, run codex-gpt6_luna-xhigh, robolab-lunaxh-limited-0928, lab; https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/fruits_green_limes_on_plate/limited/). Human review per family (PROTOCOL §2.7)."
diff --git a/tasks/robolab-fruits-on-plate3-i00-privileged/task.toml b/tasks/robolab-fruits-on-plate3-i00-privileged/task.toml
index 11099dcd96..1959913a7e 100644
--- a/tasks/robolab-fruits-on-plate3-i00-privileged/task.toml
+++ b/tasks/robolab-fruits-on-plate3-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "robolab" # NVIDIA RoboLab v0.3.1 on Isaac Lab 2.3.2 / I
robot = "franka-robotiq-2f85"
family = "fruits_on_plate3"
mode = "privileged" # privileged state, resets and primitives through the tooling (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution: RoboLab ships no expert or planner. Positive examples graded 1 by the separate verifier: robolab-fruits-on-plate3-i00__GTgmNmc (robolab-fruits-on-plate3-i00-privileged, run codex-gpt6_luna-xhigh, robolab-lunacx-0925-2038, ws; https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/fruits_on_plate3/unlimited/); robolab-fruits-on-plate3-i00-lim__zdBJXPQ (robolab-fruits-on-plate3-i00-standard, run codex-gpt6_luna-xhigh, robolab-lunaxh-limited-0928, lab; https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/fruits_on_plate3/limited/). Human review per family (PROTOCOL §2.7)."
instance_seed = 0 # scene = environment/instance.json (the full scene state after the seeded reset + its hash); identity = dataset.toml digest
diff --git a/tasks/robolab-fruits-on-plate3-i00-standard/README.md b/tasks/robolab-fruits-on-plate3-i00-standard/README.md
index b8b05beafb..58e4fe453d 100644
--- a/tasks/robolab-fruits-on-plate3-i00-standard/README.md
+++ b/tasks/robolab-fruits-on-plate3-i00-standard/README.md
@@ -73,3 +73,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
- 2026-10-10 (review round 3, board 0067): the shared service layer and the client only (`environment/sim/rcb_limited.py`, `environment/rcb_client.py`). A camera named more than once in `cameras` is rendered and returned once (N copies of one name made the service build N x N images within the frame budget, which could exhaust its memory). A service that has given up on the episode (it could not start it, or could not resume it after its crashes) answers every request with 409 and the reason, and `Robot()` / `wait_ready()` raise `EpisodeOver` at once; before, nothing answered and they waited out their timeout (up to an hour). The adapter, the collect hook, the grader, the scene and the pass rule are unchanged.
- 2026-10-10 (owner decision, board 0072): the agent's network egress is restricted to the model APIs (`[agent] network_mode = "allowlist"`, the hosts of `scripts/set_agent_egress.py`; since 2026-10-11 these include the package registries, PyPI and the Debian / Ubuntu archives, so that `pip install` and `apt-get install` work); it was open before (Harbor's default), and agents could download upstream sources, reference solutions and robot models from their own shell. The simulator image is `eai-robolab:0.1.2` (was 0.1.1): it holds the MDL materials and textures that RoboLab's asset library fetched from NVIDIA's Omniverse content bucket whenever a scene was built, so that the cameras show under the allowlist what they showed with open egress (on 0.1.1 without that download the table renders red). The scene, the frozen instance, the physics, the tooling and the pass rule are unchanged, so the task keeps its name; results from before this date were obtained with open egress.
- 2026-10-13 (owner, boards 0085 and 0087): the sim service's clock (`environment/sim/rcb_limited.py`, `environment/sim/server.py`, the docstring of `environment/rcb_client.py`). `meta["time_s"]` and `meta["time_left_s"]` count from the agent's first request to the service (any endpoint), and a restarted service keeps that moment; before, they counted from the service's own start of the episode (when its scene was ready). Where the service is ready before the agent's phase begins (the harness is installed in between: a median of 79 s over the recorded standard trials), that clock ran ahead of the agent's and `time_left_s` reached 0 while the agent still had that time. `/spec` says what the clock counts from (`budget.clock`), and the instruction's `## Time limit` tells the agent to plan by `time_left_s`. The service never ended an episode on this clock: the scene, the frozen instance, the pass rule and the grader are unchanged, so the task keeps its name and earlier results stand.
+- 2026-10-14 (board 0113, from board 0086): the success text says that the gripper must not be touching the objects at the moment of success ("at that moment"); the old wording "the gripper must touch none of them" was read by an agent as "never touch them" (board 0086). Text only: the scene, the frozen instance, RoboLab's success term and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/robolab-fruits-on-plate3-i00-standard/instruction.md b/tasks/robolab-fruits-on-plate3-i00-standard/instruction.md
index 20bdb17ebf..c34afccac4 100644
--- a/tasks/robolab-fruits-on-plate3-i00-standard/instruction.md
+++ b/tasks/robolab-fruits-on-plate3-i00-standard/instruction.md
@@ -1,7 +1,7 @@
# RoboLab · fruits_on_plate3 — STANDARD MODE (robot as a service, eai-standard/2.2)
You control a **Franka Panda arm with a Robotiq 2F-85 gripper** (NVIDIA RoboLab's DROID setup) **as if it were real
-hardware**: the simulator runs in a service you cannot open. You get one client and one episode. **Task:** Put three (3) fruits on the plate. Success: Exactly 3 of the seven fruits (both lemons, both limes, both oranges, the pomegranate) (no more, no fewer) must end on the flat beige plate, and the gripper must touch none of them. "On" means the surface supports the object (the contact force it exerts on the object points upwards, within 45 deg of vertical) and the object's centre lies over the surface's bounding-box footprint, +/- 1 cm.
+hardware**: the simulator runs in a service you cannot open. You get one client and one episode. **Task:** Put three (3) fruits on the plate. Success: Exactly 3 of the seven fruits (both lemons, both limes, both oranges, the pomegranate) (no more, no fewer) must end on the flat beige plate, and, at that moment, the gripper must touch none of them (moving them with the gripper is how the task is done; let go of them at the end). "On" means the surface supports the object (the contact force it exerts on the object points upwards, within 45 deg of vertical) and the object's centre lies over the surface's bounding-box footprint, +/- 1 cm.
## Rules (real-robot rules)
- **No reset.** The episode starts when the service comes up and ends on success (at the first control step where the
diff --git a/tasks/robolab-fruits-on-plate3-i00-standard/task.toml b/tasks/robolab-fruits-on-plate3-i00-standard/task.toml
index 84053bc89c..6ed4cb63f2 100644
--- a/tasks/robolab-fruits-on-plate3-i00-standard/task.toml
+++ b/tasks/robolab-fruits-on-plate3-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "robolab"
robot = "franka-robotiq-2f85"
family = "fruits_on_plate3"
mode = "standard" # robot as a service: RGB, depth + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution: RoboLab ships no expert or planner. Positive examples graded 1 by the separate verifier: robolab-fruits-on-plate3-i00__GTgmNmc (robolab-fruits-on-plate3-i00-privileged, run codex-gpt6_luna-xhigh, robolab-lunacx-0925-2038, ws; https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/fruits_on_plate3/unlimited/); robolab-fruits-on-plate3-i00-lim__zdBJXPQ (robolab-fruits-on-plate3-i00-standard, run codex-gpt6_luna-xhigh, robolab-lunaxh-limited-0928, lab; https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/fruits_on_plate3/limited/). Human review per family (PROTOCOL §2.7)."
diff --git a/tasks/robolab-keyboard-out-of-bin-i00-privileged/task.toml b/tasks/robolab-keyboard-out-of-bin-i00-privileged/task.toml
index f51f7d49c3..56700ac978 100644
--- a/tasks/robolab-keyboard-out-of-bin-i00-privileged/task.toml
+++ b/tasks/robolab-keyboard-out-of-bin-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "robolab" # NVIDIA RoboLab v0.3.1 on Isaac Lab 2.3.2 / I
robot = "franka-robotiq-2f85"
family = "keyboard_out_of_bin"
mode = "privileged" # privileged state, resets and primitives through the tooling (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution: RoboLab ships no expert or planner. Positive examples graded 1 by the separate verifier: robolab-keyboard-out-of-bin-i00__rG5UBUk (robolab-keyboard-out-of-bin-i00-privileged, run codex-gpt6_luna-xhigh, robolab-lunacx-0925-2038, ws; https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/keyboard_out_of_bin/unlimited/). Human review per family (PROTOCOL §2.7)."
instance_seed = 0 # scene = environment/instance.json (the full scene state after the seeded reset + its hash); identity = dataset.toml digest
diff --git a/tasks/robolab-keyboard-out-of-bin-i00-standard/task.toml b/tasks/robolab-keyboard-out-of-bin-i00-standard/task.toml
index edcb1ca6bf..f402867149 100644
--- a/tasks/robolab-keyboard-out-of-bin-i00-standard/task.toml
+++ b/tasks/robolab-keyboard-out-of-bin-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "robolab"
robot = "franka-robotiq-2f85"
family = "keyboard_out_of_bin"
mode = "standard" # robot as a service: RGB, depth + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution: RoboLab ships no expert or planner. Positive examples graded 1 by the separate verifier: robolab-keyboard-out-of-bin-i00__rG5UBUk (robolab-keyboard-out-of-bin-i00-privileged, run codex-gpt6_luna-xhigh, robolab-lunacx-0925-2038, ws; https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/keyboard_out_of_bin/unlimited/). Human review per family (PROTOCOL §2.7)."
diff --git a/tasks/robolab-larger-object-raisin-box-in-bin-i00-privileged/task.toml b/tasks/robolab-larger-object-raisin-box-in-bin-i00-privileged/task.toml
index b586b55417..e17b92bd0b 100644
--- a/tasks/robolab-larger-object-raisin-box-in-bin-i00-privileged/task.toml
+++ b/tasks/robolab-larger-object-raisin-box-in-bin-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "robolab" # NVIDIA RoboLab v0.3.1 on Isaac Lab 2.3.2 / I
robot = "franka-robotiq-2f85"
family = "larger_object_raisin_box_in_bin"
mode = "privileged" # privileged state, resets and primitives through the tooling (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution: RoboLab ships no expert or planner. Positive examples graded 1 by the separate verifier: robolab-larger-object-raisin-box__eJa8kky (robolab-larger-object-raisin-box-in-bin-i00-privileged, run codex-gpt6_luna-xhigh, robolab-lunacx-0925-2038, ws3; https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/larger_object_raisin_box_in_bin/unlimited/). Human review per family (PROTOCOL §2.7)."
instance_seed = 0 # scene = environment/instance.json (the full scene state after the seeded reset + its hash); identity = dataset.toml digest
diff --git a/tasks/robolab-larger-object-raisin-box-in-bin-i00-standard/task.toml b/tasks/robolab-larger-object-raisin-box-in-bin-i00-standard/task.toml
index 66370d16f4..6c34ccf7d9 100644
--- a/tasks/robolab-larger-object-raisin-box-in-bin-i00-standard/task.toml
+++ b/tasks/robolab-larger-object-raisin-box-in-bin-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "robolab"
robot = "franka-robotiq-2f85"
family = "larger_object_raisin_box_in_bin"
mode = "standard" # robot as a service: RGB, depth + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution: RoboLab ships no expert or planner. Positive examples graded 1 by the separate verifier: robolab-larger-object-raisin-box__eJa8kky (robolab-larger-object-raisin-box-in-bin-i00-privileged, run codex-gpt6_luna-xhigh, robolab-lunacx-0925-2038, ws3; https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/larger_object_raisin_box_in_bin/unlimited/). Human review per family (PROTOCOL §2.7)."
diff --git a/tasks/robolab-marker-in-mug-i00-privileged/task.toml b/tasks/robolab-marker-in-mug-i00-privileged/task.toml
index 7ae2288d1c..91a0b17d18 100644
--- a/tasks/robolab-marker-in-mug-i00-privileged/task.toml
+++ b/tasks/robolab-marker-in-mug-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "robolab" # NVIDIA RoboLab v0.3.1 on Isaac Lab 2.3.2 / I
robot = "franka-robotiq-2f85"
family = "marker_in_mug"
mode = "privileged" # privileged state, resets and primitives through the tooling (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution: RoboLab ships no expert or planner. Positive examples graded 1 by the separate verifier: robolab-marker-in-mug-i00__2vRfTHm (robolab-marker-in-mug-i00-privileged, run codex-gpt6_luna-xhigh, robolab-lunacx-0925-2038, ws; https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/marker_in_mug/unlimited/); robolab-marker-in-mug-i00-limite__qHzxixK (robolab-marker-in-mug-i00-standard, run codex-gpt6_luna-xhigh, robolab-lunaxh-limited-0928, lab; https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/marker_in_mug/limited/). Human review per family (PROTOCOL §2.7)."
instance_seed = 0 # scene = environment/instance.json (the full scene state after the seeded reset + its hash); identity = dataset.toml digest
diff --git a/tasks/robolab-marker-in-mug-i00-standard/task.toml b/tasks/robolab-marker-in-mug-i00-standard/task.toml
index 805d2e0630..a8a8602176 100644
--- a/tasks/robolab-marker-in-mug-i00-standard/task.toml
+++ b/tasks/robolab-marker-in-mug-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "robolab"
robot = "franka-robotiq-2f85"
family = "marker_in_mug"
mode = "standard" # robot as a service: RGB, depth + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution: RoboLab ships no expert or planner. Positive examples graded 1 by the separate verifier: robolab-marker-in-mug-i00__2vRfTHm (robolab-marker-in-mug-i00-privileged, run codex-gpt6_luna-xhigh, robolab-lunacx-0925-2038, ws; https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/marker_in_mug/unlimited/); robolab-marker-in-mug-i00-limite__qHzxixK (robolab-marker-in-mug-i00-standard, run codex-gpt6_luna-xhigh, robolab-lunaxh-limited-0928, lab; https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/marker_in_mug/limited/). Human review per family (PROTOCOL §2.7)."
diff --git a/tasks/robolab-mustard-above-raisin-i00-privileged/task.toml b/tasks/robolab-mustard-above-raisin-i00-privileged/task.toml
index 74cba70325..359763d822 100644
--- a/tasks/robolab-mustard-above-raisin-i00-privileged/task.toml
+++ b/tasks/robolab-mustard-above-raisin-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "robolab" # NVIDIA RoboLab v0.3.1 on Isaac Lab 2.3.2 / I
robot = "franka-robotiq-2f85"
family = "mustard_above_raisin"
mode = "privileged" # privileged state, resets and primitives through the tooling (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution: RoboLab ships no expert or planner. Positive examples graded 1 by the separate verifier: robolab-mustard-above-raisin-i00__innKeS2 (robolab-mustard-above-raisin-i00-privileged, run codex-gpt6_luna-xhigh, robolab-lunacx-0925-2038, ws; https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/mustard_above_raisin/unlimited/); robolab-mustard-above-raisin-i00__YmfXNsC (robolab-mustard-above-raisin-i00-standard, run codex-gpt6_luna-xhigh, robolab-lunacx-0925-2038, ws3; not published: its eai-standard/2.1 batch was withdrawn for the agent's stale wrist-camera pose; grader unaffected). Human review per family (PROTOCOL §2.7)."
instance_seed = 0 # scene = environment/instance.json (the full scene state after the seeded reset + its hash); identity = dataset.toml digest
diff --git a/tasks/robolab-mustard-above-raisin-i00-standard/task.toml b/tasks/robolab-mustard-above-raisin-i00-standard/task.toml
index 8876f18ea1..20fc7d652f 100644
--- a/tasks/robolab-mustard-above-raisin-i00-standard/task.toml
+++ b/tasks/robolab-mustard-above-raisin-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "robolab"
robot = "franka-robotiq-2f85"
family = "mustard_above_raisin"
mode = "standard" # robot as a service: RGB, depth + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution: RoboLab ships no expert or planner. Positive examples graded 1 by the separate verifier: robolab-mustard-above-raisin-i00__innKeS2 (robolab-mustard-above-raisin-i00-privileged, run codex-gpt6_luna-xhigh, robolab-lunacx-0925-2038, ws; https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/mustard_above_raisin/unlimited/); robolab-mustard-above-raisin-i00__YmfXNsC (robolab-mustard-above-raisin-i00-standard, run codex-gpt6_luna-xhigh, robolab-lunacx-0925-2038, ws3; not published: its eai-standard/2.1 batch was withdrawn for the agent's stale wrist-camera pose; grader unaffected). Human review per family (PROTOCOL §2.7)."
diff --git a/tasks/robolab-mustard-in-left-bin-i00-privileged/task.toml b/tasks/robolab-mustard-in-left-bin-i00-privileged/task.toml
index 97fd3a2cd0..5e20e02ca8 100644
--- a/tasks/robolab-mustard-in-left-bin-i00-privileged/task.toml
+++ b/tasks/robolab-mustard-in-left-bin-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "robolab" # NVIDIA RoboLab v0.3.1 on Isaac Lab 2.3.2 / I
robot = "franka-robotiq-2f85"
family = "mustard_in_left_bin"
mode = "privileged" # privileged state, resets and primitives through the tooling (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution: RoboLab ships no expert or planner. Positive examples graded 1 by the separate verifier: robolab-mustard-in-left-bin-i00__JQUwBDv (robolab-mustard-in-left-bin-i00-privileged, run codex-gpt6_luna-xhigh, robolab-lunacx-0925-2038, ws; https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/mustard_in_left_bin/unlimited/). Human review per family (PROTOCOL §2.7)."
instance_seed = 0 # scene = environment/instance.json (the full scene state after the seeded reset + its hash); identity = dataset.toml digest
diff --git a/tasks/robolab-mustard-in-left-bin-i00-standard/task.toml b/tasks/robolab-mustard-in-left-bin-i00-standard/task.toml
index d6dedc2c58..ab0e585cd2 100644
--- a/tasks/robolab-mustard-in-left-bin-i00-standard/task.toml
+++ b/tasks/robolab-mustard-in-left-bin-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "robolab"
robot = "franka-robotiq-2f85"
family = "mustard_in_left_bin"
mode = "standard" # robot as a service: RGB, depth + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution: RoboLab ships no expert or planner. Positive examples graded 1 by the separate verifier: robolab-mustard-in-left-bin-i00__JQUwBDv (robolab-mustard-in-left-bin-i00-privileged, run codex-gpt6_luna-xhigh, robolab-lunacx-0925-2038, ws; https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/mustard_in_left_bin/unlimited/). Human review per family (PROTOCOL §2.7)."
diff --git a/tasks/robolab-mustard-in-right-bin-i00-privileged/task.toml b/tasks/robolab-mustard-in-right-bin-i00-privileged/task.toml
index 9bd4c66d61..649804d90c 100644
--- a/tasks/robolab-mustard-in-right-bin-i00-privileged/task.toml
+++ b/tasks/robolab-mustard-in-right-bin-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "robolab" # NVIDIA RoboLab v0.3.1 on Isaac Lab 2.3.2 / I
robot = "franka-robotiq-2f85"
family = "mustard_in_right_bin"
mode = "privileged" # privileged state, resets and primitives through the tooling (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution: RoboLab ships no expert or planner. Positive examples graded 1 by the separate verifier: robolab-mustard-in-right-bin-i00__XRQVDHq (robolab-mustard-in-right-bin-i00-privileged, run codex-gpt6_luna-xhigh, robolab-lunacx-0925-2038, ws; https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/mustard_in_right_bin/unlimited/). Human review per family (PROTOCOL §2.7)."
instance_seed = 0 # scene = environment/instance.json (the full scene state after the seeded reset + its hash); identity = dataset.toml digest
diff --git a/tasks/robolab-mustard-in-right-bin-i00-standard/task.toml b/tasks/robolab-mustard-in-right-bin-i00-standard/task.toml
index 5a529d10da..ed2316b8aa 100644
--- a/tasks/robolab-mustard-in-right-bin-i00-standard/task.toml
+++ b/tasks/robolab-mustard-in-right-bin-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "robolab"
robot = "franka-robotiq-2f85"
family = "mustard_in_right_bin"
mode = "standard" # robot as a service: RGB, depth + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution: RoboLab ships no expert or planner. Positive examples graded 1 by the separate verifier: robolab-mustard-in-right-bin-i00__XRQVDHq (robolab-mustard-in-right-bin-i00-privileged, run codex-gpt6_luna-xhigh, robolab-lunacx-0925-2038, ws; https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/mustard_in_right_bin/unlimited/). Human review per family (PROTOCOL §2.7)."
diff --git a/tasks/robolab-pick-glasses-i00-privileged/task.toml b/tasks/robolab-pick-glasses-i00-privileged/task.toml
index db3058e3c1..f816505670 100644
--- a/tasks/robolab-pick-glasses-i00-privileged/task.toml
+++ b/tasks/robolab-pick-glasses-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "robolab" # NVIDIA RoboLab v0.3.1 on Isaac Lab 2.3.2 / I
robot = "franka-robotiq-2f85"
family = "pick_glasses"
mode = "privileged" # privileged state, resets and primitives through the tooling (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution: RoboLab ships no expert or planner. Positive examples graded 1 by the separate verifier: robolab-pick-glasses-i00__dSrLHEK (robolab-pick-glasses-i00-privileged, run codex-gpt6_luna-xhigh, robolab-lunacx-0925-2038, ws; https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/pick_glasses/unlimited/); robolab-pick-glasses-i00-limited__8hJ4rxF (robolab-pick-glasses-i00-standard, run codex-gpt6_luna-xhigh, robolab-lunaxh-limited-0928, lab; https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/pick_glasses/limited/). Human review per family (PROTOCOL §2.7)."
instance_seed = 0 # scene = environment/instance.json (the full scene state after the seeded reset + its hash); identity = dataset.toml digest
diff --git a/tasks/robolab-pick-glasses-i00-standard/task.toml b/tasks/robolab-pick-glasses-i00-standard/task.toml
index 6ebfe84938..8e397a381a 100644
--- a/tasks/robolab-pick-glasses-i00-standard/task.toml
+++ b/tasks/robolab-pick-glasses-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "robolab"
robot = "franka-robotiq-2f85"
family = "pick_glasses"
mode = "standard" # robot as a service: RGB, depth + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution: RoboLab ships no expert or planner. Positive examples graded 1 by the separate verifier: robolab-pick-glasses-i00__dSrLHEK (robolab-pick-glasses-i00-privileged, run codex-gpt6_luna-xhigh, robolab-lunacx-0925-2038, ws; https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/pick_glasses/unlimited/); robolab-pick-glasses-i00-limited__8hJ4rxF (robolab-pick-glasses-i00-standard, run codex-gpt6_luna-xhigh, robolab-lunaxh-limited-0928, lab; https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/pick_glasses/limited/). Human review per family (PROTOCOL §2.7)."
diff --git a/tasks/robolab-pick-orange-object-i00-privileged/task.toml b/tasks/robolab-pick-orange-object-i00-privileged/task.toml
index 329d06a15e..1d71c78c09 100644
--- a/tasks/robolab-pick-orange-object-i00-privileged/task.toml
+++ b/tasks/robolab-pick-orange-object-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "robolab" # NVIDIA RoboLab v0.3.1 on Isaac Lab 2.3.2 / I
robot = "franka-robotiq-2f85"
family = "pick_orange_object"
mode = "privileged" # privileged state, resets and primitives through the tooling (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution: RoboLab ships no expert or planner. Positive examples graded 1 by the separate verifier: robolab-pick-orange-object-i00__oPGJais (robolab-pick-orange-object-i00-privileged, run codex-gpt6_luna-xhigh, robolab-lunacx-0925-2038, ws; https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/pick_orange_object/unlimited/); robolab-pick-orange-object-i00-l__e2h47FK (robolab-pick-orange-object-i00-standard, run codex-gpt6_luna-xhigh, robolab-lunaxh-limited-0928, lab; https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/pick_orange_object/limited/). Human review per family (PROTOCOL §2.7)."
instance_seed = 0 # scene = environment/instance.json (the full scene state after the seeded reset + its hash); identity = dataset.toml digest
diff --git a/tasks/robolab-pick-orange-object-i00-standard/task.toml b/tasks/robolab-pick-orange-object-i00-standard/task.toml
index 418de2a35b..bc1b031111 100644
--- a/tasks/robolab-pick-orange-object-i00-standard/task.toml
+++ b/tasks/robolab-pick-orange-object-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "robolab"
robot = "franka-robotiq-2f85"
family = "pick_orange_object"
mode = "standard" # robot as a service: RGB, depth + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution: RoboLab ships no expert or planner. Positive examples graded 1 by the separate verifier: robolab-pick-orange-object-i00__oPGJais (robolab-pick-orange-object-i00-privileged, run codex-gpt6_luna-xhigh, robolab-lunacx-0925-2038, ws; https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/pick_orange_object/unlimited/); robolab-pick-orange-object-i00-l__e2h47FK (robolab-pick-orange-object-i00-standard, run codex-gpt6_luna-xhigh, robolab-lunaxh-limited-0928, lab; https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/pick_orange_object/limited/). Human review per family (PROTOCOL §2.7)."
diff --git a/tasks/robolab-pick-up-blue-pitcher-i00-privileged/task.toml b/tasks/robolab-pick-up-blue-pitcher-i00-privileged/task.toml
index 7d3a932efa..e78b1d3b54 100644
--- a/tasks/robolab-pick-up-blue-pitcher-i00-privileged/task.toml
+++ b/tasks/robolab-pick-up-blue-pitcher-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "robolab" # NVIDIA RoboLab v0.3.1 on Isaac Lab 2.3.2 / I
robot = "franka-robotiq-2f85"
family = "pick_up_blue_pitcher"
mode = "privileged" # privileged state, resets and primitives through the tooling (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution: RoboLab ships no expert or planner. Positive examples graded 1 by the separate verifier: robolab-pick-up-blue-pitcher-i00__uX7zauo (robolab-pick-up-blue-pitcher-i00-privileged, run codex-gpt6_luna-xhigh, robolab-lunacx-0925-2038, ws; https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/pick_up_blue_pitcher/unlimited/). Human review per family (PROTOCOL §2.7)."
instance_seed = 0 # scene = environment/instance.json (the full scene state after the seeded reset + its hash); identity = dataset.toml digest
diff --git a/tasks/robolab-pick-up-blue-pitcher-i00-standard/task.toml b/tasks/robolab-pick-up-blue-pitcher-i00-standard/task.toml
index 58967cf6b8..9ac307dd90 100644
--- a/tasks/robolab-pick-up-blue-pitcher-i00-standard/task.toml
+++ b/tasks/robolab-pick-up-blue-pitcher-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "robolab"
robot = "franka-robotiq-2f85"
family = "pick_up_blue_pitcher"
mode = "standard" # robot as a service: RGB, depth + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution: RoboLab ships no expert or planner. Positive examples graded 1 by the separate verifier: robolab-pick-up-blue-pitcher-i00__uX7zauo (robolab-pick-up-blue-pitcher-i00-privileged, run codex-gpt6_luna-xhigh, robolab-lunacx-0925-2038, ws; https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/pick_up_blue_pitcher/unlimited/). Human review per family (PROTOCOL §2.7)."
diff --git a/tasks/robolab-plastic-bottles-in-square-pail-i00-privileged/task.toml b/tasks/robolab-plastic-bottles-in-square-pail-i00-privileged/task.toml
index d32548f026..11270959bf 100644
--- a/tasks/robolab-plastic-bottles-in-square-pail-i00-privileged/task.toml
+++ b/tasks/robolab-plastic-bottles-in-square-pail-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "robolab" # NVIDIA RoboLab v0.3.1 on Isaac Lab 2.3.2 / I
robot = "franka-robotiq-2f85"
family = "plastic_bottles_in_square_pail"
mode = "privileged" # privileged state, resets and primitives through the tooling (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution: RoboLab ships no expert or planner. Positive examples graded 1 by the separate verifier: robolab-plastic-bottles-in-squar__Vd88MFD (robolab-plastic-bottles-in-square-pail-i00-privileged, run codex-gpt6_luna-xhigh, robolab-lunacx-0925-2038, ws3; https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/plastic_bottles_in_square_pail/unlimited/). Human review per family (PROTOCOL §2.7)."
instance_seed = 0 # scene = environment/instance.json (the full scene state after the seeded reset + its hash); identity = dataset.toml digest
diff --git a/tasks/robolab-plastic-bottles-in-square-pail-i00-standard/README.md b/tasks/robolab-plastic-bottles-in-square-pail-i00-standard/README.md
index 48b03a8b9d..58d90d63b4 100644
--- a/tasks/robolab-plastic-bottles-in-square-pail-i00-standard/README.md
+++ b/tasks/robolab-plastic-bottles-in-square-pail-i00-standard/README.md
@@ -72,3 +72,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
- 2026-10-10 (review round 3, board 0067): the shared service layer and the client only (`environment/sim/rcb_limited.py`, `environment/rcb_client.py`). A camera named more than once in `cameras` is rendered and returned once (N copies of one name made the service build N x N images within the frame budget, which could exhaust its memory). A service that has given up on the episode (it could not start it, or could not resume it after its crashes) answers every request with 409 and the reason, and `Robot()` / `wait_ready()` raise `EpisodeOver` at once; before, nothing answered and they waited out their timeout (up to an hour). The adapter, the collect hook, the grader, the scene and the pass rule are unchanged.
- 2026-10-10 (owner decision, board 0072): the agent's network egress is restricted to the model APIs (`[agent] network_mode = "allowlist"`, the hosts of `scripts/set_agent_egress.py`; since 2026-10-11 these include the package registries, PyPI and the Debian / Ubuntu archives, so that `pip install` and `apt-get install` work); it was open before (Harbor's default), and agents could download upstream sources, reference solutions and robot models from their own shell. The simulator image is `eai-robolab:0.1.2` (was 0.1.1): it holds the MDL materials and textures that RoboLab's asset library fetched from NVIDIA's Omniverse content bucket whenever a scene was built, so that the cameras show under the allowlist what they showed with open egress (on 0.1.1 without that download the table renders red). The scene, the frozen instance, the physics, the tooling and the pass rule are unchanged, so the task keeps its name; results from before this date were obtained with open egress.
- 2026-10-13 (owner, boards 0085 and 0087): the sim service's clock (`environment/sim/rcb_limited.py`, `environment/sim/server.py`, the docstring of `environment/rcb_client.py`). `meta["time_s"]` and `meta["time_left_s"]` count from the agent's first request to the service (any endpoint), and a restarted service keeps that moment; before, they counted from the service's own start of the episode (when its scene was ready). Where the service is ready before the agent's phase begins (the harness is installed in between: a median of 79 s over the recorded standard trials), that clock ran ahead of the agent's and `time_left_s` reached 0 while the agent still had that time. `/spec` says what the clock counts from (`budget.clock`), and the instruction's `## Time limit` tells the agent to plan by `time_left_s`. The service never ended an episode on this clock: the scene, the frozen instance, the pass rule and the grader are unchanged, so the task keeps its name and earlier results stand.
+- 2026-10-14 (board 0113, from board 0086): the success text says that the gripper must not be touching the objects at the moment of success ("at that moment"); the old wording "the gripper must touch none of them" was read by an agent as "never touch them" (board 0086). Text only: the scene, the frozen instance, RoboLab's success term and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/robolab-plastic-bottles-in-square-pail-i00-standard/instruction.md b/tasks/robolab-plastic-bottles-in-square-pail-i00-standard/instruction.md
index 221c39e100..441bb7926c 100644
--- a/tasks/robolab-plastic-bottles-in-square-pail-i00-standard/instruction.md
+++ b/tasks/robolab-plastic-bottles-in-square-pail-i00-standard/instruction.md
@@ -1,7 +1,7 @@
# RoboLab · plastic_bottles_in_square_pail — STANDARD MODE (robot as a service, eai-standard/2.2)
You control a **Franka Panda arm with a Robotiq 2F-85 gripper** (NVIDIA RoboLab's DROID setup) **as if it were real
-hardware**: the simulator runs in a service you cannot open. You get one client and one episode. **Task:** Put all the small plastic bottles in the square pail. Success: The three small white plastic bottles must end inside the square pail and the gripper must touch none of them. "Inside" means the centre of the object's convex hull lies within the container's convex hull, extended without limit along the container's own up axis (the open side; straight up for a container standing normally).
+hardware**: the simulator runs in a service you cannot open. You get one client and one episode. **Task:** Put all the small plastic bottles in the square pail. Success: The three small white plastic bottles must end inside the square pail and, at that moment, the gripper must touch none of them (moving them with the gripper is how the task is done; let go of them at the end). "Inside" means the centre of the object's convex hull lies within the container's convex hull, extended without limit along the container's own up axis (the open side; straight up for a container standing normally).
## Rules (real-robot rules)
- **No reset.** The episode starts when the service comes up and ends on success (at the first control step where the
diff --git a/tasks/robolab-plastic-bottles-in-square-pail-i00-standard/task.toml b/tasks/robolab-plastic-bottles-in-square-pail-i00-standard/task.toml
index e1ae1ef10c..4249126a97 100644
--- a/tasks/robolab-plastic-bottles-in-square-pail-i00-standard/task.toml
+++ b/tasks/robolab-plastic-bottles-in-square-pail-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "robolab"
robot = "franka-robotiq-2f85"
family = "plastic_bottles_in_square_pail"
mode = "standard" # robot as a service: RGB, depth + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution: RoboLab ships no expert or planner. Positive examples graded 1 by the separate verifier: robolab-plastic-bottles-in-squar__Vd88MFD (robolab-plastic-bottles-in-square-pail-i00-privileged, run codex-gpt6_luna-xhigh, robolab-lunacx-0925-2038, ws3; https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/plastic_bottles_in_square_pail/unlimited/). Human review per family (PROTOCOL §2.7)."
diff --git a/tasks/robolab-put-two-mugs-on-shelf-i00-privileged/task.toml b/tasks/robolab-put-two-mugs-on-shelf-i00-privileged/task.toml
index f3412aa406..c575b00b3c 100644
--- a/tasks/robolab-put-two-mugs-on-shelf-i00-privileged/task.toml
+++ b/tasks/robolab-put-two-mugs-on-shelf-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "robolab" # NVIDIA RoboLab v0.3.1 on Isaac Lab 2.3.2 / I
robot = "franka-robotiq-2f85"
family = "put_two_mugs_on_shelf"
mode = "privileged" # privileged state, resets and primitives through the tooling (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution: RoboLab ships no expert or planner. Positive examples graded 1 by the separate verifier: robolab-put-two-mugs-on-shelf-i0__AMuARWT (robolab-put-two-mugs-on-shelf-i00-privileged, run codex-gpt6_luna-xhigh, robolab-lunacx-0925-2038, ws3; https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/put_two_mugs_on_shelf/unlimited/). Human review per family (PROTOCOL §2.7)."
instance_seed = 0 # scene = environment/instance.json (the full scene state after the seeded reset + its hash); identity = dataset.toml digest
diff --git a/tasks/robolab-put-two-mugs-on-shelf-i00-standard/README.md b/tasks/robolab-put-two-mugs-on-shelf-i00-standard/README.md
index fe3c1a0f4a..294cfc6a6c 100644
--- a/tasks/robolab-put-two-mugs-on-shelf-i00-standard/README.md
+++ b/tasks/robolab-put-two-mugs-on-shelf-i00-standard/README.md
@@ -72,3 +72,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
- 2026-10-10 (review round 3, board 0067): the shared service layer and the client only (`environment/sim/rcb_limited.py`, `environment/rcb_client.py`). A camera named more than once in `cameras` is rendered and returned once (N copies of one name made the service build N x N images within the frame budget, which could exhaust its memory). A service that has given up on the episode (it could not start it, or could not resume it after its crashes) answers every request with 409 and the reason, and `Robot()` / `wait_ready()` raise `EpisodeOver` at once; before, nothing answered and they waited out their timeout (up to an hour). The adapter, the collect hook, the grader, the scene and the pass rule are unchanged.
- 2026-10-10 (owner decision, board 0072): the agent's network egress is restricted to the model APIs (`[agent] network_mode = "allowlist"`, the hosts of `scripts/set_agent_egress.py`; since 2026-10-11 these include the package registries, PyPI and the Debian / Ubuntu archives, so that `pip install` and `apt-get install` work); it was open before (Harbor's default), and agents could download upstream sources, reference solutions and robot models from their own shell. The simulator image is `eai-robolab:0.1.2` (was 0.1.1): it holds the MDL materials and textures that RoboLab's asset library fetched from NVIDIA's Omniverse content bucket whenever a scene was built, so that the cameras show under the allowlist what they showed with open egress (on 0.1.1 without that download the table renders red). The scene, the frozen instance, the physics, the tooling and the pass rule are unchanged, so the task keeps its name; results from before this date were obtained with open egress.
- 2026-10-13 (owner, boards 0085 and 0087): the sim service's clock (`environment/sim/rcb_limited.py`, `environment/sim/server.py`, the docstring of `environment/rcb_client.py`). `meta["time_s"]` and `meta["time_left_s"]` count from the agent's first request to the service (any endpoint), and a restarted service keeps that moment; before, they counted from the service's own start of the episode (when its scene was ready). Where the service is ready before the agent's phase begins (the harness is installed in between: a median of 79 s over the recorded standard trials), that clock ran ahead of the agent's and `time_left_s` reached 0 while the agent still had that time. `/spec` says what the clock counts from (`budget.clock`), and the instruction's `## Time limit` tells the agent to plan by `time_left_s`. The service never ended an episode on this clock: the scene, the frozen instance, the pass rule and the grader are unchanged, so the task keeps its name and earlier results stand.
+- 2026-10-14 (board 0113, from board 0086): the success text says that the gripper must not be touching the objects at the moment of success ("at that moment"); the old wording "the gripper must touch none of them" was read by an agent as "never touch them" (board 0086). Text only: the scene, the frozen instance, RoboLab's success term and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/robolab-put-two-mugs-on-shelf-i00-standard/instruction.md b/tasks/robolab-put-two-mugs-on-shelf-i00-standard/instruction.md
index 49302b206e..c0697e9762 100644
--- a/tasks/robolab-put-two-mugs-on-shelf-i00-standard/instruction.md
+++ b/tasks/robolab-put-two-mugs-on-shelf-i00-standard/instruction.md
@@ -1,7 +1,7 @@
# RoboLab · put_two_mugs_on_shelf — STANDARD MODE (robot as a service, eai-standard/2.2)
You control a **Franka Panda arm with a Robotiq 2F-85 gripper** (NVIDIA RoboLab's DROID setup) **as if it were real
-hardware**: the simulator runs in a service you cannot open. You get one client and one episode. **Task:** Put two (2) mugs on the wire shelf. Success: Exactly 2 of the three mugs (no more, no fewer) must end on the wire shelf, touching it, and the gripper must touch none of them. "Inside" means the centre of the object's convex hull lies within the container's convex hull, extended without limit along the container's own up axis (the open side; straight up for a container standing normally). (RoboLab checks "on" the wire shelf this way.)
+hardware**: the simulator runs in a service you cannot open. You get one client and one episode. **Task:** Put two (2) mugs on the wire shelf. Success: Exactly 2 of the three mugs (no more, no fewer) must end on the wire shelf, touching it, and, at that moment, the gripper must touch none of them (moving them with the gripper is how the task is done; let go of them at the end). "Inside" means the centre of the object's convex hull lies within the container's convex hull, extended without limit along the container's own up axis (the open side; straight up for a container standing normally). (RoboLab checks "on" the wire shelf this way.)
## Rules (real-robot rules)
- **No reset.** The episode starts when the service comes up and ends on success (at the first control step where the
diff --git a/tasks/robolab-put-two-mugs-on-shelf-i00-standard/task.toml b/tasks/robolab-put-two-mugs-on-shelf-i00-standard/task.toml
index 27601738a8..2fa2ec7ce2 100644
--- a/tasks/robolab-put-two-mugs-on-shelf-i00-standard/task.toml
+++ b/tasks/robolab-put-two-mugs-on-shelf-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "robolab"
robot = "franka-robotiq-2f85"
family = "put_two_mugs_on_shelf"
mode = "standard" # robot as a service: RGB, depth + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution: RoboLab ships no expert or planner. Positive examples graded 1 by the separate verifier: robolab-put-two-mugs-on-shelf-i0__AMuARWT (robolab-put-two-mugs-on-shelf-i00-privileged, run codex-gpt6_luna-xhigh, robolab-lunacx-0925-2038, ws3; https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/put_two_mugs_on_shelf/unlimited/). Human review per family (PROTOCOL §2.7)."
diff --git a/tasks/robolab-recycle-carton-i00-privileged/task.toml b/tasks/robolab-recycle-carton-i00-privileged/task.toml
index 89671b9c5f..19672bb91d 100644
--- a/tasks/robolab-recycle-carton-i00-privileged/task.toml
+++ b/tasks/robolab-recycle-carton-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "robolab" # NVIDIA RoboLab v0.3.1 on Isaac Lab 2.3.2 / I
robot = "franka-robotiq-2f85"
family = "recycle_carton"
mode = "privileged" # privileged state, resets and primitives through the tooling (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution: RoboLab ships no expert or planner. Positive examples graded 1 by the separate verifier: robolab-recycle-carton-i00__5ZZZnYE (robolab-recycle-carton-i00-privileged, run codex-gpt6_luna-xhigh, robolab-lunacx-0925-2038, ws3; https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/recycle_carton/unlimited/). Human review per family (PROTOCOL §2.7)."
instance_seed = 0 # scene = environment/instance.json (the full scene state after the seeded reset + its hash); identity = dataset.toml digest
diff --git a/tasks/robolab-recycle-carton-i00-standard/README.md b/tasks/robolab-recycle-carton-i00-standard/README.md
index f563a0b7c8..6b99457ae7 100644
--- a/tasks/robolab-recycle-carton-i00-standard/README.md
+++ b/tasks/robolab-recycle-carton-i00-standard/README.md
@@ -72,3 +72,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
- 2026-10-10 (review round 3, board 0067): the shared service layer and the client only (`environment/sim/rcb_limited.py`, `environment/rcb_client.py`). A camera named more than once in `cameras` is rendered and returned once (N copies of one name made the service build N x N images within the frame budget, which could exhaust its memory). A service that has given up on the episode (it could not start it, or could not resume it after its crashes) answers every request with 409 and the reason, and `Robot()` / `wait_ready()` raise `EpisodeOver` at once; before, nothing answered and they waited out their timeout (up to an hour). The adapter, the collect hook, the grader, the scene and the pass rule are unchanged.
- 2026-10-10 (owner decision, board 0072): the agent's network egress is restricted to the model APIs (`[agent] network_mode = "allowlist"`, the hosts of `scripts/set_agent_egress.py`; since 2026-10-11 these include the package registries, PyPI and the Debian / Ubuntu archives, so that `pip install` and `apt-get install` work); it was open before (Harbor's default), and agents could download upstream sources, reference solutions and robot models from their own shell. The simulator image is `eai-robolab:0.1.2` (was 0.1.1): it holds the MDL materials and textures that RoboLab's asset library fetched from NVIDIA's Omniverse content bucket whenever a scene was built, so that the cameras show under the allowlist what they showed with open egress (on 0.1.1 without that download the table renders red). The scene, the frozen instance, the physics, the tooling and the pass rule are unchanged, so the task keeps its name; results from before this date were obtained with open egress.
- 2026-10-13 (owner, boards 0085 and 0087): the sim service's clock (`environment/sim/rcb_limited.py`, `environment/sim/server.py`, the docstring of `environment/rcb_client.py`). `meta["time_s"]` and `meta["time_left_s"]` count from the agent's first request to the service (any endpoint), and a restarted service keeps that moment; before, they counted from the service's own start of the episode (when its scene was ready). Where the service is ready before the agent's phase begins (the harness is installed in between: a median of 79 s over the recorded standard trials), that clock ran ahead of the agent's and `time_left_s` reached 0 while the agent still had that time. `/spec` says what the clock counts from (`budget.clock`), and the instruction's `## Time limit` tells the agent to plan by `time_left_s`. The service never ended an episode on this clock: the scene, the frozen instance, the pass rule and the grader are unchanged, so the task keeps its name and earlier results stand.
+- 2026-10-14 (board 0113, from board 0086): the success text says that the gripper must not be touching the objects at the moment of success ("at that moment"); the old wording "the gripper must touch none of them" was read by an agent as "never touch them" (board 0086). Text only: the scene, the frozen instance, RoboLab's success term and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/robolab-recycle-carton-i00-standard/instruction.md b/tasks/robolab-recycle-carton-i00-standard/instruction.md
index 70ea2cba8f..efe740c3ff 100644
--- a/tasks/robolab-recycle-carton-i00-standard/instruction.md
+++ b/tasks/robolab-recycle-carton-i00-standard/instruction.md
@@ -1,7 +1,7 @@
# RoboLab · recycle_carton — STANDARD MODE (robot as a service, eai-standard/2.2)
You control a **Franka Panda arm with a Robotiq 2F-85 gripper** (NVIDIA RoboLab's DROID setup) **as if it were real
-hardware**: the simulator runs in a service you cannot open. You get one client and one episode. **Task:** Put the recyclable cartons in the grey bin. Success: The milk carton and the orange juice carton must end inside the grey bin in the centre of the table and the gripper must touch neither. "Inside" means the centre of the object's convex hull lies within the container's convex hull, extended without limit along the container's own up axis (the open side; straight up for a container standing normally).
+hardware**: the simulator runs in a service you cannot open. You get one client and one episode. **Task:** Put the recyclable cartons in the grey bin. Success: The milk carton and the orange juice carton must end inside the grey bin in the centre of the table and, at that moment, the gripper must touch neither (let go of both at the end). "Inside" means the centre of the object's convex hull lies within the container's convex hull, extended without limit along the container's own up axis (the open side; straight up for a container standing normally).
## Rules (real-robot rules)
- **No reset.** The episode starts when the service comes up and ends on success (at the first control step where the
diff --git a/tasks/robolab-recycle-carton-i00-standard/task.toml b/tasks/robolab-recycle-carton-i00-standard/task.toml
index 44d80e086a..7d6e0ddce1 100644
--- a/tasks/robolab-recycle-carton-i00-standard/task.toml
+++ b/tasks/robolab-recycle-carton-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "robolab"
robot = "franka-robotiq-2f85"
family = "recycle_carton"
mode = "standard" # robot as a service: RGB, depth + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution: RoboLab ships no expert or planner. Positive examples graded 1 by the separate verifier: robolab-recycle-carton-i00__5ZZZnYE (robolab-recycle-carton-i00-privileged, run codex-gpt6_luna-xhigh, robolab-lunacx-0925-2038, ws3; https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/recycle_carton/unlimited/). Human review per family (PROTOCOL §2.7)."
diff --git a/tasks/robolab-recycle-cartons-vertical-crate-i00-privileged/task.toml b/tasks/robolab-recycle-cartons-vertical-crate-i00-privileged/task.toml
index 565ab46e39..c820f271d7 100644
--- a/tasks/robolab-recycle-cartons-vertical-crate-i00-privileged/task.toml
+++ b/tasks/robolab-recycle-cartons-vertical-crate-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "robolab" # NVIDIA RoboLab v0.3.1 on Isaac Lab 2.3.2 / I
robot = "franka-robotiq-2f85"
family = "recycle_cartons_vertical_crate"
mode = "privileged" # privileged state, resets and primitives through the tooling (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution: RoboLab ships no expert or planner. Positive examples graded 1 by the separate verifier: robolab-recycle-cartons-vertical__gBnrYp4 (robolab-recycle-cartons-vertical-crate-i00-privileged, run codex-gpt6_luna-xhigh, robolab-lunacx-0925-2038, ws3; https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/recycle_cartons_vertical_crate/unlimited/). Human review per family (PROTOCOL §2.7)."
instance_seed = 0 # scene = environment/instance.json (the full scene state after the seeded reset + its hash); identity = dataset.toml digest
diff --git a/tasks/robolab-recycle-cartons-vertical-crate-i00-standard/README.md b/tasks/robolab-recycle-cartons-vertical-crate-i00-standard/README.md
index 0dd1a56eae..f1a0e8faea 100644
--- a/tasks/robolab-recycle-cartons-vertical-crate-i00-standard/README.md
+++ b/tasks/robolab-recycle-cartons-vertical-crate-i00-standard/README.md
@@ -72,3 +72,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
- 2026-10-10 (review round 3, board 0067): the shared service layer and the client only (`environment/sim/rcb_limited.py`, `environment/rcb_client.py`). A camera named more than once in `cameras` is rendered and returned once (N copies of one name made the service build N x N images within the frame budget, which could exhaust its memory). A service that has given up on the episode (it could not start it, or could not resume it after its crashes) answers every request with 409 and the reason, and `Robot()` / `wait_ready()` raise `EpisodeOver` at once; before, nothing answered and they waited out their timeout (up to an hour). The adapter, the collect hook, the grader, the scene and the pass rule are unchanged.
- 2026-10-10 (owner decision, board 0072): the agent's network egress is restricted to the model APIs (`[agent] network_mode = "allowlist"`, the hosts of `scripts/set_agent_egress.py`; since 2026-10-11 these include the package registries, PyPI and the Debian / Ubuntu archives, so that `pip install` and `apt-get install` work); it was open before (Harbor's default), and agents could download upstream sources, reference solutions and robot models from their own shell. The simulator image is `eai-robolab:0.1.2` (was 0.1.1): it holds the MDL materials and textures that RoboLab's asset library fetched from NVIDIA's Omniverse content bucket whenever a scene was built, so that the cameras show under the allowlist what they showed with open egress (on 0.1.1 without that download the table renders red). The scene, the frozen instance, the physics, the tooling and the pass rule are unchanged, so the task keeps its name; results from before this date were obtained with open egress.
- 2026-10-13 (owner, boards 0085 and 0087): the sim service's clock (`environment/sim/rcb_limited.py`, `environment/sim/server.py`, the docstring of `environment/rcb_client.py`). `meta["time_s"]` and `meta["time_left_s"]` count from the agent's first request to the service (any endpoint), and a restarted service keeps that moment; before, they counted from the service's own start of the episode (when its scene was ready). Where the service is ready before the agent's phase begins (the harness is installed in between: a median of 79 s over the recorded standard trials), that clock ran ahead of the agent's and `time_left_s` reached 0 while the agent still had that time. `/spec` says what the clock counts from (`budget.clock`), and the instruction's `## Time limit` tells the agent to plan by `time_left_s`. The service never ended an episode on this clock: the scene, the frozen instance, the pass rule and the grader are unchanged, so the task keeps its name and earlier results stand.
+- 2026-10-14 (board 0113, from board 0086): the success text says that the gripper must not be touching the objects at the moment of success ("at that moment"); the old wording "the gripper must touch none of them" was read by an agent as "never touch them" (board 0086). Text only: the scene, the frozen instance, RoboLab's success term and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/robolab-recycle-cartons-vertical-crate-i00-standard/instruction.md b/tasks/robolab-recycle-cartons-vertical-crate-i00-standard/instruction.md
index 0368de6fe1..4e99ad81a8 100644
--- a/tasks/robolab-recycle-cartons-vertical-crate-i00-standard/instruction.md
+++ b/tasks/robolab-recycle-cartons-vertical-crate-i00-standard/instruction.md
@@ -1,7 +1,7 @@
# RoboLab · recycle_cartons_vertical_crate — STANDARD MODE (robot as a service, eai-standard/2.2)
You control a **Franka Panda arm with a Robotiq 2F-85 gripper** (NVIDIA RoboLab's DROID setup) **as if it were real
-hardware**: the simulator runs in a service you cannot open. You get one client and one episode. **Task:** Put the cartons that can be recycled in the vertical crate. Success: The milk carton and the orange juice carton must end inside the vertical crate in the centre of the table and the gripper must touch neither. "Inside" means the centre of the object's convex hull lies within the container's convex hull, extended without limit along the container's own up axis (the open side; straight up for a container standing normally).
+hardware**: the simulator runs in a service you cannot open. You get one client and one episode. **Task:** Put the cartons that can be recycled in the vertical crate. Success: The milk carton and the orange juice carton must end inside the vertical crate in the centre of the table and, at that moment, the gripper must touch neither (let go of both at the end). "Inside" means the centre of the object's convex hull lies within the container's convex hull, extended without limit along the container's own up axis (the open side; straight up for a container standing normally).
## Rules (real-robot rules)
- **No reset.** The episode starts when the service comes up and ends on success (at the first control step where the
diff --git a/tasks/robolab-recycle-cartons-vertical-crate-i00-standard/task.toml b/tasks/robolab-recycle-cartons-vertical-crate-i00-standard/task.toml
index c4ce5af097..cebaeb8548 100644
--- a/tasks/robolab-recycle-cartons-vertical-crate-i00-standard/task.toml
+++ b/tasks/robolab-recycle-cartons-vertical-crate-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "robolab"
robot = "franka-robotiq-2f85"
family = "recycle_cartons_vertical_crate"
mode = "standard" # robot as a service: RGB, depth + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution: RoboLab ships no expert or planner. Positive examples graded 1 by the separate verifier: robolab-recycle-cartons-vertical__gBnrYp4 (robolab-recycle-cartons-vertical-crate-i00-privileged, run codex-gpt6_luna-xhigh, robolab-lunacx-0925-2038, ws3; https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/recycle_cartons_vertical_crate/unlimited/). Human review per family (PROTOCOL §2.7)."
diff --git a/tasks/robolab-red-dishes-in-bin-i00-privileged/task.toml b/tasks/robolab-red-dishes-in-bin-i00-privileged/task.toml
index f056c0dd07..5397752f36 100644
--- a/tasks/robolab-red-dishes-in-bin-i00-privileged/task.toml
+++ b/tasks/robolab-red-dishes-in-bin-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "robolab" # NVIDIA RoboLab v0.3.1 on Isaac Lab 2.3.2 / I
robot = "franka-robotiq-2f85"
family = "red_dishes_in_bin"
mode = "privileged" # privileged state, resets and primitives through the tooling (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution: RoboLab ships no expert or planner. Positive examples graded 1 by the separate verifier: robolab-red-dishes-in-bin-i00__uX2TLJ9 (robolab-red-dishes-in-bin-i00-privileged, run codex-gpt6_luna-xhigh, robolab-lunacx-0925-2038, ws3; https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/red_dishes_in_bin/unlimited/). Human review per family (PROTOCOL §2.7)."
instance_seed = 0 # scene = environment/instance.json (the full scene state after the seeded reset + its hash); identity = dataset.toml digest
diff --git a/tasks/robolab-red-dishes-in-bin-i00-standard/README.md b/tasks/robolab-red-dishes-in-bin-i00-standard/README.md
index b5fad7cfca..1028855ca4 100644
--- a/tasks/robolab-red-dishes-in-bin-i00-standard/README.md
+++ b/tasks/robolab-red-dishes-in-bin-i00-standard/README.md
@@ -72,3 +72,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
- 2026-10-10 (review round 3, board 0067): the shared service layer and the client only (`environment/sim/rcb_limited.py`, `environment/rcb_client.py`). A camera named more than once in `cameras` is rendered and returned once (N copies of one name made the service build N x N images within the frame budget, which could exhaust its memory). A service that has given up on the episode (it could not start it, or could not resume it after its crashes) answers every request with 409 and the reason, and `Robot()` / `wait_ready()` raise `EpisodeOver` at once; before, nothing answered and they waited out their timeout (up to an hour). The adapter, the collect hook, the grader, the scene and the pass rule are unchanged.
- 2026-10-10 (owner decision, board 0072): the agent's network egress is restricted to the model APIs (`[agent] network_mode = "allowlist"`, the hosts of `scripts/set_agent_egress.py`; since 2026-10-11 these include the package registries, PyPI and the Debian / Ubuntu archives, so that `pip install` and `apt-get install` work); it was open before (Harbor's default), and agents could download upstream sources, reference solutions and robot models from their own shell. The simulator image is `eai-robolab:0.1.2` (was 0.1.1): it holds the MDL materials and textures that RoboLab's asset library fetched from NVIDIA's Omniverse content bucket whenever a scene was built, so that the cameras show under the allowlist what they showed with open egress (on 0.1.1 without that download the table renders red). The scene, the frozen instance, the physics, the tooling and the pass rule are unchanged, so the task keeps its name; results from before this date were obtained with open egress.
- 2026-10-13 (owner, boards 0085 and 0087): the sim service's clock (`environment/sim/rcb_limited.py`, `environment/sim/server.py`, the docstring of `environment/rcb_client.py`). `meta["time_s"]` and `meta["time_left_s"]` count from the agent's first request to the service (any endpoint), and a restarted service keeps that moment; before, they counted from the service's own start of the episode (when its scene was ready). Where the service is ready before the agent's phase begins (the harness is installed in between: a median of 79 s over the recorded standard trials), that clock ran ahead of the agent's and `time_left_s` reached 0 while the agent still had that time. `/spec` says what the clock counts from (`budget.clock`), and the instruction's `## Time limit` tells the agent to plan by `time_left_s`. The service never ended an episode on this clock: the scene, the frozen instance, the pass rule and the grader are unchanged, so the task keeps its name and earlier results stand.
+- 2026-10-14 (board 0113, from board 0086): the success text says that the gripper must not be touching the objects at the moment of success ("at that moment"); the old wording "the gripper must touch none of them" was read by an agent as "never touch them" (board 0086). Text only: the scene, the frozen instance, RoboLab's success term and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/robolab-red-dishes-in-bin-i00-standard/instruction.md b/tasks/robolab-red-dishes-in-bin-i00-standard/instruction.md
index d866162687..0defa0704a 100644
--- a/tasks/robolab-red-dishes-in-bin-i00-standard/instruction.md
+++ b/tasks/robolab-red-dishes-in-bin-i00-standard/instruction.md
@@ -1,7 +1,7 @@
# RoboLab · red_dishes_in_bin — STANDARD MODE (robot as a service, eai-standard/2.2)
You control a **Franka Panda arm with a Robotiq 2F-85 gripper** (NVIDIA RoboLab's DROID setup) **as if it were real
-hardware**: the simulator runs in a service you cannot open. You get one client and one episode. **Task:** Put the red dishware in the grey bin. Success: The red mug and the red bowl must end inside the grey bin and the gripper must touch neither. "Inside" means the centre of the object's convex hull lies within the container's convex hull, extended without limit along the container's own up axis (the open side; straight up for a container standing normally).
+hardware**: the simulator runs in a service you cannot open. You get one client and one episode. **Task:** Put the red dishware in the grey bin. Success: The red mug and the red bowl must end inside the grey bin and, at that moment, the gripper must touch neither (let go of both at the end). "Inside" means the centre of the object's convex hull lies within the container's convex hull, extended without limit along the container's own up axis (the open side; straight up for a container standing normally).
## Rules (real-robot rules)
- **No reset.** The episode starts when the service comes up and ends on success (at the first control step where the
diff --git a/tasks/robolab-red-dishes-in-bin-i00-standard/task.toml b/tasks/robolab-red-dishes-in-bin-i00-standard/task.toml
index c837c678ac..14b6c8a79c 100644
--- a/tasks/robolab-red-dishes-in-bin-i00-standard/task.toml
+++ b/tasks/robolab-red-dishes-in-bin-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "robolab"
robot = "franka-robotiq-2f85"
family = "red_dishes_in_bin"
mode = "standard" # robot as a service: RGB, depth + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution: RoboLab ships no expert or planner. Positive examples graded 1 by the separate verifier: robolab-red-dishes-in-bin-i00__uX2TLJ9 (robolab-red-dishes-in-bin-i00-privileged, run codex-gpt6_luna-xhigh, robolab-lunacx-0925-2038, ws3; https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/red_dishes_in_bin/unlimited/). Human review per family (PROTOCOL §2.7)."
diff --git a/tasks/robolab-reorient-red-mug-i00-privileged/task.toml b/tasks/robolab-reorient-red-mug-i00-privileged/task.toml
index 92abc6a790..fde644d1ee 100644
--- a/tasks/robolab-reorient-red-mug-i00-privileged/task.toml
+++ b/tasks/robolab-reorient-red-mug-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "robolab" # NVIDIA RoboLab v0.3.1 on Isaac Lab 2.3.2 / I
robot = "franka-robotiq-2f85"
family = "reorient_red_mug"
mode = "privileged" # privileged state, resets and primitives through the tooling (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution: RoboLab ships no expert or planner. Positive examples graded 1 by the separate verifier: robolab-reorient-red-mug-i00__vbq6nDY (robolab-reorient-red-mug-i00-privileged, run codex-gpt6_luna-xhigh, robolab-lunacx-0925-2038, ws3; https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/reorient_red_mug/unlimited/); robolab-reorient-red-mug-i00-lim__PUXCHM2 (robolab-reorient-red-mug-i00-standard, run codex-gpt6_luna-xhigh, robolab-lunaxh-limited-0928, lab; https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/reorient_red_mug/limited/). Human review per family (PROTOCOL §2.7)."
instance_seed = 0 # scene = environment/instance.json (the full scene state after the seeded reset + its hash); identity = dataset.toml digest
diff --git a/tasks/robolab-reorient-red-mug-i00-standard/task.toml b/tasks/robolab-reorient-red-mug-i00-standard/task.toml
index 4d5b6b5a2b..3e7b74f824 100644
--- a/tasks/robolab-reorient-red-mug-i00-standard/task.toml
+++ b/tasks/robolab-reorient-red-mug-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "robolab"
robot = "franka-robotiq-2f85"
family = "reorient_red_mug"
mode = "standard" # robot as a service: RGB, depth + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution: RoboLab ships no expert or planner. Positive examples graded 1 by the separate verifier: robolab-reorient-red-mug-i00__vbq6nDY (robolab-reorient-red-mug-i00-privileged, run codex-gpt6_luna-xhigh, robolab-lunacx-0925-2038, ws3; https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/reorient_red_mug/unlimited/); robolab-reorient-red-mug-i00-lim__PUXCHM2 (robolab-reorient-red-mug-i00-standard, run codex-gpt6_luna-xhigh, robolab-lunaxh-limited-0928, lab; https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/reorient_red_mug/limited/). Human review per family (PROTOCOL §2.7)."
diff --git a/tasks/robolab-rubiks-cube-and-banana-i00-privileged/task.toml b/tasks/robolab-rubiks-cube-and-banana-i00-privileged/task.toml
index 78ec7801d7..851fc7d1b6 100644
--- a/tasks/robolab-rubiks-cube-and-banana-i00-privileged/task.toml
+++ b/tasks/robolab-rubiks-cube-and-banana-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "robolab" # NVIDIA RoboLab v0.3.1 on Isaac Lab 2.3.2 / I
robot = "franka-robotiq-2f85"
family = "rubiks_cube_and_banana"
mode = "privileged" # privileged state, resets and primitives through the tooling (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution: RoboLab ships no expert or planner. Positive examples graded 1 by the separate verifier: robolab-rubiks-cube-and-banana-i__URGnnbk (robolab-rubiks-cube-and-banana-i00-standard, run codex-gpt6_luna-xhigh, robolab-lunaxh-limited-rerun-0928, lab; https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/rubiks_cube_and_banana/limited/). Human review per family (PROTOCOL §2.7)."
instance_seed = 0 # scene = environment/instance.json (the full scene state after the seeded reset + its hash); identity = dataset.toml digest
diff --git a/tasks/robolab-rubiks-cube-and-banana-i00-standard/README.md b/tasks/robolab-rubiks-cube-and-banana-i00-standard/README.md
index 44babe242e..8734545ceb 100644
--- a/tasks/robolab-rubiks-cube-and-banana-i00-standard/README.md
+++ b/tasks/robolab-rubiks-cube-and-banana-i00-standard/README.md
@@ -72,3 +72,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
- 2026-10-10 (review round 3, board 0067): the shared service layer and the client only (`environment/sim/rcb_limited.py`, `environment/rcb_client.py`). A camera named more than once in `cameras` is rendered and returned once (N copies of one name made the service build N x N images within the frame budget, which could exhaust its memory). A service that has given up on the episode (it could not start it, or could not resume it after its crashes) answers every request with 409 and the reason, and `Robot()` / `wait_ready()` raise `EpisodeOver` at once; before, nothing answered and they waited out their timeout (up to an hour). The adapter, the collect hook, the grader, the scene and the pass rule are unchanged.
- 2026-10-10 (owner decision, board 0072): the agent's network egress is restricted to the model APIs (`[agent] network_mode = "allowlist"`, the hosts of `scripts/set_agent_egress.py`; since 2026-10-11 these include the package registries, PyPI and the Debian / Ubuntu archives, so that `pip install` and `apt-get install` work); it was open before (Harbor's default), and agents could download upstream sources, reference solutions and robot models from their own shell. The simulator image is `eai-robolab:0.1.2` (was 0.1.1): it holds the MDL materials and textures that RoboLab's asset library fetched from NVIDIA's Omniverse content bucket whenever a scene was built, so that the cameras show under the allowlist what they showed with open egress (on 0.1.1 without that download the table renders red). The scene, the frozen instance, the physics, the tooling and the pass rule are unchanged, so the task keeps its name; results from before this date were obtained with open egress.
- 2026-10-13 (owner, boards 0085 and 0087): the sim service's clock (`environment/sim/rcb_limited.py`, `environment/sim/server.py`, the docstring of `environment/rcb_client.py`). `meta["time_s"]` and `meta["time_left_s"]` count from the agent's first request to the service (any endpoint), and a restarted service keeps that moment; before, they counted from the service's own start of the episode (when its scene was ready). Where the service is ready before the agent's phase begins (the harness is installed in between: a median of 79 s over the recorded standard trials), that clock ran ahead of the agent's and `time_left_s` reached 0 while the agent still had that time. `/spec` says what the clock counts from (`budget.clock`), and the instruction's `## Time limit` tells the agent to plan by `time_left_s`. The service never ended an episode on this clock: the scene, the frozen instance, the pass rule and the grader are unchanged, so the task keeps its name and earlier results stand.
+- 2026-10-14 (board 0113, from board 0086): the success text says that the gripper must not be touching the objects at the moment of success ("at that moment"); the old wording "the gripper must touch none of them" was read by an agent as "never touch them" (board 0086). Text only: the scene, the frozen instance, RoboLab's success term and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/robolab-rubiks-cube-and-banana-i00-standard/instruction.md b/tasks/robolab-rubiks-cube-and-banana-i00-standard/instruction.md
index 71eee0a8e3..bbadc7f233 100644
--- a/tasks/robolab-rubiks-cube-and-banana-i00-standard/instruction.md
+++ b/tasks/robolab-rubiks-cube-and-banana-i00-standard/instruction.md
@@ -1,7 +1,7 @@
# RoboLab · rubiks_cube_and_banana — STANDARD MODE (robot as a service, eai-standard/2.2)
You control a **Franka Panda arm with a Robotiq 2F-85 gripper** (NVIDIA RoboLab's DROID setup) **as if it were real
-hardware**: the simulator runs in a service you cannot open. You get one client and one episode. **Task:** Put the cube and the banana in the bowl. Success: The Rubik's cube and the yellow banana must end inside the bowl, touching it, and the gripper must touch neither. "Inside" means the centre of the object's convex hull lies within the container's convex hull, extended without limit along the container's own up axis (the open side; straight up for a container standing normally).
+hardware**: the simulator runs in a service you cannot open. You get one client and one episode. **Task:** Put the cube and the banana in the bowl. Success: The Rubik's cube and the yellow banana must end inside the bowl, touching it, and, at that moment, the gripper must touch neither (let go of both at the end). "Inside" means the centre of the object's convex hull lies within the container's convex hull, extended without limit along the container's own up axis (the open side; straight up for a container standing normally).
## Rules (real-robot rules)
- **No reset.** The episode starts when the service comes up and ends on success (at the first control step where the
diff --git a/tasks/robolab-rubiks-cube-and-banana-i00-standard/task.toml b/tasks/robolab-rubiks-cube-and-banana-i00-standard/task.toml
index 1f68a1d2ce..cbfea11cde 100644
--- a/tasks/robolab-rubiks-cube-and-banana-i00-standard/task.toml
+++ b/tasks/robolab-rubiks-cube-and-banana-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "robolab"
robot = "franka-robotiq-2f85"
family = "rubiks_cube_and_banana"
mode = "standard" # robot as a service: RGB, depth + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution: RoboLab ships no expert or planner. Positive examples graded 1 by the separate verifier: robolab-rubiks-cube-and-banana-i__URGnnbk (robolab-rubiks-cube-and-banana-i00-standard, run codex-gpt6_luna-xhigh, robolab-lunaxh-limited-rerun-0928, lab; https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/rubiks_cube_and_banana/limited/). Human review per family (PROTOCOL §2.7)."
diff --git a/tasks/robolab-rubiks-cube-behind-bowl-i00-privileged/task.toml b/tasks/robolab-rubiks-cube-behind-bowl-i00-privileged/task.toml
index 18e05ece88..bdc994b68d 100644
--- a/tasks/robolab-rubiks-cube-behind-bowl-i00-privileged/task.toml
+++ b/tasks/robolab-rubiks-cube-behind-bowl-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "robolab" # NVIDIA RoboLab v0.3.1 on Isaac Lab 2.3.2 / I
robot = "franka-robotiq-2f85"
family = "rubiks_cube_behind_bowl"
mode = "privileged" # privileged state, resets and primitives through the tooling (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution: RoboLab ships no expert or planner. Positive examples graded 1 by the separate verifier: robolab-rubiks-cube-behind-bowl__6kpNySr (robolab-rubiks-cube-behind-bowl-i00-privileged, run codex-gpt6_luna-xhigh, robolab-lunacx-0925-2038, ws; https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/rubiks_cube_behind_bowl/unlimited/); robolab-rubiks-cube-behind-bowl__JTz2urX (robolab-rubiks-cube-behind-bowl-i00-standard, run codex-gpt6_luna-xhigh, robolab-lunaxh-limited-0928, ws3; https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/rubiks_cube_behind_bowl/limited/). Human review per family (PROTOCOL §2.7)."
instance_seed = 0 # scene = environment/instance.json (the full scene state after the seeded reset + its hash); identity = dataset.toml digest
diff --git a/tasks/robolab-rubiks-cube-behind-bowl-i00-standard/task.toml b/tasks/robolab-rubiks-cube-behind-bowl-i00-standard/task.toml
index bd028b7fc2..ac53e2dee2 100644
--- a/tasks/robolab-rubiks-cube-behind-bowl-i00-standard/task.toml
+++ b/tasks/robolab-rubiks-cube-behind-bowl-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "robolab"
robot = "franka-robotiq-2f85"
family = "rubiks_cube_behind_bowl"
mode = "standard" # robot as a service: RGB, depth + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution: RoboLab ships no expert or planner. Positive examples graded 1 by the separate verifier: robolab-rubiks-cube-behind-bowl__6kpNySr (robolab-rubiks-cube-behind-bowl-i00-privileged, run codex-gpt6_luna-xhigh, robolab-lunacx-0925-2038, ws; https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/rubiks_cube_behind_bowl/unlimited/); robolab-rubiks-cube-behind-bowl__JTz2urX (robolab-rubiks-cube-behind-bowl-i00-standard, run codex-gpt6_luna-xhigh, robolab-lunaxh-limited-0928, ws3; https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/rubiks_cube_behind_bowl/limited/). Human review per family (PROTOCOL §2.7)."
diff --git a/tasks/robolab-rubiks-cube-in-front-of-bowl-i00-privileged/task.toml b/tasks/robolab-rubiks-cube-in-front-of-bowl-i00-privileged/task.toml
index d5899e7cdc..cc8da39261 100644
--- a/tasks/robolab-rubiks-cube-in-front-of-bowl-i00-privileged/task.toml
+++ b/tasks/robolab-rubiks-cube-in-front-of-bowl-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "robolab" # NVIDIA RoboLab v0.3.1 on Isaac Lab 2.3.2 / I
robot = "franka-robotiq-2f85"
family = "rubiks_cube_in_front_of_bowl"
mode = "privileged" # privileged state, resets and primitives through the tooling (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution: RoboLab ships no expert or planner. Positive examples graded 1 by the separate verifier: robolab-rubiks-cube-in-front-of__9ygGYcP (robolab-rubiks-cube-in-front-of-bowl-i00-privileged, run codex-gpt6_luna-xhigh, robolab-lunacx-0925-2038, ws; https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/rubiks_cube_in_front_of_bowl/unlimited/); robolab-rubiks-cube-in-front-of__7zQXMPY (robolab-rubiks-cube-in-front-of-bowl-i00-standard, run codex-gpt6_luna-xhigh, robolab-lunaxh-limited-0928, lab; https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/rubiks_cube_in_front_of_bowl/limited/). Human review per family (PROTOCOL §2.7)."
instance_seed = 0 # scene = environment/instance.json (the full scene state after the seeded reset + its hash); identity = dataset.toml digest
diff --git a/tasks/robolab-rubiks-cube-in-front-of-bowl-i00-standard/task.toml b/tasks/robolab-rubiks-cube-in-front-of-bowl-i00-standard/task.toml
index 81adb255b1..5dd0ee26d8 100644
--- a/tasks/robolab-rubiks-cube-in-front-of-bowl-i00-standard/task.toml
+++ b/tasks/robolab-rubiks-cube-in-front-of-bowl-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "robolab"
robot = "franka-robotiq-2f85"
family = "rubiks_cube_in_front_of_bowl"
mode = "standard" # robot as a service: RGB, depth + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution: RoboLab ships no expert or planner. Positive examples graded 1 by the separate verifier: robolab-rubiks-cube-in-front-of__9ygGYcP (robolab-rubiks-cube-in-front-of-bowl-i00-privileged, run codex-gpt6_luna-xhigh, robolab-lunacx-0925-2038, ws; https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/rubiks_cube_in_front_of_bowl/unlimited/); robolab-rubiks-cube-in-front-of__7zQXMPY (robolab-rubiks-cube-in-front-of-bowl-i00-standard, run codex-gpt6_luna-xhigh, robolab-lunaxh-limited-0928, lab; https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/rubiks_cube_in_front_of_bowl/limited/). Human review per family (PROTOCOL §2.7)."
diff --git a/tasks/robolab-rubiks-cube-left-of-bowl-i00-privileged/task.toml b/tasks/robolab-rubiks-cube-left-of-bowl-i00-privileged/task.toml
index 6b8909f046..1830033522 100644
--- a/tasks/robolab-rubiks-cube-left-of-bowl-i00-privileged/task.toml
+++ b/tasks/robolab-rubiks-cube-left-of-bowl-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "robolab" # NVIDIA RoboLab v0.3.1 on Isaac Lab 2.3.2 / I
robot = "franka-robotiq-2f85"
family = "rubiks_cube_left_of_bowl"
mode = "privileged" # privileged state, resets and primitives through the tooling (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
oracle_evidence = "The oracle scored 1 (deterministic, instance_exact) in Harbor runs robolab-rubiks-cube-left-of-bowl__onMcaQH (laptop, job tpl11-0927-1823-oracle-robolab-rubiks-cube-left-of-bowl-i00, at eaf2aff0) and robolab-rubiks-cube-left-of-bowl__tRHn5hq (ws3, job tpl19-0927-0430-oracle-robolab-rubiks-cube-left-of-bowl-i00, at fe045825), on a byte-identical tests/grade.py, frozen instance and oracle code (tests/test.sh differs only in the command that empties /logs/verifier before grading: rm -rf, the protocol v1.0 convention); the tooling (rcb_robolab.py) changed since only off the oracle and replay path: eaf2aff0 adds other families' registry entries; b3c7ca89 adds fail-safe 2.2 helpers and was re-checked through the same tooling path (the banana_in_bowl oracle trajectory ends in the same final state sha 61afa384566be3a5 with the old and the new CLI); cde31660 flips the depth default, which run-oracle and replay pass explicitly (depth=False); one docstring line now says standard mode."
instance_seed = 0 # scene = environment/instance.json (the full scene state after the seeded reset + its hash); identity = dataset.toml digest
diff --git a/tasks/robolab-rubiks-cube-left-of-bowl-i00-standard/task.toml b/tasks/robolab-rubiks-cube-left-of-bowl-i00-standard/task.toml
index d0b5fa02bd..5db2f6be88 100644
--- a/tasks/robolab-rubiks-cube-left-of-bowl-i00-standard/task.toml
+++ b/tasks/robolab-rubiks-cube-left-of-bowl-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "robolab"
robot = "franka-robotiq-2f85"
family = "rubiks_cube_left_of_bowl"
mode = "standard" # robot as a service: RGB, depth + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "partial"
oracle_evidence = "solution/oracle_replay.py scored 1 in 6/6 recorded runs on the pre-crash-recovery service with the same verdict formula (success = live, replay and deterministic), the latest robolab-rubiks-cube-left-of-bowl__YZj6yTp (ws3, job eq-ws3-oracle-rubiks-cube-left-of-bowl, 2026-09-29, the dev/qineng 9eb18edf version); not re-run on this version (crash recovery, the collect hook's `end`, one-second chunks): static validation only, 2026-10-01. Human review per family (PROTOCOL §2.7)."
diff --git a/tasks/robolab-rubiks-cube-right-of-bowl-i00-privileged/task.toml b/tasks/robolab-rubiks-cube-right-of-bowl-i00-privileged/task.toml
index f0060b9f3f..254cefa509 100644
--- a/tasks/robolab-rubiks-cube-right-of-bowl-i00-privileged/task.toml
+++ b/tasks/robolab-rubiks-cube-right-of-bowl-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "robolab" # NVIDIA RoboLab v0.3.1 on Isaac Lab 2.3.2 / I
robot = "franka-robotiq-2f85"
family = "rubiks_cube_right_of_bowl"
mode = "privileged" # privileged state, resets and primitives through the tooling (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution: RoboLab ships no expert or planner. Positive examples graded 1 by the separate verifier: robolab-rubiks-cube-right-of-bow__W7ALYqw (robolab-rubiks-cube-right-of-bowl-i00-privileged, run codex-gpt6_luna-xhigh, robolab-lunacx-0925-2038, ws; https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/rubiks_cube_right_of_bowl/unlimited/); robolab-rubiks-cube-right-of-bow__yBJGVid (robolab-rubiks-cube-right-of-bowl-i00-standard, run codex-gpt6_luna-xhigh, robolab-lunaxh-limited-0928, lab; https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/rubiks_cube_right_of_bowl/limited/). Human review per family (PROTOCOL §2.7)."
instance_seed = 0 # scene = environment/instance.json (the full scene state after the seeded reset + its hash); identity = dataset.toml digest
diff --git a/tasks/robolab-rubiks-cube-right-of-bowl-i00-standard/task.toml b/tasks/robolab-rubiks-cube-right-of-bowl-i00-standard/task.toml
index 6e4a7382ca..6667f5be97 100644
--- a/tasks/robolab-rubiks-cube-right-of-bowl-i00-standard/task.toml
+++ b/tasks/robolab-rubiks-cube-right-of-bowl-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "robolab"
robot = "franka-robotiq-2f85"
family = "rubiks_cube_right_of_bowl"
mode = "standard" # robot as a service: RGB, depth + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution: RoboLab ships no expert or planner. Positive examples graded 1 by the separate verifier: robolab-rubiks-cube-right-of-bow__W7ALYqw (robolab-rubiks-cube-right-of-bowl-i00-privileged, run codex-gpt6_luna-xhigh, robolab-lunacx-0925-2038, ws; https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/rubiks_cube_right_of_bowl/unlimited/); robolab-rubiks-cube-right-of-bow__yBJGVid (robolab-rubiks-cube-right-of-bowl-i00-standard, run codex-gpt6_luna-xhigh, robolab-lunaxh-limited-0928, lab; https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/rubiks_cube_right_of_bowl/limited/). Human review per family (PROTOCOL §2.7)."
diff --git a/tasks/robolab-rubiks-cube-then-banana-i00-privileged/task.toml b/tasks/robolab-rubiks-cube-then-banana-i00-privileged/task.toml
index 1d00f8d094..2ec630fb1f 100644
--- a/tasks/robolab-rubiks-cube-then-banana-i00-privileged/task.toml
+++ b/tasks/robolab-rubiks-cube-then-banana-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "robolab" # NVIDIA RoboLab v0.3.1 on Isaac Lab 2.3.2 / I
robot = "franka-robotiq-2f85"
family = "rubiks_cube_then_banana"
mode = "privileged" # privileged state, resets and primitives through the tooling (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution: RoboLab ships no expert or planner. Positive examples graded 1 by the separate verifier: robolab-rubiks-cube-then-banana__ZtRBWNw (robolab-rubiks-cube-then-banana-i00-privileged, run codex-gpt6_luna-xhigh, robolab-lunacx-0925-2038, ws3; https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/rubiks_cube_then_banana/unlimited/). Human review per family (PROTOCOL §2.7)."
instance_seed = 0 # scene = environment/instance.json (the full scene state after the seeded reset + its hash); identity = dataset.toml digest
diff --git a/tasks/robolab-rubiks-cube-then-banana-i00-standard/README.md b/tasks/robolab-rubiks-cube-then-banana-i00-standard/README.md
index e962282680..9ecf3044ee 100644
--- a/tasks/robolab-rubiks-cube-then-banana-i00-standard/README.md
+++ b/tasks/robolab-rubiks-cube-then-banana-i00-standard/README.md
@@ -72,3 +72,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
- 2026-10-10 (review round 3, board 0067): the shared service layer and the client only (`environment/sim/rcb_limited.py`, `environment/rcb_client.py`). A camera named more than once in `cameras` is rendered and returned once (N copies of one name made the service build N x N images within the frame budget, which could exhaust its memory). A service that has given up on the episode (it could not start it, or could not resume it after its crashes) answers every request with 409 and the reason, and `Robot()` / `wait_ready()` raise `EpisodeOver` at once; before, nothing answered and they waited out their timeout (up to an hour). The adapter, the collect hook, the grader, the scene and the pass rule are unchanged.
- 2026-10-10 (owner decision, board 0072): the agent's network egress is restricted to the model APIs (`[agent] network_mode = "allowlist"`, the hosts of `scripts/set_agent_egress.py`; since 2026-10-11 these include the package registries, PyPI and the Debian / Ubuntu archives, so that `pip install` and `apt-get install` work); it was open before (Harbor's default), and agents could download upstream sources, reference solutions and robot models from their own shell. The simulator image is `eai-robolab:0.1.2` (was 0.1.1): it holds the MDL materials and textures that RoboLab's asset library fetched from NVIDIA's Omniverse content bucket whenever a scene was built, so that the cameras show under the allowlist what they showed with open egress (on 0.1.1 without that download the table renders red). The scene, the frozen instance, the physics, the tooling and the pass rule are unchanged, so the task keeps its name; results from before this date were obtained with open egress.
- 2026-10-13 (owner, boards 0085 and 0087): the sim service's clock (`environment/sim/rcb_limited.py`, `environment/sim/server.py`, the docstring of `environment/rcb_client.py`). `meta["time_s"]` and `meta["time_left_s"]` count from the agent's first request to the service (any endpoint), and a restarted service keeps that moment; before, they counted from the service's own start of the episode (when its scene was ready). Where the service is ready before the agent's phase begins (the harness is installed in between: a median of 79 s over the recorded standard trials), that clock ran ahead of the agent's and `time_left_s` reached 0 while the agent still had that time. `/spec` says what the clock counts from (`budget.clock`), and the instruction's `## Time limit` tells the agent to plan by `time_left_s`. The service never ended an episode on this clock: the scene, the frozen instance, the pass rule and the grader are unchanged, so the task keeps its name and earlier results stand.
+- 2026-10-14 (board 0113, from board 0086): the success text says that the gripper must not be touching the objects at the moment of success ("at that moment"); the old wording "the gripper must touch none of them" was read by an agent as "never touch them" (board 0086). Text only: the scene, the frozen instance, RoboLab's success term and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/robolab-rubiks-cube-then-banana-i00-standard/instruction.md b/tasks/robolab-rubiks-cube-then-banana-i00-standard/instruction.md
index 8274775195..6f4d1bc8c2 100644
--- a/tasks/robolab-rubiks-cube-then-banana-i00-standard/instruction.md
+++ b/tasks/robolab-rubiks-cube-then-banana-i00-standard/instruction.md
@@ -1,7 +1,7 @@
# RoboLab · rubiks_cube_then_banana — STANDARD MODE (robot as a service, eai-standard/2.2)
You control a **Franka Panda arm with a Robotiq 2F-85 gripper** (NVIDIA RoboLab's DROID setup) **as if it were real
-hardware**: the simulator runs in a service you cannot open. You get one client and one episode. **Task:** Put the cube then the banana in the bowl. Success: The Rubik's cube and then the yellow banana must end inside the bowl, both touching it, and the gripper must touch neither. Order counts: the service remembers the first moment each object is placed and a later removal does not erase it, so if the yellow banana is placed first the episode can no longer succeed. "Inside" means the centre of the object's convex hull lies within the container's convex hull, extended without limit along the container's own up axis (the open side; straight up for a container standing normally).
+hardware**: the simulator runs in a service you cannot open. You get one client and one episode. **Task:** Put the cube then the banana in the bowl. Success: The Rubik's cube and then the yellow banana must end inside the bowl, both touching it, and, at that moment, the gripper must touch neither (let go of both at the end). Order counts: the service remembers the first moment each object is placed and a later removal does not erase it, so if the yellow banana is placed first the episode can no longer succeed. "Inside" means the centre of the object's convex hull lies within the container's convex hull, extended without limit along the container's own up axis (the open side; straight up for a container standing normally).
## Rules (real-robot rules)
- **No reset.** The episode starts when the service comes up and ends on success (at the first control step where the
diff --git a/tasks/robolab-rubiks-cube-then-banana-i00-standard/task.toml b/tasks/robolab-rubiks-cube-then-banana-i00-standard/task.toml
index 278958daf5..eb4eb1ccac 100644
--- a/tasks/robolab-rubiks-cube-then-banana-i00-standard/task.toml
+++ b/tasks/robolab-rubiks-cube-then-banana-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "robolab"
robot = "franka-robotiq-2f85"
family = "rubiks_cube_then_banana"
mode = "standard" # robot as a service: RGB, depth + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution: RoboLab ships no expert or planner. Positive examples graded 1 by the separate verifier: robolab-rubiks-cube-then-banana__ZtRBWNw (robolab-rubiks-cube-then-banana-i00-privileged, run codex-gpt6_luna-xhigh, robolab-lunacx-0925-2038, ws3; https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/rubiks_cube_then_banana/unlimited/). Human review per family (PROTOCOL §2.7)."
diff --git a/tasks/robolab-rubiks-cubes-in-bin-i00-privileged/task.toml b/tasks/robolab-rubiks-cubes-in-bin-i00-privileged/task.toml
index 9b05f624c9..cbb3b1e9dd 100644
--- a/tasks/robolab-rubiks-cubes-in-bin-i00-privileged/task.toml
+++ b/tasks/robolab-rubiks-cubes-in-bin-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "robolab" # NVIDIA RoboLab v0.3.1 on Isaac Lab 2.3.2 / I
robot = "franka-robotiq-2f85"
family = "rubiks_cubes_in_bin"
mode = "privileged" # privileged state, resets and primitives through the tooling (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution: RoboLab ships no expert or planner. Positive examples graded 1 by the separate verifier: robolab-rubiks-cubes-in-bin-i00__uTLD4r8 (robolab-rubiks-cubes-in-bin-i00-privileged, run codex-gpt6_luna-xhigh, robolab-lunacx-0925-2038, ws3; https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/rubiks_cubes_in_bin/unlimited/); robolab-rubiks-cubes-in-bin-i00__Zr7BhJF (robolab-rubiks-cubes-in-bin-i00-standard, run codex-gpt6_luna-xhigh, robolab-lunaxh-limited-0928, lab; https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/rubiks_cubes_in_bin/limited/). Human review per family (PROTOCOL §2.7)."
instance_seed = 0 # scene = environment/instance.json (the full scene state after the seeded reset + its hash); identity = dataset.toml digest
diff --git a/tasks/robolab-rubiks-cubes-in-bin-i00-standard/README.md b/tasks/robolab-rubiks-cubes-in-bin-i00-standard/README.md
index cc07a80c02..8715270131 100644
--- a/tasks/robolab-rubiks-cubes-in-bin-i00-standard/README.md
+++ b/tasks/robolab-rubiks-cubes-in-bin-i00-standard/README.md
@@ -73,3 +73,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
- 2026-10-10 (review round 3, board 0067): the shared service layer and the client only (`environment/sim/rcb_limited.py`, `environment/rcb_client.py`). A camera named more than once in `cameras` is rendered and returned once (N copies of one name made the service build N x N images within the frame budget, which could exhaust its memory). A service that has given up on the episode (it could not start it, or could not resume it after its crashes) answers every request with 409 and the reason, and `Robot()` / `wait_ready()` raise `EpisodeOver` at once; before, nothing answered and they waited out their timeout (up to an hour). The adapter, the collect hook, the grader, the scene and the pass rule are unchanged.
- 2026-10-10 (owner decision, board 0072): the agent's network egress is restricted to the model APIs (`[agent] network_mode = "allowlist"`, the hosts of `scripts/set_agent_egress.py`; since 2026-10-11 these include the package registries, PyPI and the Debian / Ubuntu archives, so that `pip install` and `apt-get install` work); it was open before (Harbor's default), and agents could download upstream sources, reference solutions and robot models from their own shell. The simulator image is `eai-robolab:0.1.2` (was 0.1.1): it holds the MDL materials and textures that RoboLab's asset library fetched from NVIDIA's Omniverse content bucket whenever a scene was built, so that the cameras show under the allowlist what they showed with open egress (on 0.1.1 without that download the table renders red). The scene, the frozen instance, the physics, the tooling and the pass rule are unchanged, so the task keeps its name; results from before this date were obtained with open egress.
- 2026-10-13 (owner, boards 0085 and 0087): the sim service's clock (`environment/sim/rcb_limited.py`, `environment/sim/server.py`, the docstring of `environment/rcb_client.py`). `meta["time_s"]` and `meta["time_left_s"]` count from the agent's first request to the service (any endpoint), and a restarted service keeps that moment; before, they counted from the service's own start of the episode (when its scene was ready). Where the service is ready before the agent's phase begins (the harness is installed in between: a median of 79 s over the recorded standard trials), that clock ran ahead of the agent's and `time_left_s` reached 0 while the agent still had that time. `/spec` says what the clock counts from (`budget.clock`), and the instruction's `## Time limit` tells the agent to plan by `time_left_s`. The service never ended an episode on this clock: the scene, the frozen instance, the pass rule and the grader are unchanged, so the task keeps its name and earlier results stand.
+- 2026-10-14 (board 0113, from board 0086): the success text says that the gripper must not be touching the objects at the moment of success ("at that moment"); the old wording "the gripper must touch none of them" was read by an agent as "never touch them" (board 0086). Text only: the scene, the frozen instance, RoboLab's success term and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/robolab-rubiks-cubes-in-bin-i00-standard/instruction.md b/tasks/robolab-rubiks-cubes-in-bin-i00-standard/instruction.md
index a4ffd7dfaf..b9cd7991f8 100644
--- a/tasks/robolab-rubiks-cubes-in-bin-i00-standard/instruction.md
+++ b/tasks/robolab-rubiks-cubes-in-bin-i00-standard/instruction.md
@@ -1,7 +1,7 @@
# RoboLab · rubiks_cubes_in_bin — STANDARD MODE (robot as a service, eai-standard/2.2)
You control a **Franka Panda arm with a Robotiq 2F-85 gripper** (NVIDIA RoboLab's DROID setup) **as if it were real
-hardware**: the simulator runs in a service you cannot open. You get one client and one episode. **Task:** Sort all rubiks cubes into the bin. Success: The three Rubik's cubes must end inside the grey bin and the gripper must touch none of them. "Inside" means the centre of the object's convex hull lies within the container's convex hull, extended without limit along the container's own up axis (the open side; straight up for a container standing normally).
+hardware**: the simulator runs in a service you cannot open. You get one client and one episode. **Task:** Sort all rubiks cubes into the bin. Success: The three Rubik's cubes must end inside the grey bin and, at that moment, the gripper must touch none of them (moving them with the gripper is how the task is done; let go of them at the end). "Inside" means the centre of the object's convex hull lies within the container's convex hull, extended without limit along the container's own up axis (the open side; straight up for a container standing normally).
## Rules (real-robot rules)
- **No reset.** The episode starts when the service comes up and ends on success (at the first control step where the
diff --git a/tasks/robolab-rubiks-cubes-in-bin-i00-standard/task.toml b/tasks/robolab-rubiks-cubes-in-bin-i00-standard/task.toml
index 3e0ed77d4b..8addd2df18 100644
--- a/tasks/robolab-rubiks-cubes-in-bin-i00-standard/task.toml
+++ b/tasks/robolab-rubiks-cubes-in-bin-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "robolab"
robot = "franka-robotiq-2f85"
family = "rubiks_cubes_in_bin"
mode = "standard" # robot as a service: RGB, depth + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution: RoboLab ships no expert or planner. Positive examples graded 1 by the separate verifier: robolab-rubiks-cubes-in-bin-i00__uTLD4r8 (robolab-rubiks-cubes-in-bin-i00-privileged, run codex-gpt6_luna-xhigh, robolab-lunacx-0925-2038, ws3; https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/rubiks_cubes_in_bin/unlimited/); robolab-rubiks-cubes-in-bin-i00__Zr7BhJF (robolab-rubiks-cubes-in-bin-i00-standard, run codex-gpt6_luna-xhigh, robolab-lunaxh-limited-0928, lab; https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/rubiks_cubes_in_bin/limited/). Human review per family (PROTOCOL §2.7)."
diff --git a/tasks/robolab-spoon-in-mug-i00-privileged/task.toml b/tasks/robolab-spoon-in-mug-i00-privileged/task.toml
index 7d213b5736..e56550f236 100644
--- a/tasks/robolab-spoon-in-mug-i00-privileged/task.toml
+++ b/tasks/robolab-spoon-in-mug-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "robolab" # NVIDIA RoboLab v0.3.1 on Isaac Lab 2.3.2 / I
robot = "franka-robotiq-2f85"
family = "spoon_in_mug"
mode = "privileged" # privileged state, resets and primitives through the tooling (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution: RoboLab ships no expert or planner. Positive examples graded 1 by the separate verifier: robolab-spoon-in-mug-i00__VH7bzHd (robolab-spoon-in-mug-i00-privileged, run codex-gpt6_luna-xhigh, robolab-lunacx-0925-2038, ws; https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/spoon_in_mug/unlimited/). Human review per family (PROTOCOL §2.7)."
instance_seed = 0 # scene = environment/instance.json (the full scene state after the seeded reset + its hash); identity = dataset.toml digest
diff --git a/tasks/robolab-spoon-in-mug-i00-standard/task.toml b/tasks/robolab-spoon-in-mug-i00-standard/task.toml
index 472fbb1fbd..9239061734 100644
--- a/tasks/robolab-spoon-in-mug-i00-standard/task.toml
+++ b/tasks/robolab-spoon-in-mug-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "robolab"
robot = "franka-robotiq-2f85"
family = "spoon_in_mug"
mode = "standard" # robot as a service: RGB, depth + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution: RoboLab ships no expert or planner. Positive examples graded 1 by the separate verifier: robolab-spoon-in-mug-i00__VH7bzHd (robolab-spoon-in-mug-i00-privileged, run codex-gpt6_luna-xhigh, robolab-lunacx-0925-2038, ws; https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/spoon_in_mug/unlimited/). Human review per family (PROTOCOL §2.7)."
diff --git a/tasks/robolab-stack-yellow-on-red-i00-privileged/task.toml b/tasks/robolab-stack-yellow-on-red-i00-privileged/task.toml
index 443042889f..7c1f61d764 100644
--- a/tasks/robolab-stack-yellow-on-red-i00-privileged/task.toml
+++ b/tasks/robolab-stack-yellow-on-red-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "robolab" # NVIDIA RoboLab v0.3.1 on Isaac Lab 2.3.2 / I
robot = "franka-robotiq-2f85"
family = "stack_yellow_on_red"
mode = "privileged" # privileged state, resets and primitives through the tooling (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution: RoboLab ships no expert or planner. Positive examples graded 1 by the separate verifier: robolab-stack-yellow-on-red-i00__Wraz9PM (robolab-stack-yellow-on-red-i00-privileged, run codex-gpt6_luna-xhigh, robolab-lunacx-0925-2038, ws3; https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/stack_yellow_on_red/unlimited/); robolab-stack-yellow-on-red-i00__oCNtAax (robolab-stack-yellow-on-red-i00-standard, run codex-gpt6_luna-xhigh, robolab-lunaxh-limited-0928, lab; https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/stack_yellow_on_red/limited/). Human review per family (PROTOCOL §2.7)."
instance_seed = 0 # scene = environment/instance.json (the full scene state after the seeded reset + its hash); identity = dataset.toml digest
diff --git a/tasks/robolab-stack-yellow-on-red-i00-standard/task.toml b/tasks/robolab-stack-yellow-on-red-i00-standard/task.toml
index 38a719f804..1caa5d1b3d 100644
--- a/tasks/robolab-stack-yellow-on-red-i00-standard/task.toml
+++ b/tasks/robolab-stack-yellow-on-red-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "robolab"
robot = "franka-robotiq-2f85"
family = "stack_yellow_on_red"
mode = "standard" # robot as a service: RGB, depth + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution: RoboLab ships no expert or planner. Positive examples graded 1 by the separate verifier: robolab-stack-yellow-on-red-i00__Wraz9PM (robolab-stack-yellow-on-red-i00-privileged, run codex-gpt6_luna-xhigh, robolab-lunacx-0925-2038, ws3; https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/stack_yellow_on_red/unlimited/); robolab-stack-yellow-on-red-i00__oCNtAax (robolab-stack-yellow-on-red-i00-standard, run codex-gpt6_luna-xhigh, robolab-lunaxh-limited-0928, lab; https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/stack_yellow_on_red/limited/). Human review per family (PROTOCOL §2.7)."
diff --git a/tasks/robolab-stack3-rubiks-cube-i00-privileged/task.toml b/tasks/robolab-stack3-rubiks-cube-i00-privileged/task.toml
index bc9801fdd8..47843c0974 100644
--- a/tasks/robolab-stack3-rubiks-cube-i00-privileged/task.toml
+++ b/tasks/robolab-stack3-rubiks-cube-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "robolab" # NVIDIA RoboLab v0.3.1 on Isaac Lab 2.3.2 / I
robot = "franka-robotiq-2f85"
family = "stack3_rubiks_cube"
mode = "privileged" # privileged state, resets and primitives through the tooling (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution: RoboLab ships no expert or planner. Positive examples graded 1 by the separate verifier: robolab-stack3-rubiks-cube-i00__8hhXJ5w (robolab-stack3-rubiks-cube-i00-privileged, run codex-gpt6_luna-xhigh, robolab-lunacx-0925-2038, ws3; https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/stack3_rubiks_cube/unlimited/); robolab-stack3-rubiks-cube-i00-l__722n2cQ (robolab-stack3-rubiks-cube-i00-standard, run codex-gpt6_luna-xhigh, robolab-lunaxh-limited-0928, lab; https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/stack3_rubiks_cube/limited/). Human review per family (PROTOCOL §2.7)."
instance_seed = 0 # scene = environment/instance.json (the full scene state after the seeded reset + its hash); identity = dataset.toml digest
diff --git a/tasks/robolab-stack3-rubiks-cube-i00-standard/task.toml b/tasks/robolab-stack3-rubiks-cube-i00-standard/task.toml
index 8614695960..984f713378 100644
--- a/tasks/robolab-stack3-rubiks-cube-i00-standard/task.toml
+++ b/tasks/robolab-stack3-rubiks-cube-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "robolab"
robot = "franka-robotiq-2f85"
family = "stack3_rubiks_cube"
mode = "standard" # robot as a service: RGB, depth + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution: RoboLab ships no expert or planner. Positive examples graded 1 by the separate verifier: robolab-stack3-rubiks-cube-i00__8hhXJ5w (robolab-stack3-rubiks-cube-i00-privileged, run codex-gpt6_luna-xhigh, robolab-lunacx-0925-2038, ws3; https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/stack3_rubiks_cube/unlimited/); robolab-stack3-rubiks-cube-i00-l__722n2cQ (robolab-stack3-rubiks-cube-i00-standard, run codex-gpt6_luna-xhigh, robolab-lunaxh-limited-0928, lab; https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/stack3_rubiks_cube/limited/). Human review per family (PROTOCOL §2.7)."
diff --git a/tasks/robolab-take-measuring-spoon-out-i00-privileged/task.toml b/tasks/robolab-take-measuring-spoon-out-i00-privileged/task.toml
index 10f91f1d90..951e55bb85 100644
--- a/tasks/robolab-take-measuring-spoon-out-i00-privileged/task.toml
+++ b/tasks/robolab-take-measuring-spoon-out-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "robolab" # NVIDIA RoboLab v0.3.1 on Isaac Lab 2.3.2 / I
robot = "franka-robotiq-2f85"
family = "take_measuring_spoon_out"
mode = "privileged" # privileged state, resets and primitives through the tooling (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution: RoboLab ships no expert or planner. Positive examples graded 1 by the separate verifier: robolab-take-measuring-spoon-out__vK2T2MY (robolab-take-measuring-spoon-out-i00-privileged, run codex-gpt6_luna-xhigh, robolab-lunacx-0925-2038, ws3; https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/take_measuring_spoon_out/unlimited/); robolab-take-measuring-spoon-out__3KikKcU (robolab-take-measuring-spoon-out-i00-standard, run codex-gpt6_luna-xhigh, robolab-lunaxh-limited-0928, lab; https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/take_measuring_spoon_out/limited/). Human review per family (PROTOCOL §2.7)."
instance_seed = 0 # scene = environment/instance.json (the full scene state after the seeded reset + its hash); identity = dataset.toml digest
diff --git a/tasks/robolab-take-measuring-spoon-out-i00-standard/task.toml b/tasks/robolab-take-measuring-spoon-out-i00-standard/task.toml
index 4aa0026cc5..df173ff86d 100644
--- a/tasks/robolab-take-measuring-spoon-out-i00-standard/task.toml
+++ b/tasks/robolab-take-measuring-spoon-out-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "robolab"
robot = "franka-robotiq-2f85"
family = "take_measuring_spoon_out"
mode = "standard" # robot as a service: RGB, depth + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution: RoboLab ships no expert or planner. Positive examples graded 1 by the separate verifier: robolab-take-measuring-spoon-out__vK2T2MY (robolab-take-measuring-spoon-out-i00-privileged, run codex-gpt6_luna-xhigh, robolab-lunacx-0925-2038, ws3; https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/take_measuring_spoon_out/unlimited/); robolab-take-measuring-spoon-out__3KikKcU (robolab-take-measuring-spoon-out-i00-standard, run codex-gpt6_luna-xhigh, robolab-lunaxh-limited-0928, lab; https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/take_measuring_spoon_out/limited/). Human review per family (PROTOCOL §2.7)."
diff --git a/tasks/robolab-take-mugs-off-of-shelf-i00-privileged/task.toml b/tasks/robolab-take-mugs-off-of-shelf-i00-privileged/task.toml
index 681ed4d0a5..6ac7066b0b 100644
--- a/tasks/robolab-take-mugs-off-of-shelf-i00-privileged/task.toml
+++ b/tasks/robolab-take-mugs-off-of-shelf-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "robolab" # NVIDIA RoboLab v0.3.1 on Isaac Lab 2.3.2 / I
robot = "franka-robotiq-2f85"
family = "take_mugs_off_of_shelf"
mode = "privileged" # privileged state, resets and primitives through the tooling (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution: RoboLab ships no expert or planner. Positive examples graded 1 by the separate verifier: robolab-take-mugs-off-of-shelf-i__5HNdr8K (robolab-take-mugs-off-of-shelf-i00-privileged, run codex-gpt6_luna-xhigh, robolab-lunacx-0925-2038, ws3; https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/take_mugs_off_of_shelf/unlimited/). Human review per family (PROTOCOL §2.7)."
instance_seed = 0 # scene = environment/instance.json (the full scene state after the seeded reset + its hash); identity = dataset.toml digest
diff --git a/tasks/robolab-take-mugs-off-of-shelf-i00-standard/README.md b/tasks/robolab-take-mugs-off-of-shelf-i00-standard/README.md
index 745ff328c2..742c4549d1 100644
--- a/tasks/robolab-take-mugs-off-of-shelf-i00-standard/README.md
+++ b/tasks/robolab-take-mugs-off-of-shelf-i00-standard/README.md
@@ -72,3 +72,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
- 2026-10-10 (review round 3, board 0067): the shared service layer and the client only (`environment/sim/rcb_limited.py`, `environment/rcb_client.py`). A camera named more than once in `cameras` is rendered and returned once (N copies of one name made the service build N x N images within the frame budget, which could exhaust its memory). A service that has given up on the episode (it could not start it, or could not resume it after its crashes) answers every request with 409 and the reason, and `Robot()` / `wait_ready()` raise `EpisodeOver` at once; before, nothing answered and they waited out their timeout (up to an hour). The adapter, the collect hook, the grader, the scene and the pass rule are unchanged.
- 2026-10-10 (owner decision, board 0072): the agent's network egress is restricted to the model APIs (`[agent] network_mode = "allowlist"`, the hosts of `scripts/set_agent_egress.py`; since 2026-10-11 these include the package registries, PyPI and the Debian / Ubuntu archives, so that `pip install` and `apt-get install` work); it was open before (Harbor's default), and agents could download upstream sources, reference solutions and robot models from their own shell. The simulator image is `eai-robolab:0.1.2` (was 0.1.1): it holds the MDL materials and textures that RoboLab's asset library fetched from NVIDIA's Omniverse content bucket whenever a scene was built, so that the cameras show under the allowlist what they showed with open egress (on 0.1.1 without that download the table renders red). The scene, the frozen instance, the physics, the tooling and the pass rule are unchanged, so the task keeps its name; results from before this date were obtained with open egress.
- 2026-10-13 (owner, boards 0085 and 0087): the sim service's clock (`environment/sim/rcb_limited.py`, `environment/sim/server.py`, the docstring of `environment/rcb_client.py`). `meta["time_s"]` and `meta["time_left_s"]` count from the agent's first request to the service (any endpoint), and a restarted service keeps that moment; before, they counted from the service's own start of the episode (when its scene was ready). Where the service is ready before the agent's phase begins (the harness is installed in between: a median of 79 s over the recorded standard trials), that clock ran ahead of the agent's and `time_left_s` reached 0 while the agent still had that time. `/spec` says what the clock counts from (`budget.clock`), and the instruction's `## Time limit` tells the agent to plan by `time_left_s`. The service never ended an episode on this clock: the scene, the frozen instance, the pass rule and the grader are unchanged, so the task keeps its name and earlier results stand.
+- 2026-10-14 (board 0113, from board 0086): the success text says that the gripper must not be touching the objects at the moment of success ("at that moment"); the old wording "the gripper must touch none of them" was read by an agent as "never touch them" (board 0086). Text only: the scene, the frozen instance, RoboLab's success term and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/robolab-take-mugs-off-of-shelf-i00-standard/instruction.md b/tasks/robolab-take-mugs-off-of-shelf-i00-standard/instruction.md
index c0ad889e42..86475c9787 100644
--- a/tasks/robolab-take-mugs-off-of-shelf-i00-standard/instruction.md
+++ b/tasks/robolab-take-mugs-off-of-shelf-i00-standard/instruction.md
@@ -1,7 +1,7 @@
# RoboLab · take_mugs_off_of_shelf — STANDARD MODE (robot as a service, eai-standard/2.2)
You control a **Franka Panda arm with a Robotiq 2F-85 gripper** (NVIDIA RoboLab's DROID setup) **as if it were real
-hardware**: the simulator runs in a service you cannot open. You get one client and one episode. **Task:** Take the mugs off the shelf. Success: Both mugs on the shelf must end outside the shelf and standing on the table, and the gripper must touch neither. "Outside" means the centre of the object's convex hull is not within the container's convex hull extended along the container's own up axis. "On the table" means the table supports it (its contact force points upwards, within 45 deg of vertical).
+hardware**: the simulator runs in a service you cannot open. You get one client and one episode. **Task:** Take the mugs off the shelf. Success: Both mugs on the shelf must end outside the shelf and standing on the table, and, at that moment, the gripper must touch neither (let go of both at the end). "Outside" means the centre of the object's convex hull is not within the container's convex hull extended along the container's own up axis. "On the table" means the table supports it (its contact force points upwards, within 45 deg of vertical).
## Rules (real-robot rules)
- **No reset.** The episode starts when the service comes up and ends on success (at the first control step where the
diff --git a/tasks/robolab-take-mugs-off-of-shelf-i00-standard/task.toml b/tasks/robolab-take-mugs-off-of-shelf-i00-standard/task.toml
index 5c954dda5d..5993fb75f8 100644
--- a/tasks/robolab-take-mugs-off-of-shelf-i00-standard/task.toml
+++ b/tasks/robolab-take-mugs-off-of-shelf-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "robolab"
robot = "franka-robotiq-2f85"
family = "take_mugs_off_of_shelf"
mode = "standard" # robot as a service: RGB, depth + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution: RoboLab ships no expert or planner. Positive examples graded 1 by the separate verifier: robolab-take-mugs-off-of-shelf-i__5HNdr8K (robolab-take-mugs-off-of-shelf-i00-privileged, run codex-gpt6_luna-xhigh, robolab-lunacx-0925-2038, ws3; https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/take_mugs_off_of_shelf/unlimited/). Human review per family (PROTOCOL §2.7)."
diff --git a/tasks/robolab-take-spatula-off-shelf-i00-privileged/task.toml b/tasks/robolab-take-spatula-off-shelf-i00-privileged/task.toml
index b3d2711578..133bbd1fcb 100644
--- a/tasks/robolab-take-spatula-off-shelf-i00-privileged/task.toml
+++ b/tasks/robolab-take-spatula-off-shelf-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "robolab" # NVIDIA RoboLab v0.3.1 on Isaac Lab 2.3.2 / I
robot = "franka-robotiq-2f85"
family = "take_spatula_off_shelf"
mode = "privileged" # privileged state, resets and primitives through the tooling (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution: RoboLab ships no expert or planner. Positive examples graded 1 by the separate verifier: robolab-take-spatula-off-shelf-i__vipvPdN (robolab-take-spatula-off-shelf-i00-privileged, run codex-gpt6_luna-xhigh, robolab-lunacx-0925-2038, ws; https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/take_spatula_off_shelf/unlimited/); robolab-take-spatula-off-shelf-i__TqQfCCX (robolab-take-spatula-off-shelf-i00-standard, run codex-gpt6_luna-xhigh, robolab-lunaxh-limited-0928, ws; https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/take_spatula_off_shelf/limited/). Human review per family (PROTOCOL §2.7)."
instance_seed = 0 # scene = environment/instance.json (the full scene state after the seeded reset + its hash); identity = dataset.toml digest
diff --git a/tasks/robolab-take-spatula-off-shelf-i00-standard/task.toml b/tasks/robolab-take-spatula-off-shelf-i00-standard/task.toml
index dd90616a61..96a8a1e6e6 100644
--- a/tasks/robolab-take-spatula-off-shelf-i00-standard/task.toml
+++ b/tasks/robolab-take-spatula-off-shelf-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "robolab"
robot = "franka-robotiq-2f85"
family = "take_spatula_off_shelf"
mode = "standard" # robot as a service: RGB, depth + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution: RoboLab ships no expert or planner. Positive examples graded 1 by the separate verifier: robolab-take-spatula-off-shelf-i__vipvPdN (robolab-take-spatula-off-shelf-i00-privileged, run codex-gpt6_luna-xhigh, robolab-lunacx-0925-2038, ws; https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/take_spatula_off_shelf/unlimited/); robolab-take-spatula-off-shelf-i__TqQfCCX (robolab-take-spatula-off-shelf-i00-standard, run codex-gpt6_luna-xhigh, robolab-lunaxh-limited-0928, ws; https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/take_spatula_off_shelf/limited/). Human review per family (PROTOCOL §2.7)."
diff --git a/tasks/robolab-throw-away-snacks-i00-privileged/task.toml b/tasks/robolab-throw-away-snacks-i00-privileged/task.toml
index 4d9c467204..4e36032e14 100644
--- a/tasks/robolab-throw-away-snacks-i00-privileged/task.toml
+++ b/tasks/robolab-throw-away-snacks-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "robolab" # NVIDIA RoboLab v0.3.1 on Isaac Lab 2.3.2 / I
robot = "franka-robotiq-2f85"
family = "throw_away_snacks"
mode = "privileged" # privileged state, resets and primitives through the tooling (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution: RoboLab ships no expert or planner. Positive examples graded 1 by the separate verifier: robolab-throw-away-snacks-i00__uUNyxva (robolab-throw-away-snacks-i00-privileged, run codex-gpt6_luna-xhigh, robolab-lunacx-0925-2038, ws3; https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/throw_away_snacks/unlimited/); robolab-throw-away-snacks-i00-li__wASjqcg (robolab-throw-away-snacks-i00-standard, run codex-gpt6_luna-xhigh, robolab-lunaxh-limited-0928, lab; https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/throw_away_snacks/limited/). Human review per family (PROTOCOL §2.7)."
instance_seed = 0 # scene = environment/instance.json (the full scene state after the seeded reset + its hash); identity = dataset.toml digest
diff --git a/tasks/robolab-throw-away-snacks-i00-standard/README.md b/tasks/robolab-throw-away-snacks-i00-standard/README.md
index d43d5f53a8..44a81eb430 100644
--- a/tasks/robolab-throw-away-snacks-i00-standard/README.md
+++ b/tasks/robolab-throw-away-snacks-i00-standard/README.md
@@ -73,3 +73,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
- 2026-10-10 (review round 3, board 0067): the shared service layer and the client only (`environment/sim/rcb_limited.py`, `environment/rcb_client.py`). A camera named more than once in `cameras` is rendered and returned once (N copies of one name made the service build N x N images within the frame budget, which could exhaust its memory). A service that has given up on the episode (it could not start it, or could not resume it after its crashes) answers every request with 409 and the reason, and `Robot()` / `wait_ready()` raise `EpisodeOver` at once; before, nothing answered and they waited out their timeout (up to an hour). The adapter, the collect hook, the grader, the scene and the pass rule are unchanged.
- 2026-10-10 (owner decision, board 0072): the agent's network egress is restricted to the model APIs (`[agent] network_mode = "allowlist"`, the hosts of `scripts/set_agent_egress.py`; since 2026-10-11 these include the package registries, PyPI and the Debian / Ubuntu archives, so that `pip install` and `apt-get install` work); it was open before (Harbor's default), and agents could download upstream sources, reference solutions and robot models from their own shell. The simulator image is `eai-robolab:0.1.2` (was 0.1.1): it holds the MDL materials and textures that RoboLab's asset library fetched from NVIDIA's Omniverse content bucket whenever a scene was built, so that the cameras show under the allowlist what they showed with open egress (on 0.1.1 without that download the table renders red). The scene, the frozen instance, the physics, the tooling and the pass rule are unchanged, so the task keeps its name; results from before this date were obtained with open egress.
- 2026-10-13 (owner, boards 0085 and 0087): the sim service's clock (`environment/sim/rcb_limited.py`, `environment/sim/server.py`, the docstring of `environment/rcb_client.py`). `meta["time_s"]` and `meta["time_left_s"]` count from the agent's first request to the service (any endpoint), and a restarted service keeps that moment; before, they counted from the service's own start of the episode (when its scene was ready). Where the service is ready before the agent's phase begins (the harness is installed in between: a median of 79 s over the recorded standard trials), that clock ran ahead of the agent's and `time_left_s` reached 0 while the agent still had that time. `/spec` says what the clock counts from (`budget.clock`), and the instruction's `## Time limit` tells the agent to plan by `time_left_s`. The service never ended an episode on this clock: the scene, the frozen instance, the pass rule and the grader are unchanged, so the task keeps its name and earlier results stand.
+- 2026-10-14 (board 0113, from board 0086): the success text says that the gripper must not be touching the objects at the moment of success ("at that moment"); the old wording "the gripper must touch none of them" was read by an agent as "never touch them" (board 0086). Text only: the scene, the frozen instance, RoboLab's success term and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/robolab-throw-away-snacks-i00-standard/instruction.md b/tasks/robolab-throw-away-snacks-i00-standard/instruction.md
index d96b849d26..699482c2cd 100644
--- a/tasks/robolab-throw-away-snacks-i00-standard/instruction.md
+++ b/tasks/robolab-throw-away-snacks-i00-standard/instruction.md
@@ -1,7 +1,7 @@
# RoboLab · throw_away_snacks — STANDARD MODE (robot as a service, eai-standard/2.2)
You control a **Franka Panda arm with a Robotiq 2F-85 gripper** (NVIDIA RoboLab's DROID setup) **as if it were real
-hardware**: the simulator runs in a service you cannot open. You get one client and one episode. **Task:** Put away the snacks in the bin. Success: The red apple and the yogurt cup must end inside the plastic pail (the trash bin) and the gripper must touch neither. "Inside" means the centre of the object's convex hull lies within the container's convex hull, extended without limit along the container's own up axis (the open side; straight up for a container standing normally).
+hardware**: the simulator runs in a service you cannot open. You get one client and one episode. **Task:** Put away the snacks in the bin. Success: The red apple and the yogurt cup must end inside the plastic pail (the trash bin) and, at that moment, the gripper must touch neither (let go of both at the end). "Inside" means the centre of the object's convex hull lies within the container's convex hull, extended without limit along the container's own up axis (the open side; straight up for a container standing normally).
## Rules (real-robot rules)
- **No reset.** The episode starts when the service comes up and ends on success (at the first control step where the
diff --git a/tasks/robolab-throw-away-snacks-i00-standard/task.toml b/tasks/robolab-throw-away-snacks-i00-standard/task.toml
index ec97bd59ef..7771af7a20 100644
--- a/tasks/robolab-throw-away-snacks-i00-standard/task.toml
+++ b/tasks/robolab-throw-away-snacks-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "robolab"
robot = "franka-robotiq-2f85"
family = "throw_away_snacks"
mode = "standard" # robot as a service: RGB, depth + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution: RoboLab ships no expert or planner. Positive examples graded 1 by the separate verifier: robolab-throw-away-snacks-i00__uUNyxva (robolab-throw-away-snacks-i00-privileged, run codex-gpt6_luna-xhigh, robolab-lunacx-0925-2038, ws3; https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/throw_away_snacks/unlimited/); robolab-throw-away-snacks-i00-li__wASjqcg (robolab-throw-away-snacks-i00-standard, run codex-gpt6_luna-xhigh, robolab-lunaxh-limited-0928, lab; https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/throw_away_snacks/limited/). Human review per family (PROTOCOL §2.7)."
diff --git a/tasks/robolab-tools-picking-drill-i00-privileged/task.toml b/tasks/robolab-tools-picking-drill-i00-privileged/task.toml
index 9db88c37b4..15c22e523f 100644
--- a/tasks/robolab-tools-picking-drill-i00-privileged/task.toml
+++ b/tasks/robolab-tools-picking-drill-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "robolab" # NVIDIA RoboLab v0.3.1 on Isaac Lab 2.3.2 / I
robot = "franka-robotiq-2f85"
family = "tools_picking_drill"
mode = "privileged" # privileged state, resets and primitives through the tooling (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution: RoboLab ships no expert or planner. Positive examples graded 1 by the separate verifier: robolab-tools-picking-drill-i00__8ghAF4z (robolab-tools-picking-drill-i00-standard, run codex-gpt6_luna-xhigh, robolab-lunacx-0925-2038, ws3; not published: its eai-standard/2.1 batch was withdrawn for the agent's stale wrist-camera pose; grader unaffected). Human review per family (PROTOCOL §2.7)."
instance_seed = 0 # scene = environment/instance.json (the full scene state after the seeded reset + its hash); identity = dataset.toml digest
diff --git a/tasks/robolab-tools-picking-drill-i00-standard/README.md b/tasks/robolab-tools-picking-drill-i00-standard/README.md
index ebc1cb906a..be6dd0d8d4 100644
--- a/tasks/robolab-tools-picking-drill-i00-standard/README.md
+++ b/tasks/robolab-tools-picking-drill-i00-standard/README.md
@@ -72,3 +72,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
- 2026-10-10 (review round 3, board 0067): the shared service layer and the client only (`environment/sim/rcb_limited.py`, `environment/rcb_client.py`). A camera named more than once in `cameras` is rendered and returned once (N copies of one name made the service build N x N images within the frame budget, which could exhaust its memory). A service that has given up on the episode (it could not start it, or could not resume it after its crashes) answers every request with 409 and the reason, and `Robot()` / `wait_ready()` raise `EpisodeOver` at once; before, nothing answered and they waited out their timeout (up to an hour). The adapter, the collect hook, the grader, the scene and the pass rule are unchanged.
- 2026-10-10 (owner decision, board 0072): the agent's network egress is restricted to the model APIs (`[agent] network_mode = "allowlist"`, the hosts of `scripts/set_agent_egress.py`; since 2026-10-11 these include the package registries, PyPI and the Debian / Ubuntu archives, so that `pip install` and `apt-get install` work); it was open before (Harbor's default), and agents could download upstream sources, reference solutions and robot models from their own shell. The simulator image is `eai-robolab:0.1.2` (was 0.1.1): it holds the MDL materials and textures that RoboLab's asset library fetched from NVIDIA's Omniverse content bucket whenever a scene was built, so that the cameras show under the allowlist what they showed with open egress (on 0.1.1 without that download the table renders red). The scene, the frozen instance, the physics, the tooling and the pass rule are unchanged, so the task keeps its name; results from before this date were obtained with open egress.
- 2026-10-13 (owner, boards 0085 and 0087): the sim service's clock (`environment/sim/rcb_limited.py`, `environment/sim/server.py`, the docstring of `environment/rcb_client.py`). `meta["time_s"]` and `meta["time_left_s"]` count from the agent's first request to the service (any endpoint), and a restarted service keeps that moment; before, they counted from the service's own start of the episode (when its scene was ready). Where the service is ready before the agent's phase begins (the harness is installed in between: a median of 79 s over the recorded standard trials), that clock ran ahead of the agent's and `time_left_s` reached 0 while the agent still had that time. `/spec` says what the clock counts from (`budget.clock`), and the instruction's `## Time limit` tells the agent to plan by `time_left_s`. The service never ended an episode on this clock: the scene, the frozen instance, the pass rule and the grader are unchanged, so the task keeps its name and earlier results stand.
+- 2026-10-14 (board 0113, from board 0086): the success text explains that RoboLab's term names the left bin although the drill starts in the right bin, and that the drill has to rest on the table itself. Text only: the scene, the frozen instance, RoboLab's success term and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/robolab-tools-picking-drill-i00-standard/instruction.md b/tasks/robolab-tools-picking-drill-i00-standard/instruction.md
index 3aea731593..58611efaff 100644
--- a/tasks/robolab-tools-picking-drill-i00-standard/instruction.md
+++ b/tasks/robolab-tools-picking-drill-i00-standard/instruction.md
@@ -1,7 +1,7 @@
# RoboLab · tools_picking_drill — STANDARD MODE (robot as a service, eai-standard/2.2)
You control a **Franka Panda arm with a Robotiq 2F-85 gripper** (NVIDIA RoboLab's DROID setup) **as if it were real
-hardware**: the simulator runs in a service you cannot open. You get one client and one episode. **Task:** Select the cordless drill and put it on the table. Success: The cordless drill must end outside the left bin and resting on the table, and the gripper must no longer touch it. "Outside" means the centre of the object's convex hull is not within the container's convex hull extended along the container's own up axis. "On the table" means the table supports it (its contact force points upwards, within 45 deg of vertical).
+hardware**: the simulator runs in a service you cannot open. You get one client and one episode. **Task:** Select the cordless drill and put it on the table. Success: The cordless drill must end outside the left bin and resting on the table, and the gripper must no longer touch it. (RoboLab's term names the left bin although the drill starts in the right bin: the drill has to come out of its bin and rest on the table itself.) "Outside" means the centre of the object's convex hull is not within the container's convex hull extended along the container's own up axis. "On the table" means the table supports it (its contact force points upwards, within 45 deg of vertical).
## Rules (real-robot rules)
- **No reset.** The episode starts when the service comes up and ends on success (at the first control step where the
diff --git a/tasks/robolab-tools-picking-drill-i00-standard/task.toml b/tasks/robolab-tools-picking-drill-i00-standard/task.toml
index 453b952afc..47de96eab9 100644
--- a/tasks/robolab-tools-picking-drill-i00-standard/task.toml
+++ b/tasks/robolab-tools-picking-drill-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "robolab"
robot = "franka-robotiq-2f85"
family = "tools_picking_drill"
mode = "standard" # robot as a service: RGB, depth + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution: RoboLab ships no expert or planner. Positive examples graded 1 by the separate verifier: robolab-tools-picking-drill-i00__8ghAF4z (robolab-tools-picking-drill-i00-standard, run codex-gpt6_luna-xhigh, robolab-lunacx-0925-2038, ws3; not published: its eai-standard/2.1 batch was withdrawn for the agent's stale wrist-camera pose; grader unaffected). Human review per family (PROTOCOL §2.7)."
diff --git a/tasks/robolab-unstack-rubiks-cube-i00-privileged/task.toml b/tasks/robolab-unstack-rubiks-cube-i00-privileged/task.toml
index 9348827e33..97f44d7ec0 100644
--- a/tasks/robolab-unstack-rubiks-cube-i00-privileged/task.toml
+++ b/tasks/robolab-unstack-rubiks-cube-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "robolab" # NVIDIA RoboLab v0.3.1 on Isaac Lab 2.3.2 / I
robot = "franka-robotiq-2f85"
family = "unstack_rubiks_cube"
mode = "privileged" # privileged state, resets and primitives through the tooling (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution: RoboLab ships no expert or planner. Positive examples graded 1 by the separate verifier: robolab-unstack-rubiks-cube-i00__qwsJfKd (robolab-unstack-rubiks-cube-i00-privileged, run codex-gpt6_luna-xhigh, robolab-lunacx-0925-2038, ws3; https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/unstack_rubiks_cube/unlimited/); robolab-unstack-rubiks-cube-i00__YAR6Yci (robolab-unstack-rubiks-cube-i00-standard, run codex-gpt6_luna-xhigh, robolab-lunaxh-limited-0928, lab; https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/unstack_rubiks_cube/limited/). Human review per family (PROTOCOL §2.7)."
instance_seed = 0 # scene = environment/instance.json (the full scene state after the seeded reset + its hash); identity = dataset.toml digest
diff --git a/tasks/robolab-unstack-rubiks-cube-i00-standard/README.md b/tasks/robolab-unstack-rubiks-cube-i00-standard/README.md
index 3990549369..784a27f79c 100644
--- a/tasks/robolab-unstack-rubiks-cube-i00-standard/README.md
+++ b/tasks/robolab-unstack-rubiks-cube-i00-standard/README.md
@@ -73,3 +73,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
- 2026-10-10 (review round 3, board 0067): the shared service layer and the client only (`environment/sim/rcb_limited.py`, `environment/rcb_client.py`). A camera named more than once in `cameras` is rendered and returned once (N copies of one name made the service build N x N images within the frame budget, which could exhaust its memory). A service that has given up on the episode (it could not start it, or could not resume it after its crashes) answers every request with 409 and the reason, and `Robot()` / `wait_ready()` raise `EpisodeOver` at once; before, nothing answered and they waited out their timeout (up to an hour). The adapter, the collect hook, the grader, the scene and the pass rule are unchanged.
- 2026-10-10 (owner decision, board 0072): the agent's network egress is restricted to the model APIs (`[agent] network_mode = "allowlist"`, the hosts of `scripts/set_agent_egress.py`; since 2026-10-11 these include the package registries, PyPI and the Debian / Ubuntu archives, so that `pip install` and `apt-get install` work); it was open before (Harbor's default), and agents could download upstream sources, reference solutions and robot models from their own shell. The simulator image is `eai-robolab:0.1.2` (was 0.1.1): it holds the MDL materials and textures that RoboLab's asset library fetched from NVIDIA's Omniverse content bucket whenever a scene was built, so that the cameras show under the allowlist what they showed with open egress (on 0.1.1 without that download the table renders red). The scene, the frozen instance, the physics, the tooling and the pass rule are unchanged, so the task keeps its name; results from before this date were obtained with open egress.
- 2026-10-13 (owner, boards 0085 and 0087): the sim service's clock (`environment/sim/rcb_limited.py`, `environment/sim/server.py`, the docstring of `environment/rcb_client.py`). `meta["time_s"]` and `meta["time_left_s"]` count from the agent's first request to the service (any endpoint), and a restarted service keeps that moment; before, they counted from the service's own start of the episode (when its scene was ready). Where the service is ready before the agent's phase begins (the harness is installed in between: a median of 79 s over the recorded standard trials), that clock ran ahead of the agent's and `time_left_s` reached 0 while the agent still had that time. `/spec` says what the clock counts from (`budget.clock`), and the instruction's `## Time limit` tells the agent to plan by `time_left_s`. The service never ended an episode on this clock: the scene, the frozen instance, the pass rule and the grader are unchanged, so the task keeps its name and earlier results stand.
+- 2026-10-14 (board 0113, from board 0086): the success text says that the gripper must not be touching the objects at the moment of success ("at that moment"); the old wording "the gripper must touch none of them" was read by an agent as "never touch them" (board 0086). Text only: the scene, the frozen instance, RoboLab's success term and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/robolab-unstack-rubiks-cube-i00-standard/instruction.md b/tasks/robolab-unstack-rubiks-cube-i00-standard/instruction.md
index ac649ec3b0..b0280d0d60 100644
--- a/tasks/robolab-unstack-rubiks-cube-i00-standard/instruction.md
+++ b/tasks/robolab-unstack-rubiks-cube-i00-standard/instruction.md
@@ -1,7 +1,7 @@
# RoboLab · unstack_rubiks_cube — STANDARD MODE (robot as a service, eai-standard/2.2)
You control a **Franka Panda arm with a Robotiq 2F-85 gripper** (NVIDIA RoboLab's DROID setup) **as if it were real
-hardware**: the simulator runs in a service you cannot open. You get one client and one episode. **Task:** Unstack the rubiks cube tower. Success: The three Rubik's cubes of the tower must end outside the grey bin and standing on the table, and the gripper must touch none of them. "Outside" means the centre of the object's convex hull is not within the container's convex hull extended along the container's own up axis. "On the table" means the table supports it (its contact force points upwards, within 45 deg of vertical).
+hardware**: the simulator runs in a service you cannot open. You get one client and one episode. **Task:** Unstack the rubiks cube tower. Success: The three Rubik's cubes of the tower must end outside the grey bin and standing on the table, and, at that moment, the gripper must touch none of them (moving them with the gripper is how the task is done; let go of them at the end). "Outside" means the centre of the object's convex hull is not within the container's convex hull extended along the container's own up axis. "On the table" means the table supports it (its contact force points upwards, within 45 deg of vertical).
## Rules (real-robot rules)
- **No reset.** The episode starts when the service comes up and ends on success (at the first control step where the
diff --git a/tasks/robolab-unstack-rubiks-cube-i00-standard/task.toml b/tasks/robolab-unstack-rubiks-cube-i00-standard/task.toml
index 3de86238a0..5732da873d 100644
--- a/tasks/robolab-unstack-rubiks-cube-i00-standard/task.toml
+++ b/tasks/robolab-unstack-rubiks-cube-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "robolab"
robot = "franka-robotiq-2f85"
family = "unstack_rubiks_cube"
mode = "standard" # robot as a service: RGB, depth + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution: RoboLab ships no expert or planner. Positive examples graded 1 by the separate verifier: robolab-unstack-rubiks-cube-i00__qwsJfKd (robolab-unstack-rubiks-cube-i00-privileged, run codex-gpt6_luna-xhigh, robolab-lunacx-0925-2038, ws3; https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/unstack_rubiks_cube/unlimited/); robolab-unstack-rubiks-cube-i00__YAR6Yci (robolab-unstack-rubiks-cube-i00-standard, run codex-gpt6_luna-xhigh, robolab-lunaxh-limited-0928, lab; https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/unstack_rubiks_cube/limited/). Human review per family (PROTOCOL §2.7)."
diff --git a/tasks/robolab-utensils-in-mug-i00-privileged/task.toml b/tasks/robolab-utensils-in-mug-i00-privileged/task.toml
index 10c1d5d0c8..4ed933c42d 100644
--- a/tasks/robolab-utensils-in-mug-i00-privileged/task.toml
+++ b/tasks/robolab-utensils-in-mug-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "robolab" # NVIDIA RoboLab v0.3.1 on Isaac Lab 2.3.2 / I
robot = "franka-robotiq-2f85"
family = "utensils_in_mug"
mode = "privileged" # privileged state, resets and primitives through the tooling (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution: RoboLab ships no expert or planner. Positive examples graded 1 by the separate verifier: robolab-utensils-in-mug-i00__8k6TcYT (robolab-utensils-in-mug-i00-privileged, run codex-gpt6_luna-xhigh, robolab-lunacx-0925-2038, ws3; https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/utensils_in_mug/unlimited/). Human review per family (PROTOCOL §2.7)."
instance_seed = 0 # scene = environment/instance.json (the full scene state after the seeded reset + its hash); identity = dataset.toml digest
diff --git a/tasks/robolab-utensils-in-mug-i00-standard/README.md b/tasks/robolab-utensils-in-mug-i00-standard/README.md
index 5f98c96f2b..c4b6d01c1d 100644
--- a/tasks/robolab-utensils-in-mug-i00-standard/README.md
+++ b/tasks/robolab-utensils-in-mug-i00-standard/README.md
@@ -72,3 +72,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
- 2026-10-10 (review round 3, board 0067): the shared service layer and the client only (`environment/sim/rcb_limited.py`, `environment/rcb_client.py`). A camera named more than once in `cameras` is rendered and returned once (N copies of one name made the service build N x N images within the frame budget, which could exhaust its memory). A service that has given up on the episode (it could not start it, or could not resume it after its crashes) answers every request with 409 and the reason, and `Robot()` / `wait_ready()` raise `EpisodeOver` at once; before, nothing answered and they waited out their timeout (up to an hour). The adapter, the collect hook, the grader, the scene and the pass rule are unchanged.
- 2026-10-10 (owner decision, board 0072): the agent's network egress is restricted to the model APIs (`[agent] network_mode = "allowlist"`, the hosts of `scripts/set_agent_egress.py`; since 2026-10-11 these include the package registries, PyPI and the Debian / Ubuntu archives, so that `pip install` and `apt-get install` work); it was open before (Harbor's default), and agents could download upstream sources, reference solutions and robot models from their own shell. The simulator image is `eai-robolab:0.1.2` (was 0.1.1): it holds the MDL materials and textures that RoboLab's asset library fetched from NVIDIA's Omniverse content bucket whenever a scene was built, so that the cameras show under the allowlist what they showed with open egress (on 0.1.1 without that download the table renders red). The scene, the frozen instance, the physics, the tooling and the pass rule are unchanged, so the task keeps its name; results from before this date were obtained with open egress.
- 2026-10-13 (owner, boards 0085 and 0087): the sim service's clock (`environment/sim/rcb_limited.py`, `environment/sim/server.py`, the docstring of `environment/rcb_client.py`). `meta["time_s"]` and `meta["time_left_s"]` count from the agent's first request to the service (any endpoint), and a restarted service keeps that moment; before, they counted from the service's own start of the episode (when its scene was ready). Where the service is ready before the agent's phase begins (the harness is installed in between: a median of 79 s over the recorded standard trials), that clock ran ahead of the agent's and `time_left_s` reached 0 while the agent still had that time. `/spec` says what the clock counts from (`budget.clock`), and the instruction's `## Time limit` tells the agent to plan by `time_left_s`. The service never ended an episode on this clock: the scene, the frozen instance, the pass rule and the grader are unchanged, so the task keeps its name and earlier results stand.
+- 2026-10-14 (board 0113, from board 0086): the success text says that the gripper must not be touching the objects at the moment of success ("at that moment"); the old wording "the gripper must touch none of them" was read by an agent as "never touch them" (board 0086). Text only: the scene, the frozen instance, RoboLab's success term and the verifier are unchanged, so the task keeps its name.
diff --git a/tasks/robolab-utensils-in-mug-i00-standard/instruction.md b/tasks/robolab-utensils-in-mug-i00-standard/instruction.md
index 2b42419778..1793b478c7 100644
--- a/tasks/robolab-utensils-in-mug-i00-standard/instruction.md
+++ b/tasks/robolab-utensils-in-mug-i00-standard/instruction.md
@@ -1,7 +1,7 @@
# RoboLab · utensils_in_mug — STANDARD MODE (robot as a service, eai-standard/2.2)
You control a **Franka Panda arm with a Robotiq 2F-85 gripper** (NVIDIA RoboLab's DROID setup) **as if it were real
-hardware**: the simulator runs in a service you cannot open. You get one client and one episode. **Task:** Put the fork and spoon in the ceramicmug. Success: The fork on the bagel plate and the spoon in the red bowl must end inside the ceramic mug and the gripper must touch neither. "Inside" means the centre of the object's convex hull lies within the container's convex hull, extended without limit along the container's own up axis (the open side; straight up for a container standing normally).
+hardware**: the simulator runs in a service you cannot open. You get one client and one episode. **Task:** Put the fork and spoon in the ceramicmug. Success: The fork on the bagel plate and the spoon in the red bowl must end inside the ceramic mug and, at that moment, the gripper must touch neither (let go of both at the end). "Inside" means the centre of the object's convex hull lies within the container's convex hull, extended without limit along the container's own up axis (the open side; straight up for a container standing normally).
## Rules (real-robot rules)
- **No reset.** The episode starts when the service comes up and ends on success (at the first control step where the
diff --git a/tasks/robolab-utensils-in-mug-i00-standard/task.toml b/tasks/robolab-utensils-in-mug-i00-standard/task.toml
index fcc0c10480..c662782c85 100644
--- a/tasks/robolab-utensils-in-mug-i00-standard/task.toml
+++ b/tasks/robolab-utensils-in-mug-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "robolab"
robot = "franka-robotiq-2f85"
family = "utensils_in_mug"
mode = "standard" # robot as a service: RGB, depth + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution: RoboLab ships no expert or planner. Positive examples graded 1 by the separate verifier: robolab-utensils-in-mug-i00__8k6TcYT (robolab-utensils-in-mug-i00-privileged, run codex-gpt6_luna-xhigh, robolab-lunacx-0925-2038, ws3; https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/utensils_in_mug/unlimited/). Human review per family (PROTOCOL §2.7)."
diff --git a/tasks/robolab-white-mug-in-center-of-table-i00-privileged/task.toml b/tasks/robolab-white-mug-in-center-of-table-i00-privileged/task.toml
index d311191194..ac7e1c9d2d 100644
--- a/tasks/robolab-white-mug-in-center-of-table-i00-privileged/task.toml
+++ b/tasks/robolab-white-mug-in-center-of-table-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "robolab" # NVIDIA RoboLab v0.3.1 on Isaac Lab 2.3.2 / I
robot = "franka-robotiq-2f85"
family = "white_mug_in_center_of_table"
mode = "privileged" # privileged state, resets and primitives through the tooling (docs/PROTOCOL.md §3)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution: RoboLab ships no expert or planner. Positive examples graded 1 by the separate verifier: robolab-white-mug-in-center-of-t__Jf6tksp (robolab-white-mug-in-center-of-table-i00-standard, run codex-gpt6_luna-xhigh, robolab-lunaxh-limited-0928, lab; https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/white_mug_in_center_of_table/limited/). Human review per family (PROTOCOL §2.7)."
instance_seed = 0 # scene = environment/instance.json (the full scene state after the seeded reset + its hash); identity = dataset.toml digest
diff --git a/tasks/robolab-white-mug-in-center-of-table-i00-standard/task.toml b/tasks/robolab-white-mug-in-center-of-table-i00-standard/task.toml
index e826361247..4382d161cf 100644
--- a/tasks/robolab-white-mug-in-center-of-table-i00-standard/task.toml
+++ b/tasks/robolab-white-mug-in-center-of-table-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "robolab"
robot = "franka-robotiq-2f85"
family = "white_mug_in_center_of_table"
mode = "standard" # robot as a service: RGB, depth + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution: RoboLab ships no expert or planner. Positive examples graded 1 by the separate verifier: robolab-white-mug-in-center-of-t__Jf6tksp (robolab-white-mug-in-center-of-table-i00-standard, run codex-gpt6_luna-xhigh, robolab-lunaxh-limited-0928, lab; https://embodied-agent-interface-v2-internal.github.io/runs/robolab/codex-gpt6_luna-xhigh/white_mug_in_center_of_table/limited/). Human review per family (PROTOCOL §2.7)."
diff --git a/tasks/roboolympics-airhockey-match-i00-standard-realtime/README.md b/tasks/roboolympics-airhockey-match-i00-standard-realtime/README.md
index 8cddff1326..e017f2f1c9 100644
--- a/tasks/roboolympics-airhockey-match-i00-standard-realtime/README.md
+++ b/tasks/roboolympics-airhockey-match-i00-standard-realtime/README.md
@@ -9,6 +9,7 @@
| Family / instance | `airhockey_match` / seed 101 |
| Oracle | `full` — solution/: the sport's RoboOlympics reference player (camera images, proprio and /spec only) through the agent's client; it wins this instance (PR #106) |
| Agent time budget | 7200 s; tokens: protocol default (PROTOCOL §5) |
+| Real time | slowdown S = 10 (the world runs at 1/S of real speed: one control step every S x control_dt s of wall time; `RCB_SLOWDOWN`'s default); reset budget 0 (`RCB_RESET_BUDGET`'s default) (PROTOCOL §4b) |
| Base image | `ghcr.io/mll-lab-nu/eai-olympiad:0.2.0` |
## Task
diff --git a/tasks/roboolympics-airhockey-match-i00-standard-realtime/task.toml b/tasks/roboolympics-airhockey-match-i00-standard-realtime/task.toml
index 8dc7b7d038..e56511412b 100644
--- a/tasks/roboolympics-airhockey-match-i00-standard-realtime/task.toml
+++ b/tasks/roboolympics-airhockey-match-i00-standard-realtime/task.toml
@@ -30,10 +30,11 @@ simulator = "mujoco"
robot = "kuka-iiwa14"
family = "airhockey_match"
mode = "standard-realtime"
+dynamics = "dynamic" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
+slowdown = 10 # standard-realtime: the world runs at 1/S of real speed (docs/PROTOCOL.md §4b); RCB_SLOWDOWN's default
protocol = "eai-standard/2.3"
instance_seed = 101
opponent_level = 1
-slowdown = 10
revision = "airhockey-program-v4+players-25k+core-0.2"
play = "program"
action_level = "low"
diff --git a/tasks/roboolympics-badminton-match-i00-standard-realtime/README.md b/tasks/roboolympics-badminton-match-i00-standard-realtime/README.md
index affbd8cf63..df8134312c 100644
--- a/tasks/roboolympics-badminton-match-i00-standard-realtime/README.md
+++ b/tasks/roboolympics-badminton-match-i00-standard-realtime/README.md
@@ -9,6 +9,7 @@
| Family / instance | `badminton_match` / seed 101 |
| Oracle | `full` — solution/: the sport's RoboOlympics reference player (camera images, proprio and /spec only) through the agent's client; it wins this instance (PR #106) |
| Agent time budget | 7200 s; tokens: protocol default (PROTOCOL §5) |
+| Real time | slowdown S = 10 (the world runs at 1/S of real speed: one control step every S x control_dt s of wall time; `RCB_SLOWDOWN`'s default); reset budget 0 (`RCB_RESET_BUDGET`'s default) (PROTOCOL §4b) |
| Base image | `ghcr.io/mll-lab-nu/eai-olympiad:0.2.0` |
## Task
diff --git a/tasks/roboolympics-badminton-match-i00-standard-realtime/task.toml b/tasks/roboolympics-badminton-match-i00-standard-realtime/task.toml
index e8dce33a16..3d37675246 100644
--- a/tasks/roboolympics-badminton-match-i00-standard-realtime/task.toml
+++ b/tasks/roboolympics-badminton-match-i00-standard-realtime/task.toml
@@ -30,10 +30,11 @@ simulator = "mujoco"
robot = "franka-panda"
family = "badminton_match"
mode = "standard-realtime"
+dynamics = "dynamic" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
+slowdown = 10 # standard-realtime: the world runs at 1/S of real speed (docs/PROTOCOL.md §4b); RCB_SLOWDOWN's default
protocol = "eai-standard/2.3"
instance_seed = 101
opponent_level = 1
-slowdown = 10
revision = "badminton-wallclock-5+duel-7+core-0.2"
play = "program"
action_level = "high"
diff --git a/tasks/roboolympics-tabletennis-match-i00-standard-realtime/README.md b/tasks/roboolympics-tabletennis-match-i00-standard-realtime/README.md
index 18917d7b56..fbaad51f0d 100644
--- a/tasks/roboolympics-tabletennis-match-i00-standard-realtime/README.md
+++ b/tasks/roboolympics-tabletennis-match-i00-standard-realtime/README.md
@@ -9,6 +9,7 @@
| Family / instance | `tabletennis_match` / seed 101 |
| Oracle | `full` — solution/: the sport's RoboOlympics reference player (camera images, proprio and /spec only) through the agent's client; it wins this instance (PR #106) |
| Agent time budget | 7200 s; tokens: protocol default (PROTOCOL §5) |
+| Real time | slowdown S = 10 (the world runs at 1/S of real speed: one control step every S x control_dt s of wall time; `RCB_SLOWDOWN`'s default); reset budget 0 (`RCB_RESET_BUDGET`'s default) (PROTOCOL §4b) |
| Base image | `ghcr.io/mll-lab-nu/eai-olympiad:0.2.0` |
## Task
diff --git a/tasks/roboolympics-tabletennis-match-i00-standard-realtime/task.toml b/tasks/roboolympics-tabletennis-match-i00-standard-realtime/task.toml
index f6faf3c077..4c99f1ed96 100644
--- a/tasks/roboolympics-tabletennis-match-i00-standard-realtime/task.toml
+++ b/tasks/roboolympics-tabletennis-match-i00-standard-realtime/task.toml
@@ -30,10 +30,11 @@ simulator = "mujoco"
robot = "franka-panda"
family = "tabletennis_match"
mode = "standard-realtime"
+dynamics = "dynamic" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
+slowdown = 10 # standard-realtime: the world runs at 1/S of real speed (docs/PROTOCOL.md §4b); RCB_SLOWDOWN's default
protocol = "eai-standard/2.3"
instance_seed = 101
opponent_level = 1
-slowdown = 10
revision = "wallclock-6+duel-wallclock-4+core-0.2"
play = "program"
action_level = "high"
diff --git a/tasks/robopaint-acrylic-applesoranges-i00-privileged/task.toml b/tasks/robopaint-acrylic-applesoranges-i00-privileged/task.toml
index 21b07636c9..7270fb7e64 100644
--- a/tasks/robopaint-acrylic-applesoranges-i00-privileged/task.toml
+++ b/tasks/robopaint-acrylic-applesoranges-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "maniskill" # ManiSkill 3.0.1 / SAPIEN 3.0.3 (PhysX on the
robot = "franka-panda" # PandaStick with a paint brush: rigid handle + 15 mm collision-free tuft, TCP = the tuft's tip
family = "robopaint_acrylic_applesoranges"
mode = "privileged" # privileged (exact sheets, score, brush state) | standard (RGB only: the -standard twin)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "partial" # solution/solve.sh replays the frozen reference plan; its recorded scores: oracle_evidence
oracle_evidence = "Reference solution in solution/, unchanged from the earlier version of this task (the picture score only; its Harbor oracle and nop runs, and what changed since: docs/benchmarks/robopaint.md, Oracle evidence). Under the process rules the verifier scores it 1 (tests/grade.py with rcb_brush_strict.py, two fresh replays, ws3, 2026-10-05): path_mm 11276 of the 33900 mm budget; its strokes are sweeps. Harbor oracle and nop runs on one pair per family: docs/benchmarks/robopaint.md (Validation)."
instance_seed = 0 # scene = environment/instance.json (sheets, brush preset, paint wells, start joints, instance_sha; the exemplar is target.png)
diff --git a/tasks/robopaint-acrylic-applesoranges-i00-standard/task.toml b/tasks/robopaint-acrylic-applesoranges-i00-standard/task.toml
index d150b3cb4f..bf51e9bcd0 100644
--- a/tasks/robopaint-acrylic-applesoranges-i00-standard/task.toml
+++ b/tasks/robopaint-acrylic-applesoranges-i00-standard/task.toml
@@ -31,6 +31,7 @@ simulator = "maniskill"
robot = "franka-panda"
family = "robopaint_acrylic_applesoranges"
mode = "standard" # camera RGB + proprioception only; no sheet readout, no score, no reset, no brush model
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "partial" # solution/: the privileged oracle's joint targets, streamed through the client
oracle_evidence = "Reference solution in solution/ (streams the privileged reference's joint targets), unchanged from the earlier version of this task (the picture score only; its Harbor oracle and nop runs, and what changed since: docs/benchmarks/robopaint.md, Oracle evidence). The same trajectory scores 1 under the process rules in the verifier, whose replay CLI the collect hook runs (ws3, 2026-10-05): path_mm 11276 of the 33900 mm budget; its strokes are sweeps. Harbor oracle and nop runs on one pair per family: docs/benchmarks/robopaint.md (Validation)."
diff --git a/tasks/robopaint-acrylic-bedroom-i00-privileged/task.toml b/tasks/robopaint-acrylic-bedroom-i00-privileged/task.toml
index 8cb5d5e1ba..cdb05391c3 100644
--- a/tasks/robopaint-acrylic-bedroom-i00-privileged/task.toml
+++ b/tasks/robopaint-acrylic-bedroom-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "maniskill" # ManiSkill 3.0.1 / SAPIEN 3.0.3 (PhysX on the
robot = "franka-panda" # PandaStick with a paint brush: rigid handle + 15 mm collision-free tuft, TCP = the tuft's tip
family = "robopaint_acrylic_bedroom"
mode = "privileged" # privileged (exact sheets, score, brush state) | standard (RGB only: the -standard twin)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "partial" # solution/solve.sh replays the frozen reference plan; its recorded scores: oracle_evidence
oracle_evidence = "Reference solution in solution/, unchanged from the earlier version of this task (the picture score only; its Harbor oracle and nop runs, and what changed since: docs/benchmarks/robopaint.md, Oracle evidence). Under the process rules the verifier scores it 1 (tests/grade.py with rcb_brush_strict.py, two fresh replays, ws3, 2026-10-05): path_mm 13415 of the 40300 mm budget; its strokes are sweeps. Harbor oracle and nop runs on one pair per family: docs/benchmarks/robopaint.md (Validation)."
instance_seed = 0 # scene = environment/instance.json (sheets, brush preset, paint wells, start joints, instance_sha; the exemplar is target.png)
diff --git a/tasks/robopaint-acrylic-bedroom-i00-standard/task.toml b/tasks/robopaint-acrylic-bedroom-i00-standard/task.toml
index ce157c88a0..95da3b7644 100644
--- a/tasks/robopaint-acrylic-bedroom-i00-standard/task.toml
+++ b/tasks/robopaint-acrylic-bedroom-i00-standard/task.toml
@@ -31,6 +31,7 @@ simulator = "maniskill"
robot = "franka-panda"
family = "robopaint_acrylic_bedroom"
mode = "standard" # camera RGB + proprioception only; no sheet readout, no score, no reset, no brush model
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "partial" # solution/: the privileged oracle's joint targets, streamed through the client
oracle_evidence = "Reference solution in solution/ (streams the privileged reference's joint targets), unchanged from the earlier version of this task (the picture score only; its Harbor oracle and nop runs, and what changed since: docs/benchmarks/robopaint.md, Oracle evidence). The same trajectory scores 1 under the process rules in the verifier, whose replay CLI the collect hook runs (ws3, 2026-10-05): path_mm 13415 of the 40300 mm budget; its strokes are sweeps. Harbor oracle and nop runs on one pair per family: docs/benchmarks/robopaint.md (Validation)."
diff --git a/tasks/robopaint-acrylic-bluehorses-i00-privileged/task.toml b/tasks/robopaint-acrylic-bluehorses-i00-privileged/task.toml
index 62bc87d039..76ea49f9a6 100644
--- a/tasks/robopaint-acrylic-bluehorses-i00-privileged/task.toml
+++ b/tasks/robopaint-acrylic-bluehorses-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "maniskill" # ManiSkill 3.0.1 / SAPIEN 3.0.3 (PhysX on the
robot = "franka-panda" # PandaStick with a paint brush: rigid handle + 15 mm collision-free tuft, TCP = the tuft's tip
family = "robopaint_acrylic_bluehorses"
mode = "privileged" # privileged (exact sheets, score, brush state) | standard (RGB only: the -standard twin)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "partial" # solution/solve.sh replays the frozen reference plan; its recorded scores: oracle_evidence
oracle_evidence = "Reference solution in solution/, unchanged from the earlier version of this task (the picture score only; its Harbor oracle and nop runs, and what changed since: docs/benchmarks/robopaint.md, Oracle evidence). Under the process rules the verifier scores it 1 (tests/grade.py with rcb_brush_strict.py, two fresh replays, ws3, 2026-10-05): path_mm 11956 of the 35900 mm budget; its strokes are sweeps. Harbor oracle and nop runs on one pair per family: docs/benchmarks/robopaint.md (Validation)."
instance_seed = 0 # scene = environment/instance.json (sheets, brush preset, paint wells, start joints, instance_sha; the exemplar is target.png)
diff --git a/tasks/robopaint-acrylic-bluehorses-i00-standard/task.toml b/tasks/robopaint-acrylic-bluehorses-i00-standard/task.toml
index a1c15a73b6..cf7144aff9 100644
--- a/tasks/robopaint-acrylic-bluehorses-i00-standard/task.toml
+++ b/tasks/robopaint-acrylic-bluehorses-i00-standard/task.toml
@@ -31,6 +31,7 @@ simulator = "maniskill"
robot = "franka-panda"
family = "robopaint_acrylic_bluehorses"
mode = "standard" # camera RGB + proprioception only; no sheet readout, no score, no reset, no brush model
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "partial" # solution/: the privileged oracle's joint targets, streamed through the client
oracle_evidence = "Reference solution in solution/ (streams the privileged reference's joint targets), unchanged from the earlier version of this task (the picture score only; its Harbor oracle and nop runs, and what changed since: docs/benchmarks/robopaint.md, Oracle evidence). The same trajectory scores 1 under the process rules in the verifier, whose replay CLI the collect hook runs (ws3, 2026-10-05): path_mm 11956 of the 35900 mm budget; its strokes are sweeps. Harbor oracle and nop runs on one pair per family: docs/benchmarks/robopaint.md (Validation)."
diff --git a/tasks/robopaint-acrylic-greatwave-i00-privileged/task.toml b/tasks/robopaint-acrylic-greatwave-i00-privileged/task.toml
index e7d1da2ec3..25f2b56f62 100644
--- a/tasks/robopaint-acrylic-greatwave-i00-privileged/task.toml
+++ b/tasks/robopaint-acrylic-greatwave-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "maniskill" # ManiSkill 3.0.1 / SAPIEN 3.0.3 (PhysX on the
robot = "franka-panda" # PandaStick with a paint brush: rigid handle + 15 mm collision-free tuft, TCP = the tuft's tip
family = "robopaint_acrylic_greatwave"
mode = "privileged" # privileged (exact sheets, score, brush state) | standard (RGB only: the -standard twin)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "partial" # solution/solve.sh replays the frozen reference plan; its recorded scores: oracle_evidence
oracle_evidence = "Reference solution in solution/, unchanged from the earlier version of this task (the picture score only; its Harbor oracle and nop runs, and what changed since: docs/benchmarks/robopaint.md, Oracle evidence). Under the process rules the verifier scores it 1 (tests/grade.py with rcb_brush_strict.py, two fresh replays, ws3, 2026-10-05): path_mm 12614 of the 37900 mm budget; its strokes are sweeps. Harbor oracle and nop runs on one pair per family: docs/benchmarks/robopaint.md (Validation)."
instance_seed = 0 # scene = environment/instance.json (sheets, brush preset, paint wells, start joints, instance_sha; the exemplar is target.png)
diff --git a/tasks/robopaint-acrylic-greatwave-i00-standard/task.toml b/tasks/robopaint-acrylic-greatwave-i00-standard/task.toml
index 645fc7e789..8b44f80970 100644
--- a/tasks/robopaint-acrylic-greatwave-i00-standard/task.toml
+++ b/tasks/robopaint-acrylic-greatwave-i00-standard/task.toml
@@ -31,6 +31,7 @@ simulator = "maniskill"
robot = "franka-panda"
family = "robopaint_acrylic_greatwave"
mode = "standard" # camera RGB + proprioception only; no sheet readout, no score, no reset, no brush model
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "partial" # solution/: the privileged oracle's joint targets, streamed through the client
oracle_evidence = "Reference solution in solution/ (streams the privileged reference's joint targets), unchanged from the earlier version of this task (the picture score only; its Harbor oracle and nop runs, and what changed since: docs/benchmarks/robopaint.md, Oracle evidence). The same trajectory scores 1 under the process rules in the verifier, whose replay CLI the collect hook runs (ws3, 2026-10-05): path_mm 12614 of the 37900 mm budget; its strokes are sweeps. Harbor oracle and nop runs on one pair per family: docs/benchmarks/robopaint.md (Validation)."
diff --git a/tasks/robopaint-acrylic-kajikazawa-i00-privileged/task.toml b/tasks/robopaint-acrylic-kajikazawa-i00-privileged/task.toml
index 8fe1046612..5ce50ff48b 100644
--- a/tasks/robopaint-acrylic-kajikazawa-i00-privileged/task.toml
+++ b/tasks/robopaint-acrylic-kajikazawa-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "maniskill" # ManiSkill 3.0.1 / SAPIEN 3.0.3 (PhysX on the
robot = "franka-panda" # PandaStick with a paint brush: rigid handle + 15 mm collision-free tuft, TCP = the tuft's tip
family = "robopaint_acrylic_kajikazawa"
mode = "privileged" # privileged (exact sheets, score, brush state) | standard (RGB only: the -standard twin)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "partial" # solution/solve.sh replays the frozen reference plan; its recorded scores: oracle_evidence
oracle_evidence = "Reference solution in solution/, unchanged from the earlier version of this task (the picture score only; its Harbor oracle and nop runs, and what changed since: docs/benchmarks/robopaint.md, Oracle evidence). Under the process rules the verifier scores it 1 (tests/grade.py with rcb_brush_strict.py, two fresh replays, ws3, 2026-10-05): path_mm 14746 of the 44300 mm budget; its strokes are sweeps. Harbor oracle and nop runs on one pair per family: docs/benchmarks/robopaint.md (Validation)."
instance_seed = 0 # scene = environment/instance.json (sheets, brush preset, paint wells, start joints, instance_sha; the exemplar is target.png)
diff --git a/tasks/robopaint-acrylic-kajikazawa-i00-standard/task.toml b/tasks/robopaint-acrylic-kajikazawa-i00-standard/task.toml
index 362f9c4da2..82d9dcb401 100644
--- a/tasks/robopaint-acrylic-kajikazawa-i00-standard/task.toml
+++ b/tasks/robopaint-acrylic-kajikazawa-i00-standard/task.toml
@@ -31,6 +31,7 @@ simulator = "maniskill"
robot = "franka-panda"
family = "robopaint_acrylic_kajikazawa"
mode = "standard" # camera RGB + proprioception only; no sheet readout, no score, no reset, no brush model
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "partial" # solution/: the privileged oracle's joint targets, streamed through the client
oracle_evidence = "Reference solution in solution/ (streams the privileged reference's joint targets), unchanged from the earlier version of this task (the picture score only; its Harbor oracle and nop runs, and what changed since: docs/benchmarks/robopaint.md, Oracle evidence). The same trajectory scores 1 under the process rules in the verifier, whose replay CLI the collect hook runs (ws3, 2026-10-05): path_mm 14746 of the 44300 mm budget; its strokes are sweeps. Harbor oracle and nop runs on one pair per family: docs/benchmarks/robopaint.md (Validation)."
diff --git a/tasks/robopaint-acrylic-plumpark-i00-privileged/task.toml b/tasks/robopaint-acrylic-plumpark-i00-privileged/task.toml
index ebea3f095d..706e793cdb 100644
--- a/tasks/robopaint-acrylic-plumpark-i00-privileged/task.toml
+++ b/tasks/robopaint-acrylic-plumpark-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "maniskill" # ManiSkill 3.0.1 / SAPIEN 3.0.3 (PhysX on the
robot = "franka-panda" # PandaStick with a paint brush: rigid handle + 15 mm collision-free tuft, TCP = the tuft's tip
family = "robopaint_acrylic_plumpark"
mode = "privileged" # privileged (exact sheets, score, brush state) | standard (RGB only: the -standard twin)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "partial" # solution/solve.sh replays the frozen reference plan; its recorded scores: oracle_evidence
oracle_evidence = "Reference solution in solution/, unchanged from the earlier version of this task (the picture score only; its Harbor oracle and nop runs, and what changed since: docs/benchmarks/robopaint.md, Oracle evidence). Under the process rules the verifier scores it 1 (tests/grade.py with rcb_brush_strict.py, two fresh replays, ws3, 2026-10-05): path_mm 14099 of the 42300 mm budget; its strokes are sweeps. Harbor oracle and nop runs on one pair per family: docs/benchmarks/robopaint.md (Validation)."
instance_seed = 0 # scene = environment/instance.json (sheets, brush preset, paint wells, start joints, instance_sha; the exemplar is target.png)
diff --git a/tasks/robopaint-acrylic-plumpark-i00-standard/task.toml b/tasks/robopaint-acrylic-plumpark-i00-standard/task.toml
index ce816bb466..1e7251e186 100644
--- a/tasks/robopaint-acrylic-plumpark-i00-standard/task.toml
+++ b/tasks/robopaint-acrylic-plumpark-i00-standard/task.toml
@@ -31,6 +31,7 @@ simulator = "maniskill"
robot = "franka-panda"
family = "robopaint_acrylic_plumpark"
mode = "standard" # camera RGB + proprioception only; no sheet readout, no score, no reset, no brush model
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "partial" # solution/: the privileged oracle's joint targets, streamed through the client
oracle_evidence = "Reference solution in solution/ (streams the privileged reference's joint targets), unchanged from the earlier version of this task (the picture score only; its Harbor oracle and nop runs, and what changed since: docs/benchmarks/robopaint.md, Oracle evidence). The same trajectory scores 1 under the process rules in the verifier, whose replay CLI the collect hook runs (ws3, 2026-10-05): path_mm 14099 of the 42300 mm budget; its strokes are sweeps. Harbor oracle and nop runs on one pair per family: docs/benchmarks/robopaint.md (Validation)."
diff --git a/tasks/robopaint-acrylic-redfuji-i00-privileged/task.toml b/tasks/robopaint-acrylic-redfuji-i00-privileged/task.toml
index a6095adcda..8005781df0 100644
--- a/tasks/robopaint-acrylic-redfuji-i00-privileged/task.toml
+++ b/tasks/robopaint-acrylic-redfuji-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "maniskill" # ManiSkill 3.0.1 / SAPIEN 3.0.3 (PhysX on the
robot = "franka-panda" # PandaStick with a paint brush: rigid handle + 15 mm collision-free tuft, TCP = the tuft's tip
family = "robopaint_acrylic_redfuji"
mode = "privileged" # privileged (exact sheets, score, brush state) | standard (RGB only: the -standard twin)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "partial" # solution/solve.sh replays the frozen reference plan; its recorded scores: oracle_evidence
oracle_evidence = "Reference solution in solution/, unchanged from the earlier version of this task (the picture score only; its Harbor oracle and nop runs, and what changed since: docs/benchmarks/robopaint.md, Oracle evidence). Under the process rules the verifier scores it 1 (tests/grade.py with rcb_brush_strict.py, two fresh replays, ws3, 2026-10-05): path_mm 17225 of the 51700 mm budget; its strokes are sweeps. Harbor oracle and nop runs on one pair per family: docs/benchmarks/robopaint.md (Validation)."
instance_seed = 0 # scene = environment/instance.json (sheets, brush preset, paint wells, start joints, instance_sha; the exemplar is target.png)
diff --git a/tasks/robopaint-acrylic-redfuji-i00-standard/task.toml b/tasks/robopaint-acrylic-redfuji-i00-standard/task.toml
index 4b73e815b6..822289d09b 100644
--- a/tasks/robopaint-acrylic-redfuji-i00-standard/task.toml
+++ b/tasks/robopaint-acrylic-redfuji-i00-standard/task.toml
@@ -31,6 +31,7 @@ simulator = "maniskill"
robot = "franka-panda"
family = "robopaint_acrylic_redfuji"
mode = "standard" # camera RGB + proprioception only; no sheet readout, no score, no reset, no brush model
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "partial" # solution/: the privileged oracle's joint targets, streamed through the client
oracle_evidence = "Reference solution in solution/ (streams the privileged reference's joint targets), unchanged from the earlier version of this task (the picture score only; its Harbor oracle and nop runs, and what changed since: docs/benchmarks/robopaint.md, Oracle evidence). The same trajectory scores 1 under the process rules in the verifier, whose replay CLI the collect hook runs (ws3, 2026-10-05): path_mm 17225 of the 51700 mm budget; its strokes are sweeps. Harbor oracle and nop runs on one pair per family: docs/benchmarks/robopaint.md (Validation)."
diff --git a/tasks/robopaint-acrylic-suddenshower-i00-privileged/task.toml b/tasks/robopaint-acrylic-suddenshower-i00-privileged/task.toml
index 5b4fba8c21..52abb536c4 100644
--- a/tasks/robopaint-acrylic-suddenshower-i00-privileged/task.toml
+++ b/tasks/robopaint-acrylic-suddenshower-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "maniskill" # ManiSkill 3.0.1 / SAPIEN 3.0.3 (PhysX on the
robot = "franka-panda" # PandaStick with a paint brush: rigid handle + 15 mm collision-free tuft, TCP = the tuft's tip
family = "robopaint_acrylic_suddenshower"
mode = "privileged" # privileged (exact sheets, score, brush state) | standard (RGB only: the -standard twin)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "partial" # solution/solve.sh replays the frozen reference plan; its recorded scores: oracle_evidence
oracle_evidence = "Reference solution in solution/, unchanged from the earlier version of this task (the picture score only; its Harbor oracle and nop runs, and what changed since: docs/benchmarks/robopaint.md, Oracle evidence). Under the process rules the verifier scores it 1 (tests/grade.py with rcb_brush_strict.py, two fresh replays, ws3, 2026-10-05): path_mm 13495 of the 40500 mm budget; its strokes are sweeps. Harbor oracle and nop runs on one pair per family: docs/benchmarks/robopaint.md (Validation)."
instance_seed = 0 # scene = environment/instance.json (sheets, brush preset, paint wells, start joints, instance_sha; the exemplar is target.png)
diff --git a/tasks/robopaint-acrylic-suddenshower-i00-standard/task.toml b/tasks/robopaint-acrylic-suddenshower-i00-standard/task.toml
index 039374cacb..58478818aa 100644
--- a/tasks/robopaint-acrylic-suddenshower-i00-standard/task.toml
+++ b/tasks/robopaint-acrylic-suddenshower-i00-standard/task.toml
@@ -31,6 +31,7 @@ simulator = "maniskill"
robot = "franka-panda"
family = "robopaint_acrylic_suddenshower"
mode = "standard" # camera RGB + proprioception only; no sheet readout, no score, no reset, no brush model
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "partial" # solution/: the privileged oracle's joint targets, streamed through the client
oracle_evidence = "Reference solution in solution/ (streams the privileged reference's joint targets), unchanged from the earlier version of this task (the picture score only; its Harbor oracle and nop runs, and what changed since: docs/benchmarks/robopaint.md, Oracle evidence). The same trajectory scores 1 under the process rules in the verifier, whose replay CLI the collect hook runs (ws3, 2026-10-05): path_mm 13495 of the 40500 mm budget; its strokes are sweeps. Harbor oracle and nop runs on one pair per family: docs/benchmarks/robopaint.md (Validation)."
diff --git a/tasks/robopaint-acrylic-sunflowers-i00-privileged/task.toml b/tasks/robopaint-acrylic-sunflowers-i00-privileged/task.toml
index 4af815e21b..7e799f0129 100644
--- a/tasks/robopaint-acrylic-sunflowers-i00-privileged/task.toml
+++ b/tasks/robopaint-acrylic-sunflowers-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "maniskill" # ManiSkill 3.0.1 / SAPIEN 3.0.3 (PhysX on the
robot = "franka-panda" # PandaStick with a paint brush: rigid handle + 15 mm collision-free tuft, TCP = the tuft's tip
family = "robopaint_acrylic_sunflowers"
mode = "privileged" # privileged (exact sheets, score, brush state) | standard (RGB only: the -standard twin)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "partial" # solution/solve.sh replays the frozen reference plan; its recorded scores: oracle_evidence
oracle_evidence = "Reference solution in solution/, unchanged from the earlier version of this task (the picture score only; its Harbor oracle and nop runs, and what changed since: docs/benchmarks/robopaint.md, Oracle evidence). Under the process rules the verifier scores it 1 (tests/grade.py with rcb_brush_strict.py, two fresh replays, ws3, 2026-10-05): path_mm 13867 of the 41700 mm budget; its strokes are sweeps. Harbor oracle and nop runs on one pair per family: docs/benchmarks/robopaint.md (Validation)."
instance_seed = 0 # scene = environment/instance.json (sheets, brush preset, paint wells, start joints, instance_sha; the exemplar is target.png)
diff --git a/tasks/robopaint-acrylic-sunflowers-i00-standard/task.toml b/tasks/robopaint-acrylic-sunflowers-i00-standard/task.toml
index 9eb79e622f..e93af7379e 100644
--- a/tasks/robopaint-acrylic-sunflowers-i00-standard/task.toml
+++ b/tasks/robopaint-acrylic-sunflowers-i00-standard/task.toml
@@ -31,6 +31,7 @@ simulator = "maniskill"
robot = "franka-panda"
family = "robopaint_acrylic_sunflowers"
mode = "standard" # camera RGB + proprioception only; no sheet readout, no score, no reset, no brush model
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "partial" # solution/: the privileged oracle's joint targets, streamed through the client
oracle_evidence = "Reference solution in solution/ (streams the privileged reference's joint targets), unchanged from the earlier version of this task (the picture score only; its Harbor oracle and nop runs, and what changed since: docs/benchmarks/robopaint.md, Oracle evidence). The same trajectory scores 1 under the process rules in the verifier, whose replay CLI the collect hook runs (ws3, 2026-10-05): path_mm 13867 of the 41700 mm budget; its strokes are sweeps. Harbor oracle and nop runs on one pair per family: docs/benchmarks/robopaint.md (Validation)."
diff --git a/tasks/robopaint-acrylic-vision-i00-privileged/task.toml b/tasks/robopaint-acrylic-vision-i00-privileged/task.toml
index 33ecba2410..7695cc4ca6 100644
--- a/tasks/robopaint-acrylic-vision-i00-privileged/task.toml
+++ b/tasks/robopaint-acrylic-vision-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "maniskill" # ManiSkill 3.0.1 / SAPIEN 3.0.3 (PhysX on the
robot = "franka-panda" # PandaStick with a paint brush: rigid handle + 15 mm collision-free tuft, TCP = the tuft's tip
family = "robopaint_acrylic_vision"
mode = "privileged" # privileged (exact sheets, score, brush state) | standard (RGB only: the -standard twin)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "partial" # solution/solve.sh replays the frozen reference plan; its recorded scores: oracle_evidence
oracle_evidence = "Reference solution in solution/, unchanged from the earlier version of this task (the picture score only; its Harbor oracle and nop runs, and what changed since: docs/benchmarks/robopaint.md, Oracle evidence). Under the process rules the verifier scores it 1 (tests/grade.py with rcb_brush_strict.py, two fresh replays, ws3, 2026-10-05): path_mm 10746 of the 32300 mm budget; its strokes are sweeps. Harbor oracle and nop runs on one pair per family: docs/benchmarks/robopaint.md (Validation)."
instance_seed = 0 # scene = environment/instance.json (sheets, brush preset, paint wells, start joints, instance_sha; the exemplar is target.png)
diff --git a/tasks/robopaint-acrylic-vision-i00-standard/task.toml b/tasks/robopaint-acrylic-vision-i00-standard/task.toml
index 3de0a2b527..e324bba89a 100644
--- a/tasks/robopaint-acrylic-vision-i00-standard/task.toml
+++ b/tasks/robopaint-acrylic-vision-i00-standard/task.toml
@@ -31,6 +31,7 @@ simulator = "maniskill"
robot = "franka-panda"
family = "robopaint_acrylic_vision"
mode = "standard" # camera RGB + proprioception only; no sheet readout, no score, no reset, no brush model
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "partial" # solution/: the privileged oracle's joint targets, streamed through the client
oracle_evidence = "Reference solution in solution/ (streams the privileged reference's joint targets), unchanged from the earlier version of this task (the picture score only; its Harbor oracle and nop runs, and what changed since: docs/benchmarks/robopaint.md, Oracle evidence). The same trajectory scores 1 under the process rules in the verifier, whose replay CLI the collect hook runs (ws3, 2026-10-05): path_mm 10746 of the 32300 mm budget; its strokes are sweeps. Harbor oracle and nop runs on one pair per family: docs/benchmarks/robopaint.md (Validation)."
diff --git a/tasks/robopaint-caoshu-beiquelindanshui-i00-privileged/task.toml b/tasks/robopaint-caoshu-beiquelindanshui-i00-privileged/task.toml
index 40f9f0c297..c7c5b4b213 100644
--- a/tasks/robopaint-caoshu-beiquelindanshui-i00-privileged/task.toml
+++ b/tasks/robopaint-caoshu-beiquelindanshui-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "maniskill" # ManiSkill 3.0.1 / SAPIEN 3.0.3 (PhysX on the
robot = "franka-panda" # PandaStick with a Chinese brush: rigid handle + 30 mm collision-free tuft, TCP = the tuft's tip
family = "robopaint_caoshu_beiquelindanshui"
mode = "privileged" # privileged (exact sheets, brush state, reset, scoring service) | standard (RGB only: the -standard twin)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "partial" # solution/solve.sh replays the frozen reference plan; its recorded scores: oracle_evidence
oracle_evidence = "Reference solution in solution/ (a new family, generated on 2026-10-09 with the current tooling): the generator's self-test replayed it on the graders' copy and through the scoring service (equal metrics, success); under the process rules the verifier scores it 1 (tests/grade.py with rcb_brush_strict.py, two fresh replays): path_mm 891 and mark_width_mm 1.98, against the colouring-in limits 2680 mm and 0.59 mm. Harbor oracle and nop runs: docs/benchmarks/robopaint.md (Caoshu)."
instance_seed = 0 # scene = environment/instance.json (agent's copy: sheets, brush, ink stone, start joints, the rule, instance_sha)
diff --git a/tasks/robopaint-caoshu-beiquelindanshui-i00-standard/task.toml b/tasks/robopaint-caoshu-beiquelindanshui-i00-standard/task.toml
index a44bf40f0f..c9b3f4bd02 100644
--- a/tasks/robopaint-caoshu-beiquelindanshui-i00-standard/task.toml
+++ b/tasks/robopaint-caoshu-beiquelindanshui-i00-standard/task.toml
@@ -31,6 +31,7 @@ simulator = "maniskill"
robot = "franka-panda"
family = "robopaint_caoshu_beiquelindanshui"
mode = "standard" # camera RGB + proprioception only; no sheet readout, no score, no reset, no brush model
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "partial" # solution/: the privileged oracle's joint targets, streamed through the client
oracle_evidence = "Reference solution in solution/ (streams the privileged reference's joint targets; a new family, generated on 2026-10-09 with the current tooling): the same trajectory scores 1 under the process rules in the verifier, whose replay CLI the collect hook runs: path_mm 891 and mark_width_mm 1.98, against the colouring-in limits 2680 mm and 0.59 mm. Harbor oracle and nop runs: docs/benchmarks/robopaint.md (Caoshu)."
diff --git a/tasks/robopaint-caoshu-nangong-i00-privileged/task.toml b/tasks/robopaint-caoshu-nangong-i00-privileged/task.toml
index 5f55b6167f..69c85635d4 100644
--- a/tasks/robopaint-caoshu-nangong-i00-privileged/task.toml
+++ b/tasks/robopaint-caoshu-nangong-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "maniskill" # ManiSkill 3.0.1 / SAPIEN 3.0.3 (PhysX on the
robot = "franka-panda" # PandaStick with a Chinese brush: rigid handle + 30 mm collision-free tuft, TCP = the tuft's tip
family = "robopaint_caoshu_nangong"
mode = "privileged" # privileged (exact sheets, brush state, reset, scoring service) | standard (RGB only: the -standard twin)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "partial" # solution/solve.sh replays the frozen reference plan; its recorded scores: oracle_evidence
oracle_evidence = "Reference solution in solution/ (a new family, generated on 2026-10-09 with the current tooling): the generator's self-test replayed it on the graders' copy and through the scoring service (equal metrics, success); under the process rules the verifier scores it 1 (tests/grade.py with rcb_brush_strict.py, two fresh replays): path_mm 890 and mark_width_mm 3.11, against the colouring-in limits 2680 mm and 0.93 mm. Harbor oracle and nop runs: docs/benchmarks/robopaint.md (Caoshu)."
instance_seed = 0 # scene = environment/instance.json (agent's copy: sheets, brush, ink stone, start joints, the rule, instance_sha)
diff --git a/tasks/robopaint-caoshu-nangong-i00-standard/task.toml b/tasks/robopaint-caoshu-nangong-i00-standard/task.toml
index ed3ef1c9fd..f6bfe1b322 100644
--- a/tasks/robopaint-caoshu-nangong-i00-standard/task.toml
+++ b/tasks/robopaint-caoshu-nangong-i00-standard/task.toml
@@ -31,6 +31,7 @@ simulator = "maniskill"
robot = "franka-panda"
family = "robopaint_caoshu_nangong"
mode = "standard" # camera RGB + proprioception only; no sheet readout, no score, no reset, no brush model
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "partial" # solution/: the privileged oracle's joint targets, streamed through the client
oracle_evidence = "Reference solution in solution/ (streams the privileged reference's joint targets; a new family, generated on 2026-10-09 with the current tooling): the same trajectory scores 1 under the process rules in the verifier, whose replay CLI the collect hook runs: path_mm 890 and mark_width_mm 3.11, against the colouring-in limits 2680 mm and 0.93 mm. Harbor oracle and nop runs: docs/benchmarks/robopaint.md (Caoshu)."
diff --git a/tasks/robopaint-caoshu-siwushaonian-i00-privileged/task.toml b/tasks/robopaint-caoshu-siwushaonian-i00-privileged/task.toml
index 85298a1700..a972da1cb2 100644
--- a/tasks/robopaint-caoshu-siwushaonian-i00-privileged/task.toml
+++ b/tasks/robopaint-caoshu-siwushaonian-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "maniskill" # ManiSkill 3.0.1 / SAPIEN 3.0.3 (PhysX on the
robot = "franka-panda" # PandaStick with a Chinese brush: rigid handle + 30 mm collision-free tuft, TCP = the tuft's tip
family = "robopaint_caoshu_siwushaonian"
mode = "privileged" # privileged (exact sheets, brush state, reset, scoring service) | standard (RGB only: the -standard twin)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "partial" # solution/solve.sh replays the frozen reference plan; its recorded scores: oracle_evidence
oracle_evidence = "Reference solution in solution/ (a new family, generated on 2026-10-09 with the current tooling): the generator's self-test replayed it on the graders' copy and through the scoring service (equal metrics, success); under the process rules the verifier scores it 1 (tests/grade.py with rcb_brush_strict.py, two fresh replays): path_mm 641 and mark_width_mm 2.66, against the colouring-in limits 1930 mm and 0.79 mm. Harbor oracle and nop runs: docs/benchmarks/robopaint.md (Caoshu)."
instance_seed = 0 # scene = environment/instance.json (agent's copy: sheets, brush, ink stone, start joints, the rule, instance_sha)
diff --git a/tasks/robopaint-caoshu-siwushaonian-i00-standard/task.toml b/tasks/robopaint-caoshu-siwushaonian-i00-standard/task.toml
index 9a17e32eee..3cc238c8b8 100644
--- a/tasks/robopaint-caoshu-siwushaonian-i00-standard/task.toml
+++ b/tasks/robopaint-caoshu-siwushaonian-i00-standard/task.toml
@@ -31,6 +31,7 @@ simulator = "maniskill"
robot = "franka-panda"
family = "robopaint_caoshu_siwushaonian"
mode = "standard" # camera RGB + proprioception only; no sheet readout, no score, no reset, no brush model
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "partial" # solution/: the privileged oracle's joint targets, streamed through the client
oracle_evidence = "Reference solution in solution/ (streams the privileged reference's joint targets; a new family, generated on 2026-10-09 with the current tooling): the same trajectory scores 1 under the process rules in the verifier, whose replay CLI the collect hook runs: path_mm 641 and mark_width_mm 2.66, against the colouring-in limits 1930 mm and 0.79 mm. Harbor oracle and nop runs: docs/benchmarks/robopaint.md (Caoshu)."
diff --git a/tasks/robopaint-caoshu-wen-i00-privileged/task.toml b/tasks/robopaint-caoshu-wen-i00-privileged/task.toml
index 914102e5ae..6f10e6f302 100644
--- a/tasks/robopaint-caoshu-wen-i00-privileged/task.toml
+++ b/tasks/robopaint-caoshu-wen-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "maniskill" # ManiSkill 3.0.1 / SAPIEN 3.0.3 (PhysX on the
robot = "franka-panda" # PandaStick with a Chinese brush: rigid handle + 30 mm collision-free tuft, TCP = the tuft's tip
family = "robopaint_caoshu_wen"
mode = "privileged" # privileged (exact sheets, brush state, reset, scoring service) | standard (RGB only: the -standard twin)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "partial" # solution/solve.sh replays the frozen reference plan; its recorded scores: oracle_evidence
oracle_evidence = "Reference solution in solution/ (a new family, generated on 2026-10-09 with the current tooling): the generator's self-test replayed it on the graders' copy and through the scoring service (equal metrics, success); under the process rules the verifier scores it 1 (tests/grade.py with rcb_brush_strict.py, two fresh replays): path_mm 269 and mark_width_mm 4.57, against the colouring-in limits 810 mm and 1.37 mm. Harbor oracle and nop runs: docs/benchmarks/robopaint.md (Caoshu)."
instance_seed = 0 # scene = environment/instance.json (agent's copy: sheets, brush, ink stone, start joints, the rule, instance_sha)
diff --git a/tasks/robopaint-caoshu-wen-i00-standard/task.toml b/tasks/robopaint-caoshu-wen-i00-standard/task.toml
index ba8f0174c9..3fb9cb76f3 100644
--- a/tasks/robopaint-caoshu-wen-i00-standard/task.toml
+++ b/tasks/robopaint-caoshu-wen-i00-standard/task.toml
@@ -31,6 +31,7 @@ simulator = "maniskill"
robot = "franka-panda"
family = "robopaint_caoshu_wen"
mode = "standard" # camera RGB + proprioception only; no sheet readout, no score, no reset, no brush model
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "partial" # solution/: the privileged oracle's joint targets, streamed through the client
oracle_evidence = "Reference solution in solution/ (streams the privileged reference's joint targets; a new family, generated on 2026-10-09 with the current tooling): the same trajectory scores 1 under the process rules in the verifier, whose replay CLI the collect hook runs: path_mm 269 and mark_width_mm 4.57, against the colouring-in limits 810 mm and 1.37 mm. Harbor oracle and nop runs: docs/benchmarks/robopaint.md (Caoshu)."
diff --git a/tasks/robopaint-caoshu-zhen-i00-privileged/task.toml b/tasks/robopaint-caoshu-zhen-i00-privileged/task.toml
index 046cd9a3b7..c7cfb4f58c 100644
--- a/tasks/robopaint-caoshu-zhen-i00-privileged/task.toml
+++ b/tasks/robopaint-caoshu-zhen-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "maniskill" # ManiSkill 3.0.1 / SAPIEN 3.0.3 (PhysX on the
robot = "franka-panda" # PandaStick with a Chinese brush: rigid handle + 30 mm collision-free tuft, TCP = the tuft's tip
family = "robopaint_caoshu_zhen"
mode = "privileged" # privileged (exact sheets, brush state, reset, scoring service) | standard (RGB only: the -standard twin)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "partial" # solution/solve.sh replays the frozen reference plan; its recorded scores: oracle_evidence
oracle_evidence = "Reference solution in solution/ (a new family, generated on 2026-10-09 with the current tooling): the generator's self-test replayed it on the graders' copy and through the scoring service (equal metrics, success); under the process rules the verifier scores it 1 (tests/grade.py with rcb_brush_strict.py, two fresh replays): path_mm 534 and mark_width_mm 5.86, against the colouring-in limits 1610 mm and 1.75 mm. Harbor oracle and nop runs: docs/benchmarks/robopaint.md (Caoshu)."
instance_seed = 0 # scene = environment/instance.json (agent's copy: sheets, brush, ink stone, start joints, the rule, instance_sha)
diff --git a/tasks/robopaint-caoshu-zhen-i00-standard/task.toml b/tasks/robopaint-caoshu-zhen-i00-standard/task.toml
index 594209b9a3..8e3370823e 100644
--- a/tasks/robopaint-caoshu-zhen-i00-standard/task.toml
+++ b/tasks/robopaint-caoshu-zhen-i00-standard/task.toml
@@ -31,6 +31,7 @@ simulator = "maniskill"
robot = "franka-panda"
family = "robopaint_caoshu_zhen"
mode = "standard" # camera RGB + proprioception only; no sheet readout, no score, no reset, no brush model
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "partial" # solution/: the privileged oracle's joint targets, streamed through the client
oracle_evidence = "Reference solution in solution/ (streams the privileged reference's joint targets; a new family, generated on 2026-10-09 with the current tooling): the same trajectory scores 1 under the process rules in the verifier, whose replay CLI the collect hook runs: path_mm 534 and mark_width_mm 5.86, against the colouring-in limits 1610 mm and 1.75 mm. Harbor oracle and nop runs: docs/benchmarks/robopaint.md (Caoshu)."
diff --git a/tasks/robopaint-caoshu-zhongtiangechuifen-i00-privileged/task.toml b/tasks/robopaint-caoshu-zhongtiangechuifen-i00-privileged/task.toml
index ad6c54fc7b..8dcf2ac9ac 100644
--- a/tasks/robopaint-caoshu-zhongtiangechuifen-i00-privileged/task.toml
+++ b/tasks/robopaint-caoshu-zhongtiangechuifen-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "maniskill" # ManiSkill 3.0.1 / SAPIEN 3.0.3 (PhysX on the
robot = "franka-panda" # PandaStick with a Chinese brush: rigid handle + 30 mm collision-free tuft, TCP = the tuft's tip
family = "robopaint_caoshu_zhongtiangechuifen"
mode = "privileged" # privileged (exact sheets, brush state, reset, scoring service) | standard (RGB only: the -standard twin)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "partial" # solution/solve.sh replays the frozen reference plan; its recorded scores: oracle_evidence
oracle_evidence = "Reference solution in solution/ (a new family, generated on 2026-10-09 with the current tooling): the generator's self-test replayed it on the graders' copy and through the scoring service (equal metrics, success); under the process rules the verifier scores it 1 (tests/grade.py with rcb_brush_strict.py, two fresh replays): path_mm 525 and mark_width_mm 1.91, against the colouring-in limits 1580 mm and 0.57 mm. Harbor oracle and nop runs: docs/benchmarks/robopaint.md (Caoshu)."
instance_seed = 0 # scene = environment/instance.json (agent's copy: sheets, brush, ink stone, start joints, the rule, instance_sha)
diff --git a/tasks/robopaint-caoshu-zhongtiangechuifen-i00-standard/task.toml b/tasks/robopaint-caoshu-zhongtiangechuifen-i00-standard/task.toml
index c225e08e2a..ef1f1dedc3 100644
--- a/tasks/robopaint-caoshu-zhongtiangechuifen-i00-standard/task.toml
+++ b/tasks/robopaint-caoshu-zhongtiangechuifen-i00-standard/task.toml
@@ -31,6 +31,7 @@ simulator = "maniskill"
robot = "franka-panda"
family = "robopaint_caoshu_zhongtiangechuifen"
mode = "standard" # camera RGB + proprioception only; no sheet readout, no score, no reset, no brush model
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "partial" # solution/: the privileged oracle's joint targets, streamed through the client
oracle_evidence = "Reference solution in solution/ (streams the privileged reference's joint targets; a new family, generated on 2026-10-09 with the current tooling): the same trajectory scores 1 under the process rules in the verifier, whose replay CLI the collect hook runs: path_mm 525 and mark_width_mm 1.91, against the colouring-in limits 1580 mm and 0.57 mm. Harbor oracle and nop runs: docs/benchmarks/robopaint.md (Caoshu)."
diff --git a/tasks/robopaint-kaishu-linfengguilong-i00-privileged/task.toml b/tasks/robopaint-kaishu-linfengguilong-i00-privileged/task.toml
index b4f5653135..4a84f02fc3 100644
--- a/tasks/robopaint-kaishu-linfengguilong-i00-privileged/task.toml
+++ b/tasks/robopaint-kaishu-linfengguilong-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "maniskill" # ManiSkill 3.0.1 / SAPIEN 3.0.3 (PhysX on the
robot = "franka-panda" # PandaStick with a Chinese brush: rigid handle + 30 mm collision-free tuft, TCP = the tuft's tip
family = "robopaint_kaishu_linfengguilong"
mode = "privileged" # privileged (exact sheets, brush state, reset, scoring service) | standard (RGB only: the -standard twin)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "partial" # solution/solve.sh replays the frozen reference plan; its recorded scores: oracle_evidence
oracle_evidence = "Reference solution in solution/, made for this task on 2026-10-08 (no earlier version) and re-made on 2026-10-09 by the calligraphy planner of solution/README.md (docs/benchmarks/robopaint.md, Targets). The verifier scores it 1 under the process rules (tests/grade.py with rcb_brush_strict.py, two fresh replays): path_mm 2828 and mark_width_mm 3.07, against the colouring-in limits 7880 mm and 0.99 mm. Harbor oracle 1 and nop 0 on both modes of this pair on its current files (ws3, 2026-10-09): docs/benchmarks/robopaint.md (Validation)."
instance_seed = 0 # scene = environment/instance.json (agent's copy: sheets, brush, ink stone, start joints, the rule, instance_sha)
diff --git a/tasks/robopaint-kaishu-linfengguilong-i00-standard/task.toml b/tasks/robopaint-kaishu-linfengguilong-i00-standard/task.toml
index 0061140170..90f39a92e4 100644
--- a/tasks/robopaint-kaishu-linfengguilong-i00-standard/task.toml
+++ b/tasks/robopaint-kaishu-linfengguilong-i00-standard/task.toml
@@ -31,6 +31,7 @@ simulator = "maniskill"
robot = "franka-panda"
family = "robopaint_kaishu_linfengguilong"
mode = "standard" # camera RGB + proprioception only; no sheet readout, no score, no reset, no brush model
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "partial" # solution/: the privileged oracle's joint targets, streamed through the client
oracle_evidence = "Reference solution in solution/ (streams the privileged reference's joint targets), made for this task on 2026-10-08 (no earlier version) and re-made on 2026-10-09 (docs/benchmarks/robopaint.md, Targets). The same trajectory scores 1 under the process rules in the verifier, whose replay CLI the collect hook runs: path_mm 2828 and mark_width_mm 3.07, against the colouring-in limits 7880 mm and 0.99 mm. Harbor oracle 1 and nop 0 on both modes of this pair on its current files (ws3, 2026-10-09): docs/benchmarks/robopaint.md (Validation)."
diff --git a/tasks/robopaint-kaishu-yongzibafa-i00-privileged/task.toml b/tasks/robopaint-kaishu-yongzibafa-i00-privileged/task.toml
index 9b190ada0a..8b316b5ee9 100644
--- a/tasks/robopaint-kaishu-yongzibafa-i00-privileged/task.toml
+++ b/tasks/robopaint-kaishu-yongzibafa-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "maniskill" # ManiSkill 3.0.1 / SAPIEN 3.0.3 (PhysX on the
robot = "franka-panda" # PandaStick with a Chinese brush: rigid handle + 30 mm collision-free tuft, TCP = the tuft's tip
family = "robopaint_kaishu_yongzibafa"
mode = "privileged" # privileged (exact sheets, brush state, reset, scoring service) | standard (RGB only: the -standard twin)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "partial" # solution/solve.sh replays the frozen reference plan; its recorded scores: oracle_evidence
oracle_evidence = "Reference solution in solution/, re-made on 2026-10-09 by the calligraphy planner of solution/README.md (README, Change Log; docs/benchmarks/robopaint.md, Targets). The verifier scores it 1 under the process rules (tests/grade.py with rcb_brush_strict.py, two fresh replays): path_mm 1270 and mark_width_mm 3.71, against the colouring-in limits 3230 mm and 1.34 mm. Harbor oracle 1 and nop 0 on both modes of this pair on its current files (ws3, 2026-10-09): docs/benchmarks/robopaint.md (Validation)."
instance_seed = 0 # scene = environment/instance.json (agent's copy: sheets, brush, ink stone, start joints, the rule, instance_sha)
diff --git a/tasks/robopaint-kaishu-yongzibafa-i00-standard/task.toml b/tasks/robopaint-kaishu-yongzibafa-i00-standard/task.toml
index aaec561935..308fbb9ca6 100644
--- a/tasks/robopaint-kaishu-yongzibafa-i00-standard/task.toml
+++ b/tasks/robopaint-kaishu-yongzibafa-i00-standard/task.toml
@@ -31,6 +31,7 @@ simulator = "maniskill"
robot = "franka-panda"
family = "robopaint_kaishu_yongzibafa"
mode = "standard" # camera RGB + proprioception only; no sheet readout, no score, no reset, no brush model
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "partial" # solution/: the privileged oracle's joint targets, streamed through the client
oracle_evidence = "Reference solution in solution/ (streams the privileged reference's joint targets), re-made on 2026-10-09 (README, Change Log; docs/benchmarks/robopaint.md, Targets). The same trajectory scores 1 under the process rules in the verifier, whose replay CLI the collect hook runs: path_mm 1270 and mark_width_mm 3.71, against the colouring-in limits 3230 mm and 1.34 mm. Harbor oracle 1 and nop 0 on both modes of this pair on its current files (ws3, 2026-10-09): docs/benchmarks/robopaint.md (Validation)."
diff --git a/tasks/robopaint-lettering-aim-i00-privileged/task.toml b/tasks/robopaint-lettering-aim-i00-privileged/task.toml
index b9250cca98..0528e16ae7 100644
--- a/tasks/robopaint-lettering-aim-i00-privileged/task.toml
+++ b/tasks/robopaint-lettering-aim-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "maniskill" # ManiSkill 3.0.1 / SAPIEN 3.0.3 (PhysX on the
robot = "franka-panda" # PandaStick with a dip pen: rigid holder + a collision-free 6 mm flexible point; TCP = the point
family = "robopaint_lettering_aim"
mode = "privileged" # privileged (exact sheets, pen state, reset, scoring service) | standard (RGB only: the -standard twin)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "partial" # solution/solve.sh replays the frozen reference plan; its recorded scores: oracle_evidence
oracle_evidence = "Reference solution in solution/ (the lettering family redesigned on 2026-10-09: 臨摹 of engraved round hand; generated with the current tooling on the native eai-robopaint:0.2.0 runtime): the generator's self-test replayed it on the graders' copy and through the scoring service (equal metrics, success); under the process rules the verifier scores it 1 (tests/grade.py with rcb_brush_strict.py, two fresh replays): path_mm 998 and mark_width_mm 0.43, against the colouring-in limits 3000 mm and 0.12 mm. Harbor oracle and nop runs: docs/benchmarks/robopaint.md (Lettering)."
instance_seed = 0 # scene = environment/instance.json (agent's copy: sheets, pen, ink well, start joints, the rule, instance_sha)
diff --git a/tasks/robopaint-lettering-aim-i00-standard/task.toml b/tasks/robopaint-lettering-aim-i00-standard/task.toml
index cc30b03fe0..4821373c73 100644
--- a/tasks/robopaint-lettering-aim-i00-standard/task.toml
+++ b/tasks/robopaint-lettering-aim-i00-standard/task.toml
@@ -31,6 +31,7 @@ simulator = "maniskill"
robot = "franka-panda"
family = "robopaint_lettering_aim"
mode = "standard" # camera RGB + proprioception only; no sheet readout, no score, no reset, no pen model
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "partial" # solution/: the privileged oracle's joint targets, streamed through the client
oracle_evidence = "Reference solution in solution/ (streams the privileged reference's joint targets; the lettering family redesigned on 2026-10-09, generated with the current tooling on the native eai-robopaint:0.2.0 runtime): the same trajectory scores 1 under the process rules in the verifier, whose replay CLI the collect hook runs: path_mm 998 and mark_width_mm 0.43, against the colouring-in limits 3000 mm and 0.12 mm. Harbor oracle and nop runs: docs/benchmarks/robopaint.md (Lettering)."
diff --git a/tasks/robopaint-lettering-business-i00-privileged/task.toml b/tasks/robopaint-lettering-business-i00-privileged/task.toml
index 9ea798996a..5039b77e94 100644
--- a/tasks/robopaint-lettering-business-i00-privileged/task.toml
+++ b/tasks/robopaint-lettering-business-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "maniskill" # ManiSkill 3.0.1 / SAPIEN 3.0.3 (PhysX on the
robot = "franka-panda" # PandaStick with a dip pen: rigid holder + a collision-free 6 mm flexible point; TCP = the point
family = "robopaint_lettering_business"
mode = "privileged" # privileged (exact sheets, pen state, reset, scoring service) | standard (RGB only: the -standard twin)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "partial" # solution/solve.sh replays the frozen reference plan; its recorded scores: oracle_evidence
oracle_evidence = "Reference solution in solution/ (the lettering family redesigned on 2026-10-09: 臨摹 of engraved round hand; generated with the current tooling on the native eai-robopaint:0.2.0 runtime): the generator's self-test replayed it on the graders' copy and through the scoring service (equal metrics, success); under the process rules the verifier scores it 1 (tests/grade.py with rcb_brush_strict.py, two fresh replays): path_mm 1240 and mark_width_mm 0.37, against the colouring-in limits 3730 mm and 0.10 mm. Harbor oracle and nop runs: docs/benchmarks/robopaint.md (Lettering)."
instance_seed = 0 # scene = environment/instance.json (agent's copy: sheets, pen, ink well, start joints, the rule, instance_sha)
diff --git a/tasks/robopaint-lettering-business-i00-standard/task.toml b/tasks/robopaint-lettering-business-i00-standard/task.toml
index 5d9c90dec7..6dbf263d45 100644
--- a/tasks/robopaint-lettering-business-i00-standard/task.toml
+++ b/tasks/robopaint-lettering-business-i00-standard/task.toml
@@ -31,6 +31,7 @@ simulator = "maniskill"
robot = "franka-panda"
family = "robopaint_lettering_business"
mode = "standard" # camera RGB + proprioception only; no sheet readout, no score, no reset, no pen model
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "partial" # solution/: the privileged oracle's joint targets, streamed through the client
oracle_evidence = "Reference solution in solution/ (streams the privileged reference's joint targets; the lettering family redesigned on 2026-10-09, generated with the current tooling on the native eai-robopaint:0.2.0 runtime): the same trajectory scores 1 under the process rules in the verifier, whose replay CLI the collect hook runs: path_mm 1240 and mark_width_mm 0.37, against the colouring-in limits 3730 mm and 0.10 mm. Harbor oracle and nop runs: docs/benchmarks/robopaint.md (Lettering)."
diff --git a/tasks/robopaint-lettering-commendation-i00-privileged/task.toml b/tasks/robopaint-lettering-commendation-i00-privileged/task.toml
index 0c3b4496eb..02f32e63fd 100644
--- a/tasks/robopaint-lettering-commendation-i00-privileged/task.toml
+++ b/tasks/robopaint-lettering-commendation-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "maniskill" # ManiSkill 3.0.1 / SAPIEN 3.0.3 (PhysX on the
robot = "franka-panda" # PandaStick with a dip pen: rigid holder + a collision-free 6 mm flexible point; TCP = the point
family = "robopaint_lettering_commendation"
mode = "privileged" # privileged (exact sheets, pen state, reset, scoring service) | standard (RGB only: the -standard twin)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "partial" # solution/solve.sh replays the frozen reference plan; its recorded scores: oracle_evidence
oracle_evidence = "Reference solution in solution/ (the lettering family redesigned on 2026-10-09: 臨摹 of engraved round hand; generated with the current tooling on the native eai-robopaint:0.2.0 runtime): the generator's self-test replayed it on the graders' copy and through the scoring service (equal metrics, success); under the process rules the verifier scores it 1 (tests/grade.py with rcb_brush_strict.py, two fresh replays): path_mm 1079 and mark_width_mm 0.41, against the colouring-in limits 3240 mm and 0.12 mm. Harbor oracle and nop runs: docs/benchmarks/robopaint.md (Lettering)."
instance_seed = 0 # scene = environment/instance.json (agent's copy: sheets, pen, ink well, start joints, the rule, instance_sha)
diff --git a/tasks/robopaint-lettering-commendation-i00-standard/task.toml b/tasks/robopaint-lettering-commendation-i00-standard/task.toml
index d907949129..9db06b79e1 100644
--- a/tasks/robopaint-lettering-commendation-i00-standard/task.toml
+++ b/tasks/robopaint-lettering-commendation-i00-standard/task.toml
@@ -31,6 +31,7 @@ simulator = "maniskill"
robot = "franka-panda"
family = "robopaint_lettering_commendation"
mode = "standard" # camera RGB + proprioception only; no sheet readout, no score, no reset, no pen model
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "partial" # solution/: the privileged oracle's joint targets, streamed through the client
oracle_evidence = "Reference solution in solution/ (streams the privileged reference's joint targets; the lettering family redesigned on 2026-10-09, generated with the current tooling on the native eai-robopaint:0.2.0 runtime): the same trajectory scores 1 under the process rules in the verifier, whose replay CLI the collect hook runs: path_mm 1079 and mark_width_mm 0.41, against the colouring-in limits 3240 mm and 0.12 mm. Harbor oracle and nop runs: docs/benchmarks/robopaint.md (Lettering)."
diff --git a/tasks/robopaint-lettering-diligence-i00-privileged/task.toml b/tasks/robopaint-lettering-diligence-i00-privileged/task.toml
index e84d7fb3e5..bd9e0e4c3e 100644
--- a/tasks/robopaint-lettering-diligence-i00-privileged/task.toml
+++ b/tasks/robopaint-lettering-diligence-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "maniskill" # ManiSkill 3.0.1 / SAPIEN 3.0.3 (PhysX on the
robot = "franka-panda" # PandaStick with a dip pen: rigid holder + a collision-free 6 mm flexible point; TCP = the point
family = "robopaint_lettering_diligence"
mode = "privileged" # privileged (exact sheets, pen state, reset, scoring service) | standard (RGB only: the -standard twin)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "partial" # solution/solve.sh replays the frozen reference plan; its recorded scores: oracle_evidence
oracle_evidence = "Reference solution in solution/ (the lettering family redesigned on 2026-10-09: 臨摹 of engraved round hand; generated with the current tooling on the native eai-robopaint:0.2.0 runtime): the generator's self-test replayed it on the graders' copy and through the scoring service (equal metrics, success); under the process rules the verifier scores it 1 (tests/grade.py with rcb_brush_strict.py, two fresh replays): path_mm 1343 and mark_width_mm 0.36, against the colouring-in limits 4040 mm and 0.10 mm. Harbor oracle and nop runs: docs/benchmarks/robopaint.md (Lettering)."
instance_seed = 0 # scene = environment/instance.json (agent's copy: sheets, pen, ink well, start joints, the rule, instance_sha)
diff --git a/tasks/robopaint-lettering-diligence-i00-standard/task.toml b/tasks/robopaint-lettering-diligence-i00-standard/task.toml
index f55f9d3db6..ac722a4264 100644
--- a/tasks/robopaint-lettering-diligence-i00-standard/task.toml
+++ b/tasks/robopaint-lettering-diligence-i00-standard/task.toml
@@ -31,6 +31,7 @@ simulator = "maniskill"
robot = "franka-panda"
family = "robopaint_lettering_diligence"
mode = "standard" # camera RGB + proprioception only; no sheet readout, no score, no reset, no pen model
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "partial" # solution/: the privileged oracle's joint targets, streamed through the client
oracle_evidence = "Reference solution in solution/ (streams the privileged reference's joint targets; the lettering family redesigned on 2026-10-09, generated with the current tooling on the native eai-robopaint:0.2.0 runtime): the same trajectory scores 1 under the process rules in the verifier, whose replay CLI the collect hook runs: path_mm 1343 and mark_width_mm 0.36, against the colouring-in limits 4040 mm and 0.10 mm. Harbor oracle and nop runs: docs/benchmarks/robopaint.md (Lettering)."
diff --git a/tasks/robopaint-lettering-liberty-i00-privileged/task.toml b/tasks/robopaint-lettering-liberty-i00-privileged/task.toml
index 0312228d19..9b2aff3890 100644
--- a/tasks/robopaint-lettering-liberty-i00-privileged/task.toml
+++ b/tasks/robopaint-lettering-liberty-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "maniskill" # ManiSkill 3.0.1 / SAPIEN 3.0.3 (PhysX on the
robot = "franka-panda" # PandaStick with a dip pen: rigid holder + a collision-free 6 mm flexible point; TCP = the point
family = "robopaint_lettering_liberty"
mode = "privileged" # privileged (exact sheets, pen state, reset, scoring service) | standard (RGB only: the -standard twin)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "partial" # solution/solve.sh replays the frozen reference plan; its recorded scores: oracle_evidence
oracle_evidence = "Reference solution in solution/ (the lettering family redesigned on 2026-10-09: 臨摹 of engraved round hand; generated with the current tooling on the native eai-robopaint:0.2.0 runtime): the generator's self-test replayed it on the graders' copy and through the scoring service (equal metrics, success); under the process rules the verifier scores it 1 (tests/grade.py with rcb_brush_strict.py, two fresh replays): path_mm 2644 and mark_width_mm 0.47, against the colouring-in limits 7940 mm and 0.13 mm. Harbor oracle and nop runs: docs/benchmarks/robopaint.md (Lettering)."
instance_seed = 0 # scene = environment/instance.json (agent's copy: sheets, pen, ink well, start joints, the rule, instance_sha)
diff --git a/tasks/robopaint-lettering-liberty-i00-standard/task.toml b/tasks/robopaint-lettering-liberty-i00-standard/task.toml
index 0001b29b3c..0fa2860b56 100644
--- a/tasks/robopaint-lettering-liberty-i00-standard/task.toml
+++ b/tasks/robopaint-lettering-liberty-i00-standard/task.toml
@@ -31,6 +31,7 @@ simulator = "maniskill"
robot = "franka-panda"
family = "robopaint_lettering_liberty"
mode = "standard" # camera RGB + proprioception only; no sheet readout, no score, no reset, no pen model
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "partial" # solution/: the privileged oracle's joint targets, streamed through the client
oracle_evidence = "Reference solution in solution/ (streams the privileged reference's joint targets; the lettering family redesigned on 2026-10-09, generated with the current tooling on the native eai-robopaint:0.2.0 runtime): the same trajectory scores 1 under the process rules in the verifier, whose replay CLI the collect hook runs: path_mm 2644 and mark_width_mm 0.47, against the colouring-in limits 7940 mm and 0.13 mm. Harbor oracle and nop runs: docs/benchmarks/robopaint.md (Lettering)."
diff --git a/tasks/robopaint-lettering-virtue-i00-privileged/task.toml b/tasks/robopaint-lettering-virtue-i00-privileged/task.toml
index 6b6220481e..7a3ea7757a 100644
--- a/tasks/robopaint-lettering-virtue-i00-privileged/task.toml
+++ b/tasks/robopaint-lettering-virtue-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "maniskill" # ManiSkill 3.0.1 / SAPIEN 3.0.3 (PhysX on the
robot = "franka-panda" # PandaStick with a dip pen: rigid holder + a collision-free 6 mm flexible point; TCP = the point
family = "robopaint_lettering_virtue"
mode = "privileged" # privileged (exact sheets, pen state, reset, scoring service) | standard (RGB only: the -standard twin)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "partial" # solution/solve.sh replays the frozen reference plan; its recorded scores: oracle_evidence
oracle_evidence = "Reference solution in solution/ (the lettering family redesigned on 2026-10-09: 臨摹 of engraved round hand; generated with the current tooling on the native eai-robopaint:0.2.0 runtime): the generator's self-test replayed it on the graders' copy and through the scoring service (equal metrics, success); under the process rules the verifier scores it 1 (tests/grade.py with rcb_brush_strict.py, two fresh replays): path_mm 2288 and mark_width_mm 0.42, against the colouring-in limits 6870 mm and 0.12 mm. Harbor oracle and nop runs: docs/benchmarks/robopaint.md (Lettering)."
instance_seed = 0 # scene = environment/instance.json (agent's copy: sheets, pen, ink well, start joints, the rule, instance_sha)
diff --git a/tasks/robopaint-lettering-virtue-i00-standard/task.toml b/tasks/robopaint-lettering-virtue-i00-standard/task.toml
index 6e379f1b53..5551b5c0aa 100644
--- a/tasks/robopaint-lettering-virtue-i00-standard/task.toml
+++ b/tasks/robopaint-lettering-virtue-i00-standard/task.toml
@@ -31,6 +31,7 @@ simulator = "maniskill"
robot = "franka-panda"
family = "robopaint_lettering_virtue"
mode = "standard" # camera RGB + proprioception only; no sheet readout, no score, no reset, no pen model
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "partial" # solution/: the privileged oracle's joint targets, streamed through the client
oracle_evidence = "Reference solution in solution/ (streams the privileged reference's joint targets; the lettering family redesigned on 2026-10-09, generated with the current tooling on the native eai-robopaint:0.2.0 runtime): the same trajectory scores 1 under the process rules in the verifier, whose replay CLI the collect hook runs: path_mm 2288 and mark_width_mm 0.42, against the colouring-in limits 6870 mm and 0.12 mm. Harbor oracle and nop runs: docs/benchmarks/robopaint.md (Lettering)."
diff --git a/tasks/robopaint-oil-bathers-i01-privileged/task.toml b/tasks/robopaint-oil-bathers-i01-privileged/task.toml
index 3894b40c1a..ea4416bfd6 100644
--- a/tasks/robopaint-oil-bathers-i01-privileged/task.toml
+++ b/tasks/robopaint-oil-bathers-i01-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "maniskill" # ManiSkill 3.0.1 / SAPIEN 3.0.3 (PhysX on the
robot = "franka-panda" # PandaStick with a paint brush: rigid handle + 15 mm collision-free tuft, TCP = the tuft's tip
family = "robopaint_oil_bathers"
mode = "privileged" # privileged (exact sheets, score, brush state) | standard (RGB only: the -standard twin)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "partial" # solution/solve.sh replays the frozen reference plan; its recorded scores: oracle_evidence
oracle_evidence = "Reference solution in solution/: the hi-res version's stroke program for the pointed brush, executed by the robot at generation time (2026-10-09; README Change Log). The verifier scores it 1 (tests/grade.py with rcb_brush_strict.py, two fresh replays, the eai-robopaint:0.2.0 runtime, 2026-10-09): path_mm 21192 of the 63600 mm budget; its strokes are sweeps. Harbor oracle and nop runs: docs/benchmarks/robopaint.md (Oil: the pointed brush)."
instance_seed = 0 # scene = environment/instance.json (sheets, brush preset, paint wells, start joints, instance_sha; the exemplar is target.png)
diff --git a/tasks/robopaint-oil-bathers-i01-standard/task.toml b/tasks/robopaint-oil-bathers-i01-standard/task.toml
index a083675eb0..9d66e4a439 100644
--- a/tasks/robopaint-oil-bathers-i01-standard/task.toml
+++ b/tasks/robopaint-oil-bathers-i01-standard/task.toml
@@ -31,6 +31,7 @@ simulator = "maniskill"
robot = "franka-panda"
family = "robopaint_oil_bathers"
mode = "standard" # camera RGB + proprioception only; no sheet readout, no score, no reset, no brush model
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "partial" # solution/: the privileged oracle's joint targets, streamed through the client
oracle_evidence = "Reference solution in solution/ (streams the privileged reference's joint targets; the hi-res version of 2026-10-09, README Change Log). The same trajectory scores 1 under the process rules in the verifier, whose replay CLI the collect hook runs (2026-10-09): path_mm 21192 of the 63600 mm budget; its strokes are sweeps. Harbor oracle and nop runs: docs/benchmarks/robopaint.md (Oil: the pointed brush)."
diff --git a/tasks/robopaint-oil-cafeterrace-i01-privileged/task.toml b/tasks/robopaint-oil-cafeterrace-i01-privileged/task.toml
index 7e31490006..be80e812ae 100644
--- a/tasks/robopaint-oil-cafeterrace-i01-privileged/task.toml
+++ b/tasks/robopaint-oil-cafeterrace-i01-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "maniskill" # ManiSkill 3.0.1 / SAPIEN 3.0.3 (PhysX on the
robot = "franka-panda" # PandaStick with a paint brush: rigid handle + 15 mm collision-free tuft, TCP = the tuft's tip
family = "robopaint_oil_cafeterrace"
mode = "privileged" # privileged (exact sheets, score, brush state) | standard (RGB only: the -standard twin)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "partial" # solution/solve.sh replays the frozen reference plan; its recorded scores: oracle_evidence
oracle_evidence = "Reference solution in solution/: the hi-res version's stroke program for the pointed brush, executed by the robot at generation time (2026-10-09; README Change Log). The verifier scores it 1 (tests/grade.py with rcb_brush_strict.py, two fresh replays, the eai-robopaint:0.2.0 runtime, 2026-10-09): path_mm 22864 of the 68600 mm budget; its strokes are sweeps. Harbor oracle and nop runs: docs/benchmarks/robopaint.md (Oil: the pointed brush)."
instance_seed = 0 # scene = environment/instance.json (sheets, brush preset, paint wells, start joints, instance_sha; the exemplar is target.png)
diff --git a/tasks/robopaint-oil-cafeterrace-i01-standard/task.toml b/tasks/robopaint-oil-cafeterrace-i01-standard/task.toml
index 746df66673..8623e40f06 100644
--- a/tasks/robopaint-oil-cafeterrace-i01-standard/task.toml
+++ b/tasks/robopaint-oil-cafeterrace-i01-standard/task.toml
@@ -31,6 +31,7 @@ simulator = "maniskill"
robot = "franka-panda"
family = "robopaint_oil_cafeterrace"
mode = "standard" # camera RGB + proprioception only; no sheet readout, no score, no reset, no brush model
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "partial" # solution/: the privileged oracle's joint targets, streamed through the client
oracle_evidence = "Reference solution in solution/ (streams the privileged reference's joint targets; the hi-res version of 2026-10-09, README Change Log). The same trajectory scores 1 under the process rules in the verifier, whose replay CLI the collect hook runs (2026-10-09): path_mm 22864 of the 68600 mm budget; its strokes are sweeps. Harbor oracle and nop runs: docs/benchmarks/robopaint.md (Oil: the pointed brush)."
diff --git a/tasks/robopaint-oil-cypresses-i01-privileged/task.toml b/tasks/robopaint-oil-cypresses-i01-privileged/task.toml
index aac5bc1571..050015773a 100644
--- a/tasks/robopaint-oil-cypresses-i01-privileged/task.toml
+++ b/tasks/robopaint-oil-cypresses-i01-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "maniskill" # ManiSkill 3.0.1 / SAPIEN 3.0.3 (PhysX on the
robot = "franka-panda" # PandaStick with a paint brush: rigid handle + 15 mm collision-free tuft, TCP = the tuft's tip
family = "robopaint_oil_cypresses"
mode = "privileged" # privileged (exact sheets, score, brush state) | standard (RGB only: the -standard twin)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "partial" # solution/solve.sh replays the frozen reference plan; its recorded scores: oracle_evidence
oracle_evidence = "Reference solution in solution/: the hi-res version's stroke program for the pointed brush, executed by the robot at generation time (2026-10-09; README Change Log). The verifier scores it 1 (tests/grade.py with rcb_brush_strict.py, two fresh replays, the eai-robopaint:0.2.0 runtime, 2026-10-09): path_mm 17170 of the 51600 mm budget; its strokes are sweeps. Harbor oracle and nop runs: docs/benchmarks/robopaint.md (Oil: the pointed brush)."
instance_seed = 0 # scene = environment/instance.json (sheets, brush preset, paint wells, start joints, instance_sha; the exemplar is target.png)
diff --git a/tasks/robopaint-oil-cypresses-i01-standard/task.toml b/tasks/robopaint-oil-cypresses-i01-standard/task.toml
index 74970d0258..4830b1f140 100644
--- a/tasks/robopaint-oil-cypresses-i01-standard/task.toml
+++ b/tasks/robopaint-oil-cypresses-i01-standard/task.toml
@@ -31,6 +31,7 @@ simulator = "maniskill"
robot = "franka-panda"
family = "robopaint_oil_cypresses"
mode = "standard" # camera RGB + proprioception only; no sheet readout, no score, no reset, no brush model
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "partial" # solution/: the privileged oracle's joint targets, streamed through the client
oracle_evidence = "Reference solution in solution/ (streams the privileged reference's joint targets; the hi-res version of 2026-10-09, README Change Log). The same trajectory scores 1 under the process rules in the verifier, whose replay CLI the collect hook runs (2026-10-09): path_mm 17170 of the 51600 mm budget; its strokes are sweeps. Harbor oracle and nop runs: docs/benchmarks/robopaint.md (Oil: the pointed brush)."
diff --git a/tasks/robopaint-oil-irises-i01-privileged/task.toml b/tasks/robopaint-oil-irises-i01-privileged/task.toml
index 10ace53537..cae698424e 100644
--- a/tasks/robopaint-oil-irises-i01-privileged/task.toml
+++ b/tasks/robopaint-oil-irises-i01-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "maniskill" # ManiSkill 3.0.1 / SAPIEN 3.0.3 (PhysX on the
robot = "franka-panda" # PandaStick with a paint brush: rigid handle + 15 mm collision-free tuft, TCP = the tuft's tip
family = "robopaint_oil_irises"
mode = "privileged" # privileged (exact sheets, score, brush state) | standard (RGB only: the -standard twin)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "partial" # solution/solve.sh replays the frozen reference plan; its recorded scores: oracle_evidence
oracle_evidence = "Reference solution in solution/: the hi-res version's stroke program for the pointed brush, executed by the robot at generation time (2026-10-09; README Change Log). The verifier scores it 1 (tests/grade.py with rcb_brush_strict.py, two fresh replays, the eai-robopaint:0.2.0 runtime, 2026-10-09): path_mm 18503 of the 55600 mm budget; its strokes are sweeps. Harbor oracle and nop runs: docs/benchmarks/robopaint.md (Oil: the pointed brush)."
instance_seed = 0 # scene = environment/instance.json (sheets, brush preset, paint wells, start joints, instance_sha; the exemplar is target.png)
diff --git a/tasks/robopaint-oil-irises-i01-standard/task.toml b/tasks/robopaint-oil-irises-i01-standard/task.toml
index 8104220dbe..6d678cbd0e 100644
--- a/tasks/robopaint-oil-irises-i01-standard/task.toml
+++ b/tasks/robopaint-oil-irises-i01-standard/task.toml
@@ -31,6 +31,7 @@ simulator = "maniskill"
robot = "franka-panda"
family = "robopaint_oil_irises"
mode = "standard" # camera RGB + proprioception only; no sheet readout, no score, no reset, no brush model
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "partial" # solution/: the privileged oracle's joint targets, streamed through the client
oracle_evidence = "Reference solution in solution/ (streams the privileged reference's joint targets; the hi-res version of 2026-10-09, README Change Log). The same trajectory scores 1 under the process rules in the verifier, whose replay CLI the collect hook runs (2026-10-09): path_mm 18503 of the 55600 mm budget; its strokes are sweeps. Harbor oracle and nop runs: docs/benchmarks/robopaint.md (Oil: the pointed brush)."
diff --git a/tasks/robopaint-oil-milkmaid-i01-privileged/task.toml b/tasks/robopaint-oil-milkmaid-i01-privileged/task.toml
index 3ad1f40976..76cca5fb62 100644
--- a/tasks/robopaint-oil-milkmaid-i01-privileged/task.toml
+++ b/tasks/robopaint-oil-milkmaid-i01-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "maniskill" # ManiSkill 3.0.1 / SAPIEN 3.0.3 (PhysX on the
robot = "franka-panda" # PandaStick with a paint brush: rigid handle + 15 mm collision-free tuft, TCP = the tuft's tip
family = "robopaint_oil_milkmaid"
mode = "privileged" # privileged (exact sheets, score, brush state) | standard (RGB only: the -standard twin)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "partial" # solution/solve.sh replays the frozen reference plan; its recorded scores: oracle_evidence
oracle_evidence = "Reference solution in solution/: the hi-res version's stroke program for the pointed brush, executed by the robot at generation time (2026-10-09; README Change Log). The verifier scores it 1 (tests/grade.py with rcb_brush_strict.py, two fresh replays, the eai-robopaint:0.2.0 runtime, 2026-10-09): path_mm 23964 of the 71900 mm budget; its strokes are sweeps. Harbor oracle and nop runs: docs/benchmarks/robopaint.md (Oil: the pointed brush)."
instance_seed = 0 # scene = environment/instance.json (sheets, brush preset, paint wells, start joints, instance_sha; the exemplar is target.png)
diff --git a/tasks/robopaint-oil-milkmaid-i01-standard/task.toml b/tasks/robopaint-oil-milkmaid-i01-standard/task.toml
index 2e75a52b45..870bc60d7c 100644
--- a/tasks/robopaint-oil-milkmaid-i01-standard/task.toml
+++ b/tasks/robopaint-oil-milkmaid-i01-standard/task.toml
@@ -31,6 +31,7 @@ simulator = "maniskill"
robot = "franka-panda"
family = "robopaint_oil_milkmaid"
mode = "standard" # camera RGB + proprioception only; no sheet readout, no score, no reset, no brush model
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "partial" # solution/: the privileged oracle's joint targets, streamed through the client
oracle_evidence = "Reference solution in solution/ (streams the privileged reference's joint targets; the hi-res version of 2026-10-09, README Change Log). The same trajectory scores 1 under the process rules in the verifier, whose replay CLI the collect hook runs (2026-10-09): path_mm 23964 of the 71900 mm budget; its strokes are sweeps. Harbor oracle and nop runs: docs/benchmarks/robopaint.md (Oil: the pointed brush)."
diff --git a/tasks/robopaint-oil-monalisa-i01-privileged/task.toml b/tasks/robopaint-oil-monalisa-i01-privileged/task.toml
index 9a0d55f71e..7da0405046 100644
--- a/tasks/robopaint-oil-monalisa-i01-privileged/task.toml
+++ b/tasks/robopaint-oil-monalisa-i01-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "maniskill" # ManiSkill 3.0.1 / SAPIEN 3.0.3 (PhysX on the
robot = "franka-panda" # PandaStick with a paint brush: rigid handle + 15 mm collision-free tuft, TCP = the tuft's tip
family = "robopaint_oil_monalisa"
mode = "privileged" # privileged (exact sheets, score, brush state) | standard (RGB only: the -standard twin)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "partial" # solution/solve.sh replays the frozen reference plan; its recorded scores: oracle_evidence
oracle_evidence = "Reference solution in solution/: the hi-res version's stroke program for the pointed brush, executed by the robot at generation time (2026-10-09; README Change Log). The verifier scores it 1 (tests/grade.py with rcb_brush_strict.py, two fresh replays, the eai-robopaint:0.2.0 runtime, 2026-10-09): path_mm 23292 of the 69900 mm budget; its strokes are sweeps. Harbor oracle and nop runs: docs/benchmarks/robopaint.md (Oil: the pointed brush)."
instance_seed = 0 # scene = environment/instance.json (sheets, brush preset, paint wells, start joints, instance_sha; the exemplar is target.png)
diff --git a/tasks/robopaint-oil-monalisa-i01-standard/task.toml b/tasks/robopaint-oil-monalisa-i01-standard/task.toml
index d8e1c092c8..4b2a85eac5 100644
--- a/tasks/robopaint-oil-monalisa-i01-standard/task.toml
+++ b/tasks/robopaint-oil-monalisa-i01-standard/task.toml
@@ -31,6 +31,7 @@ simulator = "maniskill"
robot = "franka-panda"
family = "robopaint_oil_monalisa"
mode = "standard" # camera RGB + proprioception only; no sheet readout, no score, no reset, no brush model
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "partial" # solution/: the privileged oracle's joint targets, streamed through the client
oracle_evidence = "Reference solution in solution/ (streams the privileged reference's joint targets; the hi-res version of 2026-10-09, README Change Log). The same trajectory scores 1 under the process rules in the verifier, whose replay CLI the collect hook runs (2026-10-09): path_mm 23292 of the 69900 mm budget; its strokes are sweeps. Harbor oracle and nop runs: docs/benchmarks/robopaint.md (Oil: the pointed brush)."
diff --git a/tasks/robopaint-oil-pearlearring-i01-privileged/task.toml b/tasks/robopaint-oil-pearlearring-i01-privileged/task.toml
index 53ee62153e..dcf74d7df5 100644
--- a/tasks/robopaint-oil-pearlearring-i01-privileged/task.toml
+++ b/tasks/robopaint-oil-pearlearring-i01-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "maniskill" # ManiSkill 3.0.1 / SAPIEN 3.0.3 (PhysX on the
robot = "franka-panda" # PandaStick with a paint brush: rigid handle + 15 mm collision-free tuft, TCP = the tuft's tip
family = "robopaint_oil_pearlearring"
mode = "privileged" # privileged (exact sheets, score, brush state) | standard (RGB only: the -standard twin)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "partial" # solution/solve.sh replays the frozen reference plan; its recorded scores: oracle_evidence
oracle_evidence = "Reference solution in solution/: the hi-res version's stroke program for the pointed brush, executed by the robot at generation time (2026-10-09; README Change Log). The verifier scores it 1 (tests/grade.py with rcb_brush_strict.py, two fresh replays, the eai-robopaint:0.2.0 runtime, 2026-10-09): path_mm 28809 of the 86500 mm budget; its strokes are sweeps. Harbor oracle and nop runs: docs/benchmarks/robopaint.md (Oil: the pointed brush)."
instance_seed = 0 # scene = environment/instance.json (sheets, brush preset, paint wells, start joints, instance_sha; the exemplar is target.png)
diff --git a/tasks/robopaint-oil-pearlearring-i01-standard/task.toml b/tasks/robopaint-oil-pearlearring-i01-standard/task.toml
index 2711e11ab6..d586cc7925 100644
--- a/tasks/robopaint-oil-pearlearring-i01-standard/task.toml
+++ b/tasks/robopaint-oil-pearlearring-i01-standard/task.toml
@@ -31,6 +31,7 @@ simulator = "maniskill"
robot = "franka-panda"
family = "robopaint_oil_pearlearring"
mode = "standard" # camera RGB + proprioception only; no sheet readout, no score, no reset, no brush model
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "partial" # solution/: the privileged oracle's joint targets, streamed through the client
oracle_evidence = "Reference solution in solution/ (streams the privileged reference's joint targets; the hi-res version of 2026-10-09, README Change Log). The same trajectory scores 1 under the process rules in the verifier, whose replay CLI the collect hook runs (2026-10-09): path_mm 28809 of the 86500 mm budget; its strokes are sweeps. Harbor oracle and nop runs: docs/benchmarks/robopaint.md (Oil: the pointed brush)."
diff --git a/tasks/robopaint-oil-saintevictoire-i01-privileged/task.toml b/tasks/robopaint-oil-saintevictoire-i01-privileged/task.toml
index ed0d41ec7e..98df0a8743 100644
--- a/tasks/robopaint-oil-saintevictoire-i01-privileged/task.toml
+++ b/tasks/robopaint-oil-saintevictoire-i01-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "maniskill" # ManiSkill 3.0.1 / SAPIEN 3.0.3 (PhysX on the
robot = "franka-panda" # PandaStick with a paint brush: rigid handle + 15 mm collision-free tuft, TCP = the tuft's tip
family = "robopaint_oil_saintevictoire"
mode = "privileged" # privileged (exact sheets, score, brush state) | standard (RGB only: the -standard twin)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "partial" # solution/solve.sh replays the frozen reference plan; its recorded scores: oracle_evidence
oracle_evidence = "Reference solution in solution/: the hi-res version's stroke program for the pointed brush, executed by the robot at generation time (2026-10-09; README Change Log). The verifier scores it 1 (tests/grade.py with rcb_brush_strict.py, two fresh replays, the eai-robopaint:0.2.0 runtime, 2026-10-09): path_mm 19229 of the 57700 mm budget; its strokes are sweeps. Harbor oracle and nop runs: docs/benchmarks/robopaint.md (Oil: the pointed brush)."
instance_seed = 0 # scene = environment/instance.json (sheets, brush preset, paint wells, start joints, instance_sha; the exemplar is target.png)
diff --git a/tasks/robopaint-oil-saintevictoire-i01-standard/task.toml b/tasks/robopaint-oil-saintevictoire-i01-standard/task.toml
index ab822373d2..675f6a4a31 100644
--- a/tasks/robopaint-oil-saintevictoire-i01-standard/task.toml
+++ b/tasks/robopaint-oil-saintevictoire-i01-standard/task.toml
@@ -31,6 +31,7 @@ simulator = "maniskill"
robot = "franka-panda"
family = "robopaint_oil_saintevictoire"
mode = "standard" # camera RGB + proprioception only; no sheet readout, no score, no reset, no brush model
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "partial" # solution/: the privileged oracle's joint targets, streamed through the client
oracle_evidence = "Reference solution in solution/ (streams the privileged reference's joint targets; the hi-res version of 2026-10-09, README Change Log). The same trajectory scores 1 under the process rules in the verifier, whose replay CLI the collect hook runs (2026-10-09): path_mm 19229 of the 57700 mm budget; its strokes are sweeps. Harbor oracle and nop runs: docs/benchmarks/robopaint.md (Oil: the pointed brush)."
diff --git a/tasks/robopaint-oil-starrynight-i01-privileged/task.toml b/tasks/robopaint-oil-starrynight-i01-privileged/task.toml
index 75414549da..4b038c552d 100644
--- a/tasks/robopaint-oil-starrynight-i01-privileged/task.toml
+++ b/tasks/robopaint-oil-starrynight-i01-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "maniskill" # ManiSkill 3.0.1 / SAPIEN 3.0.3 (PhysX on the
robot = "franka-panda" # PandaStick with a paint brush: rigid handle + 15 mm collision-free tuft, TCP = the tuft's tip
family = "robopaint_oil_starrynight"
mode = "privileged" # privileged (exact sheets, score, brush state) | standard (RGB only: the -standard twin)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "partial" # solution/solve.sh replays the frozen reference plan; its recorded scores: oracle_evidence
oracle_evidence = "Reference solution in solution/: the hi-res version's stroke program for the pointed brush, executed by the robot at generation time (2026-10-09; README Change Log). The verifier scores it 1 (tests/grade.py with rcb_brush_strict.py, two fresh replays, the eai-robopaint:0.2.0 runtime, 2026-10-09): path_mm 20747 of the 62300 mm budget; its strokes are sweeps. Harbor oracle and nop runs: docs/benchmarks/robopaint.md (Oil: the pointed brush)."
instance_seed = 0 # scene = environment/instance.json (sheets, brush preset, paint wells, start joints, instance_sha; the exemplar is target.png)
diff --git a/tasks/robopaint-oil-starrynight-i01-standard/task.toml b/tasks/robopaint-oil-starrynight-i01-standard/task.toml
index d5422877e6..db5c866b2a 100644
--- a/tasks/robopaint-oil-starrynight-i01-standard/task.toml
+++ b/tasks/robopaint-oil-starrynight-i01-standard/task.toml
@@ -31,6 +31,7 @@ simulator = "maniskill"
robot = "franka-panda"
family = "robopaint_oil_starrynight"
mode = "standard" # camera RGB + proprioception only; no sheet readout, no score, no reset, no brush model
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "partial" # solution/: the privileged oracle's joint targets, streamed through the client
oracle_evidence = "Reference solution in solution/ (streams the privileged reference's joint targets; the hi-res version of 2026-10-09, README Change Log). The same trajectory scores 1 under the process rules in the verifier, whose replay CLI the collect hook runs (2026-10-09): path_mm 20747 of the 62300 mm budget; its strokes are sweeps. Harbor oracle and nop runs: docs/benchmarks/robopaint.md (Oil: the pointed brush)."
diff --git a/tasks/robopaint-oil-temeraire-i01-privileged/task.toml b/tasks/robopaint-oil-temeraire-i01-privileged/task.toml
index 4d0098ec4b..466442eb24 100644
--- a/tasks/robopaint-oil-temeraire-i01-privileged/task.toml
+++ b/tasks/robopaint-oil-temeraire-i01-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "maniskill" # ManiSkill 3.0.1 / SAPIEN 3.0.3 (PhysX on the
robot = "franka-panda" # PandaStick with a paint brush: rigid handle + 15 mm collision-free tuft, TCP = the tuft's tip
family = "robopaint_oil_temeraire"
mode = "privileged" # privileged (exact sheets, score, brush state) | standard (RGB only: the -standard twin)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "partial" # solution/solve.sh replays the frozen reference plan; its recorded scores: oracle_evidence
oracle_evidence = "Reference solution in solution/: the hi-res version's stroke program for the pointed brush, executed by the robot at generation time (2026-10-09; README Change Log). The verifier scores it 1 (tests/grade.py with rcb_brush_strict.py, two fresh replays, the eai-robopaint:0.2.0 runtime, 2026-10-09): path_mm 17125 of the 51400 mm budget; its strokes are sweeps. Harbor oracle and nop runs: docs/benchmarks/robopaint.md (Oil: the pointed brush)."
instance_seed = 0 # scene = environment/instance.json (sheets, brush preset, paint wells, start joints, instance_sha; the exemplar is target.png)
diff --git a/tasks/robopaint-oil-temeraire-i01-standard/task.toml b/tasks/robopaint-oil-temeraire-i01-standard/task.toml
index 3a2c9e7e35..45fab69460 100644
--- a/tasks/robopaint-oil-temeraire-i01-standard/task.toml
+++ b/tasks/robopaint-oil-temeraire-i01-standard/task.toml
@@ -31,6 +31,7 @@ simulator = "maniskill"
robot = "franka-panda"
family = "robopaint_oil_temeraire"
mode = "standard" # camera RGB + proprioception only; no sheet readout, no score, no reset, no brush model
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "partial" # solution/: the privileged oracle's joint targets, streamed through the client
oracle_evidence = "Reference solution in solution/ (streams the privileged reference's joint targets; the hi-res version of 2026-10-09, README Change Log). The same trajectory scores 1 under the process rules in the verifier, whose replay CLI the collect hook runs (2026-10-09): path_mm 17125 of the 51400 mm budget; its strokes are sweeps. Harbor oracle and nop runs: docs/benchmarks/robopaint.md (Oil: the pointed brush)."
diff --git a/tasks/robopaint-oil-wanderer-i01-privileged/task.toml b/tasks/robopaint-oil-wanderer-i01-privileged/task.toml
index bb3015a4f4..a5ef68ac73 100644
--- a/tasks/robopaint-oil-wanderer-i01-privileged/task.toml
+++ b/tasks/robopaint-oil-wanderer-i01-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "maniskill" # ManiSkill 3.0.1 / SAPIEN 3.0.3 (PhysX on the
robot = "franka-panda" # PandaStick with a paint brush: rigid handle + 15 mm collision-free tuft, TCP = the tuft's tip
family = "robopaint_oil_wanderer"
mode = "privileged" # privileged (exact sheets, score, brush state) | standard (RGB only: the -standard twin)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "partial" # solution/solve.sh replays the frozen reference plan; its recorded scores: oracle_evidence
oracle_evidence = "Reference solution in solution/: the hi-res version's stroke program for the pointed brush, executed by the robot at generation time (2026-10-09; README Change Log). The verifier scores it 1 (tests/grade.py with rcb_brush_strict.py, two fresh replays, the eai-robopaint:0.2.0 runtime, 2026-10-09): path_mm 21539 of the 64700 mm budget; its strokes are sweeps. Harbor oracle and nop runs: docs/benchmarks/robopaint.md (Oil: the pointed brush)."
instance_seed = 0 # scene = environment/instance.json (sheets, brush preset, paint wells, start joints, instance_sha; the exemplar is target.png)
diff --git a/tasks/robopaint-oil-wanderer-i01-standard/task.toml b/tasks/robopaint-oil-wanderer-i01-standard/task.toml
index c907e0f0ed..a31eb012e6 100644
--- a/tasks/robopaint-oil-wanderer-i01-standard/task.toml
+++ b/tasks/robopaint-oil-wanderer-i01-standard/task.toml
@@ -31,6 +31,7 @@ simulator = "maniskill"
robot = "franka-panda"
family = "robopaint_oil_wanderer"
mode = "standard" # camera RGB + proprioception only; no sheet readout, no score, no reset, no brush model
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "partial" # solution/: the privileged oracle's joint targets, streamed through the client
oracle_evidence = "Reference solution in solution/ (streams the privileged reference's joint targets; the hi-res version of 2026-10-09, README Change Log). The same trajectory scores 1 under the process rules in the verifier, whose replay CLI the collect hook runs (2026-10-09): path_mm 21539 of the 64700 mm budget; its strokes are sweeps. Harbor oracle and nop runs: docs/benchmarks/robopaint.md (Oil: the pointed brush)."
diff --git a/tasks/robopaint-xingshu-feng-i00-privileged/task.toml b/tasks/robopaint-xingshu-feng-i00-privileged/task.toml
index 2b2dba1afe..ee0328a7e3 100644
--- a/tasks/robopaint-xingshu-feng-i00-privileged/task.toml
+++ b/tasks/robopaint-xingshu-feng-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "maniskill" # ManiSkill 3.0.1 / SAPIEN 3.0.3 (PhysX on the
robot = "franka-panda" # PandaStick with a Chinese brush: rigid handle + 30 mm collision-free tuft, TCP = the tuft's tip
family = "robopaint_xingshu_feng"
mode = "privileged" # privileged (exact sheets, brush state, reset, scoring service) | standard (RGB only: the -standard twin)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "partial" # solution/solve.sh replays the frozen reference plan; its recorded scores: oracle_evidence
oracle_evidence = "Reference solution in solution/, re-made on 2026-10-09 by the calligraphy planner of solution/README.md (README, Change Log; docs/benchmarks/robopaint.md, Targets). The verifier scores it 1 under the process rules (tests/grade.py with rcb_brush_strict.py, two fresh replays): path_mm 1181 and mark_width_mm 5.00, against the colouring-in limits 2150 mm and 2.42 mm. Harbor oracle 1 and nop 0 on both modes of this pair on its current files (ws3, 2026-10-09): docs/benchmarks/robopaint.md (Validation)."
instance_seed = 0 # scene = environment/instance.json (agent's copy: sheets, brush, ink stone, start joints, the rule, instance_sha)
diff --git a/tasks/robopaint-xingshu-feng-i00-standard/task.toml b/tasks/robopaint-xingshu-feng-i00-standard/task.toml
index 9c049fcd78..12f30f4ba7 100644
--- a/tasks/robopaint-xingshu-feng-i00-standard/task.toml
+++ b/tasks/robopaint-xingshu-feng-i00-standard/task.toml
@@ -31,6 +31,7 @@ simulator = "maniskill"
robot = "franka-panda"
family = "robopaint_xingshu_feng"
mode = "standard" # camera RGB + proprioception only; no sheet readout, no score, no reset, no brush model
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "partial" # solution/: the privileged oracle's joint targets, streamed through the client
oracle_evidence = "Reference solution in solution/ (streams the privileged reference's joint targets), re-made on 2026-10-09 (README, Change Log; docs/benchmarks/robopaint.md, Targets). The same trajectory scores 1 under the process rules in the verifier, whose replay CLI the collect hook runs: path_mm 1181 and mark_width_mm 5.00, against the colouring-in limits 2150 mm and 2.42 mm. Harbor oracle 1 and nop 0 on both modes of this pair on its current files (ws3, 2026-10-09): docs/benchmarks/robopaint.md (Validation)."
diff --git a/tasks/robopaint-xingshu-he-i00-privileged/task.toml b/tasks/robopaint-xingshu-he-i00-privileged/task.toml
index 0e5c486732..8ad56d1de4 100644
--- a/tasks/robopaint-xingshu-he-i00-privileged/task.toml
+++ b/tasks/robopaint-xingshu-he-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "maniskill" # ManiSkill 3.0.1 / SAPIEN 3.0.3 (PhysX on the
robot = "franka-panda" # PandaStick with a Chinese brush: rigid handle + 30 mm collision-free tuft, TCP = the tuft's tip
family = "robopaint_xingshu_he"
mode = "privileged" # privileged (exact sheets, brush state, reset, scoring service) | standard (RGB only: the -standard twin)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "partial" # solution/solve.sh replays the frozen reference plan; its recorded scores: oracle_evidence
oracle_evidence = "Reference solution in solution/, re-made on 2026-10-09 by the calligraphy planner of solution/README.md (README, Change Log; docs/benchmarks/robopaint.md, Targets). The verifier scores it 1 under the process rules (tests/grade.py with rcb_brush_strict.py, two fresh replays): path_mm 927 and mark_width_mm 4.56, against the colouring-in limits 1850 mm and 1.93 mm. Harbor oracle 1 and nop 0 on both modes of this pair on its current files (ws3, 2026-10-09): docs/benchmarks/robopaint.md (Validation)."
instance_seed = 0 # scene = environment/instance.json (agent's copy: sheets, brush, ink stone, start joints, the rule, instance_sha)
diff --git a/tasks/robopaint-xingshu-he-i00-standard/task.toml b/tasks/robopaint-xingshu-he-i00-standard/task.toml
index 5d30e76762..c34dd10883 100644
--- a/tasks/robopaint-xingshu-he-i00-standard/task.toml
+++ b/tasks/robopaint-xingshu-he-i00-standard/task.toml
@@ -31,6 +31,7 @@ simulator = "maniskill"
robot = "franka-panda"
family = "robopaint_xingshu_he"
mode = "standard" # camera RGB + proprioception only; no sheet readout, no score, no reset, no brush model
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "partial" # solution/: the privileged oracle's joint targets, streamed through the client
oracle_evidence = "Reference solution in solution/ (streams the privileged reference's joint targets), re-made on 2026-10-09 (README, Change Log; docs/benchmarks/robopaint.md, Targets). The same trajectory scores 1 under the process rules in the verifier, whose replay CLI the collect hook runs: path_mm 927 and mark_width_mm 4.56, against the colouring-in limits 1850 mm and 1.93 mm. Harbor oracle 1 and nop 0 on both modes of this pair on its current files (ws3, 2026-10-09): docs/benchmarks/robopaint.md (Validation)."
diff --git a/tasks/robopaint-xingshu-huifenghechang-i00-privileged/task.toml b/tasks/robopaint-xingshu-huifenghechang-i00-privileged/task.toml
index c8b2290330..6c1a4438e6 100644
--- a/tasks/robopaint-xingshu-huifenghechang-i00-privileged/task.toml
+++ b/tasks/robopaint-xingshu-huifenghechang-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "maniskill" # ManiSkill 3.0.1 / SAPIEN 3.0.3 (PhysX on the
robot = "franka-panda" # PandaStick with a Chinese brush: rigid handle + 30 mm collision-free tuft, TCP = the tuft's tip
family = "robopaint_xingshu_huifenghechang"
mode = "privileged" # privileged (exact sheets, brush state, reset, scoring service) | standard (RGB only: the -standard twin)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "partial" # solution/solve.sh replays the frozen reference plan; its recorded scores: oracle_evidence
oracle_evidence = "Reference solution in solution/, re-made on 2026-10-09 by the calligraphy planner of solution/README.md (README, Change Log; docs/benchmarks/robopaint.md, Targets). The verifier scores it 1 under the process rules (tests/grade.py with rcb_brush_strict.py, two fresh replays): path_mm 2483 and mark_width_mm 3.04, against the colouring-in limits 6060 mm and 1.15 mm. Harbor oracle 1 and nop 0 on both modes of this pair on its current files (ws3, 2026-10-09): docs/benchmarks/robopaint.md (Validation)."
instance_seed = 0 # scene = environment/instance.json (agent's copy: sheets, brush, ink stone, start joints, the rule, instance_sha)
diff --git a/tasks/robopaint-xingshu-huifenghechang-i00-standard/task.toml b/tasks/robopaint-xingshu-huifenghechang-i00-standard/task.toml
index 63346144bb..e6d790d32d 100644
--- a/tasks/robopaint-xingshu-huifenghechang-i00-standard/task.toml
+++ b/tasks/robopaint-xingshu-huifenghechang-i00-standard/task.toml
@@ -31,6 +31,7 @@ simulator = "maniskill"
robot = "franka-panda"
family = "robopaint_xingshu_huifenghechang"
mode = "standard" # camera RGB + proprioception only; no sheet readout, no score, no reset, no brush model
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "partial" # solution/: the privileged oracle's joint targets, streamed through the client
oracle_evidence = "Reference solution in solution/ (streams the privileged reference's joint targets), re-made on 2026-10-09 (README, Change Log; docs/benchmarks/robopaint.md, Targets). The same trajectory scores 1 under the process rules in the verifier, whose replay CLI the collect hook runs: path_mm 2483 and mark_width_mm 3.04, against the colouring-in limits 6060 mm and 1.15 mm. Harbor oracle 1 and nop 0 on both modes of this pair on its current files (ws3, 2026-10-09): docs/benchmarks/robopaint.md (Validation)."
diff --git a/tasks/robopaint-xingshu-maolinxiuzhu-i00-privileged/task.toml b/tasks/robopaint-xingshu-maolinxiuzhu-i00-privileged/task.toml
index 2e55bec27b..5e3680378d 100644
--- a/tasks/robopaint-xingshu-maolinxiuzhu-i00-privileged/task.toml
+++ b/tasks/robopaint-xingshu-maolinxiuzhu-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "maniskill" # ManiSkill 3.0.1 / SAPIEN 3.0.3 (PhysX on the
robot = "franka-panda" # PandaStick with a Chinese brush: rigid handle + 30 mm collision-free tuft, TCP = the tuft's tip
family = "robopaint_xingshu_maolinxiuzhu"
mode = "privileged" # privileged (exact sheets, brush state, reset, scoring service) | standard (RGB only: the -standard twin)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "partial" # solution/solve.sh replays the frozen reference plan; its recorded scores: oracle_evidence
oracle_evidence = "Reference solution in solution/, re-made on 2026-10-09 by the calligraphy planner of solution/README.md (README, Change Log; docs/benchmarks/robopaint.md, Targets). The verifier scores it 1 under the process rules (tests/grade.py with rcb_brush_strict.py, two fresh replays): path_mm 1412 and mark_width_mm 2.79, against the colouring-in limits 3160 mm and 1.12 mm. Harbor oracle 1 and nop 0 on both modes of this pair on its current files (ws3, 2026-10-09): docs/benchmarks/robopaint.md (Validation)."
instance_seed = 0 # scene = environment/instance.json (agent's copy: sheets, brush, ink stone, start joints, the rule, instance_sha)
diff --git a/tasks/robopaint-xingshu-maolinxiuzhu-i00-standard/task.toml b/tasks/robopaint-xingshu-maolinxiuzhu-i00-standard/task.toml
index 9dd4fc12c7..2194dc9f1a 100644
--- a/tasks/robopaint-xingshu-maolinxiuzhu-i00-standard/task.toml
+++ b/tasks/robopaint-xingshu-maolinxiuzhu-i00-standard/task.toml
@@ -31,6 +31,7 @@ simulator = "maniskill"
robot = "franka-panda"
family = "robopaint_xingshu_maolinxiuzhu"
mode = "standard" # camera RGB + proprioception only; no sheet readout, no score, no reset, no brush model
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "partial" # solution/: the privileged oracle's joint targets, streamed through the client
oracle_evidence = "Reference solution in solution/ (streams the privileged reference's joint targets), re-made on 2026-10-09 (README, Change Log; docs/benchmarks/robopaint.md, Targets). The same trajectory scores 1 under the process rules in the verifier, whose replay CLI the collect hook runs: path_mm 1412 and mark_width_mm 2.79, against the colouring-in limits 3160 mm and 1.12 mm. Harbor oracle 1 and nop 0 on both modes of this pair on its current files (ws3, 2026-10-09): docs/benchmarks/robopaint.md (Validation)."
diff --git a/tasks/robopaint-xingshu-nian-i00-privileged/task.toml b/tasks/robopaint-xingshu-nian-i00-privileged/task.toml
index 1958702630..e7022a4fc4 100644
--- a/tasks/robopaint-xingshu-nian-i00-privileged/task.toml
+++ b/tasks/robopaint-xingshu-nian-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "maniskill" # ManiSkill 3.0.1 / SAPIEN 3.0.3 (PhysX on the
robot = "franka-panda" # PandaStick with a Chinese brush: rigid handle + 30 mm collision-free tuft, TCP = the tuft's tip
family = "robopaint_xingshu_nian"
mode = "privileged" # privileged (exact sheets, brush state, reset, scoring service) | standard (RGB only: the -standard twin)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "partial" # solution/solve.sh replays the frozen reference plan; its recorded scores: oracle_evidence
oracle_evidence = "Reference solution in solution/, re-made on 2026-10-09 by the calligraphy planner of solution/README.md (README, Change Log; docs/benchmarks/robopaint.md, Targets). The verifier scores it 1 under the process rules (tests/grade.py with rcb_brush_strict.py, two fresh replays): path_mm 1013 and mark_width_mm 4.59, against the colouring-in limits 1780 mm and 2.20 mm. Harbor oracle 1 and nop 0 on both modes of this pair on its current files (ws3, 2026-10-09): docs/benchmarks/robopaint.md (Validation)."
instance_seed = 0 # scene = environment/instance.json (agent's copy: sheets, brush, ink stone, start joints, the rule, instance_sha)
diff --git a/tasks/robopaint-xingshu-nian-i00-standard/task.toml b/tasks/robopaint-xingshu-nian-i00-standard/task.toml
index abd876f50c..8b02275910 100644
--- a/tasks/robopaint-xingshu-nian-i00-standard/task.toml
+++ b/tasks/robopaint-xingshu-nian-i00-standard/task.toml
@@ -31,6 +31,7 @@ simulator = "maniskill"
robot = "franka-panda"
family = "robopaint_xingshu_nian"
mode = "standard" # camera RGB + proprioception only; no sheet readout, no score, no reset, no brush model
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "partial" # solution/: the privileged oracle's joint targets, streamed through the client
oracle_evidence = "Reference solution in solution/ (streams the privileged reference's joint targets), re-made on 2026-10-09 (README, Change Log; docs/benchmarks/robopaint.md, Targets). The same trajectory scores 1 under the process rules in the verifier, whose replay CLI the collect hook runs: path_mm 1013 and mark_width_mm 4.59, against the colouring-in limits 1780 mm and 2.20 mm. Harbor oracle 1 and nop 0 on both modes of this pair on its current files (ws3, 2026-10-09): docs/benchmarks/robopaint.md (Validation)."
diff --git a/tasks/robopaint-xingshu-shan-i00-privileged/task.toml b/tasks/robopaint-xingshu-shan-i00-privileged/task.toml
index 251d8f0548..1f53f18e81 100644
--- a/tasks/robopaint-xingshu-shan-i00-privileged/task.toml
+++ b/tasks/robopaint-xingshu-shan-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "maniskill" # ManiSkill 3.0.1 / SAPIEN 3.0.3 (PhysX on the
robot = "franka-panda" # PandaStick with a Chinese brush: rigid handle + 30 mm collision-free tuft, TCP = the tuft's tip
family = "robopaint_xingshu_shan"
mode = "privileged" # privileged (exact sheets, brush state, reset, scoring service) | standard (RGB only: the -standard twin)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "partial" # solution/solve.sh replays the frozen reference plan; its recorded scores: oracle_evidence
oracle_evidence = "Reference solution in solution/, re-made on 2026-10-09 by the calligraphy planner of solution/README.md (README, Change Log; docs/benchmarks/robopaint.md, Targets). The verifier scores it 1 under the process rules (tests/grade.py with rcb_brush_strict.py, two fresh replays): path_mm 608 and mark_width_mm 6.75, against the colouring-in limits 1200 mm and 3.06 mm. Harbor oracle 1 and nop 0 on both modes of this pair on its current files (ws3, 2026-10-09): docs/benchmarks/robopaint.md (Validation)."
instance_seed = 0 # scene = environment/instance.json (agent's copy: sheets, brush, ink stone, start joints, the rule, instance_sha)
diff --git a/tasks/robopaint-xingshu-shan-i00-standard/task.toml b/tasks/robopaint-xingshu-shan-i00-standard/task.toml
index a64f447042..7ae46ff64f 100644
--- a/tasks/robopaint-xingshu-shan-i00-standard/task.toml
+++ b/tasks/robopaint-xingshu-shan-i00-standard/task.toml
@@ -31,6 +31,7 @@ simulator = "maniskill"
robot = "franka-panda"
family = "robopaint_xingshu_shan"
mode = "standard" # camera RGB + proprioception only; no sheet readout, no score, no reset, no brush model
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "partial" # solution/: the privileged oracle's joint targets, streamed through the client
oracle_evidence = "Reference solution in solution/ (streams the privileged reference's joint targets), re-made on 2026-10-09 (README, Change Log; docs/benchmarks/robopaint.md, Targets). The same trajectory scores 1 under the process rules in the verifier, whose replay CLI the collect hook runs: path_mm 608 and mark_width_mm 6.75, against the colouring-in limits 1200 mm and 3.06 mm. Harbor oracle 1 and nop 0 on both modes of this pair on its current files (ws3, 2026-10-09): docs/benchmarks/robopaint.md (Validation)."
diff --git a/tasks/robopaint-xingshu-xian-i00-privileged/task.toml b/tasks/robopaint-xingshu-xian-i00-privileged/task.toml
index 46be74b8b8..590980f2e8 100644
--- a/tasks/robopaint-xingshu-xian-i00-privileged/task.toml
+++ b/tasks/robopaint-xingshu-xian-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "maniskill" # ManiSkill 3.0.1 / SAPIEN 3.0.3 (PhysX on the
robot = "franka-panda" # PandaStick with a Chinese brush: rigid handle + 30 mm collision-free tuft, TCP = the tuft's tip
family = "robopaint_xingshu_xian"
mode = "privileged" # privileged (exact sheets, brush state, reset, scoring service) | standard (RGB only: the -standard twin)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "partial" # solution/solve.sh replays the frozen reference plan; its recorded scores: oracle_evidence
oracle_evidence = "Reference solution in solution/, re-made on 2026-10-09 by the calligraphy planner of solution/README.md (README, Change Log; docs/benchmarks/robopaint.md, Targets). The verifier scores it 1 under the process rules (tests/grade.py with rcb_brush_strict.py, two fresh replays): path_mm 1234 and mark_width_mm 4.59, against the colouring-in limits 2630 mm and 1.88 mm. Harbor oracle 1 and nop 0 on both modes of this pair on its current files (ws3, 2026-10-09): docs/benchmarks/robopaint.md (Validation)."
instance_seed = 0 # scene = environment/instance.json (agent's copy: sheets, brush, ink stone, start joints, the rule, instance_sha)
diff --git a/tasks/robopaint-xingshu-xian-i00-standard/task.toml b/tasks/robopaint-xingshu-xian-i00-standard/task.toml
index be1cc501aa..79fbf51a22 100644
--- a/tasks/robopaint-xingshu-xian-i00-standard/task.toml
+++ b/tasks/robopaint-xingshu-xian-i00-standard/task.toml
@@ -31,6 +31,7 @@ simulator = "maniskill"
robot = "franka-panda"
family = "robopaint_xingshu_xian"
mode = "standard" # camera RGB + proprioception only; no sheet readout, no score, no reset, no brush model
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "partial" # solution/: the privileged oracle's joint targets, streamed through the client
oracle_evidence = "Reference solution in solution/ (streams the privileged reference's joint targets), re-made on 2026-10-09 (README, Change Log; docs/benchmarks/robopaint.md, Targets). The same trajectory scores 1 under the process rules in the verifier, whose replay CLI the collect hook runs: path_mm 1234 and mark_width_mm 4.59, against the colouring-in limits 2630 mm and 1.88 mm. Harbor oracle 1 and nop 0 on both modes of this pair on its current files (ws3, 2026-10-09): docs/benchmarks/robopaint.md (Validation)."
diff --git a/tasks/robopaint-xingshu-yong-i00-privileged/task.toml b/tasks/robopaint-xingshu-yong-i00-privileged/task.toml
index e4f6c78f90..a436bd2114 100644
--- a/tasks/robopaint-xingshu-yong-i00-privileged/task.toml
+++ b/tasks/robopaint-xingshu-yong-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "maniskill" # ManiSkill 3.0.1 / SAPIEN 3.0.3 (PhysX on the
robot = "franka-panda" # PandaStick with a Chinese brush: rigid handle + 30 mm collision-free tuft, TCP = the tuft's tip
family = "robopaint_xingshu_yong"
mode = "privileged" # privileged (exact sheets, brush state, reset, scoring service) | standard (RGB only: the -standard twin)
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "partial" # solution/solve.sh replays the frozen reference plan; its recorded scores: oracle_evidence
oracle_evidence = "Reference solution in solution/, re-made on 2026-10-09 by the calligraphy planner of solution/README.md (README, Change Log; docs/benchmarks/robopaint.md, Targets). The verifier scores it 1 under the process rules (tests/grade.py with rcb_brush_strict.py, two fresh replays): path_mm 726 and mark_width_mm 5.14, against the colouring-in limits 1450 mm and 2.25 mm. Harbor oracle 1 and nop 0 on both modes of this pair on its current files (ws3, 2026-10-09): docs/benchmarks/robopaint.md (Validation)."
instance_seed = 0 # scene = environment/instance.json (agent's copy: sheets, brush, ink stone, start joints, the rule, instance_sha)
diff --git a/tasks/robopaint-xingshu-yong-i00-standard/task.toml b/tasks/robopaint-xingshu-yong-i00-standard/task.toml
index 974058460f..12f256d2a4 100644
--- a/tasks/robopaint-xingshu-yong-i00-standard/task.toml
+++ b/tasks/robopaint-xingshu-yong-i00-standard/task.toml
@@ -31,6 +31,7 @@ simulator = "maniskill"
robot = "franka-panda"
family = "robopaint_xingshu_yong"
mode = "standard" # camera RGB + proprioception only; no sheet readout, no score, no reset, no brush model
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "partial" # solution/: the privileged oracle's joint targets, streamed through the client
oracle_evidence = "Reference solution in solution/ (streams the privileged reference's joint targets), re-made on 2026-10-09 (README, Change Log; docs/benchmarks/robopaint.md, Targets). The same trajectory scores 1 under the process rules in the verifier, whose replay CLI the collect hook runs: path_mm 726 and mark_width_mm 5.14, against the colouring-in limits 1450 mm and 2.25 mm. Harbor oracle 1 and nop 0 on both modes of this pair on its current files (ws3, 2026-10-09): docs/benchmarks/robopaint.md (Validation)."
diff --git a/tasks/robotwin_2_0_adjust_bottle-privileged/README.md b/tasks/robotwin_2_0_adjust_bottle-privileged/README.md
index eafe901b38..696eaae0c5 100644
--- a/tasks/robotwin_2_0_adjust_bottle-privileged/README.md
+++ b/tasks/robotwin_2_0_adjust_bottle-privileged/README.md
@@ -56,3 +56,4 @@ Pick up the bottle on the table headup with the correct arm.
- 2026-10-13 (board 0086): the tooling refuses a trajectory the verifier cannot grade (`environment/rcb_traj.py`, `tests/rcb_traj.py`). `Session.save_trajectory` and `Session.replay` raise `ValueError` for more than 200,000 rows (`rcb_traj.MAX_T`, the verifier's `MAX_T` in `tests/grade.py`: 800 s of physics), and `record_stop` warns on stderr. Before, the tool saved and replayed such a file ("success") and the verifier then graded it `bad_trajectory`; the instruction has stated the cap since 2026-10-10. The verifier, the scene, the instance and the pass rule are unchanged: every trajectory is graded as before.
; `solution/` rebuilds the scene with upstream's `load_actors` to get them back for the expert. None of them is read by `check_success`. The scene, the frozen instance (same initial-state hash), the tooling, the verifier and the pass rule are unchanged, so the task keeps its name; privileged results from before this date were obtained with these values reachable.
- 2026-10-13 (board 0100, owner decision): base image `eai-robotwin` 0.1.2. The image's scrub of RoboTwin's expert kept `load_actors` whole, so values it computes only for the removed expert stayed reachable: `env.left_target_pose`, `env.right_target_pose` (poses the expert moves to or places at), and the instruction listed them among the scene objects. They are gone: the env no longer has them, `/opt/RoboTwin/envs/adjust_bottle.py` no longer shows them, and the scene-object list no longer names them; `solution/` rebuilds the scene with upstream's `load_actors` to get them back for the expert. None of them is read by `check_success`. The scene, the frozen instance (same initial-state hash), the tooling, the verifier and the pass rule are unchanged, so the task keeps its name; privileged results from before this date were obtained with these values reachable.
+- 2026-10-14 (board 0113): the tooling only (`environment/rcb_traj.py`, `tests/rcb_traj.py`): while it builds the scene, `glob.glob` returns its matches sorted, so that a task which lists files to draw an object depends on its seed alone, not on the host's directory order (four RoboTwin tasks did; this one lists no files while it builds its scene). Its scene, frozen instance (same initial-state hash), verifier and pass rule are unchanged.
diff --git a/tasks/robotwin_2_0_adjust_bottle-privileged/environment/rcb_traj.py b/tasks/robotwin_2_0_adjust_bottle-privileged/environment/rcb_traj.py
index 565f6c8d62..5794e7e6de 100644
--- a/tasks/robotwin_2_0_adjust_bottle-privileged/environment/rcb_traj.py
+++ b/tasks/robotwin_2_0_adjust_bottle-privileged/environment/rcb_traj.py
@@ -113,6 +113,26 @@ BASE_ARGS: dict[str, Any] = {
}
+class _SortedGlob:
+ """While a scene is built, glob.glob / glob.iglob return their matches sorted. Four RoboTwin tasks (place_a2b_left,
+ place_a2b_right, place_object_scale, place_object_stand: envs/<task>.py load_actors) list an object's model files
+ with glob.glob and then draw one with the seeded np.random.choice; glob returns the file system's directory order,
+ which differs between hosts (and image layers), so one seed built a different object on another host and the frozen
+ instance's hash check failed there (board 0086 claim 5; never rebuilt off the reference workstation). Sorted, the
+ draw depends on the seed alone. No other task of ours lists files while it builds its scene."""
+
+ def __enter__(self):
+ import glob
+ self._glob, self._orig = glob, (glob.glob, glob.iglob)
+ glob.glob = lambda *a, **k: sorted(self._orig[0](*a, **k))
+ glob.iglob = lambda *a, **k: iter(sorted(self._orig[1](*a, **k)))
+ return self
+
+ def __exit__(self, *exc):
+ self._glob.glob, self._glob.iglob = self._orig
+ return False
+
+
def _chdir_root() -> None:
# RoboTwin resolves assets/config paths relative to the repo root and via ROOT_PATH
if os.getcwd() != ROBOTWIN_ROOT and os.path.isdir(ROBOTWIN_ROOT):
@@ -218,7 +238,8 @@ class Session:
if self.enable_render:
sapien.render.set_camera_shader_dir(self.render_shader)
args = _build_args(self.task, self.embodiment, self.seed)
- self.env.setup_demo(**args)
+ with _SortedGlob(): # the object draw depends on the seed alone, not the host's directory order
+ self.env.setup_demo(**args)
if self.enable_render: # setup_scene() switches the shader to ray tracing; put it back for cheap frames
sapien.render.set_camera_shader_dir(self.render_shader)
self._raw_scene = self.env.scene
diff --git a/tasks/robotwin_2_0_adjust_bottle-privileged/task.toml b/tasks/robotwin_2_0_adjust_bottle-privileged/task.toml
index 403f57355a..a021b7786a 100644
--- a/tasks/robotwin_2_0_adjust_bottle-privileged/task.toml
+++ b/tasks/robotwin_2_0_adjust_bottle-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "robotwin" # RoboTwin 2.0 = SAPIEN 3 / PhysX + CuRobo
robot = "aloha-agilex"
family = "adjust_bottle"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 60417 # scene = environment/instance.json (RoboTwin seed + expected initial-state hash); identity = dataset.toml digest
expert_time_estimate_hours = 0.5
diff --git a/tasks/robotwin_2_0_adjust_bottle-privileged/tests/rcb_traj.py b/tasks/robotwin_2_0_adjust_bottle-privileged/tests/rcb_traj.py
index 565f6c8d62..5794e7e6de 100644
--- a/tasks/robotwin_2_0_adjust_bottle-privileged/tests/rcb_traj.py
+++ b/tasks/robotwin_2_0_adjust_bottle-privileged/tests/rcb_traj.py
@@ -113,6 +113,26 @@ BASE_ARGS: dict[str, Any] = {
}
+class _SortedGlob:
+ """While a scene is built, glob.glob / glob.iglob return their matches sorted. Four RoboTwin tasks (place_a2b_left,
+ place_a2b_right, place_object_scale, place_object_stand: envs/<task>.py load_actors) list an object's model files
+ with glob.glob and then draw one with the seeded np.random.choice; glob returns the file system's directory order,
+ which differs between hosts (and image layers), so one seed built a different object on another host and the frozen
+ instance's hash check failed there (board 0086 claim 5; never rebuilt off the reference workstation). Sorted, the
+ draw depends on the seed alone. No other task of ours lists files while it builds its scene."""
+
+ def __enter__(self):
+ import glob
+ self._glob, self._orig = glob, (glob.glob, glob.iglob)
+ glob.glob = lambda *a, **k: sorted(self._orig[0](*a, **k))
+ glob.iglob = lambda *a, **k: iter(sorted(self._orig[1](*a, **k)))
+ return self
+
+ def __exit__(self, *exc):
+ self._glob.glob, self._glob.iglob = self._orig
+ return False
+
+
def _chdir_root() -> None:
# RoboTwin resolves assets/config paths relative to the repo root and via ROOT_PATH
if os.getcwd() != ROBOTWIN_ROOT and os.path.isdir(ROBOTWIN_ROOT):
@@ -218,7 +238,8 @@ class Session:
if self.enable_render:
sapien.render.set_camera_shader_dir(self.render_shader)
args = _build_args(self.task, self.embodiment, self.seed)
- self.env.setup_demo(**args)
+ with _SortedGlob(): # the object draw depends on the seed alone, not the host's directory order
+ self.env.setup_demo(**args)
if self.enable_render: # setup_scene() switches the shader to ray tracing; put it back for cheap frames
sapien.render.set_camera_shader_dir(self.render_shader)
self._raw_scene = self.env.scene
diff --git a/tasks/robotwin_2_0_adjust_bottle-standard/README.md b/tasks/robotwin_2_0_adjust_bottle-standard/README.md
index 886726d067..d138fe26b9 100644
--- a/tasks/robotwin_2_0_adjust_bottle-standard/README.md
+++ b/tasks/robotwin_2_0_adjust_bottle-standard/README.md
@@ -53,3 +53,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
- 2026-10-10 (owner decision, board 0072): the agent's network egress is restricted to the model APIs (`[agent] network_mode = "allowlist"`, the hosts of `scripts/set_agent_egress.py`; since 2026-10-11 these include the package registries, PyPI and the Debian / Ubuntu archives, so that `pip install` and `apt-get install` work); it was open before (Harbor's default), and agents could download upstream sources, reference solutions and robot models from their own shell. The scene, the frozen instance, the tooling and the pass rule are unchanged, so the task keeps its name; results from before this date were obtained with open egress.
- 2026-10-13 (board 0086): the service's copy of the tooling only (`environment/sim/rcb_traj.py`, one line of `environment/sim/server.py`). The tooling now refuses to save or replay a trajectory of more than 200,000 rows, the privileged verifier's cap; a standard episode has no row cap, so the service lifts it for its own session (`S.max_rows = None`). Nothing an agent sees and nothing the service grades changes.
- 2026-10-13 (board 0100): base image `eai-robotwin` 0.1.2, whose scrub of RoboTwin's expert also removes the reference poses that `load_actors` computed only for the expert (here `left_target_pose`, `right_target_pose`). The standard-mode agent never reaches the env, so nothing it observes changes; the scene, the instance and the judge are unchanged.
+- 2026-10-14 (board 0113): the tooling only (`environment/sim/rcb_traj.py`): while it builds the scene, `glob.glob` returns its matches sorted, so that a task which lists files to draw an object depends on its seed alone, not on the host's directory order (four RoboTwin tasks did; this one lists no files while it builds its scene). Its scene, frozen instance (same initial-state hash), verifier and pass rule are unchanged.
diff --git a/tasks/robotwin_2_0_adjust_bottle-standard/environment/sim/rcb_traj.py b/tasks/robotwin_2_0_adjust_bottle-standard/environment/sim/rcb_traj.py
index 565f6c8d62..5794e7e6de 100644
--- a/tasks/robotwin_2_0_adjust_bottle-standard/environment/sim/rcb_traj.py
+++ b/tasks/robotwin_2_0_adjust_bottle-standard/environment/sim/rcb_traj.py
@@ -113,6 +113,26 @@ BASE_ARGS: dict[str, Any] = {
}
+class _SortedGlob:
+ """While a scene is built, glob.glob / glob.iglob return their matches sorted. Four RoboTwin tasks (place_a2b_left,
+ place_a2b_right, place_object_scale, place_object_stand: envs/<task>.py load_actors) list an object's model files
+ with glob.glob and then draw one with the seeded np.random.choice; glob returns the file system's directory order,
+ which differs between hosts (and image layers), so one seed built a different object on another host and the frozen
+ instance's hash check failed there (board 0086 claim 5; never rebuilt off the reference workstation). Sorted, the
+ draw depends on the seed alone. No other task of ours lists files while it builds its scene."""
+
+ def __enter__(self):
+ import glob
+ self._glob, self._orig = glob, (glob.glob, glob.iglob)
+ glob.glob = lambda *a, **k: sorted(self._orig[0](*a, **k))
+ glob.iglob = lambda *a, **k: iter(sorted(self._orig[1](*a, **k)))
+ return self
+
+ def __exit__(self, *exc):
+ self._glob.glob, self._glob.iglob = self._orig
+ return False
+
+
def _chdir_root() -> None:
# RoboTwin resolves assets/config paths relative to the repo root and via ROOT_PATH
if os.getcwd() != ROBOTWIN_ROOT and os.path.isdir(ROBOTWIN_ROOT):
@@ -218,7 +238,8 @@ class Session:
if self.enable_render:
sapien.render.set_camera_shader_dir(self.render_shader)
args = _build_args(self.task, self.embodiment, self.seed)
- self.env.setup_demo(**args)
+ with _SortedGlob(): # the object draw depends on the seed alone, not the host's directory order
+ self.env.setup_demo(**args)
if self.enable_render: # setup_scene() switches the shader to ray tracing; put it back for cheap frames
sapien.render.set_camera_shader_dir(self.render_shader)
self._raw_scene = self.env.scene
diff --git a/tasks/robotwin_2_0_adjust_bottle-standard/task.toml b/tasks/robotwin_2_0_adjust_bottle-standard/task.toml
index 3a8eb674fa..4854ff4987 100644
--- a/tasks/robotwin_2_0_adjust_bottle-standard/task.toml
+++ b/tasks/robotwin_2_0_adjust_bottle-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "robotwin"
robot = "aloha-agilex"
family = "adjust_bottle"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 60417
diff --git a/tasks/robotwin_2_0_beat_block_hammer-privileged/README.md b/tasks/robotwin_2_0_beat_block_hammer-privileged/README.md
index aa17f64221..306e1c1fb0 100644
--- a/tasks/robotwin_2_0_beat_block_hammer-privileged/README.md
+++ b/tasks/robotwin_2_0_beat_block_hammer-privileged/README.md
@@ -55,3 +55,4 @@ There is a hammer and a block on the table, use the arm to grab the hammer and b
- 2026-10-10 (owner decision, board 0072): the agent's network egress is restricted to the model APIs (`[agent] network_mode = "allowlist"`, the hosts of `scripts/set_agent_egress.py`; since 2026-10-11 these include the package registries, PyPI and the Debian / Ubuntu archives, so that `pip install` and `apt-get install` work); it was open before (Harbor's default), and agents could download upstream sources, reference solutions and robot models from their own shell. The scene, the frozen instance, the tooling and the pass rule are unchanged, so the task keeps its name; results from before this date were obtained with open egress.
- 2026-10-13 (board 0086): the tooling refuses a trajectory the verifier cannot grade (`environment/rcb_traj.py`, `tests/rcb_traj.py`). `Session.save_trajectory` and `Session.replay` raise `ValueError` for more than 200,000 rows (`rcb_traj.MAX_T`, the verifier's `MAX_T` in `tests/grade.py`: 800 s of physics), and `record_stop` warns on stderr. Before, the tool saved and replayed such a file ("success") and the verifier then graded it `bad_trajectory`; the instruction has stated the cap since 2026-10-10. The verifier, the scene, the instance and the pass rule are unchanged: every trajectory is graded as before.
- 2026-10-13 (board 0100): base image `eai-robotwin` 0.1.2. Its scrub of RoboTwin's expert also removes the reference poses that 18 tasks' `load_actors` computed only for the expert, and the expert's comments above the scrubbed method bodies (`images/robotwin/scrub_oracles.py`). This task's `load_actors` set no such value: its scene, instance, tooling, verifier and pass rule are unchanged.
+- 2026-10-14 (board 0113): the tooling only (`environment/rcb_traj.py`, `tests/rcb_traj.py`): while it builds the scene, `glob.glob` returns its matches sorted, so that a task which lists files to draw an object depends on its seed alone, not on the host's directory order (four RoboTwin tasks did; this one lists no files while it builds its scene). Its scene, frozen instance (same initial-state hash), verifier and pass rule are unchanged.
diff --git a/tasks/robotwin_2_0_beat_block_hammer-privileged/environment/rcb_traj.py b/tasks/robotwin_2_0_beat_block_hammer-privileged/environment/rcb_traj.py
index 583de6869f..33f2a446b9 100644
--- a/tasks/robotwin_2_0_beat_block_hammer-privileged/environment/rcb_traj.py
+++ b/tasks/robotwin_2_0_beat_block_hammer-privileged/environment/rcb_traj.py
@@ -113,6 +113,26 @@ BASE_ARGS: dict[str, Any] = {
}
+class _SortedGlob:
+ """While a scene is built, glob.glob / glob.iglob return their matches sorted. Four RoboTwin tasks (place_a2b_left,
+ place_a2b_right, place_object_scale, place_object_stand: envs/<task>.py load_actors) list an object's model files
+ with glob.glob and then draw one with the seeded np.random.choice; glob returns the file system's directory order,
+ which differs between hosts (and image layers), so one seed built a different object on another host and the frozen
+ instance's hash check failed there (board 0086 claim 5; never rebuilt off the reference workstation). Sorted, the
+ draw depends on the seed alone. No other task of ours lists files while it builds its scene."""
+
+ def __enter__(self):
+ import glob
+ self._glob, self._orig = glob, (glob.glob, glob.iglob)
+ glob.glob = lambda *a, **k: sorted(self._orig[0](*a, **k))
+ glob.iglob = lambda *a, **k: iter(sorted(self._orig[1](*a, **k)))
+ return self
+
+ def __exit__(self, *exc):
+ self._glob.glob, self._glob.iglob = self._orig
+ return False
+
+
def _chdir_root() -> None:
# RoboTwin resolves assets/config paths relative to the repo root and via ROOT_PATH
if os.getcwd() != ROBOTWIN_ROOT and os.path.isdir(ROBOTWIN_ROOT):
@@ -218,7 +238,8 @@ class Session:
if self.enable_render:
sapien.render.set_camera_shader_dir(self.render_shader)
args = _build_args(self.task, self.embodiment, self.seed)
- self.env.setup_demo(**args)
+ with _SortedGlob(): # the object draw depends on the seed alone, not the host's directory order
+ self.env.setup_demo(**args)
if self.enable_render: # setup_scene() switches the shader to ray tracing; put it back for cheap frames
sapien.render.set_camera_shader_dir(self.render_shader)
self._raw_scene = self.env.scene
diff --git a/tasks/robotwin_2_0_beat_block_hammer-privileged/task.toml b/tasks/robotwin_2_0_beat_block_hammer-privileged/task.toml
index cacd9dceb8..11887f3d88 100644
--- a/tasks/robotwin_2_0_beat_block_hammer-privileged/task.toml
+++ b/tasks/robotwin_2_0_beat_block_hammer-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "robotwin" # RoboTwin 2.0 = SAPIEN 3 / PhysX + CuRobo
robot = "aloha-agilex"
family = "beat_block_hammer"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 28831 # scene = environment/instance.json (RoboTwin seed + expected initial-state hash); identity = dataset.toml digest
expert_time_estimate_hours = 0.5
diff --git a/tasks/robotwin_2_0_beat_block_hammer-privileged/tests/rcb_traj.py b/tasks/robotwin_2_0_beat_block_hammer-privileged/tests/rcb_traj.py
index 583de6869f..33f2a446b9 100644
--- a/tasks/robotwin_2_0_beat_block_hammer-privileged/tests/rcb_traj.py
+++ b/tasks/robotwin_2_0_beat_block_hammer-privileged/tests/rcb_traj.py
@@ -113,6 +113,26 @@ BASE_ARGS: dict[str, Any] = {
}
+class _SortedGlob:
+ """While a scene is built, glob.glob / glob.iglob return their matches sorted. Four RoboTwin tasks (place_a2b_left,
+ place_a2b_right, place_object_scale, place_object_stand: envs/<task>.py load_actors) list an object's model files
+ with glob.glob and then draw one with the seeded np.random.choice; glob returns the file system's directory order,
+ which differs between hosts (and image layers), so one seed built a different object on another host and the frozen
+ instance's hash check failed there (board 0086 claim 5; never rebuilt off the reference workstation). Sorted, the
+ draw depends on the seed alone. No other task of ours lists files while it builds its scene."""
+
+ def __enter__(self):
+ import glob
+ self._glob, self._orig = glob, (glob.glob, glob.iglob)
+ glob.glob = lambda *a, **k: sorted(self._orig[0](*a, **k))
+ glob.iglob = lambda *a, **k: iter(sorted(self._orig[1](*a, **k)))
+ return self
+
+ def __exit__(self, *exc):
+ self._glob.glob, self._glob.iglob = self._orig
+ return False
+
+
def _chdir_root() -> None:
# RoboTwin resolves assets/config paths relative to the repo root and via ROOT_PATH
if os.getcwd() != ROBOTWIN_ROOT and os.path.isdir(ROBOTWIN_ROOT):
@@ -218,7 +238,8 @@ class Session:
if self.enable_render:
sapien.render.set_camera_shader_dir(self.render_shader)
args = _build_args(self.task, self.embodiment, self.seed)
- self.env.setup_demo(**args)
+ with _SortedGlob(): # the object draw depends on the seed alone, not the host's directory order
+ self.env.setup_demo(**args)
if self.enable_render: # setup_scene() switches the shader to ray tracing; put it back for cheap frames
sapien.render.set_camera_shader_dir(self.render_shader)
self._raw_scene = self.env.scene
diff --git a/tasks/robotwin_2_0_beat_block_hammer-standard/README.md b/tasks/robotwin_2_0_beat_block_hammer-standard/README.md
index b5fd0a773b..07960feb8e 100644
--- a/tasks/robotwin_2_0_beat_block_hammer-standard/README.md
+++ b/tasks/robotwin_2_0_beat_block_hammer-standard/README.md
@@ -53,3 +53,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
- 2026-10-10 (owner decision, board 0072): the agent's network egress is restricted to the model APIs (`[agent] network_mode = "allowlist"`, the hosts of `scripts/set_agent_egress.py`; since 2026-10-11 these include the package registries, PyPI and the Debian / Ubuntu archives, so that `pip install` and `apt-get install` work); it was open before (Harbor's default), and agents could download upstream sources, reference solutions and robot models from their own shell. The scene, the frozen instance, the tooling and the pass rule are unchanged, so the task keeps its name; results from before this date were obtained with open egress.
- 2026-10-13 (board 0086): the service's copy of the tooling only (`environment/sim/rcb_traj.py`, one line of `environment/sim/server.py`). The tooling now refuses to save or replay a trajectory of more than 200,000 rows, the privileged verifier's cap; a standard episode has no row cap, so the service lifts it for its own session (`S.max_rows = None`). Nothing an agent sees and nothing the service grades changes.
- 2026-10-13 (board 0100): base image `eai-robotwin` 0.1.2, whose scrub of RoboTwin's expert also removes the reference poses that `load_actors` computed only for the expert. The standard-mode agent never reaches the env, so nothing it observes changes; the scene, the instance and the judge are unchanged.
+- 2026-10-14 (board 0113): the tooling only (`environment/sim/rcb_traj.py`): while it builds the scene, `glob.glob` returns its matches sorted, so that a task which lists files to draw an object depends on its seed alone, not on the host's directory order (four RoboTwin tasks did; this one lists no files while it builds its scene). Its scene, frozen instance (same initial-state hash), verifier and pass rule are unchanged.
diff --git a/tasks/robotwin_2_0_beat_block_hammer-standard/environment/sim/rcb_traj.py b/tasks/robotwin_2_0_beat_block_hammer-standard/environment/sim/rcb_traj.py
index 583de6869f..33f2a446b9 100644
--- a/tasks/robotwin_2_0_beat_block_hammer-standard/environment/sim/rcb_traj.py
+++ b/tasks/robotwin_2_0_beat_block_hammer-standard/environment/sim/rcb_traj.py
@@ -113,6 +113,26 @@ BASE_ARGS: dict[str, Any] = {
}
+class _SortedGlob:
+ """While a scene is built, glob.glob / glob.iglob return their matches sorted. Four RoboTwin tasks (place_a2b_left,
+ place_a2b_right, place_object_scale, place_object_stand: envs/<task>.py load_actors) list an object's model files
+ with glob.glob and then draw one with the seeded np.random.choice; glob returns the file system's directory order,
+ which differs between hosts (and image layers), so one seed built a different object on another host and the frozen
+ instance's hash check failed there (board 0086 claim 5; never rebuilt off the reference workstation). Sorted, the
+ draw depends on the seed alone. No other task of ours lists files while it builds its scene."""
+
+ def __enter__(self):
+ import glob
+ self._glob, self._orig = glob, (glob.glob, glob.iglob)
+ glob.glob = lambda *a, **k: sorted(self._orig[0](*a, **k))
+ glob.iglob = lambda *a, **k: iter(sorted(self._orig[1](*a, **k)))
+ return self
+
+ def __exit__(self, *exc):
+ self._glob.glob, self._glob.iglob = self._orig
+ return False
+
+
def _chdir_root() -> None:
# RoboTwin resolves assets/config paths relative to the repo root and via ROOT_PATH
if os.getcwd() != ROBOTWIN_ROOT and os.path.isdir(ROBOTWIN_ROOT):
@@ -218,7 +238,8 @@ class Session:
if self.enable_render:
sapien.render.set_camera_shader_dir(self.render_shader)
args = _build_args(self.task, self.embodiment, self.seed)
- self.env.setup_demo(**args)
+ with _SortedGlob(): # the object draw depends on the seed alone, not the host's directory order
+ self.env.setup_demo(**args)
if self.enable_render: # setup_scene() switches the shader to ray tracing; put it back for cheap frames
sapien.render.set_camera_shader_dir(self.render_shader)
self._raw_scene = self.env.scene
diff --git a/tasks/robotwin_2_0_beat_block_hammer-standard/task.toml b/tasks/robotwin_2_0_beat_block_hammer-standard/task.toml
index e7da0ac608..681ce179a8 100644
--- a/tasks/robotwin_2_0_beat_block_hammer-standard/task.toml
+++ b/tasks/robotwin_2_0_beat_block_hammer-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "robotwin"
robot = "aloha-agilex"
family = "beat_block_hammer"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 28831
diff --git a/tasks/robotwin_2_0_blocks_ranking_rgb-privileged/README.md b/tasks/robotwin_2_0_blocks_ranking_rgb-privileged/README.md
index 4fa600b582..fb76167085 100644
--- a/tasks/robotwin_2_0_blocks_ranking_rgb-privileged/README.md
+++ b/tasks/robotwin_2_0_blocks_ranking_rgb-privileged/README.md
@@ -56,3 +56,4 @@ Place the red block, green block, and blue block in the order of red, green, and
- 2026-10-13 (board 0086): the tooling refuses a trajectory the verifier cannot grade (`environment/rcb_traj.py`, `tests/rcb_traj.py`). `Session.save_trajectory` and `Session.replay` raise `ValueError` for more than 200,000 rows (`rcb_traj.MAX_T`, the verifier's `MAX_T` in `tests/grade.py`: 800 s of physics), and `record_stop` warns on stderr. Before, the tool saved and replayed such a file ("success") and the verifier then graded it `bad_trajectory`; the instruction has stated the cap since 2026-10-10. The verifier, the scene, the instance and the pass rule are unchanged: every trajectory is graded as before.
; `solution/` rebuilds the scene with upstream's `load_actors` to get them back for the expert. None of them is read by `check_success`. The scene, the frozen instance (same initial-state hash), the tooling, the verifier and the pass rule are unchanged, so the task keeps its name; privileged results from before this date were obtained with these values reachable.
- 2026-10-13 (board 0100, owner decision): base image `eai-robotwin` 0.1.2. The image's scrub of RoboTwin's expert kept `load_actors` whole, so values it computes only for the removed expert stayed reachable: `env.block1_target_pose`, `env.block2_target_pose`, `env.block3_target_pose` (poses the expert moves to or places at), and the instruction listed them among the scene objects. They are gone: the env no longer has them, `/opt/RoboTwin/envs/blocks_ranking_rgb.py` no longer shows them, and the scene-object list no longer names them; `solution/` rebuilds the scene with upstream's `load_actors` to get them back for the expert. None of them is read by `check_success`. The scene, the frozen instance (same initial-state hash), the tooling, the verifier and the pass rule are unchanged, so the task keeps its name; privileged results from before this date were obtained with these values reachable.
+- 2026-10-14 (board 0113): the tooling only (`environment/rcb_traj.py`, `tests/rcb_traj.py`): while it builds the scene, `glob.glob` returns its matches sorted, so that a task which lists files to draw an object depends on its seed alone, not on the host's directory order (four RoboTwin tasks did; this one lists no files while it builds its scene). Its scene, frozen instance (same initial-state hash), verifier and pass rule are unchanged.
diff --git a/tasks/robotwin_2_0_blocks_ranking_rgb-privileged/environment/rcb_traj.py b/tasks/robotwin_2_0_blocks_ranking_rgb-privileged/environment/rcb_traj.py
index bb08c9ac7d..5cbb5005fb 100644
--- a/tasks/robotwin_2_0_blocks_ranking_rgb-privileged/environment/rcb_traj.py
+++ b/tasks/robotwin_2_0_blocks_ranking_rgb-privileged/environment/rcb_traj.py
@@ -113,6 +113,26 @@ BASE_ARGS: dict[str, Any] = {
}
+class _SortedGlob:
+ """While a scene is built, glob.glob / glob.iglob return their matches sorted. Four RoboTwin tasks (place_a2b_left,
+ place_a2b_right, place_object_scale, place_object_stand: envs/<task>.py load_actors) list an object's model files
+ with glob.glob and then draw one with the seeded np.random.choice; glob returns the file system's directory order,
+ which differs between hosts (and image layers), so one seed built a different object on another host and the frozen
+ instance's hash check failed there (board 0086 claim 5; never rebuilt off the reference workstation). Sorted, the
+ draw depends on the seed alone. No other task of ours lists files while it builds its scene."""
+
+ def __enter__(self):
+ import glob
+ self._glob, self._orig = glob, (glob.glob, glob.iglob)
+ glob.glob = lambda *a, **k: sorted(self._orig[0](*a, **k))
+ glob.iglob = lambda *a, **k: iter(sorted(self._orig[1](*a, **k)))
+ return self
+
+ def __exit__(self, *exc):
+ self._glob.glob, self._glob.iglob = self._orig
+ return False
+
+
def _chdir_root() -> None:
# RoboTwin resolves assets/config paths relative to the repo root and via ROOT_PATH
if os.getcwd() != ROBOTWIN_ROOT and os.path.isdir(ROBOTWIN_ROOT):
@@ -218,7 +238,8 @@ class Session:
if self.enable_render:
sapien.render.set_camera_shader_dir(self.render_shader)
args = _build_args(self.task, self.embodiment, self.seed)
- self.env.setup_demo(**args)
+ with _SortedGlob(): # the object draw depends on the seed alone, not the host's directory order
+ self.env.setup_demo(**args)
if self.enable_render: # setup_scene() switches the shader to ray tracing; put it back for cheap frames
sapien.render.set_camera_shader_dir(self.render_shader)
self._raw_scene = self.env.scene
diff --git a/tasks/robotwin_2_0_blocks_ranking_rgb-privileged/task.toml b/tasks/robotwin_2_0_blocks_ranking_rgb-privileged/task.toml
index 514b39a14b..99b57b622d 100644
--- a/tasks/robotwin_2_0_blocks_ranking_rgb-privileged/task.toml
+++ b/tasks/robotwin_2_0_blocks_ranking_rgb-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "robotwin" # RoboTwin 2.0 = SAPIEN 3 / PhysX + CuRobo
robot = "aloha-agilex"
family = "blocks_ranking_rgb"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 60417 # scene = environment/instance.json (RoboTwin seed + expected initial-state hash); identity = dataset.toml digest
expert_time_estimate_hours = 0.5
diff --git a/tasks/robotwin_2_0_blocks_ranking_rgb-privileged/tests/rcb_traj.py b/tasks/robotwin_2_0_blocks_ranking_rgb-privileged/tests/rcb_traj.py
index bb08c9ac7d..5cbb5005fb 100644
--- a/tasks/robotwin_2_0_blocks_ranking_rgb-privileged/tests/rcb_traj.py
+++ b/tasks/robotwin_2_0_blocks_ranking_rgb-privileged/tests/rcb_traj.py
@@ -113,6 +113,26 @@ BASE_ARGS: dict[str, Any] = {
}
+class _SortedGlob:
+ """While a scene is built, glob.glob / glob.iglob return their matches sorted. Four RoboTwin tasks (place_a2b_left,
+ place_a2b_right, place_object_scale, place_object_stand: envs/<task>.py load_actors) list an object's model files
+ with glob.glob and then draw one with the seeded np.random.choice; glob returns the file system's directory order,
+ which differs between hosts (and image layers), so one seed built a different object on another host and the frozen
+ instance's hash check failed there (board 0086 claim 5; never rebuilt off the reference workstation). Sorted, the
+ draw depends on the seed alone. No other task of ours lists files while it builds its scene."""
+
+ def __enter__(self):
+ import glob
+ self._glob, self._orig = glob, (glob.glob, glob.iglob)
+ glob.glob = lambda *a, **k: sorted(self._orig[0](*a, **k))
+ glob.iglob = lambda *a, **k: iter(sorted(self._orig[1](*a, **k)))
+ return self
+
+ def __exit__(self, *exc):
+ self._glob.glob, self._glob.iglob = self._orig
+ return False
+
+
def _chdir_root() -> None:
# RoboTwin resolves assets/config paths relative to the repo root and via ROOT_PATH
if os.getcwd() != ROBOTWIN_ROOT and os.path.isdir(ROBOTWIN_ROOT):
@@ -218,7 +238,8 @@ class Session:
if self.enable_render:
sapien.render.set_camera_shader_dir(self.render_shader)
args = _build_args(self.task, self.embodiment, self.seed)
- self.env.setup_demo(**args)
+ with _SortedGlob(): # the object draw depends on the seed alone, not the host's directory order
+ self.env.setup_demo(**args)
if self.enable_render: # setup_scene() switches the shader to ray tracing; put it back for cheap frames
sapien.render.set_camera_shader_dir(self.render_shader)
self._raw_scene = self.env.scene
diff --git a/tasks/robotwin_2_0_blocks_ranking_rgb-standard/README.md b/tasks/robotwin_2_0_blocks_ranking_rgb-standard/README.md
index 8205233228..75c1bea705 100644
--- a/tasks/robotwin_2_0_blocks_ranking_rgb-standard/README.md
+++ b/tasks/robotwin_2_0_blocks_ranking_rgb-standard/README.md
@@ -53,3 +53,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
- 2026-10-10 (owner decision, board 0072): the agent's network egress is restricted to the model APIs (`[agent] network_mode = "allowlist"`, the hosts of `scripts/set_agent_egress.py`; since 2026-10-11 these include the package registries, PyPI and the Debian / Ubuntu archives, so that `pip install` and `apt-get install` work); it was open before (Harbor's default), and agents could download upstream sources, reference solutions and robot models from their own shell. The scene, the frozen instance, the tooling and the pass rule are unchanged, so the task keeps its name; results from before this date were obtained with open egress.
- 2026-10-13 (board 0086): the service's copy of the tooling only (`environment/sim/rcb_traj.py`, one line of `environment/sim/server.py`). The tooling now refuses to save or replay a trajectory of more than 200,000 rows, the privileged verifier's cap; a standard episode has no row cap, so the service lifts it for its own session (`S.max_rows = None`). Nothing an agent sees and nothing the service grades changes.
- 2026-10-13 (board 0100): base image `eai-robotwin` 0.1.2, whose scrub of RoboTwin's expert also removes the reference poses that `load_actors` computed only for the expert (here `block1_target_pose`, `block2_target_pose`, `block3_target_pose`). The standard-mode agent never reaches the env, so nothing it observes changes; the scene, the instance and the judge are unchanged.
+- 2026-10-14 (board 0113): the tooling only (`environment/sim/rcb_traj.py`): while it builds the scene, `glob.glob` returns its matches sorted, so that a task which lists files to draw an object depends on its seed alone, not on the host's directory order (four RoboTwin tasks did; this one lists no files while it builds its scene). Its scene, frozen instance (same initial-state hash), verifier and pass rule are unchanged.
diff --git a/tasks/robotwin_2_0_blocks_ranking_rgb-standard/environment/sim/rcb_traj.py b/tasks/robotwin_2_0_blocks_ranking_rgb-standard/environment/sim/rcb_traj.py
index bb08c9ac7d..5cbb5005fb 100644
--- a/tasks/robotwin_2_0_blocks_ranking_rgb-standard/environment/sim/rcb_traj.py
+++ b/tasks/robotwin_2_0_blocks_ranking_rgb-standard/environment/sim/rcb_traj.py
@@ -113,6 +113,26 @@ BASE_ARGS: dict[str, Any] = {
}
+class _SortedGlob:
+ """While a scene is built, glob.glob / glob.iglob return their matches sorted. Four RoboTwin tasks (place_a2b_left,
+ place_a2b_right, place_object_scale, place_object_stand: envs/<task>.py load_actors) list an object's model files
+ with glob.glob and then draw one with the seeded np.random.choice; glob returns the file system's directory order,
+ which differs between hosts (and image layers), so one seed built a different object on another host and the frozen
+ instance's hash check failed there (board 0086 claim 5; never rebuilt off the reference workstation). Sorted, the
+ draw depends on the seed alone. No other task of ours lists files while it builds its scene."""
+
+ def __enter__(self):
+ import glob
+ self._glob, self._orig = glob, (glob.glob, glob.iglob)
+ glob.glob = lambda *a, **k: sorted(self._orig[0](*a, **k))
+ glob.iglob = lambda *a, **k: iter(sorted(self._orig[1](*a, **k)))
+ return self
+
+ def __exit__(self, *exc):
+ self._glob.glob, self._glob.iglob = self._orig
+ return False
+
+
def _chdir_root() -> None:
# RoboTwin resolves assets/config paths relative to the repo root and via ROOT_PATH
if os.getcwd() != ROBOTWIN_ROOT and os.path.isdir(ROBOTWIN_ROOT):
@@ -218,7 +238,8 @@ class Session:
if self.enable_render:
sapien.render.set_camera_shader_dir(self.render_shader)
args = _build_args(self.task, self.embodiment, self.seed)
- self.env.setup_demo(**args)
+ with _SortedGlob(): # the object draw depends on the seed alone, not the host's directory order
+ self.env.setup_demo(**args)
if self.enable_render: # setup_scene() switches the shader to ray tracing; put it back for cheap frames
sapien.render.set_camera_shader_dir(self.render_shader)
self._raw_scene = self.env.scene
diff --git a/tasks/robotwin_2_0_blocks_ranking_rgb-standard/task.toml b/tasks/robotwin_2_0_blocks_ranking_rgb-standard/task.toml
index d01898904a..cf7aeef7d6 100644
--- a/tasks/robotwin_2_0_blocks_ranking_rgb-standard/task.toml
+++ b/tasks/robotwin_2_0_blocks_ranking_rgb-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "robotwin"
robot = "aloha-agilex"
family = "blocks_ranking_rgb"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 60417
diff --git a/tasks/robotwin_2_0_blocks_ranking_size-privileged/README.md b/tasks/robotwin_2_0_blocks_ranking_size-privileged/README.md
index 50c97676c4..d9d6900f4d 100644
--- a/tasks/robotwin_2_0_blocks_ranking_size-privileged/README.md
+++ b/tasks/robotwin_2_0_blocks_ranking_size-privileged/README.md
@@ -56,3 +56,4 @@ There are three blocks on the table, the color of the blocks is random, move the
- 2026-10-13 (board 0086): the tooling refuses a trajectory the verifier cannot grade (`environment/rcb_traj.py`, `tests/rcb_traj.py`). `Session.save_trajectory` and `Session.replay` raise `ValueError` for more than 200,000 rows (`rcb_traj.MAX_T`, the verifier's `MAX_T` in `tests/grade.py`: 800 s of physics), and `record_stop` warns on stderr. Before, the tool saved and replayed such a file ("success") and the verifier then graded it `bad_trajectory`; the instruction has stated the cap since 2026-10-10. The verifier, the scene, the instance and the pass rule are unchanged: every trajectory is graded as before.
; `solution/` rebuilds the scene with upstream's `load_actors` to get them back for the expert. None of them is read by `check_success`. The scene, the frozen instance (same initial-state hash), the tooling, the verifier and the pass rule are unchanged, so the task keeps its name; privileged results from before this date were obtained with these values reachable.
- 2026-10-13 (board 0100, owner decision): base image `eai-robotwin` 0.1.2. The image's scrub of RoboTwin's expert kept `load_actors` whole, so values it computes only for the removed expert stayed reachable: `env.block1_target_pose`, `env.block2_target_pose`, `env.block3_target_pose` (poses the expert moves to or places at), and the instruction listed them among the scene objects. They are gone: the env no longer has them, `/opt/RoboTwin/envs/blocks_ranking_size.py` no longer shows them, and the scene-object list no longer names them; `solution/` rebuilds the scene with upstream's `load_actors` to get them back for the expert. None of them is read by `check_success`. The scene, the frozen instance (same initial-state hash), the tooling, the verifier and the pass rule are unchanged, so the task keeps its name; privileged results from before this date were obtained with these values reachable.
+- 2026-10-14 (board 0113): the tooling only (`environment/rcb_traj.py`, `tests/rcb_traj.py`): while it builds the scene, `glob.glob` returns its matches sorted, so that a task which lists files to draw an object depends on its seed alone, not on the host's directory order (four RoboTwin tasks did; this one lists no files while it builds its scene). Its scene, frozen instance (same initial-state hash), verifier and pass rule are unchanged.
diff --git a/tasks/robotwin_2_0_blocks_ranking_size-privileged/environment/rcb_traj.py b/tasks/robotwin_2_0_blocks_ranking_size-privileged/environment/rcb_traj.py
index 1e11195a6a..e792a3c9d9 100644
--- a/tasks/robotwin_2_0_blocks_ranking_size-privileged/environment/rcb_traj.py
+++ b/tasks/robotwin_2_0_blocks_ranking_size-privileged/environment/rcb_traj.py
@@ -113,6 +113,26 @@ BASE_ARGS: dict[str, Any] = {
}
+class _SortedGlob:
+ """While a scene is built, glob.glob / glob.iglob return their matches sorted. Four RoboTwin tasks (place_a2b_left,
+ place_a2b_right, place_object_scale, place_object_stand: envs/<task>.py load_actors) list an object's model files
+ with glob.glob and then draw one with the seeded np.random.choice; glob returns the file system's directory order,
+ which differs between hosts (and image layers), so one seed built a different object on another host and the frozen
+ instance's hash check failed there (board 0086 claim 5; never rebuilt off the reference workstation). Sorted, the
+ draw depends on the seed alone. No other task of ours lists files while it builds its scene."""
+
+ def __enter__(self):
+ import glob
+ self._glob, self._orig = glob, (glob.glob, glob.iglob)
+ glob.glob = lambda *a, **k: sorted(self._orig[0](*a, **k))
+ glob.iglob = lambda *a, **k: iter(sorted(self._orig[1](*a, **k)))
+ return self
+
+ def __exit__(self, *exc):
+ self._glob.glob, self._glob.iglob = self._orig
+ return False
+
+
def _chdir_root() -> None:
# RoboTwin resolves assets/config paths relative to the repo root and via ROOT_PATH
if os.getcwd() != ROBOTWIN_ROOT and os.path.isdir(ROBOTWIN_ROOT):
@@ -218,7 +238,8 @@ class Session:
if self.enable_render:
sapien.render.set_camera_shader_dir(self.render_shader)
args = _build_args(self.task, self.embodiment, self.seed)
- self.env.setup_demo(**args)
+ with _SortedGlob(): # the object draw depends on the seed alone, not the host's directory order
+ self.env.setup_demo(**args)
if self.enable_render: # setup_scene() switches the shader to ray tracing; put it back for cheap frames
sapien.render.set_camera_shader_dir(self.render_shader)
self._raw_scene = self.env.scene
diff --git a/tasks/robotwin_2_0_blocks_ranking_size-privileged/task.toml b/tasks/robotwin_2_0_blocks_ranking_size-privileged/task.toml
index 9b8813c342..2a3c5ff9b9 100644
--- a/tasks/robotwin_2_0_blocks_ranking_size-privileged/task.toml
+++ b/tasks/robotwin_2_0_blocks_ranking_size-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "robotwin" # RoboTwin 2.0 = SAPIEN 3 / PhysX + CuRobo
robot = "aloha-agilex"
family = "blocks_ranking_size"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 60417 # scene = environment/instance.json (RoboTwin seed + expected initial-state hash); identity = dataset.toml digest
expert_time_estimate_hours = 0.5
diff --git a/tasks/robotwin_2_0_blocks_ranking_size-privileged/tests/rcb_traj.py b/tasks/robotwin_2_0_blocks_ranking_size-privileged/tests/rcb_traj.py
index 1e11195a6a..e792a3c9d9 100644
--- a/tasks/robotwin_2_0_blocks_ranking_size-privileged/tests/rcb_traj.py
+++ b/tasks/robotwin_2_0_blocks_ranking_size-privileged/tests/rcb_traj.py
@@ -113,6 +113,26 @@ BASE_ARGS: dict[str, Any] = {
}
+class _SortedGlob:
+ """While a scene is built, glob.glob / glob.iglob return their matches sorted. Four RoboTwin tasks (place_a2b_left,
+ place_a2b_right, place_object_scale, place_object_stand: envs/<task>.py load_actors) list an object's model files
+ with glob.glob and then draw one with the seeded np.random.choice; glob returns the file system's directory order,
+ which differs between hosts (and image layers), so one seed built a different object on another host and the frozen
+ instance's hash check failed there (board 0086 claim 5; never rebuilt off the reference workstation). Sorted, the
+ draw depends on the seed alone. No other task of ours lists files while it builds its scene."""
+
+ def __enter__(self):
+ import glob
+ self._glob, self._orig = glob, (glob.glob, glob.iglob)
+ glob.glob = lambda *a, **k: sorted(self._orig[0](*a, **k))
+ glob.iglob = lambda *a, **k: iter(sorted(self._orig[1](*a, **k)))
+ return self
+
+ def __exit__(self, *exc):
+ self._glob.glob, self._glob.iglob = self._orig
+ return False
+
+
def _chdir_root() -> None:
# RoboTwin resolves assets/config paths relative to the repo root and via ROOT_PATH
if os.getcwd() != ROBOTWIN_ROOT and os.path.isdir(ROBOTWIN_ROOT):
@@ -218,7 +238,8 @@ class Session:
if self.enable_render:
sapien.render.set_camera_shader_dir(self.render_shader)
args = _build_args(self.task, self.embodiment, self.seed)
- self.env.setup_demo(**args)
+ with _SortedGlob(): # the object draw depends on the seed alone, not the host's directory order
+ self.env.setup_demo(**args)
if self.enable_render: # setup_scene() switches the shader to ray tracing; put it back for cheap frames
sapien.render.set_camera_shader_dir(self.render_shader)
self._raw_scene = self.env.scene
diff --git a/tasks/robotwin_2_0_blocks_ranking_size-standard/README.md b/tasks/robotwin_2_0_blocks_ranking_size-standard/README.md
index db07b813be..978b33addf 100644
--- a/tasks/robotwin_2_0_blocks_ranking_size-standard/README.md
+++ b/tasks/robotwin_2_0_blocks_ranking_size-standard/README.md
@@ -53,3 +53,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
- 2026-10-10 (owner decision, board 0072): the agent's network egress is restricted to the model APIs (`[agent] network_mode = "allowlist"`, the hosts of `scripts/set_agent_egress.py`; since 2026-10-11 these include the package registries, PyPI and the Debian / Ubuntu archives, so that `pip install` and `apt-get install` work); it was open before (Harbor's default), and agents could download upstream sources, reference solutions and robot models from their own shell. The scene, the frozen instance, the tooling and the pass rule are unchanged, so the task keeps its name; results from before this date were obtained with open egress.
- 2026-10-13 (board 0086): the service's copy of the tooling only (`environment/sim/rcb_traj.py`, one line of `environment/sim/server.py`). The tooling now refuses to save or replay a trajectory of more than 200,000 rows, the privileged verifier's cap; a standard episode has no row cap, so the service lifts it for its own session (`S.max_rows = None`). Nothing an agent sees and nothing the service grades changes.
- 2026-10-13 (board 0100): base image `eai-robotwin` 0.1.2, whose scrub of RoboTwin's expert also removes the reference poses that `load_actors` computed only for the expert (here `block1_target_pose`, `block2_target_pose`, `block3_target_pose`). The standard-mode agent never reaches the env, so nothing it observes changes; the scene, the instance and the judge are unchanged.
+- 2026-10-14 (board 0113): the tooling only (`environment/sim/rcb_traj.py`): while it builds the scene, `glob.glob` returns its matches sorted, so that a task which lists files to draw an object depends on its seed alone, not on the host's directory order (four RoboTwin tasks did; this one lists no files while it builds its scene). Its scene, frozen instance (same initial-state hash), verifier and pass rule are unchanged.
diff --git a/tasks/robotwin_2_0_blocks_ranking_size-standard/environment/sim/rcb_traj.py b/tasks/robotwin_2_0_blocks_ranking_size-standard/environment/sim/rcb_traj.py
index 1e11195a6a..e792a3c9d9 100644
--- a/tasks/robotwin_2_0_blocks_ranking_size-standard/environment/sim/rcb_traj.py
+++ b/tasks/robotwin_2_0_blocks_ranking_size-standard/environment/sim/rcb_traj.py
@@ -113,6 +113,26 @@ BASE_ARGS: dict[str, Any] = {
}
+class _SortedGlob:
+ """While a scene is built, glob.glob / glob.iglob return their matches sorted. Four RoboTwin tasks (place_a2b_left,
+ place_a2b_right, place_object_scale, place_object_stand: envs/<task>.py load_actors) list an object's model files
+ with glob.glob and then draw one with the seeded np.random.choice; glob returns the file system's directory order,
+ which differs between hosts (and image layers), so one seed built a different object on another host and the frozen
+ instance's hash check failed there (board 0086 claim 5; never rebuilt off the reference workstation). Sorted, the
+ draw depends on the seed alone. No other task of ours lists files while it builds its scene."""
+
+ def __enter__(self):
+ import glob
+ self._glob, self._orig = glob, (glob.glob, glob.iglob)
+ glob.glob = lambda *a, **k: sorted(self._orig[0](*a, **k))
+ glob.iglob = lambda *a, **k: iter(sorted(self._orig[1](*a, **k)))
+ return self
+
+ def __exit__(self, *exc):
+ self._glob.glob, self._glob.iglob = self._orig
+ return False
+
+
def _chdir_root() -> None:
# RoboTwin resolves assets/config paths relative to the repo root and via ROOT_PATH
if os.getcwd() != ROBOTWIN_ROOT and os.path.isdir(ROBOTWIN_ROOT):
@@ -218,7 +238,8 @@ class Session:
if self.enable_render:
sapien.render.set_camera_shader_dir(self.render_shader)
args = _build_args(self.task, self.embodiment, self.seed)
- self.env.setup_demo(**args)
+ with _SortedGlob(): # the object draw depends on the seed alone, not the host's directory order
+ self.env.setup_demo(**args)
if self.enable_render: # setup_scene() switches the shader to ray tracing; put it back for cheap frames
sapien.render.set_camera_shader_dir(self.render_shader)
self._raw_scene = self.env.scene
diff --git a/tasks/robotwin_2_0_blocks_ranking_size-standard/task.toml b/tasks/robotwin_2_0_blocks_ranking_size-standard/task.toml
index 733d25185c..eaf2cfdcee 100644
--- a/tasks/robotwin_2_0_blocks_ranking_size-standard/task.toml
+++ b/tasks/robotwin_2_0_blocks_ranking_size-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "robotwin"
robot = "aloha-agilex"
family = "blocks_ranking_size"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 60417
diff --git a/tasks/robotwin_2_0_dump_bin_bigbin-privileged/README.md b/tasks/robotwin_2_0_dump_bin_bigbin-privileged/README.md
index 6bad67c19d..b3884e992f 100644
--- a/tasks/robotwin_2_0_dump_bin_bigbin-privileged/README.md
+++ b/tasks/robotwin_2_0_dump_bin_bigbin-privileged/README.md
@@ -55,3 +55,4 @@ Grab the small bin and pour the balls into the big bin.
- 2026-10-10 (owner decision, board 0072): the agent's network egress is restricted to the model APIs (`[agent] network_mode = "allowlist"`, the hosts of `scripts/set_agent_egress.py`; since 2026-10-11 these include the package registries, PyPI and the Debian / Ubuntu archives, so that `pip install` and `apt-get install` work); it was open before (Harbor's default), and agents could download upstream sources, reference solutions and robot models from their own shell. The scene, the frozen instance, the tooling and the pass rule are unchanged, so the task keeps its name; results from before this date were obtained with open egress.
- 2026-10-13 (board 0086): the tooling refuses a trajectory the verifier cannot grade (`environment/rcb_traj.py`, `tests/rcb_traj.py`). `Session.save_trajectory` and `Session.replay` raise `ValueError` for more than 200,000 rows (`rcb_traj.MAX_T`, the verifier's `MAX_T` in `tests/grade.py`: 800 s of physics), and `record_stop` warns on stderr. Before, the tool saved and replayed such a file ("success") and the verifier then graded it `bad_trajectory`; the instruction has stated the cap since 2026-10-10. The verifier, the scene, the instance and the pass rule are unchanged: every trajectory is graded as before.
- 2026-10-13 (board 0100, owner decision): base image `eai-robotwin` 0.1.2. The image's scrub of RoboTwin's expert kept `load_actors` whole, so values it computes only for the removed expert stayed reachable: `env.middle_pose`, `env.pour_actions` (the expert's whole pouring motion as a ready-made action list, and the pose it regrasps the bin at), and the instruction listed them among the scene objects. They are gone: the env no longer has them, `/opt/RoboTwin/envs/dump_bin_bigbin.py` no longer shows them, and the scene-object list no longer names them; `solution/` rebuilds the scene with upstream's `load_actors` to get them back for the expert. None of them is read by `check_success`. The scene, the frozen instance (same initial-state hash), the tooling, the verifier and the pass rule are unchanged, so the task keeps its name; privileged results from before this date were obtained with these values reachable.
+- 2026-10-14 (board 0113): the tooling only (`environment/rcb_traj.py`, `tests/rcb_traj.py`): while it builds the scene, `glob.glob` returns its matches sorted, so that a task which lists files to draw an object depends on its seed alone, not on the host's directory order (four RoboTwin tasks did; this one lists no files while it builds its scene). Its scene, frozen instance (same initial-state hash), verifier and pass rule are unchanged.
diff --git a/tasks/robotwin_2_0_dump_bin_bigbin-privileged/environment/rcb_traj.py b/tasks/robotwin_2_0_dump_bin_bigbin-privileged/environment/rcb_traj.py
index 185d5c1ffb..7e828ae538 100644
--- a/tasks/robotwin_2_0_dump_bin_bigbin-privileged/environment/rcb_traj.py
+++ b/tasks/robotwin_2_0_dump_bin_bigbin-privileged/environment/rcb_traj.py
@@ -113,6 +113,26 @@ BASE_ARGS: dict[str, Any] = {
}
+class _SortedGlob:
+ """While a scene is built, glob.glob / glob.iglob return their matches sorted. Four RoboTwin tasks (place_a2b_left,
+ place_a2b_right, place_object_scale, place_object_stand: envs/<task>.py load_actors) list an object's model files
+ with glob.glob and then draw one with the seeded np.random.choice; glob returns the file system's directory order,
+ which differs between hosts (and image layers), so one seed built a different object on another host and the frozen
+ instance's hash check failed there (board 0086 claim 5; never rebuilt off the reference workstation). Sorted, the
+ draw depends on the seed alone. No other task of ours lists files while it builds its scene."""
+
+ def __enter__(self):
+ import glob
+ self._glob, self._orig = glob, (glob.glob, glob.iglob)
+ glob.glob = lambda *a, **k: sorted(self._orig[0](*a, **k))
+ glob.iglob = lambda *a, **k: iter(sorted(self._orig[1](*a, **k)))
+ return self
+
+ def __exit__(self, *exc):
+ self._glob.glob, self._glob.iglob = self._orig
+ return False
+
+
def _chdir_root() -> None:
# RoboTwin resolves assets/config paths relative to the repo root and via ROOT_PATH
if os.getcwd() != ROBOTWIN_ROOT and os.path.isdir(ROBOTWIN_ROOT):
@@ -218,7 +238,8 @@ class Session:
if self.enable_render:
sapien.render.set_camera_shader_dir(self.render_shader)
args = _build_args(self.task, self.embodiment, self.seed)
- self.env.setup_demo(**args)
+ with _SortedGlob(): # the object draw depends on the seed alone, not the host's directory order
+ self.env.setup_demo(**args)
if self.enable_render: # setup_scene() switches the shader to ray tracing; put it back for cheap frames
sapien.render.set_camera_shader_dir(self.render_shader)
self._raw_scene = self.env.scene
diff --git a/tasks/robotwin_2_0_dump_bin_bigbin-privileged/task.toml b/tasks/robotwin_2_0_dump_bin_bigbin-privileged/task.toml
index d49a90859d..b00d99f058 100644
--- a/tasks/robotwin_2_0_dump_bin_bigbin-privileged/task.toml
+++ b/tasks/robotwin_2_0_dump_bin_bigbin-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "robotwin" # RoboTwin 2.0 = SAPIEN 3 / PhysX + CuRobo
robot = "aloha-agilex"
family = "dump_bin_bigbin"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 60417 # scene = environment/instance.json (RoboTwin seed + expected initial-state hash); identity = dataset.toml digest
expert_time_estimate_hours = 0.5
diff --git a/tasks/robotwin_2_0_dump_bin_bigbin-privileged/tests/rcb_traj.py b/tasks/robotwin_2_0_dump_bin_bigbin-privileged/tests/rcb_traj.py
index 185d5c1ffb..7e828ae538 100644
--- a/tasks/robotwin_2_0_dump_bin_bigbin-privileged/tests/rcb_traj.py
+++ b/tasks/robotwin_2_0_dump_bin_bigbin-privileged/tests/rcb_traj.py
@@ -113,6 +113,26 @@ BASE_ARGS: dict[str, Any] = {
}
+class _SortedGlob:
+ """While a scene is built, glob.glob / glob.iglob return their matches sorted. Four RoboTwin tasks (place_a2b_left,
+ place_a2b_right, place_object_scale, place_object_stand: envs/<task>.py load_actors) list an object's model files
+ with glob.glob and then draw one with the seeded np.random.choice; glob returns the file system's directory order,
+ which differs between hosts (and image layers), so one seed built a different object on another host and the frozen
+ instance's hash check failed there (board 0086 claim 5; never rebuilt off the reference workstation). Sorted, the
+ draw depends on the seed alone. No other task of ours lists files while it builds its scene."""
+
+ def __enter__(self):
+ import glob
+ self._glob, self._orig = glob, (glob.glob, glob.iglob)
+ glob.glob = lambda *a, **k: sorted(self._orig[0](*a, **k))
+ glob.iglob = lambda *a, **k: iter(sorted(self._orig[1](*a, **k)))
+ return self
+
+ def __exit__(self, *exc):
+ self._glob.glob, self._glob.iglob = self._orig
+ return False
+
+
def _chdir_root() -> None:
# RoboTwin resolves assets/config paths relative to the repo root and via ROOT_PATH
if os.getcwd() != ROBOTWIN_ROOT and os.path.isdir(ROBOTWIN_ROOT):
@@ -218,7 +238,8 @@ class Session:
if self.enable_render:
sapien.render.set_camera_shader_dir(self.render_shader)
args = _build_args(self.task, self.embodiment, self.seed)
- self.env.setup_demo(**args)
+ with _SortedGlob(): # the object draw depends on the seed alone, not the host's directory order
+ self.env.setup_demo(**args)
if self.enable_render: # setup_scene() switches the shader to ray tracing; put it back for cheap frames
sapien.render.set_camera_shader_dir(self.render_shader)
self._raw_scene = self.env.scene
diff --git a/tasks/robotwin_2_0_dump_bin_bigbin-standard/README.md b/tasks/robotwin_2_0_dump_bin_bigbin-standard/README.md
index 5beb3d1e99..f942cc4b51 100644
--- a/tasks/robotwin_2_0_dump_bin_bigbin-standard/README.md
+++ b/tasks/robotwin_2_0_dump_bin_bigbin-standard/README.md
@@ -53,3 +53,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
- 2026-10-10 (owner decision, board 0072): the agent's network egress is restricted to the model APIs (`[agent] network_mode = "allowlist"`, the hosts of `scripts/set_agent_egress.py`; since 2026-10-11 these include the package registries, PyPI and the Debian / Ubuntu archives, so that `pip install` and `apt-get install` work); it was open before (Harbor's default), and agents could download upstream sources, reference solutions and robot models from their own shell. The scene, the frozen instance, the tooling and the pass rule are unchanged, so the task keeps its name; results from before this date were obtained with open egress.
- 2026-10-13 (board 0086): the service's copy of the tooling only (`environment/sim/rcb_traj.py`, one line of `environment/sim/server.py`). The tooling now refuses to save or replay a trajectory of more than 200,000 rows, the privileged verifier's cap; a standard episode has no row cap, so the service lifts it for its own session (`S.max_rows = None`). Nothing an agent sees and nothing the service grades changes.
- 2026-10-13 (board 0100): base image `eai-robotwin` 0.1.2, whose scrub of RoboTwin's expert also removes the reference poses that `load_actors` computed only for the expert (here `middle_pose`, `pour_actions`). The standard-mode agent never reaches the env, so nothing it observes changes; the scene, the instance and the judge are unchanged.
+- 2026-10-14 (board 0113): the tooling only (`environment/sim/rcb_traj.py`): while it builds the scene, `glob.glob` returns its matches sorted, so that a task which lists files to draw an object depends on its seed alone, not on the host's directory order (four RoboTwin tasks did; this one lists no files while it builds its scene). Its scene, frozen instance (same initial-state hash), verifier and pass rule are unchanged.
diff --git a/tasks/robotwin_2_0_dump_bin_bigbin-standard/environment/sim/rcb_traj.py b/tasks/robotwin_2_0_dump_bin_bigbin-standard/environment/sim/rcb_traj.py
index 185d5c1ffb..7e828ae538 100644
--- a/tasks/robotwin_2_0_dump_bin_bigbin-standard/environment/sim/rcb_traj.py
+++ b/tasks/robotwin_2_0_dump_bin_bigbin-standard/environment/sim/rcb_traj.py
@@ -113,6 +113,26 @@ BASE_ARGS: dict[str, Any] = {
}
+class _SortedGlob:
+ """While a scene is built, glob.glob / glob.iglob return their matches sorted. Four RoboTwin tasks (place_a2b_left,
+ place_a2b_right, place_object_scale, place_object_stand: envs/<task>.py load_actors) list an object's model files
+ with glob.glob and then draw one with the seeded np.random.choice; glob returns the file system's directory order,
+ which differs between hosts (and image layers), so one seed built a different object on another host and the frozen
+ instance's hash check failed there (board 0086 claim 5; never rebuilt off the reference workstation). Sorted, the
+ draw depends on the seed alone. No other task of ours lists files while it builds its scene."""
+
+ def __enter__(self):
+ import glob
+ self._glob, self._orig = glob, (glob.glob, glob.iglob)
+ glob.glob = lambda *a, **k: sorted(self._orig[0](*a, **k))
+ glob.iglob = lambda *a, **k: iter(sorted(self._orig[1](*a, **k)))
+ return self
+
+ def __exit__(self, *exc):
+ self._glob.glob, self._glob.iglob = self._orig
+ return False
+
+
def _chdir_root() -> None:
# RoboTwin resolves assets/config paths relative to the repo root and via ROOT_PATH
if os.getcwd() != ROBOTWIN_ROOT and os.path.isdir(ROBOTWIN_ROOT):
@@ -218,7 +238,8 @@ class Session:
if self.enable_render:
sapien.render.set_camera_shader_dir(self.render_shader)
args = _build_args(self.task, self.embodiment, self.seed)
- self.env.setup_demo(**args)
+ with _SortedGlob(): # the object draw depends on the seed alone, not the host's directory order
+ self.env.setup_demo(**args)
if self.enable_render: # setup_scene() switches the shader to ray tracing; put it back for cheap frames
sapien.render.set_camera_shader_dir(self.render_shader)
self._raw_scene = self.env.scene
diff --git a/tasks/robotwin_2_0_dump_bin_bigbin-standard/task.toml b/tasks/robotwin_2_0_dump_bin_bigbin-standard/task.toml
index 32e88ca692..5207b8dc5b 100644
--- a/tasks/robotwin_2_0_dump_bin_bigbin-standard/task.toml
+++ b/tasks/robotwin_2_0_dump_bin_bigbin-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "robotwin"
robot = "aloha-agilex"
family = "dump_bin_bigbin"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 60417
diff --git a/tasks/robotwin_2_0_grab_roller-privileged/README.md b/tasks/robotwin_2_0_grab_roller-privileged/README.md
index 11b05b4567..70d643610a 100644
--- a/tasks/robotwin_2_0_grab_roller-privileged/README.md
+++ b/tasks/robotwin_2_0_grab_roller-privileged/README.md
@@ -55,3 +55,4 @@ Use both arms to grab the roller on the table.
- 2026-10-10 (owner decision, board 0072): the agent's network egress is restricted to the model APIs (`[agent] network_mode = "allowlist"`, the hosts of `scripts/set_agent_egress.py`; since 2026-10-11 these include the package registries, PyPI and the Debian / Ubuntu archives, so that `pip install` and `apt-get install` work); it was open before (Harbor's default), and agents could download upstream sources, reference solutions and robot models from their own shell. The scene, the frozen instance, the tooling and the pass rule are unchanged, so the task keeps its name; results from before this date were obtained with open egress.
- 2026-10-13 (board 0086): the tooling refuses a trajectory the verifier cannot grade (`environment/rcb_traj.py`, `tests/rcb_traj.py`). `Session.save_trajectory` and `Session.replay` raise `ValueError` for more than 200,000 rows (`rcb_traj.MAX_T`, the verifier's `MAX_T` in `tests/grade.py`: 800 s of physics), and `record_stop` warns on stderr. Before, the tool saved and replayed such a file ("success") and the verifier then graded it `bad_trajectory`; the instruction has stated the cap since 2026-10-10. The verifier, the scene, the instance and the pass rule are unchanged: every trajectory is graded as before.
- 2026-10-13 (board 0100): base image `eai-robotwin` 0.1.2. Its scrub of RoboTwin's expert also removes the reference poses that 18 tasks' `load_actors` computed only for the expert, and the expert's comments above the scrubbed method bodies (`images/robotwin/scrub_oracles.py`). This task's `load_actors` set no such value: its scene, instance, tooling, verifier and pass rule are unchanged.
+- 2026-10-14 (board 0113): the tooling only (`environment/rcb_traj.py`, `tests/rcb_traj.py`): while it builds the scene, `glob.glob` returns its matches sorted, so that a task which lists files to draw an object depends on its seed alone, not on the host's directory order (four RoboTwin tasks did; this one lists no files while it builds its scene). Its scene, frozen instance (same initial-state hash), verifier and pass rule are unchanged.
diff --git a/tasks/robotwin_2_0_grab_roller-privileged/environment/rcb_traj.py b/tasks/robotwin_2_0_grab_roller-privileged/environment/rcb_traj.py
index 69b38b81e0..c2de9c76df 100644
--- a/tasks/robotwin_2_0_grab_roller-privileged/environment/rcb_traj.py
+++ b/tasks/robotwin_2_0_grab_roller-privileged/environment/rcb_traj.py
@@ -113,6 +113,26 @@ BASE_ARGS: dict[str, Any] = {
}
+class _SortedGlob:
+ """While a scene is built, glob.glob / glob.iglob return their matches sorted. Four RoboTwin tasks (place_a2b_left,
+ place_a2b_right, place_object_scale, place_object_stand: envs/<task>.py load_actors) list an object's model files
+ with glob.glob and then draw one with the seeded np.random.choice; glob returns the file system's directory order,
+ which differs between hosts (and image layers), so one seed built a different object on another host and the frozen
+ instance's hash check failed there (board 0086 claim 5; never rebuilt off the reference workstation). Sorted, the
+ draw depends on the seed alone. No other task of ours lists files while it builds its scene."""
+
+ def __enter__(self):
+ import glob
+ self._glob, self._orig = glob, (glob.glob, glob.iglob)
+ glob.glob = lambda *a, **k: sorted(self._orig[0](*a, **k))
+ glob.iglob = lambda *a, **k: iter(sorted(self._orig[1](*a, **k)))
+ return self
+
+ def __exit__(self, *exc):
+ self._glob.glob, self._glob.iglob = self._orig
+ return False
+
+
def _chdir_root() -> None:
# RoboTwin resolves assets/config paths relative to the repo root and via ROOT_PATH
if os.getcwd() != ROBOTWIN_ROOT and os.path.isdir(ROBOTWIN_ROOT):
@@ -218,7 +238,8 @@ class Session:
if self.enable_render:
sapien.render.set_camera_shader_dir(self.render_shader)
args = _build_args(self.task, self.embodiment, self.seed)
- self.env.setup_demo(**args)
+ with _SortedGlob(): # the object draw depends on the seed alone, not the host's directory order
+ self.env.setup_demo(**args)
if self.enable_render: # setup_scene() switches the shader to ray tracing; put it back for cheap frames
sapien.render.set_camera_shader_dir(self.render_shader)
self._raw_scene = self.env.scene
diff --git a/tasks/robotwin_2_0_grab_roller-privileged/task.toml b/tasks/robotwin_2_0_grab_roller-privileged/task.toml
index e4cde7be1a..5c164ed825 100644
--- a/tasks/robotwin_2_0_grab_roller-privileged/task.toml
+++ b/tasks/robotwin_2_0_grab_roller-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "robotwin" # RoboTwin 2.0 = SAPIEN 3 / PhysX + CuRobo
robot = "aloha-agilex"
family = "grab_roller"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 60417 # scene = environment/instance.json (RoboTwin seed + expected initial-state hash); identity = dataset.toml digest
expert_time_estimate_hours = 0.5
diff --git a/tasks/robotwin_2_0_grab_roller-privileged/tests/rcb_traj.py b/tasks/robotwin_2_0_grab_roller-privileged/tests/rcb_traj.py
index 69b38b81e0..c2de9c76df 100644
--- a/tasks/robotwin_2_0_grab_roller-privileged/tests/rcb_traj.py
+++ b/tasks/robotwin_2_0_grab_roller-privileged/tests/rcb_traj.py
@@ -113,6 +113,26 @@ BASE_ARGS: dict[str, Any] = {
}
+class _SortedGlob:
+ """While a scene is built, glob.glob / glob.iglob return their matches sorted. Four RoboTwin tasks (place_a2b_left,
+ place_a2b_right, place_object_scale, place_object_stand: envs/<task>.py load_actors) list an object's model files
+ with glob.glob and then draw one with the seeded np.random.choice; glob returns the file system's directory order,
+ which differs between hosts (and image layers), so one seed built a different object on another host and the frozen
+ instance's hash check failed there (board 0086 claim 5; never rebuilt off the reference workstation). Sorted, the
+ draw depends on the seed alone. No other task of ours lists files while it builds its scene."""
+
+ def __enter__(self):
+ import glob
+ self._glob, self._orig = glob, (glob.glob, glob.iglob)
+ glob.glob = lambda *a, **k: sorted(self._orig[0](*a, **k))
+ glob.iglob = lambda *a, **k: iter(sorted(self._orig[1](*a, **k)))
+ return self
+
+ def __exit__(self, *exc):
+ self._glob.glob, self._glob.iglob = self._orig
+ return False
+
+
def _chdir_root() -> None:
# RoboTwin resolves assets/config paths relative to the repo root and via ROOT_PATH
if os.getcwd() != ROBOTWIN_ROOT and os.path.isdir(ROBOTWIN_ROOT):
@@ -218,7 +238,8 @@ class Session:
if self.enable_render:
sapien.render.set_camera_shader_dir(self.render_shader)
args = _build_args(self.task, self.embodiment, self.seed)
- self.env.setup_demo(**args)
+ with _SortedGlob(): # the object draw depends on the seed alone, not the host's directory order
+ self.env.setup_demo(**args)
if self.enable_render: # setup_scene() switches the shader to ray tracing; put it back for cheap frames
sapien.render.set_camera_shader_dir(self.render_shader)
self._raw_scene = self.env.scene
diff --git a/tasks/robotwin_2_0_grab_roller-standard/README.md b/tasks/robotwin_2_0_grab_roller-standard/README.md
index cfffd1a5ab..38502d0db9 100644
--- a/tasks/robotwin_2_0_grab_roller-standard/README.md
+++ b/tasks/robotwin_2_0_grab_roller-standard/README.md
@@ -53,3 +53,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
- 2026-10-10 (owner decision, board 0072): the agent's network egress is restricted to the model APIs (`[agent] network_mode = "allowlist"`, the hosts of `scripts/set_agent_egress.py`; since 2026-10-11 these include the package registries, PyPI and the Debian / Ubuntu archives, so that `pip install` and `apt-get install` work); it was open before (Harbor's default), and agents could download upstream sources, reference solutions and robot models from their own shell. The scene, the frozen instance, the tooling and the pass rule are unchanged, so the task keeps its name; results from before this date were obtained with open egress.
- 2026-10-13 (board 0086): the service's copy of the tooling only (`environment/sim/rcb_traj.py`, one line of `environment/sim/server.py`). The tooling now refuses to save or replay a trajectory of more than 200,000 rows, the privileged verifier's cap; a standard episode has no row cap, so the service lifts it for its own session (`S.max_rows = None`). Nothing an agent sees and nothing the service grades changes.
- 2026-10-13 (board 0100): base image `eai-robotwin` 0.1.2, whose scrub of RoboTwin's expert also removes the reference poses that `load_actors` computed only for the expert. The standard-mode agent never reaches the env, so nothing it observes changes; the scene, the instance and the judge are unchanged.
+- 2026-10-14 (board 0113): the tooling only (`environment/sim/rcb_traj.py`): while it builds the scene, `glob.glob` returns its matches sorted, so that a task which lists files to draw an object depends on its seed alone, not on the host's directory order (four RoboTwin tasks did; this one lists no files while it builds its scene). Its scene, frozen instance (same initial-state hash), verifier and pass rule are unchanged.
diff --git a/tasks/robotwin_2_0_grab_roller-standard/environment/sim/rcb_traj.py b/tasks/robotwin_2_0_grab_roller-standard/environment/sim/rcb_traj.py
index 69b38b81e0..c2de9c76df 100644
--- a/tasks/robotwin_2_0_grab_roller-standard/environment/sim/rcb_traj.py
+++ b/tasks/robotwin_2_0_grab_roller-standard/environment/sim/rcb_traj.py
@@ -113,6 +113,26 @@ BASE_ARGS: dict[str, Any] = {
}
+class _SortedGlob:
+ """While a scene is built, glob.glob / glob.iglob return their matches sorted. Four RoboTwin tasks (place_a2b_left,
+ place_a2b_right, place_object_scale, place_object_stand: envs/<task>.py load_actors) list an object's model files
+ with glob.glob and then draw one with the seeded np.random.choice; glob returns the file system's directory order,
+ which differs between hosts (and image layers), so one seed built a different object on another host and the frozen
+ instance's hash check failed there (board 0086 claim 5; never rebuilt off the reference workstation). Sorted, the
+ draw depends on the seed alone. No other task of ours lists files while it builds its scene."""
+
+ def __enter__(self):
+ import glob
+ self._glob, self._orig = glob, (glob.glob, glob.iglob)
+ glob.glob = lambda *a, **k: sorted(self._orig[0](*a, **k))
+ glob.iglob = lambda *a, **k: iter(sorted(self._orig[1](*a, **k)))
+ return self
+
+ def __exit__(self, *exc):
+ self._glob.glob, self._glob.iglob = self._orig
+ return False
+
+
def _chdir_root() -> None:
# RoboTwin resolves assets/config paths relative to the repo root and via ROOT_PATH
if os.getcwd() != ROBOTWIN_ROOT and os.path.isdir(ROBOTWIN_ROOT):
@@ -218,7 +238,8 @@ class Session:
if self.enable_render:
sapien.render.set_camera_shader_dir(self.render_shader)
args = _build_args(self.task, self.embodiment, self.seed)
- self.env.setup_demo(**args)
+ with _SortedGlob(): # the object draw depends on the seed alone, not the host's directory order
+ self.env.setup_demo(**args)
if self.enable_render: # setup_scene() switches the shader to ray tracing; put it back for cheap frames
sapien.render.set_camera_shader_dir(self.render_shader)
self._raw_scene = self.env.scene
diff --git a/tasks/robotwin_2_0_grab_roller-standard/task.toml b/tasks/robotwin_2_0_grab_roller-standard/task.toml
index 69988696ab..0052c499b6 100644
--- a/tasks/robotwin_2_0_grab_roller-standard/task.toml
+++ b/tasks/robotwin_2_0_grab_roller-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "robotwin"
robot = "aloha-agilex"
family = "grab_roller"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 60417
diff --git a/tasks/robotwin_2_0_handover_block-privileged/README.md b/tasks/robotwin_2_0_handover_block-privileged/README.md
index 8d1819b83e..e79e4d07c7 100644
--- a/tasks/robotwin_2_0_handover_block-privileged/README.md
+++ b/tasks/robotwin_2_0_handover_block-privileged/README.md
@@ -56,3 +56,4 @@ Use the left arm to grasp the red block on the table, handover it to the right a
- 2026-10-13 (board 0086): the tooling refuses a trajectory the verifier cannot grade (`environment/rcb_traj.py`, `tests/rcb_traj.py`). `Session.save_trajectory` and `Session.replay` raise `ValueError` for more than 200,000 rows (`rcb_traj.MAX_T`, the verifier's `MAX_T` in `tests/grade.py`: 800 s of physics), and `record_stop` warns on stderr. Before, the tool saved and replayed such a file ("success") and the verifier then graded it `bad_trajectory`; the instruction has stated the cap since 2026-10-10. The verifier, the scene, the instance and the pass rule are unchanged: every trajectory is graded as before.
; `solution/` rebuilds the scene with upstream's `load_actors` to get them back for the expert. None of them is read by `check_success`. The scene, the frozen instance (same initial-state hash), the tooling, the verifier and the pass rule are unchanged, so the task keeps its name; privileged results from before this date were obtained with these values reachable.
- 2026-10-13 (board 0100, owner decision): base image `eai-robotwin` 0.1.2. The image's scrub of RoboTwin's expert kept `load_actors` whole, so values it computes only for the removed expert stayed reachable: `env.block_middle_pose` (poses the expert moves to or places at), and the instruction listed them among the scene objects. They are gone: the env no longer has them, `/opt/RoboTwin/envs/handover_block.py` no longer shows them, and the scene-object list no longer names them; `solution/` rebuilds the scene with upstream's `load_actors` to get them back for the expert. None of them is read by `check_success`. The scene, the frozen instance (same initial-state hash), the tooling, the verifier and the pass rule are unchanged, so the task keeps its name; privileged results from before this date were obtained with these values reachable.
+- 2026-10-14 (board 0113): the tooling only (`environment/rcb_traj.py`, `tests/rcb_traj.py`): while it builds the scene, `glob.glob` returns its matches sorted, so that a task which lists files to draw an object depends on its seed alone, not on the host's directory order (four RoboTwin tasks did; this one lists no files while it builds its scene). Its scene, frozen instance (same initial-state hash), verifier and pass rule are unchanged.
diff --git a/tasks/robotwin_2_0_handover_block-privileged/environment/rcb_traj.py b/tasks/robotwin_2_0_handover_block-privileged/environment/rcb_traj.py
index 4d7ffd3f41..da51a35e82 100644
--- a/tasks/robotwin_2_0_handover_block-privileged/environment/rcb_traj.py
+++ b/tasks/robotwin_2_0_handover_block-privileged/environment/rcb_traj.py
@@ -113,6 +113,26 @@ BASE_ARGS: dict[str, Any] = {
}
+class _SortedGlob:
+ """While a scene is built, glob.glob / glob.iglob return their matches sorted. Four RoboTwin tasks (place_a2b_left,
+ place_a2b_right, place_object_scale, place_object_stand: envs/<task>.py load_actors) list an object's model files
+ with glob.glob and then draw one with the seeded np.random.choice; glob returns the file system's directory order,
+ which differs between hosts (and image layers), so one seed built a different object on another host and the frozen
+ instance's hash check failed there (board 0086 claim 5; never rebuilt off the reference workstation). Sorted, the
+ draw depends on the seed alone. No other task of ours lists files while it builds its scene."""
+
+ def __enter__(self):
+ import glob
+ self._glob, self._orig = glob, (glob.glob, glob.iglob)
+ glob.glob = lambda *a, **k: sorted(self._orig[0](*a, **k))
+ glob.iglob = lambda *a, **k: iter(sorted(self._orig[1](*a, **k)))
+ return self
+
+ def __exit__(self, *exc):
+ self._glob.glob, self._glob.iglob = self._orig
+ return False
+
+
def _chdir_root() -> None:
# RoboTwin resolves assets/config paths relative to the repo root and via ROOT_PATH
if os.getcwd() != ROBOTWIN_ROOT and os.path.isdir(ROBOTWIN_ROOT):
@@ -218,7 +238,8 @@ class Session:
if self.enable_render:
sapien.render.set_camera_shader_dir(self.render_shader)
args = _build_args(self.task, self.embodiment, self.seed)
- self.env.setup_demo(**args)
+ with _SortedGlob(): # the object draw depends on the seed alone, not the host's directory order
+ self.env.setup_demo(**args)
if self.enable_render: # setup_scene() switches the shader to ray tracing; put it back for cheap frames
sapien.render.set_camera_shader_dir(self.render_shader)
self._raw_scene = self.env.scene
diff --git a/tasks/robotwin_2_0_handover_block-privileged/task.toml b/tasks/robotwin_2_0_handover_block-privileged/task.toml
index 4f08d6131d..c53b7410cd 100644
--- a/tasks/robotwin_2_0_handover_block-privileged/task.toml
+++ b/tasks/robotwin_2_0_handover_block-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "robotwin" # RoboTwin 2.0 = SAPIEN 3 / PhysX + CuRobo
robot = "aloha-agilex"
family = "handover_block"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 60417 # scene = environment/instance.json (RoboTwin seed + expected initial-state hash); identity = dataset.toml digest
expert_time_estimate_hours = 0.5
diff --git a/tasks/robotwin_2_0_handover_block-privileged/tests/rcb_traj.py b/tasks/robotwin_2_0_handover_block-privileged/tests/rcb_traj.py
index 4d7ffd3f41..da51a35e82 100644
--- a/tasks/robotwin_2_0_handover_block-privileged/tests/rcb_traj.py
+++ b/tasks/robotwin_2_0_handover_block-privileged/tests/rcb_traj.py
@@ -113,6 +113,26 @@ BASE_ARGS: dict[str, Any] = {
}
+class _SortedGlob:
+ """While a scene is built, glob.glob / glob.iglob return their matches sorted. Four RoboTwin tasks (place_a2b_left,
+ place_a2b_right, place_object_scale, place_object_stand: envs/<task>.py load_actors) list an object's model files
+ with glob.glob and then draw one with the seeded np.random.choice; glob returns the file system's directory order,
+ which differs between hosts (and image layers), so one seed built a different object on another host and the frozen
+ instance's hash check failed there (board 0086 claim 5; never rebuilt off the reference workstation). Sorted, the
+ draw depends on the seed alone. No other task of ours lists files while it builds its scene."""
+
+ def __enter__(self):
+ import glob
+ self._glob, self._orig = glob, (glob.glob, glob.iglob)
+ glob.glob = lambda *a, **k: sorted(self._orig[0](*a, **k))
+ glob.iglob = lambda *a, **k: iter(sorted(self._orig[1](*a, **k)))
+ return self
+
+ def __exit__(self, *exc):
+ self._glob.glob, self._glob.iglob = self._orig
+ return False
+
+
def _chdir_root() -> None:
# RoboTwin resolves assets/config paths relative to the repo root and via ROOT_PATH
if os.getcwd() != ROBOTWIN_ROOT and os.path.isdir(ROBOTWIN_ROOT):
@@ -218,7 +238,8 @@ class Session:
if self.enable_render:
sapien.render.set_camera_shader_dir(self.render_shader)
args = _build_args(self.task, self.embodiment, self.seed)
- self.env.setup_demo(**args)
+ with _SortedGlob(): # the object draw depends on the seed alone, not the host's directory order
+ self.env.setup_demo(**args)
if self.enable_render: # setup_scene() switches the shader to ray tracing; put it back for cheap frames
sapien.render.set_camera_shader_dir(self.render_shader)
self._raw_scene = self.env.scene
diff --git a/tasks/robotwin_2_0_handover_block-standard/README.md b/tasks/robotwin_2_0_handover_block-standard/README.md
index eea130bcc2..c7143d7287 100644
--- a/tasks/robotwin_2_0_handover_block-standard/README.md
+++ b/tasks/robotwin_2_0_handover_block-standard/README.md
@@ -53,3 +53,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
- 2026-10-10 (owner decision, board 0072): the agent's network egress is restricted to the model APIs (`[agent] network_mode = "allowlist"`, the hosts of `scripts/set_agent_egress.py`; since 2026-10-11 these include the package registries, PyPI and the Debian / Ubuntu archives, so that `pip install` and `apt-get install` work); it was open before (Harbor's default), and agents could download upstream sources, reference solutions and robot models from their own shell. The scene, the frozen instance, the tooling and the pass rule are unchanged, so the task keeps its name; results from before this date were obtained with open egress.
- 2026-10-13 (board 0086): the service's copy of the tooling only (`environment/sim/rcb_traj.py`, one line of `environment/sim/server.py`). The tooling now refuses to save or replay a trajectory of more than 200,000 rows, the privileged verifier's cap; a standard episode has no row cap, so the service lifts it for its own session (`S.max_rows = None`). Nothing an agent sees and nothing the service grades changes.
- 2026-10-13 (board 0100): base image `eai-robotwin` 0.1.2, whose scrub of RoboTwin's expert also removes the reference poses that `load_actors` computed only for the expert (here `block_middle_pose`). The standard-mode agent never reaches the env, so nothing it observes changes; the scene, the instance and the judge are unchanged.
+- 2026-10-14 (board 0113): the tooling only (`environment/sim/rcb_traj.py`): while it builds the scene, `glob.glob` returns its matches sorted, so that a task which lists files to draw an object depends on its seed alone, not on the host's directory order (four RoboTwin tasks did; this one lists no files while it builds its scene). Its scene, frozen instance (same initial-state hash), verifier and pass rule are unchanged.
diff --git a/tasks/robotwin_2_0_handover_block-standard/environment/sim/rcb_traj.py b/tasks/robotwin_2_0_handover_block-standard/environment/sim/rcb_traj.py
index 4d7ffd3f41..da51a35e82 100644
--- a/tasks/robotwin_2_0_handover_block-standard/environment/sim/rcb_traj.py
+++ b/tasks/robotwin_2_0_handover_block-standard/environment/sim/rcb_traj.py
@@ -113,6 +113,26 @@ BASE_ARGS: dict[str, Any] = {
}
+class _SortedGlob:
+ """While a scene is built, glob.glob / glob.iglob return their matches sorted. Four RoboTwin tasks (place_a2b_left,
+ place_a2b_right, place_object_scale, place_object_stand: envs/<task>.py load_actors) list an object's model files
+ with glob.glob and then draw one with the seeded np.random.choice; glob returns the file system's directory order,
+ which differs between hosts (and image layers), so one seed built a different object on another host and the frozen
+ instance's hash check failed there (board 0086 claim 5; never rebuilt off the reference workstation). Sorted, the
+ draw depends on the seed alone. No other task of ours lists files while it builds its scene."""
+
+ def __enter__(self):
+ import glob
+ self._glob, self._orig = glob, (glob.glob, glob.iglob)
+ glob.glob = lambda *a, **k: sorted(self._orig[0](*a, **k))
+ glob.iglob = lambda *a, **k: iter(sorted(self._orig[1](*a, **k)))
+ return self
+
+ def __exit__(self, *exc):
+ self._glob.glob, self._glob.iglob = self._orig
+ return False
+
+
def _chdir_root() -> None:
# RoboTwin resolves assets/config paths relative to the repo root and via ROOT_PATH
if os.getcwd() != ROBOTWIN_ROOT and os.path.isdir(ROBOTWIN_ROOT):
@@ -218,7 +238,8 @@ class Session:
if self.enable_render:
sapien.render.set_camera_shader_dir(self.render_shader)
args = _build_args(self.task, self.embodiment, self.seed)
- self.env.setup_demo(**args)
+ with _SortedGlob(): # the object draw depends on the seed alone, not the host's directory order
+ self.env.setup_demo(**args)
if self.enable_render: # setup_scene() switches the shader to ray tracing; put it back for cheap frames
sapien.render.set_camera_shader_dir(self.render_shader)
self._raw_scene = self.env.scene
diff --git a/tasks/robotwin_2_0_handover_block-standard/task.toml b/tasks/robotwin_2_0_handover_block-standard/task.toml
index f811a193a8..7b9e851288 100644
--- a/tasks/robotwin_2_0_handover_block-standard/task.toml
+++ b/tasks/robotwin_2_0_handover_block-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "robotwin"
robot = "aloha-agilex"
family = "handover_block"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 60417
diff --git a/tasks/robotwin_2_0_handover_mic-privileged/README.md b/tasks/robotwin_2_0_handover_mic-privileged/README.md
index 93463aa9c4..eae6c4fc35 100644
--- a/tasks/robotwin_2_0_handover_mic-privileged/README.md
+++ b/tasks/robotwin_2_0_handover_mic-privileged/README.md
@@ -56,3 +56,4 @@ Use one arm to grasp the microphone on the table and handover it to the other ar
- 2026-10-13 (board 0086): the tooling refuses a trajectory the verifier cannot grade (`environment/rcb_traj.py`, `tests/rcb_traj.py`). `Session.save_trajectory` and `Session.replay` raise `ValueError` for more than 200,000 rows (`rcb_traj.MAX_T`, the verifier's `MAX_T` in `tests/grade.py`: 800 s of physics), and `record_stop` warns on stderr. Before, the tool saved and replayed such a file ("success") and the verifier then graded it `bad_trajectory`; the instruction has stated the cap since 2026-10-10. The verifier, the scene, the instance and the pass rule are unchanged: every trajectory is graded as before.
; `solution/` rebuilds the scene with upstream's `load_actors` to get them back for the expert. None of them is read by `check_success`. The scene, the frozen instance (same initial-state hash), the tooling, the verifier and the pass rule are unchanged, so the task keeps its name; privileged results from before this date were obtained with these values reachable.
- 2026-10-13 (board 0100, owner decision): base image `eai-robotwin` 0.1.2. The image's scrub of RoboTwin's expert kept `load_actors` whole, so values it computes only for the removed expert stayed reachable: `env.handover_middle_pose` (poses the expert moves to or places at), and the instruction listed them among the scene objects. They are gone: the env no longer has them, `/opt/RoboTwin/envs/handover_mic.py` no longer shows them, and the scene-object list no longer names them; `solution/` rebuilds the scene with upstream's `load_actors` to get them back for the expert. None of them is read by `check_success`. The scene, the frozen instance (same initial-state hash), the tooling, the verifier and the pass rule are unchanged, so the task keeps its name; privileged results from before this date were obtained with these values reachable.
+- 2026-10-14 (board 0113): the tooling only (`environment/rcb_traj.py`, `tests/rcb_traj.py`): while it builds the scene, `glob.glob` returns its matches sorted, so that a task which lists files to draw an object depends on its seed alone, not on the host's directory order (four RoboTwin tasks did; this one lists no files while it builds its scene). Its scene, frozen instance (same initial-state hash), verifier and pass rule are unchanged.
diff --git a/tasks/robotwin_2_0_handover_mic-privileged/environment/rcb_traj.py b/tasks/robotwin_2_0_handover_mic-privileged/environment/rcb_traj.py
index 9ef5c00b48..34d8838d43 100644
--- a/tasks/robotwin_2_0_handover_mic-privileged/environment/rcb_traj.py
+++ b/tasks/robotwin_2_0_handover_mic-privileged/environment/rcb_traj.py
@@ -113,6 +113,26 @@ BASE_ARGS: dict[str, Any] = {
}
+class _SortedGlob:
+ """While a scene is built, glob.glob / glob.iglob return their matches sorted. Four RoboTwin tasks (place_a2b_left,
+ place_a2b_right, place_object_scale, place_object_stand: envs/<task>.py load_actors) list an object's model files
+ with glob.glob and then draw one with the seeded np.random.choice; glob returns the file system's directory order,
+ which differs between hosts (and image layers), so one seed built a different object on another host and the frozen
+ instance's hash check failed there (board 0086 claim 5; never rebuilt off the reference workstation). Sorted, the
+ draw depends on the seed alone. No other task of ours lists files while it builds its scene."""
+
+ def __enter__(self):
+ import glob
+ self._glob, self._orig = glob, (glob.glob, glob.iglob)
+ glob.glob = lambda *a, **k: sorted(self._orig[0](*a, **k))
+ glob.iglob = lambda *a, **k: iter(sorted(self._orig[1](*a, **k)))
+ return self
+
+ def __exit__(self, *exc):
+ self._glob.glob, self._glob.iglob = self._orig
+ return False
+
+
def _chdir_root() -> None:
# RoboTwin resolves assets/config paths relative to the repo root and via ROOT_PATH
if os.getcwd() != ROBOTWIN_ROOT and os.path.isdir(ROBOTWIN_ROOT):
@@ -218,7 +238,8 @@ class Session:
if self.enable_render:
sapien.render.set_camera_shader_dir(self.render_shader)
args = _build_args(self.task, self.embodiment, self.seed)
- self.env.setup_demo(**args)
+ with _SortedGlob(): # the object draw depends on the seed alone, not the host's directory order
+ self.env.setup_demo(**args)
if self.enable_render: # setup_scene() switches the shader to ray tracing; put it back for cheap frames
sapien.render.set_camera_shader_dir(self.render_shader)
self._raw_scene = self.env.scene
diff --git a/tasks/robotwin_2_0_handover_mic-privileged/task.toml b/tasks/robotwin_2_0_handover_mic-privileged/task.toml
index 0669f78526..9623d67a63 100644
--- a/tasks/robotwin_2_0_handover_mic-privileged/task.toml
+++ b/tasks/robotwin_2_0_handover_mic-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "robotwin" # RoboTwin 2.0 = SAPIEN 3 / PhysX + CuRobo
robot = "aloha-agilex"
family = "handover_mic"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 28831 # scene = environment/instance.json (RoboTwin seed + expected initial-state hash); identity = dataset.toml digest
expert_time_estimate_hours = 0.5
diff --git a/tasks/robotwin_2_0_handover_mic-privileged/tests/rcb_traj.py b/tasks/robotwin_2_0_handover_mic-privileged/tests/rcb_traj.py
index 9ef5c00b48..34d8838d43 100644
--- a/tasks/robotwin_2_0_handover_mic-privileged/tests/rcb_traj.py
+++ b/tasks/robotwin_2_0_handover_mic-privileged/tests/rcb_traj.py
@@ -113,6 +113,26 @@ BASE_ARGS: dict[str, Any] = {
}
+class _SortedGlob:
+ """While a scene is built, glob.glob / glob.iglob return their matches sorted. Four RoboTwin tasks (place_a2b_left,
+ place_a2b_right, place_object_scale, place_object_stand: envs/<task>.py load_actors) list an object's model files
+ with glob.glob and then draw one with the seeded np.random.choice; glob returns the file system's directory order,
+ which differs between hosts (and image layers), so one seed built a different object on another host and the frozen
+ instance's hash check failed there (board 0086 claim 5; never rebuilt off the reference workstation). Sorted, the
+ draw depends on the seed alone. No other task of ours lists files while it builds its scene."""
+
+ def __enter__(self):
+ import glob
+ self._glob, self._orig = glob, (glob.glob, glob.iglob)
+ glob.glob = lambda *a, **k: sorted(self._orig[0](*a, **k))
+ glob.iglob = lambda *a, **k: iter(sorted(self._orig[1](*a, **k)))
+ return self
+
+ def __exit__(self, *exc):
+ self._glob.glob, self._glob.iglob = self._orig
+ return False
+
+
def _chdir_root() -> None:
# RoboTwin resolves assets/config paths relative to the repo root and via ROOT_PATH
if os.getcwd() != ROBOTWIN_ROOT and os.path.isdir(ROBOTWIN_ROOT):
@@ -218,7 +238,8 @@ class Session:
if self.enable_render:
sapien.render.set_camera_shader_dir(self.render_shader)
args = _build_args(self.task, self.embodiment, self.seed)
- self.env.setup_demo(**args)
+ with _SortedGlob(): # the object draw depends on the seed alone, not the host's directory order
+ self.env.setup_demo(**args)
if self.enable_render: # setup_scene() switches the shader to ray tracing; put it back for cheap frames
sapien.render.set_camera_shader_dir(self.render_shader)
self._raw_scene = self.env.scene
diff --git a/tasks/robotwin_2_0_handover_mic-standard/README.md b/tasks/robotwin_2_0_handover_mic-standard/README.md
index 2eab7916a5..8ab0b4af3c 100644
--- a/tasks/robotwin_2_0_handover_mic-standard/README.md
+++ b/tasks/robotwin_2_0_handover_mic-standard/README.md
@@ -53,3 +53,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
- 2026-10-10 (owner decision, board 0072): the agent's network egress is restricted to the model APIs (`[agent] network_mode = "allowlist"`, the hosts of `scripts/set_agent_egress.py`; since 2026-10-11 these include the package registries, PyPI and the Debian / Ubuntu archives, so that `pip install` and `apt-get install` work); it was open before (Harbor's default), and agents could download upstream sources, reference solutions and robot models from their own shell. The scene, the frozen instance, the tooling and the pass rule are unchanged, so the task keeps its name; results from before this date were obtained with open egress.
- 2026-10-13 (board 0086): the service's copy of the tooling only (`environment/sim/rcb_traj.py`, one line of `environment/sim/server.py`). The tooling now refuses to save or replay a trajectory of more than 200,000 rows, the privileged verifier's cap; a standard episode has no row cap, so the service lifts it for its own session (`S.max_rows = None`). Nothing an agent sees and nothing the service grades changes.
- 2026-10-13 (board 0100): base image `eai-robotwin` 0.1.2, whose scrub of RoboTwin's expert also removes the reference poses that `load_actors` computed only for the expert (here `handover_middle_pose`). The standard-mode agent never reaches the env, so nothing it observes changes; the scene, the instance and the judge are unchanged.
+- 2026-10-14 (board 0113): the tooling only (`environment/sim/rcb_traj.py`): while it builds the scene, `glob.glob` returns its matches sorted, so that a task which lists files to draw an object depends on its seed alone, not on the host's directory order (four RoboTwin tasks did; this one lists no files while it builds its scene). Its scene, frozen instance (same initial-state hash), verifier and pass rule are unchanged.
diff --git a/tasks/robotwin_2_0_handover_mic-standard/environment/sim/rcb_traj.py b/tasks/robotwin_2_0_handover_mic-standard/environment/sim/rcb_traj.py
index 9ef5c00b48..34d8838d43 100644
--- a/tasks/robotwin_2_0_handover_mic-standard/environment/sim/rcb_traj.py
+++ b/tasks/robotwin_2_0_handover_mic-standard/environment/sim/rcb_traj.py
@@ -113,6 +113,26 @@ BASE_ARGS: dict[str, Any] = {
}
+class _SortedGlob:
+ """While a scene is built, glob.glob / glob.iglob return their matches sorted. Four RoboTwin tasks (place_a2b_left,
+ place_a2b_right, place_object_scale, place_object_stand: envs/<task>.py load_actors) list an object's model files
+ with glob.glob and then draw one with the seeded np.random.choice; glob returns the file system's directory order,
+ which differs between hosts (and image layers), so one seed built a different object on another host and the frozen
+ instance's hash check failed there (board 0086 claim 5; never rebuilt off the reference workstation). Sorted, the
+ draw depends on the seed alone. No other task of ours lists files while it builds its scene."""
+
+ def __enter__(self):
+ import glob
+ self._glob, self._orig = glob, (glob.glob, glob.iglob)
+ glob.glob = lambda *a, **k: sorted(self._orig[0](*a, **k))
+ glob.iglob = lambda *a, **k: iter(sorted(self._orig[1](*a, **k)))
+ return self
+
+ def __exit__(self, *exc):
+ self._glob.glob, self._glob.iglob = self._orig
+ return False
+
+
def _chdir_root() -> None:
# RoboTwin resolves assets/config paths relative to the repo root and via ROOT_PATH
if os.getcwd() != ROBOTWIN_ROOT and os.path.isdir(ROBOTWIN_ROOT):
@@ -218,7 +238,8 @@ class Session:
if self.enable_render:
sapien.render.set_camera_shader_dir(self.render_shader)
args = _build_args(self.task, self.embodiment, self.seed)
- self.env.setup_demo(**args)
+ with _SortedGlob(): # the object draw depends on the seed alone, not the host's directory order
+ self.env.setup_demo(**args)
if self.enable_render: # setup_scene() switches the shader to ray tracing; put it back for cheap frames
sapien.render.set_camera_shader_dir(self.render_shader)
self._raw_scene = self.env.scene
diff --git a/tasks/robotwin_2_0_handover_mic-standard/task.toml b/tasks/robotwin_2_0_handover_mic-standard/task.toml
index ddd4356183..7e5d7ea3d8 100644
--- a/tasks/robotwin_2_0_handover_mic-standard/task.toml
+++ b/tasks/robotwin_2_0_handover_mic-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "robotwin"
robot = "aloha-agilex"
family = "handover_mic"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 28831
diff --git a/tasks/robotwin_2_0_hanging_mug-privileged/README.md b/tasks/robotwin_2_0_hanging_mug-privileged/README.md
index db1660cbef..7717e3478e 100644
--- a/tasks/robotwin_2_0_hanging_mug-privileged/README.md
+++ b/tasks/robotwin_2_0_hanging_mug-privileged/README.md
@@ -56,3 +56,4 @@ Use left arm to pick the mug on the table, rotate the mug and put the mug down i
- 2026-10-13 (board 0086): the tooling refuses a trajectory the verifier cannot grade (`environment/rcb_traj.py`, `tests/rcb_traj.py`). `Session.save_trajectory` and `Session.replay` raise `ValueError` for more than 200,000 rows (`rcb_traj.MAX_T`, the verifier's `MAX_T` in `tests/grade.py`: 800 s of physics), and `record_stop` warns on stderr. Before, the tool saved and replayed such a file ("success") and the verifier then graded it `bad_trajectory`; the instruction has stated the cap since 2026-10-10. The verifier, the scene, the instance and the pass rule are unchanged: every trajectory is graded as before.
; `solution/` rebuilds the scene with upstream's `load_actors` to get them back for the expert. None of them is read by `check_success`. The scene, the frozen instance (same initial-state hash), the tooling, the verifier and the pass rule are unchanged, so the task keeps its name; privileged results from before this date were obtained with these values reachable.
- 2026-10-13 (board 0100, owner decision): base image `eai-robotwin` 0.1.2. The image's scrub of RoboTwin's expert kept `load_actors` whole, so values it computes only for the removed expert stayed reachable: `env.middle_pos` (poses the expert moves to or places at), and the instruction listed them among the scene objects. They are gone: the env no longer has them, `/opt/RoboTwin/envs/hanging_mug.py` no longer shows them, and the scene-object list no longer names them; `solution/` rebuilds the scene with upstream's `load_actors` to get them back for the expert. None of them is read by `check_success`. The scene, the frozen instance (same initial-state hash), the tooling, the verifier and the pass rule are unchanged, so the task keeps its name; privileged results from before this date were obtained with these values reachable.
+- 2026-10-14 (board 0113): the tooling only (`environment/rcb_traj.py`, `tests/rcb_traj.py`): while it builds the scene, `glob.glob` returns its matches sorted, so that a task which lists files to draw an object depends on its seed alone, not on the host's directory order (four RoboTwin tasks did; this one lists no files while it builds its scene). Its scene, frozen instance (same initial-state hash), verifier and pass rule are unchanged.
diff --git a/tasks/robotwin_2_0_hanging_mug-privileged/environment/rcb_traj.py b/tasks/robotwin_2_0_hanging_mug-privileged/environment/rcb_traj.py
index ad8b5025c7..7be71e97f9 100644
--- a/tasks/robotwin_2_0_hanging_mug-privileged/environment/rcb_traj.py
+++ b/tasks/robotwin_2_0_hanging_mug-privileged/environment/rcb_traj.py
@@ -113,6 +113,26 @@ BASE_ARGS: dict[str, Any] = {
}
+class _SortedGlob:
+ """While a scene is built, glob.glob / glob.iglob return their matches sorted. Four RoboTwin tasks (place_a2b_left,
+ place_a2b_right, place_object_scale, place_object_stand: envs/<task>.py load_actors) list an object's model files
+ with glob.glob and then draw one with the seeded np.random.choice; glob returns the file system's directory order,
+ which differs between hosts (and image layers), so one seed built a different object on another host and the frozen
+ instance's hash check failed there (board 0086 claim 5; never rebuilt off the reference workstation). Sorted, the
+ draw depends on the seed alone. No other task of ours lists files while it builds its scene."""
+
+ def __enter__(self):
+ import glob
+ self._glob, self._orig = glob, (glob.glob, glob.iglob)
+ glob.glob = lambda *a, **k: sorted(self._orig[0](*a, **k))
+ glob.iglob = lambda *a, **k: iter(sorted(self._orig[1](*a, **k)))
+ return self
+
+ def __exit__(self, *exc):
+ self._glob.glob, self._glob.iglob = self._orig
+ return False
+
+
def _chdir_root() -> None:
# RoboTwin resolves assets/config paths relative to the repo root and via ROOT_PATH
if os.getcwd() != ROBOTWIN_ROOT and os.path.isdir(ROBOTWIN_ROOT):
@@ -218,7 +238,8 @@ class Session:
if self.enable_render:
sapien.render.set_camera_shader_dir(self.render_shader)
args = _build_args(self.task, self.embodiment, self.seed)
- self.env.setup_demo(**args)
+ with _SortedGlob(): # the object draw depends on the seed alone, not the host's directory order
+ self.env.setup_demo(**args)
if self.enable_render: # setup_scene() switches the shader to ray tracing; put it back for cheap frames
sapien.render.set_camera_shader_dir(self.render_shader)
self._raw_scene = self.env.scene
diff --git a/tasks/robotwin_2_0_hanging_mug-privileged/task.toml b/tasks/robotwin_2_0_hanging_mug-privileged/task.toml
index a7d679daf3..ec33e41b32 100644
--- a/tasks/robotwin_2_0_hanging_mug-privileged/task.toml
+++ b/tasks/robotwin_2_0_hanging_mug-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "robotwin" # RoboTwin 2.0 = SAPIEN 3 / PhysX + CuRobo
robot = "aloha-agilex"
family = "hanging_mug"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 60417 # scene = environment/instance.json (RoboTwin seed + expected initial-state hash); identity = dataset.toml digest
expert_time_estimate_hours = 0.5
diff --git a/tasks/robotwin_2_0_hanging_mug-privileged/tests/rcb_traj.py b/tasks/robotwin_2_0_hanging_mug-privileged/tests/rcb_traj.py
index ad8b5025c7..7be71e97f9 100644
--- a/tasks/robotwin_2_0_hanging_mug-privileged/tests/rcb_traj.py
+++ b/tasks/robotwin_2_0_hanging_mug-privileged/tests/rcb_traj.py
@@ -113,6 +113,26 @@ BASE_ARGS: dict[str, Any] = {
}
+class _SortedGlob:
+ """While a scene is built, glob.glob / glob.iglob return their matches sorted. Four RoboTwin tasks (place_a2b_left,
+ place_a2b_right, place_object_scale, place_object_stand: envs/<task>.py load_actors) list an object's model files
+ with glob.glob and then draw one with the seeded np.random.choice; glob returns the file system's directory order,
+ which differs between hosts (and image layers), so one seed built a different object on another host and the frozen
+ instance's hash check failed there (board 0086 claim 5; never rebuilt off the reference workstation). Sorted, the
+ draw depends on the seed alone. No other task of ours lists files while it builds its scene."""
+
+ def __enter__(self):
+ import glob
+ self._glob, self._orig = glob, (glob.glob, glob.iglob)
+ glob.glob = lambda *a, **k: sorted(self._orig[0](*a, **k))
+ glob.iglob = lambda *a, **k: iter(sorted(self._orig[1](*a, **k)))
+ return self
+
+ def __exit__(self, *exc):
+ self._glob.glob, self._glob.iglob = self._orig
+ return False
+
+
def _chdir_root() -> None:
# RoboTwin resolves assets/config paths relative to the repo root and via ROOT_PATH
if os.getcwd() != ROBOTWIN_ROOT and os.path.isdir(ROBOTWIN_ROOT):
@@ -218,7 +238,8 @@ class Session:
if self.enable_render:
sapien.render.set_camera_shader_dir(self.render_shader)
args = _build_args(self.task, self.embodiment, self.seed)
- self.env.setup_demo(**args)
+ with _SortedGlob(): # the object draw depends on the seed alone, not the host's directory order
+ self.env.setup_demo(**args)
if self.enable_render: # setup_scene() switches the shader to ray tracing; put it back for cheap frames
sapien.render.set_camera_shader_dir(self.render_shader)
self._raw_scene = self.env.scene
diff --git a/tasks/robotwin_2_0_hanging_mug-standard/README.md b/tasks/robotwin_2_0_hanging_mug-standard/README.md
index 7463e201aa..5ca5038653 100644
--- a/tasks/robotwin_2_0_hanging_mug-standard/README.md
+++ b/tasks/robotwin_2_0_hanging_mug-standard/README.md
@@ -53,3 +53,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
- 2026-10-10 (owner decision, board 0072): the agent's network egress is restricted to the model APIs (`[agent] network_mode = "allowlist"`, the hosts of `scripts/set_agent_egress.py`; since 2026-10-11 these include the package registries, PyPI and the Debian / Ubuntu archives, so that `pip install` and `apt-get install` work); it was open before (Harbor's default), and agents could download upstream sources, reference solutions and robot models from their own shell. The scene, the frozen instance, the tooling and the pass rule are unchanged, so the task keeps its name; results from before this date were obtained with open egress.
- 2026-10-13 (board 0086): the service's copy of the tooling only (`environment/sim/rcb_traj.py`, one line of `environment/sim/server.py`). The tooling now refuses to save or replay a trajectory of more than 200,000 rows, the privileged verifier's cap; a standard episode has no row cap, so the service lifts it for its own session (`S.max_rows = None`). Nothing an agent sees and nothing the service grades changes.
- 2026-10-13 (board 0100): base image `eai-robotwin` 0.1.2, whose scrub of RoboTwin's expert also removes the reference poses that `load_actors` computed only for the expert (here `middle_pos`). The standard-mode agent never reaches the env, so nothing it observes changes; the scene, the instance and the judge are unchanged.
+- 2026-10-14 (board 0113): the tooling only (`environment/sim/rcb_traj.py`): while it builds the scene, `glob.glob` returns its matches sorted, so that a task which lists files to draw an object depends on its seed alone, not on the host's directory order (four RoboTwin tasks did; this one lists no files while it builds its scene). Its scene, frozen instance (same initial-state hash), verifier and pass rule are unchanged.
diff --git a/tasks/robotwin_2_0_hanging_mug-standard/environment/sim/rcb_traj.py b/tasks/robotwin_2_0_hanging_mug-standard/environment/sim/rcb_traj.py
index ad8b5025c7..7be71e97f9 100644
--- a/tasks/robotwin_2_0_hanging_mug-standard/environment/sim/rcb_traj.py
+++ b/tasks/robotwin_2_0_hanging_mug-standard/environment/sim/rcb_traj.py
@@ -113,6 +113,26 @@ BASE_ARGS: dict[str, Any] = {
}
+class _SortedGlob:
+ """While a scene is built, glob.glob / glob.iglob return their matches sorted. Four RoboTwin tasks (place_a2b_left,
+ place_a2b_right, place_object_scale, place_object_stand: envs/<task>.py load_actors) list an object's model files
+ with glob.glob and then draw one with the seeded np.random.choice; glob returns the file system's directory order,
+ which differs between hosts (and image layers), so one seed built a different object on another host and the frozen
+ instance's hash check failed there (board 0086 claim 5; never rebuilt off the reference workstation). Sorted, the
+ draw depends on the seed alone. No other task of ours lists files while it builds its scene."""
+
+ def __enter__(self):
+ import glob
+ self._glob, self._orig = glob, (glob.glob, glob.iglob)
+ glob.glob = lambda *a, **k: sorted(self._orig[0](*a, **k))
+ glob.iglob = lambda *a, **k: iter(sorted(self._orig[1](*a, **k)))
+ return self
+
+ def __exit__(self, *exc):
+ self._glob.glob, self._glob.iglob = self._orig
+ return False
+
+
def _chdir_root() -> None:
# RoboTwin resolves assets/config paths relative to the repo root and via ROOT_PATH
if os.getcwd() != ROBOTWIN_ROOT and os.path.isdir(ROBOTWIN_ROOT):
@@ -218,7 +238,8 @@ class Session:
if self.enable_render:
sapien.render.set_camera_shader_dir(self.render_shader)
args = _build_args(self.task, self.embodiment, self.seed)
- self.env.setup_demo(**args)
+ with _SortedGlob(): # the object draw depends on the seed alone, not the host's directory order
+ self.env.setup_demo(**args)
if self.enable_render: # setup_scene() switches the shader to ray tracing; put it back for cheap frames
sapien.render.set_camera_shader_dir(self.render_shader)
self._raw_scene = self.env.scene
diff --git a/tasks/robotwin_2_0_hanging_mug-standard/task.toml b/tasks/robotwin_2_0_hanging_mug-standard/task.toml
index 6f426c5526..3e8e8667ec 100644
--- a/tasks/robotwin_2_0_hanging_mug-standard/task.toml
+++ b/tasks/robotwin_2_0_hanging_mug-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "robotwin"
robot = "aloha-agilex"
family = "hanging_mug"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 60417
diff --git a/tasks/robotwin_2_0_lift_pot-privileged/README.md b/tasks/robotwin_2_0_lift_pot-privileged/README.md
index 61a995f468..98ad0aa8dd 100644
--- a/tasks/robotwin_2_0_lift_pot-privileged/README.md
+++ b/tasks/robotwin_2_0_lift_pot-privileged/README.md
@@ -55,3 +55,4 @@ Use arms to lift the pot.
- 2026-10-10 (owner decision, board 0072): the agent's network egress is restricted to the model APIs (`[agent] network_mode = "allowlist"`, the hosts of `scripts/set_agent_egress.py`; since 2026-10-11 these include the package registries, PyPI and the Debian / Ubuntu archives, so that `pip install` and `apt-get install` work); it was open before (Harbor's default), and agents could download upstream sources, reference solutions and robot models from their own shell. The scene, the frozen instance, the tooling and the pass rule are unchanged, so the task keeps its name; results from before this date were obtained with open egress.
- 2026-10-13 (board 0086): the tooling refuses a trajectory the verifier cannot grade (`environment/rcb_traj.py`, `tests/rcb_traj.py`). `Session.save_trajectory` and `Session.replay` raise `ValueError` for more than 200,000 rows (`rcb_traj.MAX_T`, the verifier's `MAX_T` in `tests/grade.py`: 800 s of physics), and `record_stop` warns on stderr. Before, the tool saved and replayed such a file ("success") and the verifier then graded it `bad_trajectory`; the instruction has stated the cap since 2026-10-10. The verifier, the scene, the instance and the pass rule are unchanged: every trajectory is graded as before.
- 2026-10-13 (board 0100): base image `eai-robotwin` 0.1.2. Its scrub of RoboTwin's expert also removes the reference poses that 18 tasks' `load_actors` computed only for the expert, and the expert's comments above the scrubbed method bodies (`images/robotwin/scrub_oracles.py`). This task's `load_actors` set no such value: its scene, instance, tooling, verifier and pass rule are unchanged.
+- 2026-10-14 (board 0113): the tooling only (`environment/rcb_traj.py`, `tests/rcb_traj.py`): while it builds the scene, `glob.glob` returns its matches sorted, so that a task which lists files to draw an object depends on its seed alone, not on the host's directory order (four RoboTwin tasks did; this one lists no files while it builds its scene). Its scene, frozen instance (same initial-state hash), verifier and pass rule are unchanged.
diff --git a/tasks/robotwin_2_0_lift_pot-privileged/environment/rcb_traj.py b/tasks/robotwin_2_0_lift_pot-privileged/environment/rcb_traj.py
index bf72715729..f650fbacc6 100644
--- a/tasks/robotwin_2_0_lift_pot-privileged/environment/rcb_traj.py
+++ b/tasks/robotwin_2_0_lift_pot-privileged/environment/rcb_traj.py
@@ -113,6 +113,26 @@ BASE_ARGS: dict[str, Any] = {
}
+class _SortedGlob:
+ """While a scene is built, glob.glob / glob.iglob return their matches sorted. Four RoboTwin tasks (place_a2b_left,
+ place_a2b_right, place_object_scale, place_object_stand: envs/<task>.py load_actors) list an object's model files
+ with glob.glob and then draw one with the seeded np.random.choice; glob returns the file system's directory order,
+ which differs between hosts (and image layers), so one seed built a different object on another host and the frozen
+ instance's hash check failed there (board 0086 claim 5; never rebuilt off the reference workstation). Sorted, the
+ draw depends on the seed alone. No other task of ours lists files while it builds its scene."""
+
+ def __enter__(self):
+ import glob
+ self._glob, self._orig = glob, (glob.glob, glob.iglob)
+ glob.glob = lambda *a, **k: sorted(self._orig[0](*a, **k))
+ glob.iglob = lambda *a, **k: iter(sorted(self._orig[1](*a, **k)))
+ return self
+
+ def __exit__(self, *exc):
+ self._glob.glob, self._glob.iglob = self._orig
+ return False
+
+
def _chdir_root() -> None:
# RoboTwin resolves assets/config paths relative to the repo root and via ROOT_PATH
if os.getcwd() != ROBOTWIN_ROOT and os.path.isdir(ROBOTWIN_ROOT):
@@ -218,7 +238,8 @@ class Session:
if self.enable_render:
sapien.render.set_camera_shader_dir(self.render_shader)
args = _build_args(self.task, self.embodiment, self.seed)
- self.env.setup_demo(**args)
+ with _SortedGlob(): # the object draw depends on the seed alone, not the host's directory order
+ self.env.setup_demo(**args)
if self.enable_render: # setup_scene() switches the shader to ray tracing; put it back for cheap frames
sapien.render.set_camera_shader_dir(self.render_shader)
self._raw_scene = self.env.scene
diff --git a/tasks/robotwin_2_0_lift_pot-privileged/task.toml b/tasks/robotwin_2_0_lift_pot-privileged/task.toml
index 1306284340..9932d76e93 100644
--- a/tasks/robotwin_2_0_lift_pot-privileged/task.toml
+++ b/tasks/robotwin_2_0_lift_pot-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "robotwin" # RoboTwin 2.0 = SAPIEN 3 / PhysX + CuRobo
robot = "aloha-agilex"
family = "lift_pot"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 28831 # scene = environment/instance.json (RoboTwin seed + expected initial-state hash); identity = dataset.toml digest
expert_time_estimate_hours = 0.5
diff --git a/tasks/robotwin_2_0_lift_pot-privileged/tests/rcb_traj.py b/tasks/robotwin_2_0_lift_pot-privileged/tests/rcb_traj.py
index bf72715729..f650fbacc6 100644
--- a/tasks/robotwin_2_0_lift_pot-privileged/tests/rcb_traj.py
+++ b/tasks/robotwin_2_0_lift_pot-privileged/tests/rcb_traj.py
@@ -113,6 +113,26 @@ BASE_ARGS: dict[str, Any] = {
}
+class _SortedGlob:
+ """While a scene is built, glob.glob / glob.iglob return their matches sorted. Four RoboTwin tasks (place_a2b_left,
+ place_a2b_right, place_object_scale, place_object_stand: envs/<task>.py load_actors) list an object's model files
+ with glob.glob and then draw one with the seeded np.random.choice; glob returns the file system's directory order,
+ which differs between hosts (and image layers), so one seed built a different object on another host and the frozen
+ instance's hash check failed there (board 0086 claim 5; never rebuilt off the reference workstation). Sorted, the
+ draw depends on the seed alone. No other task of ours lists files while it builds its scene."""
+
+ def __enter__(self):
+ import glob
+ self._glob, self._orig = glob, (glob.glob, glob.iglob)
+ glob.glob = lambda *a, **k: sorted(self._orig[0](*a, **k))
+ glob.iglob = lambda *a, **k: iter(sorted(self._orig[1](*a, **k)))
+ return self
+
+ def __exit__(self, *exc):
+ self._glob.glob, self._glob.iglob = self._orig
+ return False
+
+
def _chdir_root() -> None:
# RoboTwin resolves assets/config paths relative to the repo root and via ROOT_PATH
if os.getcwd() != ROBOTWIN_ROOT and os.path.isdir(ROBOTWIN_ROOT):
@@ -218,7 +238,8 @@ class Session:
if self.enable_render:
sapien.render.set_camera_shader_dir(self.render_shader)
args = _build_args(self.task, self.embodiment, self.seed)
- self.env.setup_demo(**args)
+ with _SortedGlob(): # the object draw depends on the seed alone, not the host's directory order
+ self.env.setup_demo(**args)
if self.enable_render: # setup_scene() switches the shader to ray tracing; put it back for cheap frames
sapien.render.set_camera_shader_dir(self.render_shader)
self._raw_scene = self.env.scene
diff --git a/tasks/robotwin_2_0_lift_pot-standard/README.md b/tasks/robotwin_2_0_lift_pot-standard/README.md
index 5209cb505b..19a72dff93 100644
--- a/tasks/robotwin_2_0_lift_pot-standard/README.md
+++ b/tasks/robotwin_2_0_lift_pot-standard/README.md
@@ -53,3 +53,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
- 2026-10-10 (owner decision, board 0072): the agent's network egress is restricted to the model APIs (`[agent] network_mode = "allowlist"`, the hosts of `scripts/set_agent_egress.py`; since 2026-10-11 these include the package registries, PyPI and the Debian / Ubuntu archives, so that `pip install` and `apt-get install` work); it was open before (Harbor's default), and agents could download upstream sources, reference solutions and robot models from their own shell. The scene, the frozen instance, the tooling and the pass rule are unchanged, so the task keeps its name; results from before this date were obtained with open egress.
- 2026-10-13 (board 0086): the service's copy of the tooling only (`environment/sim/rcb_traj.py`, one line of `environment/sim/server.py`). The tooling now refuses to save or replay a trajectory of more than 200,000 rows, the privileged verifier's cap; a standard episode has no row cap, so the service lifts it for its own session (`S.max_rows = None`). Nothing an agent sees and nothing the service grades changes.
- 2026-10-13 (board 0100): base image `eai-robotwin` 0.1.2, whose scrub of RoboTwin's expert also removes the reference poses that `load_actors` computed only for the expert. The standard-mode agent never reaches the env, so nothing it observes changes; the scene, the instance and the judge are unchanged.
+- 2026-10-14 (board 0113): the tooling only (`environment/sim/rcb_traj.py`): while it builds the scene, `glob.glob` returns its matches sorted, so that a task which lists files to draw an object depends on its seed alone, not on the host's directory order (four RoboTwin tasks did; this one lists no files while it builds its scene). Its scene, frozen instance (same initial-state hash), verifier and pass rule are unchanged.
diff --git a/tasks/robotwin_2_0_lift_pot-standard/environment/sim/rcb_traj.py b/tasks/robotwin_2_0_lift_pot-standard/environment/sim/rcb_traj.py
index bf72715729..f650fbacc6 100644
--- a/tasks/robotwin_2_0_lift_pot-standard/environment/sim/rcb_traj.py
+++ b/tasks/robotwin_2_0_lift_pot-standard/environment/sim/rcb_traj.py
@@ -113,6 +113,26 @@ BASE_ARGS: dict[str, Any] = {
}
+class _SortedGlob:
+ """While a scene is built, glob.glob / glob.iglob return their matches sorted. Four RoboTwin tasks (place_a2b_left,
+ place_a2b_right, place_object_scale, place_object_stand: envs/<task>.py load_actors) list an object's model files
+ with glob.glob and then draw one with the seeded np.random.choice; glob returns the file system's directory order,
+ which differs between hosts (and image layers), so one seed built a different object on another host and the frozen
+ instance's hash check failed there (board 0086 claim 5; never rebuilt off the reference workstation). Sorted, the
+ draw depends on the seed alone. No other task of ours lists files while it builds its scene."""
+
+ def __enter__(self):
+ import glob
+ self._glob, self._orig = glob, (glob.glob, glob.iglob)
+ glob.glob = lambda *a, **k: sorted(self._orig[0](*a, **k))
+ glob.iglob = lambda *a, **k: iter(sorted(self._orig[1](*a, **k)))
+ return self
+
+ def __exit__(self, *exc):
+ self._glob.glob, self._glob.iglob = self._orig
+ return False
+
+
def _chdir_root() -> None:
# RoboTwin resolves assets/config paths relative to the repo root and via ROOT_PATH
if os.getcwd() != ROBOTWIN_ROOT and os.path.isdir(ROBOTWIN_ROOT):
@@ -218,7 +238,8 @@ class Session:
if self.enable_render:
sapien.render.set_camera_shader_dir(self.render_shader)
args = _build_args(self.task, self.embodiment, self.seed)
- self.env.setup_demo(**args)
+ with _SortedGlob(): # the object draw depends on the seed alone, not the host's directory order
+ self.env.setup_demo(**args)
if self.enable_render: # setup_scene() switches the shader to ray tracing; put it back for cheap frames
sapien.render.set_camera_shader_dir(self.render_shader)
self._raw_scene = self.env.scene
diff --git a/tasks/robotwin_2_0_lift_pot-standard/task.toml b/tasks/robotwin_2_0_lift_pot-standard/task.toml
index 18a40c9ebc..902571caca 100644
--- a/tasks/robotwin_2_0_lift_pot-standard/task.toml
+++ b/tasks/robotwin_2_0_lift_pot-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "robotwin"
robot = "aloha-agilex"
family = "lift_pot"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 28831
diff --git a/tasks/robotwin_2_0_move_can_pot-privileged/README.md b/tasks/robotwin_2_0_move_can_pot-privileged/README.md
index d36bdea9fc..872c048012 100644
--- a/tasks/robotwin_2_0_move_can_pot-privileged/README.md
+++ b/tasks/robotwin_2_0_move_can_pot-privileged/README.md
@@ -56,3 +56,4 @@ There is a can and a pot on the table, use one arm to pick up the can and move i
- 2026-10-13 (board 0086): the tooling refuses a trajectory the verifier cannot grade (`environment/rcb_traj.py`, `tests/rcb_traj.py`). `Session.save_trajectory` and `Session.replay` raise `ValueError` for more than 200,000 rows (`rcb_traj.MAX_T`, the verifier's `MAX_T` in `tests/grade.py`: 800 s of physics), and `record_stop` warns on stderr. Before, the tool saved and replayed such a file ("success") and the verifier then graded it `bad_trajectory`; the instruction has stated the cap since 2026-10-10. The verifier, the scene, the instance and the pass rule are unchanged: every trajectory is graded as before.
; `solution/` rebuilds the scene with upstream's `load_actors` to get them back for the expert. None of them is read by `check_success`. The scene, the frozen instance (same initial-state hash), the tooling, the verifier and the pass rule are unchanged, so the task keeps its name; privileged results from before this date were obtained with these values reachable.
- 2026-10-13 (board 0100, owner decision): base image `eai-robotwin` 0.1.2. The image's scrub of RoboTwin's expert kept `load_actors` whole, so values it computes only for the removed expert stayed reachable: `env.target_pose` (poses the expert moves to or places at), and the instruction listed them among the scene objects. They are gone: the env no longer has them, `/opt/RoboTwin/envs/move_can_pot.py` no longer shows them, and the scene-object list no longer names them; `solution/` rebuilds the scene with upstream's `load_actors` to get them back for the expert. None of them is read by `check_success`. The scene, the frozen instance (same initial-state hash), the tooling, the verifier and the pass rule are unchanged, so the task keeps its name; privileged results from before this date were obtained with these values reachable.
+- 2026-10-14 (board 0113): the tooling only (`environment/rcb_traj.py`, `tests/rcb_traj.py`): while it builds the scene, `glob.glob` returns its matches sorted, so that a task which lists files to draw an object depends on its seed alone, not on the host's directory order (four RoboTwin tasks did; this one lists no files while it builds its scene). Its scene, frozen instance (same initial-state hash), verifier and pass rule are unchanged.
diff --git a/tasks/robotwin_2_0_move_can_pot-privileged/environment/rcb_traj.py b/tasks/robotwin_2_0_move_can_pot-privileged/environment/rcb_traj.py
index 9ddd80960b..ca2d2369e6 100644
--- a/tasks/robotwin_2_0_move_can_pot-privileged/environment/rcb_traj.py
+++ b/tasks/robotwin_2_0_move_can_pot-privileged/environment/rcb_traj.py
@@ -113,6 +113,26 @@ BASE_ARGS: dict[str, Any] = {
}
+class _SortedGlob:
+ """While a scene is built, glob.glob / glob.iglob return their matches sorted. Four RoboTwin tasks (place_a2b_left,
+ place_a2b_right, place_object_scale, place_object_stand: envs/<task>.py load_actors) list an object's model files
+ with glob.glob and then draw one with the seeded np.random.choice; glob returns the file system's directory order,
+ which differs between hosts (and image layers), so one seed built a different object on another host and the frozen
+ instance's hash check failed there (board 0086 claim 5; never rebuilt off the reference workstation). Sorted, the
+ draw depends on the seed alone. No other task of ours lists files while it builds its scene."""
+
+ def __enter__(self):
+ import glob
+ self._glob, self._orig = glob, (glob.glob, glob.iglob)
+ glob.glob = lambda *a, **k: sorted(self._orig[0](*a, **k))
+ glob.iglob = lambda *a, **k: iter(sorted(self._orig[1](*a, **k)))
+ return self
+
+ def __exit__(self, *exc):
+ self._glob.glob, self._glob.iglob = self._orig
+ return False
+
+
def _chdir_root() -> None:
# RoboTwin resolves assets/config paths relative to the repo root and via ROOT_PATH
if os.getcwd() != ROBOTWIN_ROOT and os.path.isdir(ROBOTWIN_ROOT):
@@ -218,7 +238,8 @@ class Session:
if self.enable_render:
sapien.render.set_camera_shader_dir(self.render_shader)
args = _build_args(self.task, self.embodiment, self.seed)
- self.env.setup_demo(**args)
+ with _SortedGlob(): # the object draw depends on the seed alone, not the host's directory order
+ self.env.setup_demo(**args)
if self.enable_render: # setup_scene() switches the shader to ray tracing; put it back for cheap frames
sapien.render.set_camera_shader_dir(self.render_shader)
self._raw_scene = self.env.scene
diff --git a/tasks/robotwin_2_0_move_can_pot-privileged/task.toml b/tasks/robotwin_2_0_move_can_pot-privileged/task.toml
index 8f0ab6f597..a2bf744cfb 100644
--- a/tasks/robotwin_2_0_move_can_pot-privileged/task.toml
+++ b/tasks/robotwin_2_0_move_can_pot-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "robotwin" # RoboTwin 2.0 = SAPIEN 3 / PhysX + CuRobo
robot = "aloha-agilex"
family = "move_can_pot"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 60417 # scene = environment/instance.json (RoboTwin seed + expected initial-state hash); identity = dataset.toml digest
expert_time_estimate_hours = 0.5
diff --git a/tasks/robotwin_2_0_move_can_pot-privileged/tests/rcb_traj.py b/tasks/robotwin_2_0_move_can_pot-privileged/tests/rcb_traj.py
index 9ddd80960b..ca2d2369e6 100644
--- a/tasks/robotwin_2_0_move_can_pot-privileged/tests/rcb_traj.py
+++ b/tasks/robotwin_2_0_move_can_pot-privileged/tests/rcb_traj.py
@@ -113,6 +113,26 @@ BASE_ARGS: dict[str, Any] = {
}
+class _SortedGlob:
+ """While a scene is built, glob.glob / glob.iglob return their matches sorted. Four RoboTwin tasks (place_a2b_left,
+ place_a2b_right, place_object_scale, place_object_stand: envs/<task>.py load_actors) list an object's model files
+ with glob.glob and then draw one with the seeded np.random.choice; glob returns the file system's directory order,
+ which differs between hosts (and image layers), so one seed built a different object on another host and the frozen
+ instance's hash check failed there (board 0086 claim 5; never rebuilt off the reference workstation). Sorted, the
+ draw depends on the seed alone. No other task of ours lists files while it builds its scene."""
+
+ def __enter__(self):
+ import glob
+ self._glob, self._orig = glob, (glob.glob, glob.iglob)
+ glob.glob = lambda *a, **k: sorted(self._orig[0](*a, **k))
+ glob.iglob = lambda *a, **k: iter(sorted(self._orig[1](*a, **k)))
+ return self
+
+ def __exit__(self, *exc):
+ self._glob.glob, self._glob.iglob = self._orig
+ return False
+
+
def _chdir_root() -> None:
# RoboTwin resolves assets/config paths relative to the repo root and via ROOT_PATH
if os.getcwd() != ROBOTWIN_ROOT and os.path.isdir(ROBOTWIN_ROOT):
@@ -218,7 +238,8 @@ class Session:
if self.enable_render:
sapien.render.set_camera_shader_dir(self.render_shader)
args = _build_args(self.task, self.embodiment, self.seed)
- self.env.setup_demo(**args)
+ with _SortedGlob(): # the object draw depends on the seed alone, not the host's directory order
+ self.env.setup_demo(**args)
if self.enable_render: # setup_scene() switches the shader to ray tracing; put it back for cheap frames
sapien.render.set_camera_shader_dir(self.render_shader)
self._raw_scene = self.env.scene
diff --git a/tasks/robotwin_2_0_move_can_pot-standard/README.md b/tasks/robotwin_2_0_move_can_pot-standard/README.md
index d9ed8ae068..a10bbc60fe 100644
--- a/tasks/robotwin_2_0_move_can_pot-standard/README.md
+++ b/tasks/robotwin_2_0_move_can_pot-standard/README.md
@@ -53,3 +53,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
- 2026-10-10 (owner decision, board 0072): the agent's network egress is restricted to the model APIs (`[agent] network_mode = "allowlist"`, the hosts of `scripts/set_agent_egress.py`; since 2026-10-11 these include the package registries, PyPI and the Debian / Ubuntu archives, so that `pip install` and `apt-get install` work); it was open before (Harbor's default), and agents could download upstream sources, reference solutions and robot models from their own shell. The scene, the frozen instance, the tooling and the pass rule are unchanged, so the task keeps its name; results from before this date were obtained with open egress.
- 2026-10-13 (board 0086): the service's copy of the tooling only (`environment/sim/rcb_traj.py`, one line of `environment/sim/server.py`). The tooling now refuses to save or replay a trajectory of more than 200,000 rows, the privileged verifier's cap; a standard episode has no row cap, so the service lifts it for its own session (`S.max_rows = None`). Nothing an agent sees and nothing the service grades changes.
- 2026-10-13 (board 0100): base image `eai-robotwin` 0.1.2, whose scrub of RoboTwin's expert also removes the reference poses that `load_actors` computed only for the expert (here `target_pose`). The standard-mode agent never reaches the env, so nothing it observes changes; the scene, the instance and the judge are unchanged.
+- 2026-10-14 (board 0113): the tooling only (`environment/sim/rcb_traj.py`): while it builds the scene, `glob.glob` returns its matches sorted, so that a task which lists files to draw an object depends on its seed alone, not on the host's directory order (four RoboTwin tasks did; this one lists no files while it builds its scene). Its scene, frozen instance (same initial-state hash), verifier and pass rule are unchanged.
diff --git a/tasks/robotwin_2_0_move_can_pot-standard/environment/sim/rcb_traj.py b/tasks/robotwin_2_0_move_can_pot-standard/environment/sim/rcb_traj.py
index 9ddd80960b..ca2d2369e6 100644
--- a/tasks/robotwin_2_0_move_can_pot-standard/environment/sim/rcb_traj.py
+++ b/tasks/robotwin_2_0_move_can_pot-standard/environment/sim/rcb_traj.py
@@ -113,6 +113,26 @@ BASE_ARGS: dict[str, Any] = {
}
+class _SortedGlob:
+ """While a scene is built, glob.glob / glob.iglob return their matches sorted. Four RoboTwin tasks (place_a2b_left,
+ place_a2b_right, place_object_scale, place_object_stand: envs/<task>.py load_actors) list an object's model files
+ with glob.glob and then draw one with the seeded np.random.choice; glob returns the file system's directory order,
+ which differs between hosts (and image layers), so one seed built a different object on another host and the frozen
+ instance's hash check failed there (board 0086 claim 5; never rebuilt off the reference workstation). Sorted, the
+ draw depends on the seed alone. No other task of ours lists files while it builds its scene."""
+
+ def __enter__(self):
+ import glob
+ self._glob, self._orig = glob, (glob.glob, glob.iglob)
+ glob.glob = lambda *a, **k: sorted(self._orig[0](*a, **k))
+ glob.iglob = lambda *a, **k: iter(sorted(self._orig[1](*a, **k)))
+ return self
+
+ def __exit__(self, *exc):
+ self._glob.glob, self._glob.iglob = self._orig
+ return False
+
+
def _chdir_root() -> None:
# RoboTwin resolves assets/config paths relative to the repo root and via ROOT_PATH
if os.getcwd() != ROBOTWIN_ROOT and os.path.isdir(ROBOTWIN_ROOT):
@@ -218,7 +238,8 @@ class Session:
if self.enable_render:
sapien.render.set_camera_shader_dir(self.render_shader)
args = _build_args(self.task, self.embodiment, self.seed)
- self.env.setup_demo(**args)
+ with _SortedGlob(): # the object draw depends on the seed alone, not the host's directory order
+ self.env.setup_demo(**args)
if self.enable_render: # setup_scene() switches the shader to ray tracing; put it back for cheap frames
sapien.render.set_camera_shader_dir(self.render_shader)
self._raw_scene = self.env.scene
diff --git a/tasks/robotwin_2_0_move_can_pot-standard/task.toml b/tasks/robotwin_2_0_move_can_pot-standard/task.toml
index 6683ce7983..8e0bef37a2 100644
--- a/tasks/robotwin_2_0_move_can_pot-standard/task.toml
+++ b/tasks/robotwin_2_0_move_can_pot-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "robotwin"
robot = "aloha-agilex"
family = "move_can_pot"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 60417
diff --git a/tasks/robotwin_2_0_move_pillbottle_pad-privileged/README.md b/tasks/robotwin_2_0_move_pillbottle_pad-privileged/README.md
index a8243e1cad..76e364f874 100644
--- a/tasks/robotwin_2_0_move_pillbottle_pad-privileged/README.md
+++ b/tasks/robotwin_2_0_move_pillbottle_pad-privileged/README.md
@@ -55,3 +55,4 @@ Use one arm to pick the pillbottle and place it onto the pad.
- 2026-10-10 (owner decision, board 0072): the agent's network egress is restricted to the model APIs (`[agent] network_mode = "allowlist"`, the hosts of `scripts/set_agent_egress.py`; since 2026-10-11 these include the package registries, PyPI and the Debian / Ubuntu archives, so that `pip install` and `apt-get install` work); it was open before (Harbor's default), and agents could download upstream sources, reference solutions and robot models from their own shell. The scene, the frozen instance, the tooling and the pass rule are unchanged, so the task keeps its name; results from before this date were obtained with open egress.
- 2026-10-13 (board 0086): the tooling refuses a trajectory the verifier cannot grade (`environment/rcb_traj.py`, `tests/rcb_traj.py`). `Session.save_trajectory` and `Session.replay` raise `ValueError` for more than 200,000 rows (`rcb_traj.MAX_T`, the verifier's `MAX_T` in `tests/grade.py`: 800 s of physics), and `record_stop` warns on stderr. Before, the tool saved and replayed such a file ("success") and the verifier then graded it `bad_trajectory`; the instruction has stated the cap since 2026-10-10. The verifier, the scene, the instance and the pass rule are unchanged: every trajectory is graded as before.
- 2026-10-13 (board 0100): base image `eai-robotwin` 0.1.2. Its scrub of RoboTwin's expert also removes the reference poses that 18 tasks' `load_actors` computed only for the expert, and the expert's comments above the scrubbed method bodies (`images/robotwin/scrub_oracles.py`). This task's `load_actors` set no such value: its scene, instance, tooling, verifier and pass rule are unchanged.
+- 2026-10-14 (board 0113): the tooling only (`environment/rcb_traj.py`, `tests/rcb_traj.py`): while it builds the scene, `glob.glob` returns its matches sorted, so that a task which lists files to draw an object depends on its seed alone, not on the host's directory order (four RoboTwin tasks did; this one lists no files while it builds its scene). Its scene, frozen instance (same initial-state hash), verifier and pass rule are unchanged.
diff --git a/tasks/robotwin_2_0_move_pillbottle_pad-privileged/environment/rcb_traj.py b/tasks/robotwin_2_0_move_pillbottle_pad-privileged/environment/rcb_traj.py
index 1eb1fbd39a..75eaa73e28 100644
--- a/tasks/robotwin_2_0_move_pillbottle_pad-privileged/environment/rcb_traj.py
+++ b/tasks/robotwin_2_0_move_pillbottle_pad-privileged/environment/rcb_traj.py
@@ -113,6 +113,26 @@ BASE_ARGS: dict[str, Any] = {
}
+class _SortedGlob:
+ """While a scene is built, glob.glob / glob.iglob return their matches sorted. Four RoboTwin tasks (place_a2b_left,
+ place_a2b_right, place_object_scale, place_object_stand: envs/<task>.py load_actors) list an object's model files
+ with glob.glob and then draw one with the seeded np.random.choice; glob returns the file system's directory order,
+ which differs between hosts (and image layers), so one seed built a different object on another host and the frozen
+ instance's hash check failed there (board 0086 claim 5; never rebuilt off the reference workstation). Sorted, the
+ draw depends on the seed alone. No other task of ours lists files while it builds its scene."""
+
+ def __enter__(self):
+ import glob
+ self._glob, self._orig = glob, (glob.glob, glob.iglob)
+ glob.glob = lambda *a, **k: sorted(self._orig[0](*a, **k))
+ glob.iglob = lambda *a, **k: iter(sorted(self._orig[1](*a, **k)))
+ return self
+
+ def __exit__(self, *exc):
+ self._glob.glob, self._glob.iglob = self._orig
+ return False
+
+
def _chdir_root() -> None:
# RoboTwin resolves assets/config paths relative to the repo root and via ROOT_PATH
if os.getcwd() != ROBOTWIN_ROOT and os.path.isdir(ROBOTWIN_ROOT):
@@ -218,7 +238,8 @@ class Session:
if self.enable_render:
sapien.render.set_camera_shader_dir(self.render_shader)
args = _build_args(self.task, self.embodiment, self.seed)
- self.env.setup_demo(**args)
+ with _SortedGlob(): # the object draw depends on the seed alone, not the host's directory order
+ self.env.setup_demo(**args)
if self.enable_render: # setup_scene() switches the shader to ray tracing; put it back for cheap frames
sapien.render.set_camera_shader_dir(self.render_shader)
self._raw_scene = self.env.scene
diff --git a/tasks/robotwin_2_0_move_pillbottle_pad-privileged/task.toml b/tasks/robotwin_2_0_move_pillbottle_pad-privileged/task.toml
index dd74b364ea..214b825821 100644
--- a/tasks/robotwin_2_0_move_pillbottle_pad-privileged/task.toml
+++ b/tasks/robotwin_2_0_move_pillbottle_pad-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "robotwin" # RoboTwin 2.0 = SAPIEN 3 / PhysX + CuRobo
robot = "aloha-agilex"
family = "move_pillbottle_pad"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 60417 # scene = environment/instance.json (RoboTwin seed + expected initial-state hash); identity = dataset.toml digest
expert_time_estimate_hours = 0.5
diff --git a/tasks/robotwin_2_0_move_pillbottle_pad-privileged/tests/rcb_traj.py b/tasks/robotwin_2_0_move_pillbottle_pad-privileged/tests/rcb_traj.py
index 1eb1fbd39a..75eaa73e28 100644
--- a/tasks/robotwin_2_0_move_pillbottle_pad-privileged/tests/rcb_traj.py
+++ b/tasks/robotwin_2_0_move_pillbottle_pad-privileged/tests/rcb_traj.py
@@ -113,6 +113,26 @@ BASE_ARGS: dict[str, Any] = {
}
+class _SortedGlob:
+ """While a scene is built, glob.glob / glob.iglob return their matches sorted. Four RoboTwin tasks (place_a2b_left,
+ place_a2b_right, place_object_scale, place_object_stand: envs/<task>.py load_actors) list an object's model files
+ with glob.glob and then draw one with the seeded np.random.choice; glob returns the file system's directory order,
+ which differs between hosts (and image layers), so one seed built a different object on another host and the frozen
+ instance's hash check failed there (board 0086 claim 5; never rebuilt off the reference workstation). Sorted, the
+ draw depends on the seed alone. No other task of ours lists files while it builds its scene."""
+
+ def __enter__(self):
+ import glob
+ self._glob, self._orig = glob, (glob.glob, glob.iglob)
+ glob.glob = lambda *a, **k: sorted(self._orig[0](*a, **k))
+ glob.iglob = lambda *a, **k: iter(sorted(self._orig[1](*a, **k)))
+ return self
+
+ def __exit__(self, *exc):
+ self._glob.glob, self._glob.iglob = self._orig
+ return False
+
+
def _chdir_root() -> None:
# RoboTwin resolves assets/config paths relative to the repo root and via ROOT_PATH
if os.getcwd() != ROBOTWIN_ROOT and os.path.isdir(ROBOTWIN_ROOT):
@@ -218,7 +238,8 @@ class Session:
if self.enable_render:
sapien.render.set_camera_shader_dir(self.render_shader)
args = _build_args(self.task, self.embodiment, self.seed)
- self.env.setup_demo(**args)
+ with _SortedGlob(): # the object draw depends on the seed alone, not the host's directory order
+ self.env.setup_demo(**args)
if self.enable_render: # setup_scene() switches the shader to ray tracing; put it back for cheap frames
sapien.render.set_camera_shader_dir(self.render_shader)
self._raw_scene = self.env.scene
diff --git a/tasks/robotwin_2_0_move_pillbottle_pad-standard/README.md b/tasks/robotwin_2_0_move_pillbottle_pad-standard/README.md
index 6f81368843..76052c5b26 100644
--- a/tasks/robotwin_2_0_move_pillbottle_pad-standard/README.md
+++ b/tasks/robotwin_2_0_move_pillbottle_pad-standard/README.md
@@ -53,3 +53,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
- 2026-10-10 (owner decision, board 0072): the agent's network egress is restricted to the model APIs (`[agent] network_mode = "allowlist"`, the hosts of `scripts/set_agent_egress.py`; since 2026-10-11 these include the package registries, PyPI and the Debian / Ubuntu archives, so that `pip install` and `apt-get install` work); it was open before (Harbor's default), and agents could download upstream sources, reference solutions and robot models from their own shell. The scene, the frozen instance, the tooling and the pass rule are unchanged, so the task keeps its name; results from before this date were obtained with open egress.
- 2026-10-13 (board 0086): the service's copy of the tooling only (`environment/sim/rcb_traj.py`, one line of `environment/sim/server.py`). The tooling now refuses to save or replay a trajectory of more than 200,000 rows, the privileged verifier's cap; a standard episode has no row cap, so the service lifts it for its own session (`S.max_rows = None`). Nothing an agent sees and nothing the service grades changes.
- 2026-10-13 (board 0100): base image `eai-robotwin` 0.1.2, whose scrub of RoboTwin's expert also removes the reference poses that `load_actors` computed only for the expert. The standard-mode agent never reaches the env, so nothing it observes changes; the scene, the instance and the judge are unchanged.
+- 2026-10-14 (board 0113): the tooling only (`environment/sim/rcb_traj.py`): while it builds the scene, `glob.glob` returns its matches sorted, so that a task which lists files to draw an object depends on its seed alone, not on the host's directory order (four RoboTwin tasks did; this one lists no files while it builds its scene). Its scene, frozen instance (same initial-state hash), verifier and pass rule are unchanged.
diff --git a/tasks/robotwin_2_0_move_pillbottle_pad-standard/environment/sim/rcb_traj.py b/tasks/robotwin_2_0_move_pillbottle_pad-standard/environment/sim/rcb_traj.py
index 1eb1fbd39a..75eaa73e28 100644
--- a/tasks/robotwin_2_0_move_pillbottle_pad-standard/environment/sim/rcb_traj.py
+++ b/tasks/robotwin_2_0_move_pillbottle_pad-standard/environment/sim/rcb_traj.py
@@ -113,6 +113,26 @@ BASE_ARGS: dict[str, Any] = {
}
+class _SortedGlob:
+ """While a scene is built, glob.glob / glob.iglob return their matches sorted. Four RoboTwin tasks (place_a2b_left,
+ place_a2b_right, place_object_scale, place_object_stand: envs/<task>.py load_actors) list an object's model files
+ with glob.glob and then draw one with the seeded np.random.choice; glob returns the file system's directory order,
+ which differs between hosts (and image layers), so one seed built a different object on another host and the frozen
+ instance's hash check failed there (board 0086 claim 5; never rebuilt off the reference workstation). Sorted, the
+ draw depends on the seed alone. No other task of ours lists files while it builds its scene."""
+
+ def __enter__(self):
+ import glob
+ self._glob, self._orig = glob, (glob.glob, glob.iglob)
+ glob.glob = lambda *a, **k: sorted(self._orig[0](*a, **k))
+ glob.iglob = lambda *a, **k: iter(sorted(self._orig[1](*a, **k)))
+ return self
+
+ def __exit__(self, *exc):
+ self._glob.glob, self._glob.iglob = self._orig
+ return False
+
+
def _chdir_root() -> None:
# RoboTwin resolves assets/config paths relative to the repo root and via ROOT_PATH
if os.getcwd() != ROBOTWIN_ROOT and os.path.isdir(ROBOTWIN_ROOT):
@@ -218,7 +238,8 @@ class Session:
if self.enable_render:
sapien.render.set_camera_shader_dir(self.render_shader)
args = _build_args(self.task, self.embodiment, self.seed)
- self.env.setup_demo(**args)
+ with _SortedGlob(): # the object draw depends on the seed alone, not the host's directory order
+ self.env.setup_demo(**args)
if self.enable_render: # setup_scene() switches the shader to ray tracing; put it back for cheap frames
sapien.render.set_camera_shader_dir(self.render_shader)
self._raw_scene = self.env.scene
diff --git a/tasks/robotwin_2_0_move_pillbottle_pad-standard/task.toml b/tasks/robotwin_2_0_move_pillbottle_pad-standard/task.toml
index bed7ec62af..5e91665e49 100644
--- a/tasks/robotwin_2_0_move_pillbottle_pad-standard/task.toml
+++ b/tasks/robotwin_2_0_move_pillbottle_pad-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "robotwin"
robot = "aloha-agilex"
family = "move_pillbottle_pad"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 60417
diff --git a/tasks/robotwin_2_0_move_playingcard_away-privileged/README.md b/tasks/robotwin_2_0_move_playingcard_away-privileged/README.md
index 7760f3c101..d30ebb1149 100644
--- a/tasks/robotwin_2_0_move_playingcard_away-privileged/README.md
+++ b/tasks/robotwin_2_0_move_playingcard_away-privileged/README.md
@@ -55,3 +55,4 @@ Use the arm to pick up the playing card and move it away from the table. For exa
- 2026-10-10 (owner decision, board 0072): the agent's network egress is restricted to the model APIs (`[agent] network_mode = "allowlist"`, the hosts of `scripts/set_agent_egress.py`; since 2026-10-11 these include the package registries, PyPI and the Debian / Ubuntu archives, so that `pip install` and `apt-get install` work); it was open before (Harbor's default), and agents could download upstream sources, reference solutions and robot models from their own shell. The scene, the frozen instance, the tooling and the pass rule are unchanged, so the task keeps its name; results from before this date were obtained with open egress.
- 2026-10-13 (board 0086): the tooling refuses a trajectory the verifier cannot grade (`environment/rcb_traj.py`, `tests/rcb_traj.py`). `Session.save_trajectory` and `Session.replay` raise `ValueError` for more than 200,000 rows (`rcb_traj.MAX_T`, the verifier's `MAX_T` in `tests/grade.py`: 800 s of physics), and `record_stop` warns on stderr. Before, the tool saved and replayed such a file ("success") and the verifier then graded it `bad_trajectory`; the instruction has stated the cap since 2026-10-10. The verifier, the scene, the instance and the pass rule are unchanged: every trajectory is graded as before.
- 2026-10-13 (board 0100): base image `eai-robotwin` 0.1.2. Its scrub of RoboTwin's expert also removes the reference poses that 18 tasks' `load_actors` computed only for the expert, and the expert's comments above the scrubbed method bodies (`images/robotwin/scrub_oracles.py`). This task's `load_actors` set no such value: its scene, instance, tooling, verifier and pass rule are unchanged.
+- 2026-10-14 (board 0113): the tooling only (`environment/rcb_traj.py`, `tests/rcb_traj.py`): while it builds the scene, `glob.glob` returns its matches sorted, so that a task which lists files to draw an object depends on its seed alone, not on the host's directory order (four RoboTwin tasks did; this one lists no files while it builds its scene). Its scene, frozen instance (same initial-state hash), verifier and pass rule are unchanged.
diff --git a/tasks/robotwin_2_0_move_playingcard_away-privileged/environment/rcb_traj.py b/tasks/robotwin_2_0_move_playingcard_away-privileged/environment/rcb_traj.py
index a69766de6b..43d3acdab3 100644
--- a/tasks/robotwin_2_0_move_playingcard_away-privileged/environment/rcb_traj.py
+++ b/tasks/robotwin_2_0_move_playingcard_away-privileged/environment/rcb_traj.py
@@ -113,6 +113,26 @@ BASE_ARGS: dict[str, Any] = {
}
+class _SortedGlob:
+ """While a scene is built, glob.glob / glob.iglob return their matches sorted. Four RoboTwin tasks (place_a2b_left,
+ place_a2b_right, place_object_scale, place_object_stand: envs/<task>.py load_actors) list an object's model files
+ with glob.glob and then draw one with the seeded np.random.choice; glob returns the file system's directory order,
+ which differs between hosts (and image layers), so one seed built a different object on another host and the frozen
+ instance's hash check failed there (board 0086 claim 5; never rebuilt off the reference workstation). Sorted, the
+ draw depends on the seed alone. No other task of ours lists files while it builds its scene."""
+
+ def __enter__(self):
+ import glob
+ self._glob, self._orig = glob, (glob.glob, glob.iglob)
+ glob.glob = lambda *a, **k: sorted(self._orig[0](*a, **k))
+ glob.iglob = lambda *a, **k: iter(sorted(self._orig[1](*a, **k)))
+ return self
+
+ def __exit__(self, *exc):
+ self._glob.glob, self._glob.iglob = self._orig
+ return False
+
+
def _chdir_root() -> None:
# RoboTwin resolves assets/config paths relative to the repo root and via ROOT_PATH
if os.getcwd() != ROBOTWIN_ROOT and os.path.isdir(ROBOTWIN_ROOT):
@@ -218,7 +238,8 @@ class Session:
if self.enable_render:
sapien.render.set_camera_shader_dir(self.render_shader)
args = _build_args(self.task, self.embodiment, self.seed)
- self.env.setup_demo(**args)
+ with _SortedGlob(): # the object draw depends on the seed alone, not the host's directory order
+ self.env.setup_demo(**args)
if self.enable_render: # setup_scene() switches the shader to ray tracing; put it back for cheap frames
sapien.render.set_camera_shader_dir(self.render_shader)
self._raw_scene = self.env.scene
diff --git a/tasks/robotwin_2_0_move_playingcard_away-privileged/task.toml b/tasks/robotwin_2_0_move_playingcard_away-privileged/task.toml
index 17b1b74b88..7f3f5e5752 100644
--- a/tasks/robotwin_2_0_move_playingcard_away-privileged/task.toml
+++ b/tasks/robotwin_2_0_move_playingcard_away-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "robotwin" # RoboTwin 2.0 = SAPIEN 3 / PhysX + CuRobo
robot = "aloha-agilex"
family = "move_playingcard_away"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 60417 # scene = environment/instance.json (RoboTwin seed + expected initial-state hash); identity = dataset.toml digest
expert_time_estimate_hours = 0.5
diff --git a/tasks/robotwin_2_0_move_playingcard_away-privileged/tests/rcb_traj.py b/tasks/robotwin_2_0_move_playingcard_away-privileged/tests/rcb_traj.py
index a69766de6b..43d3acdab3 100644
--- a/tasks/robotwin_2_0_move_playingcard_away-privileged/tests/rcb_traj.py
+++ b/tasks/robotwin_2_0_move_playingcard_away-privileged/tests/rcb_traj.py
@@ -113,6 +113,26 @@ BASE_ARGS: dict[str, Any] = {
}
+class _SortedGlob:
+ """While a scene is built, glob.glob / glob.iglob return their matches sorted. Four RoboTwin tasks (place_a2b_left,
+ place_a2b_right, place_object_scale, place_object_stand: envs/<task>.py load_actors) list an object's model files
+ with glob.glob and then draw one with the seeded np.random.choice; glob returns the file system's directory order,
+ which differs between hosts (and image layers), so one seed built a different object on another host and the frozen
+ instance's hash check failed there (board 0086 claim 5; never rebuilt off the reference workstation). Sorted, the
+ draw depends on the seed alone. No other task of ours lists files while it builds its scene."""
+
+ def __enter__(self):
+ import glob
+ self._glob, self._orig = glob, (glob.glob, glob.iglob)
+ glob.glob = lambda *a, **k: sorted(self._orig[0](*a, **k))
+ glob.iglob = lambda *a, **k: iter(sorted(self._orig[1](*a, **k)))
+ return self
+
+ def __exit__(self, *exc):
+ self._glob.glob, self._glob.iglob = self._orig
+ return False
+
+
def _chdir_root() -> None:
# RoboTwin resolves assets/config paths relative to the repo root and via ROOT_PATH
if os.getcwd() != ROBOTWIN_ROOT and os.path.isdir(ROBOTWIN_ROOT):
@@ -218,7 +238,8 @@ class Session:
if self.enable_render:
sapien.render.set_camera_shader_dir(self.render_shader)
args = _build_args(self.task, self.embodiment, self.seed)
- self.env.setup_demo(**args)
+ with _SortedGlob(): # the object draw depends on the seed alone, not the host's directory order
+ self.env.setup_demo(**args)
if self.enable_render: # setup_scene() switches the shader to ray tracing; put it back for cheap frames
sapien.render.set_camera_shader_dir(self.render_shader)
self._raw_scene = self.env.scene
diff --git a/tasks/robotwin_2_0_move_playingcard_away-standard/README.md b/tasks/robotwin_2_0_move_playingcard_away-standard/README.md
index dc8941a360..471285fbbf 100644
--- a/tasks/robotwin_2_0_move_playingcard_away-standard/README.md
+++ b/tasks/robotwin_2_0_move_playingcard_away-standard/README.md
@@ -53,3 +53,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
- 2026-10-10 (owner decision, board 0072): the agent's network egress is restricted to the model APIs (`[agent] network_mode = "allowlist"`, the hosts of `scripts/set_agent_egress.py`; since 2026-10-11 these include the package registries, PyPI and the Debian / Ubuntu archives, so that `pip install` and `apt-get install` work); it was open before (Harbor's default), and agents could download upstream sources, reference solutions and robot models from their own shell. The scene, the frozen instance, the tooling and the pass rule are unchanged, so the task keeps its name; results from before this date were obtained with open egress.
- 2026-10-13 (board 0086): the service's copy of the tooling only (`environment/sim/rcb_traj.py`, one line of `environment/sim/server.py`). The tooling now refuses to save or replay a trajectory of more than 200,000 rows, the privileged verifier's cap; a standard episode has no row cap, so the service lifts it for its own session (`S.max_rows = None`). Nothing an agent sees and nothing the service grades changes.
- 2026-10-13 (board 0100): base image `eai-robotwin` 0.1.2, whose scrub of RoboTwin's expert also removes the reference poses that `load_actors` computed only for the expert. The standard-mode agent never reaches the env, so nothing it observes changes; the scene, the instance and the judge are unchanged.
+- 2026-10-14 (board 0113): the tooling only (`environment/sim/rcb_traj.py`): while it builds the scene, `glob.glob` returns its matches sorted, so that a task which lists files to draw an object depends on its seed alone, not on the host's directory order (four RoboTwin tasks did; this one lists no files while it builds its scene). Its scene, frozen instance (same initial-state hash), verifier and pass rule are unchanged.
diff --git a/tasks/robotwin_2_0_move_playingcard_away-standard/environment/sim/rcb_traj.py b/tasks/robotwin_2_0_move_playingcard_away-standard/environment/sim/rcb_traj.py
index a69766de6b..43d3acdab3 100644
--- a/tasks/robotwin_2_0_move_playingcard_away-standard/environment/sim/rcb_traj.py
+++ b/tasks/robotwin_2_0_move_playingcard_away-standard/environment/sim/rcb_traj.py
@@ -113,6 +113,26 @@ BASE_ARGS: dict[str, Any] = {
}
+class _SortedGlob:
+ """While a scene is built, glob.glob / glob.iglob return their matches sorted. Four RoboTwin tasks (place_a2b_left,
+ place_a2b_right, place_object_scale, place_object_stand: envs/<task>.py load_actors) list an object's model files
+ with glob.glob and then draw one with the seeded np.random.choice; glob returns the file system's directory order,
+ which differs between hosts (and image layers), so one seed built a different object on another host and the frozen
+ instance's hash check failed there (board 0086 claim 5; never rebuilt off the reference workstation). Sorted, the
+ draw depends on the seed alone. No other task of ours lists files while it builds its scene."""
+
+ def __enter__(self):
+ import glob
+ self._glob, self._orig = glob, (glob.glob, glob.iglob)
+ glob.glob = lambda *a, **k: sorted(self._orig[0](*a, **k))
+ glob.iglob = lambda *a, **k: iter(sorted(self._orig[1](*a, **k)))
+ return self
+
+ def __exit__(self, *exc):
+ self._glob.glob, self._glob.iglob = self._orig
+ return False
+
+
def _chdir_root() -> None:
# RoboTwin resolves assets/config paths relative to the repo root and via ROOT_PATH
if os.getcwd() != ROBOTWIN_ROOT and os.path.isdir(ROBOTWIN_ROOT):
@@ -218,7 +238,8 @@ class Session:
if self.enable_render:
sapien.render.set_camera_shader_dir(self.render_shader)
args = _build_args(self.task, self.embodiment, self.seed)
- self.env.setup_demo(**args)
+ with _SortedGlob(): # the object draw depends on the seed alone, not the host's directory order
+ self.env.setup_demo(**args)
if self.enable_render: # setup_scene() switches the shader to ray tracing; put it back for cheap frames
sapien.render.set_camera_shader_dir(self.render_shader)
self._raw_scene = self.env.scene
diff --git a/tasks/robotwin_2_0_move_playingcard_away-standard/task.toml b/tasks/robotwin_2_0_move_playingcard_away-standard/task.toml
index b49e8e3e8c..a6cdeea5f7 100644
--- a/tasks/robotwin_2_0_move_playingcard_away-standard/task.toml
+++ b/tasks/robotwin_2_0_move_playingcard_away-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "robotwin"
robot = "aloha-agilex"
family = "move_playingcard_away"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 60417
diff --git a/tasks/robotwin_2_0_move_stapler_pad-privileged/README.md b/tasks/robotwin_2_0_move_stapler_pad-privileged/README.md
index 77922c1f60..8be539d9de 100644
--- a/tasks/robotwin_2_0_move_stapler_pad-privileged/README.md
+++ b/tasks/robotwin_2_0_move_stapler_pad-privileged/README.md
@@ -56,3 +56,4 @@ Use appropriate arm to move the stapler to a colored mat.
- 2026-10-13 (board 0086): the tooling refuses a trajectory the verifier cannot grade (`environment/rcb_traj.py`, `tests/rcb_traj.py`). `Session.save_trajectory` and `Session.replay` raise `ValueError` for more than 200,000 rows (`rcb_traj.MAX_T`, the verifier's `MAX_T` in `tests/grade.py`: 800 s of physics), and `record_stop` warns on stderr. Before, the tool saved and replayed such a file ("success") and the verifier then graded it `bad_trajectory`; the instruction has stated the cap since 2026-10-10. The verifier, the scene, the instance and the pass rule are unchanged: every trajectory is graded as before.
; `solution/` rebuilds the scene with upstream's `load_actors` to get them back for the expert. None of them is read by `check_success`. The scene, the frozen instance (same initial-state hash), the tooling, the verifier and the pass rule are unchanged, so the task keeps its name; privileged results from before this date were obtained with these values reachable.
- 2026-10-13 (board 0100, owner decision): base image `eai-robotwin` 0.1.2. The image's scrub of RoboTwin's expert kept `load_actors` whole, so values it computes only for the removed expert stayed reachable: `env.pad_pose` (poses the expert moves to or places at), and the instruction listed them among the scene objects. They are gone: the env no longer has them, `/opt/RoboTwin/envs/move_stapler_pad.py` no longer shows them, and the scene-object list no longer names them; `solution/` rebuilds the scene with upstream's `load_actors` to get them back for the expert. None of them is read by `check_success`. The scene, the frozen instance (same initial-state hash), the tooling, the verifier and the pass rule are unchanged, so the task keeps its name; privileged results from before this date were obtained with these values reachable.
+- 2026-10-14 (board 0113): the tooling only (`environment/rcb_traj.py`, `tests/rcb_traj.py`): while it builds the scene, `glob.glob` returns its matches sorted, so that a task which lists files to draw an object depends on its seed alone, not on the host's directory order (four RoboTwin tasks did; this one lists no files while it builds its scene). Its scene, frozen instance (same initial-state hash), verifier and pass rule are unchanged.
diff --git a/tasks/robotwin_2_0_move_stapler_pad-privileged/environment/rcb_traj.py b/tasks/robotwin_2_0_move_stapler_pad-privileged/environment/rcb_traj.py
index 4a6575a485..43c2d6d3c0 100644
--- a/tasks/robotwin_2_0_move_stapler_pad-privileged/environment/rcb_traj.py
+++ b/tasks/robotwin_2_0_move_stapler_pad-privileged/environment/rcb_traj.py
@@ -113,6 +113,26 @@ BASE_ARGS: dict[str, Any] = {
}
+class _SortedGlob:
+ """While a scene is built, glob.glob / glob.iglob return their matches sorted. Four RoboTwin tasks (place_a2b_left,
+ place_a2b_right, place_object_scale, place_object_stand: envs/<task>.py load_actors) list an object's model files
+ with glob.glob and then draw one with the seeded np.random.choice; glob returns the file system's directory order,
+ which differs between hosts (and image layers), so one seed built a different object on another host and the frozen
+ instance's hash check failed there (board 0086 claim 5; never rebuilt off the reference workstation). Sorted, the
+ draw depends on the seed alone. No other task of ours lists files while it builds its scene."""
+
+ def __enter__(self):
+ import glob
+ self._glob, self._orig = glob, (glob.glob, glob.iglob)
+ glob.glob = lambda *a, **k: sorted(self._orig[0](*a, **k))
+ glob.iglob = lambda *a, **k: iter(sorted(self._orig[1](*a, **k)))
+ return self
+
+ def __exit__(self, *exc):
+ self._glob.glob, self._glob.iglob = self._orig
+ return False
+
+
def _chdir_root() -> None:
# RoboTwin resolves assets/config paths relative to the repo root and via ROOT_PATH
if os.getcwd() != ROBOTWIN_ROOT and os.path.isdir(ROBOTWIN_ROOT):
@@ -218,7 +238,8 @@ class Session:
if self.enable_render:
sapien.render.set_camera_shader_dir(self.render_shader)
args = _build_args(self.task, self.embodiment, self.seed)
- self.env.setup_demo(**args)
+ with _SortedGlob(): # the object draw depends on the seed alone, not the host's directory order
+ self.env.setup_demo(**args)
if self.enable_render: # setup_scene() switches the shader to ray tracing; put it back for cheap frames
sapien.render.set_camera_shader_dir(self.render_shader)
self._raw_scene = self.env.scene
diff --git a/tasks/robotwin_2_0_move_stapler_pad-privileged/task.toml b/tasks/robotwin_2_0_move_stapler_pad-privileged/task.toml
index 9aeaaf89af..f7c98dd32a 100644
--- a/tasks/robotwin_2_0_move_stapler_pad-privileged/task.toml
+++ b/tasks/robotwin_2_0_move_stapler_pad-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "robotwin" # RoboTwin 2.0 = SAPIEN 3 / PhysX + CuRobo
robot = "aloha-agilex"
family = "move_stapler_pad"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 60417 # scene = environment/instance.json (RoboTwin seed + expected initial-state hash); identity = dataset.toml digest
expert_time_estimate_hours = 0.5
diff --git a/tasks/robotwin_2_0_move_stapler_pad-privileged/tests/rcb_traj.py b/tasks/robotwin_2_0_move_stapler_pad-privileged/tests/rcb_traj.py
index 4a6575a485..43c2d6d3c0 100644
--- a/tasks/robotwin_2_0_move_stapler_pad-privileged/tests/rcb_traj.py
+++ b/tasks/robotwin_2_0_move_stapler_pad-privileged/tests/rcb_traj.py
@@ -113,6 +113,26 @@ BASE_ARGS: dict[str, Any] = {
}
+class _SortedGlob:
+ """While a scene is built, glob.glob / glob.iglob return their matches sorted. Four RoboTwin tasks (place_a2b_left,
+ place_a2b_right, place_object_scale, place_object_stand: envs/<task>.py load_actors) list an object's model files
+ with glob.glob and then draw one with the seeded np.random.choice; glob returns the file system's directory order,
+ which differs between hosts (and image layers), so one seed built a different object on another host and the frozen
+ instance's hash check failed there (board 0086 claim 5; never rebuilt off the reference workstation). Sorted, the
+ draw depends on the seed alone. No other task of ours lists files while it builds its scene."""
+
+ def __enter__(self):
+ import glob
+ self._glob, self._orig = glob, (glob.glob, glob.iglob)
+ glob.glob = lambda *a, **k: sorted(self._orig[0](*a, **k))
+ glob.iglob = lambda *a, **k: iter(sorted(self._orig[1](*a, **k)))
+ return self
+
+ def __exit__(self, *exc):
+ self._glob.glob, self._glob.iglob = self._orig
+ return False
+
+
def _chdir_root() -> None:
# RoboTwin resolves assets/config paths relative to the repo root and via ROOT_PATH
if os.getcwd() != ROBOTWIN_ROOT and os.path.isdir(ROBOTWIN_ROOT):
@@ -218,7 +238,8 @@ class Session:
if self.enable_render:
sapien.render.set_camera_shader_dir(self.render_shader)
args = _build_args(self.task, self.embodiment, self.seed)
- self.env.setup_demo(**args)
+ with _SortedGlob(): # the object draw depends on the seed alone, not the host's directory order
+ self.env.setup_demo(**args)
if self.enable_render: # setup_scene() switches the shader to ray tracing; put it back for cheap frames
sapien.render.set_camera_shader_dir(self.render_shader)
self._raw_scene = self.env.scene
diff --git a/tasks/robotwin_2_0_move_stapler_pad-standard/README.md b/tasks/robotwin_2_0_move_stapler_pad-standard/README.md
index 0127a46c50..1f022cf322 100644
--- a/tasks/robotwin_2_0_move_stapler_pad-standard/README.md
+++ b/tasks/robotwin_2_0_move_stapler_pad-standard/README.md
@@ -53,3 +53,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
- 2026-10-10 (owner decision, board 0072): the agent's network egress is restricted to the model APIs (`[agent] network_mode = "allowlist"`, the hosts of `scripts/set_agent_egress.py`; since 2026-10-11 these include the package registries, PyPI and the Debian / Ubuntu archives, so that `pip install` and `apt-get install` work); it was open before (Harbor's default), and agents could download upstream sources, reference solutions and robot models from their own shell. The scene, the frozen instance, the tooling and the pass rule are unchanged, so the task keeps its name; results from before this date were obtained with open egress.
- 2026-10-13 (board 0086): the service's copy of the tooling only (`environment/sim/rcb_traj.py`, one line of `environment/sim/server.py`). The tooling now refuses to save or replay a trajectory of more than 200,000 rows, the privileged verifier's cap; a standard episode has no row cap, so the service lifts it for its own session (`S.max_rows = None`). Nothing an agent sees and nothing the service grades changes.
- 2026-10-13 (board 0100): base image `eai-robotwin` 0.1.2, whose scrub of RoboTwin's expert also removes the reference poses that `load_actors` computed only for the expert (here `pad_pose`). The standard-mode agent never reaches the env, so nothing it observes changes; the scene, the instance and the judge are unchanged.
+- 2026-10-14 (board 0113): the tooling only (`environment/sim/rcb_traj.py`): while it builds the scene, `glob.glob` returns its matches sorted, so that a task which lists files to draw an object depends on its seed alone, not on the host's directory order (four RoboTwin tasks did; this one lists no files while it builds its scene). Its scene, frozen instance (same initial-state hash), verifier and pass rule are unchanged.
diff --git a/tasks/robotwin_2_0_move_stapler_pad-standard/environment/sim/rcb_traj.py b/tasks/robotwin_2_0_move_stapler_pad-standard/environment/sim/rcb_traj.py
index 4a6575a485..43c2d6d3c0 100644
--- a/tasks/robotwin_2_0_move_stapler_pad-standard/environment/sim/rcb_traj.py
+++ b/tasks/robotwin_2_0_move_stapler_pad-standard/environment/sim/rcb_traj.py
@@ -113,6 +113,26 @@ BASE_ARGS: dict[str, Any] = {
}
+class _SortedGlob:
+ """While a scene is built, glob.glob / glob.iglob return their matches sorted. Four RoboTwin tasks (place_a2b_left,
+ place_a2b_right, place_object_scale, place_object_stand: envs/<task>.py load_actors) list an object's model files
+ with glob.glob and then draw one with the seeded np.random.choice; glob returns the file system's directory order,
+ which differs between hosts (and image layers), so one seed built a different object on another host and the frozen
+ instance's hash check failed there (board 0086 claim 5; never rebuilt off the reference workstation). Sorted, the
+ draw depends on the seed alone. No other task of ours lists files while it builds its scene."""
+
+ def __enter__(self):
+ import glob
+ self._glob, self._orig = glob, (glob.glob, glob.iglob)
+ glob.glob = lambda *a, **k: sorted(self._orig[0](*a, **k))
+ glob.iglob = lambda *a, **k: iter(sorted(self._orig[1](*a, **k)))
+ return self
+
+ def __exit__(self, *exc):
+ self._glob.glob, self._glob.iglob = self._orig
+ return False
+
+
def _chdir_root() -> None:
# RoboTwin resolves assets/config paths relative to the repo root and via ROOT_PATH
if os.getcwd() != ROBOTWIN_ROOT and os.path.isdir(ROBOTWIN_ROOT):
@@ -218,7 +238,8 @@ class Session:
if self.enable_render:
sapien.render.set_camera_shader_dir(self.render_shader)
args = _build_args(self.task, self.embodiment, self.seed)
- self.env.setup_demo(**args)
+ with _SortedGlob(): # the object draw depends on the seed alone, not the host's directory order
+ self.env.setup_demo(**args)
if self.enable_render: # setup_scene() switches the shader to ray tracing; put it back for cheap frames
sapien.render.set_camera_shader_dir(self.render_shader)
self._raw_scene = self.env.scene
diff --git a/tasks/robotwin_2_0_move_stapler_pad-standard/task.toml b/tasks/robotwin_2_0_move_stapler_pad-standard/task.toml
index 827be1e386..e8d6404d91 100644
--- a/tasks/robotwin_2_0_move_stapler_pad-standard/task.toml
+++ b/tasks/robotwin_2_0_move_stapler_pad-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "robotwin"
robot = "aloha-agilex"
family = "move_stapler_pad"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 60417
diff --git a/tasks/robotwin_2_0_open_laptop-privileged/README.md b/tasks/robotwin_2_0_open_laptop-privileged/README.md
index 40ef12cfc5..694f5b5530 100644
--- a/tasks/robotwin_2_0_open_laptop-privileged/README.md
+++ b/tasks/robotwin_2_0_open_laptop-privileged/README.md
@@ -55,3 +55,4 @@ Use one arm to open the laptop.
- 2026-10-10 (owner decision, board 0072): the agent's network egress is restricted to the model APIs (`[agent] network_mode = "allowlist"`, the hosts of `scripts/set_agent_egress.py`; since 2026-10-11 these include the package registries, PyPI and the Debian / Ubuntu archives, so that `pip install` and `apt-get install` work); it was open before (Harbor's default), and agents could download upstream sources, reference solutions and robot models from their own shell. The scene, the frozen instance, the tooling and the pass rule are unchanged, so the task keeps its name; results from before this date were obtained with open egress.
- 2026-10-13 (board 0086): the tooling refuses a trajectory the verifier cannot grade (`environment/rcb_traj.py`, `tests/rcb_traj.py`). `Session.save_trajectory` and `Session.replay` raise `ValueError` for more than 200,000 rows (`rcb_traj.MAX_T`, the verifier's `MAX_T` in `tests/grade.py`: 800 s of physics), and `record_stop` warns on stderr. Before, the tool saved and replayed such a file ("success") and the verifier then graded it `bad_trajectory`; the instruction has stated the cap since 2026-10-10. The verifier, the scene, the instance and the pass rule are unchanged: every trajectory is graded as before.
- 2026-10-13 (board 0100): base image `eai-robotwin` 0.1.2. Its scrub of RoboTwin's expert also removes the reference poses that 18 tasks' `load_actors` computed only for the expert, and the expert's comments above the scrubbed method bodies (`images/robotwin/scrub_oracles.py`). This task's `load_actors` set no such value: its scene, instance, tooling, verifier and pass rule are unchanged.
+- 2026-10-14 (board 0113): the tooling only (`environment/rcb_traj.py`, `tests/rcb_traj.py`): while it builds the scene, `glob.glob` returns its matches sorted, so that a task which lists files to draw an object depends on its seed alone, not on the host's directory order (four RoboTwin tasks did; this one lists no files while it builds its scene). Its scene, frozen instance (same initial-state hash), verifier and pass rule are unchanged.
diff --git a/tasks/robotwin_2_0_open_laptop-privileged/environment/rcb_traj.py b/tasks/robotwin_2_0_open_laptop-privileged/environment/rcb_traj.py
index f703007af1..ec5227b46a 100644
--- a/tasks/robotwin_2_0_open_laptop-privileged/environment/rcb_traj.py
+++ b/tasks/robotwin_2_0_open_laptop-privileged/environment/rcb_traj.py
@@ -113,6 +113,26 @@ BASE_ARGS: dict[str, Any] = {
}
+class _SortedGlob:
+ """While a scene is built, glob.glob / glob.iglob return their matches sorted. Four RoboTwin tasks (place_a2b_left,
+ place_a2b_right, place_object_scale, place_object_stand: envs/<task>.py load_actors) list an object's model files
+ with glob.glob and then draw one with the seeded np.random.choice; glob returns the file system's directory order,
+ which differs between hosts (and image layers), so one seed built a different object on another host and the frozen
+ instance's hash check failed there (board 0086 claim 5; never rebuilt off the reference workstation). Sorted, the
+ draw depends on the seed alone. No other task of ours lists files while it builds its scene."""
+
+ def __enter__(self):
+ import glob
+ self._glob, self._orig = glob, (glob.glob, glob.iglob)
+ glob.glob = lambda *a, **k: sorted(self._orig[0](*a, **k))
+ glob.iglob = lambda *a, **k: iter(sorted(self._orig[1](*a, **k)))
+ return self
+
+ def __exit__(self, *exc):
+ self._glob.glob, self._glob.iglob = self._orig
+ return False
+
+
def _chdir_root() -> None:
# RoboTwin resolves assets/config paths relative to the repo root and via ROOT_PATH
if os.getcwd() != ROBOTWIN_ROOT and os.path.isdir(ROBOTWIN_ROOT):
@@ -218,7 +238,8 @@ class Session:
if self.enable_render:
sapien.render.set_camera_shader_dir(self.render_shader)
args = _build_args(self.task, self.embodiment, self.seed)
- self.env.setup_demo(**args)
+ with _SortedGlob(): # the object draw depends on the seed alone, not the host's directory order
+ self.env.setup_demo(**args)
if self.enable_render: # setup_scene() switches the shader to ray tracing; put it back for cheap frames
sapien.render.set_camera_shader_dir(self.render_shader)
self._raw_scene = self.env.scene
diff --git a/tasks/robotwin_2_0_open_laptop-privileged/task.toml b/tasks/robotwin_2_0_open_laptop-privileged/task.toml
index 7a20a99ff1..be4aa7e9b3 100644
--- a/tasks/robotwin_2_0_open_laptop-privileged/task.toml
+++ b/tasks/robotwin_2_0_open_laptop-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "robotwin" # RoboTwin 2.0 = SAPIEN 3 / PhysX + CuRobo
robot = "aloha-agilex"
family = "open_laptop"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 60417 # scene = environment/instance.json (RoboTwin seed + expected initial-state hash); identity = dataset.toml digest
expert_time_estimate_hours = 0.5
diff --git a/tasks/robotwin_2_0_open_laptop-privileged/tests/rcb_traj.py b/tasks/robotwin_2_0_open_laptop-privileged/tests/rcb_traj.py
index f703007af1..ec5227b46a 100644
--- a/tasks/robotwin_2_0_open_laptop-privileged/tests/rcb_traj.py
+++ b/tasks/robotwin_2_0_open_laptop-privileged/tests/rcb_traj.py
@@ -113,6 +113,26 @@ BASE_ARGS: dict[str, Any] = {
}
+class _SortedGlob:
+ """While a scene is built, glob.glob / glob.iglob return their matches sorted. Four RoboTwin tasks (place_a2b_left,
+ place_a2b_right, place_object_scale, place_object_stand: envs/<task>.py load_actors) list an object's model files
+ with glob.glob and then draw one with the seeded np.random.choice; glob returns the file system's directory order,
+ which differs between hosts (and image layers), so one seed built a different object on another host and the frozen
+ instance's hash check failed there (board 0086 claim 5; never rebuilt off the reference workstation). Sorted, the
+ draw depends on the seed alone. No other task of ours lists files while it builds its scene."""
+
+ def __enter__(self):
+ import glob
+ self._glob, self._orig = glob, (glob.glob, glob.iglob)
+ glob.glob = lambda *a, **k: sorted(self._orig[0](*a, **k))
+ glob.iglob = lambda *a, **k: iter(sorted(self._orig[1](*a, **k)))
+ return self
+
+ def __exit__(self, *exc):
+ self._glob.glob, self._glob.iglob = self._orig
+ return False
+
+
def _chdir_root() -> None:
# RoboTwin resolves assets/config paths relative to the repo root and via ROOT_PATH
if os.getcwd() != ROBOTWIN_ROOT and os.path.isdir(ROBOTWIN_ROOT):
@@ -218,7 +238,8 @@ class Session:
if self.enable_render:
sapien.render.set_camera_shader_dir(self.render_shader)
args = _build_args(self.task, self.embodiment, self.seed)
- self.env.setup_demo(**args)
+ with _SortedGlob(): # the object draw depends on the seed alone, not the host's directory order
+ self.env.setup_demo(**args)
if self.enable_render: # setup_scene() switches the shader to ray tracing; put it back for cheap frames
sapien.render.set_camera_shader_dir(self.render_shader)
self._raw_scene = self.env.scene
diff --git a/tasks/robotwin_2_0_open_laptop-standard/README.md b/tasks/robotwin_2_0_open_laptop-standard/README.md
index 1f2cda4c6d..2f50110642 100644
--- a/tasks/robotwin_2_0_open_laptop-standard/README.md
+++ b/tasks/robotwin_2_0_open_laptop-standard/README.md
@@ -53,3 +53,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
- 2026-10-10 (owner decision, board 0072): the agent's network egress is restricted to the model APIs (`[agent] network_mode = "allowlist"`, the hosts of `scripts/set_agent_egress.py`; since 2026-10-11 these include the package registries, PyPI and the Debian / Ubuntu archives, so that `pip install` and `apt-get install` work); it was open before (Harbor's default), and agents could download upstream sources, reference solutions and robot models from their own shell. The scene, the frozen instance, the tooling and the pass rule are unchanged, so the task keeps its name; results from before this date were obtained with open egress.
- 2026-10-13 (board 0086): the service's copy of the tooling only (`environment/sim/rcb_traj.py`, one line of `environment/sim/server.py`). The tooling now refuses to save or replay a trajectory of more than 200,000 rows, the privileged verifier's cap; a standard episode has no row cap, so the service lifts it for its own session (`S.max_rows = None`). Nothing an agent sees and nothing the service grades changes.
- 2026-10-13 (board 0100): base image `eai-robotwin` 0.1.2, whose scrub of RoboTwin's expert also removes the reference poses that `load_actors` computed only for the expert. The standard-mode agent never reaches the env, so nothing it observes changes; the scene, the instance and the judge are unchanged.
+- 2026-10-14 (board 0113): the tooling only (`environment/sim/rcb_traj.py`): while it builds the scene, `glob.glob` returns its matches sorted, so that a task which lists files to draw an object depends on its seed alone, not on the host's directory order (four RoboTwin tasks did; this one lists no files while it builds its scene). Its scene, frozen instance (same initial-state hash), verifier and pass rule are unchanged.
diff --git a/tasks/robotwin_2_0_open_laptop-standard/environment/sim/rcb_traj.py b/tasks/robotwin_2_0_open_laptop-standard/environment/sim/rcb_traj.py
index f703007af1..ec5227b46a 100644
--- a/tasks/robotwin_2_0_open_laptop-standard/environment/sim/rcb_traj.py
+++ b/tasks/robotwin_2_0_open_laptop-standard/environment/sim/rcb_traj.py
@@ -113,6 +113,26 @@ BASE_ARGS: dict[str, Any] = {
}
+class _SortedGlob:
+ """While a scene is built, glob.glob / glob.iglob return their matches sorted. Four RoboTwin tasks (place_a2b_left,
+ place_a2b_right, place_object_scale, place_object_stand: envs/<task>.py load_actors) list an object's model files
+ with glob.glob and then draw one with the seeded np.random.choice; glob returns the file system's directory order,
+ which differs between hosts (and image layers), so one seed built a different object on another host and the frozen
+ instance's hash check failed there (board 0086 claim 5; never rebuilt off the reference workstation). Sorted, the
+ draw depends on the seed alone. No other task of ours lists files while it builds its scene."""
+
+ def __enter__(self):
+ import glob
+ self._glob, self._orig = glob, (glob.glob, glob.iglob)
+ glob.glob = lambda *a, **k: sorted(self._orig[0](*a, **k))
+ glob.iglob = lambda *a, **k: iter(sorted(self._orig[1](*a, **k)))
+ return self
+
+ def __exit__(self, *exc):
+ self._glob.glob, self._glob.iglob = self._orig
+ return False
+
+
def _chdir_root() -> None:
# RoboTwin resolves assets/config paths relative to the repo root and via ROOT_PATH
if os.getcwd() != ROBOTWIN_ROOT and os.path.isdir(ROBOTWIN_ROOT):
@@ -218,7 +238,8 @@ class Session:
if self.enable_render:
sapien.render.set_camera_shader_dir(self.render_shader)
args = _build_args(self.task, self.embodiment, self.seed)
- self.env.setup_demo(**args)
+ with _SortedGlob(): # the object draw depends on the seed alone, not the host's directory order
+ self.env.setup_demo(**args)
if self.enable_render: # setup_scene() switches the shader to ray tracing; put it back for cheap frames
sapien.render.set_camera_shader_dir(self.render_shader)
self._raw_scene = self.env.scene
diff --git a/tasks/robotwin_2_0_open_laptop-standard/task.toml b/tasks/robotwin_2_0_open_laptop-standard/task.toml
index c57cd7fc1c..4f65725fa3 100644
--- a/tasks/robotwin_2_0_open_laptop-standard/task.toml
+++ b/tasks/robotwin_2_0_open_laptop-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "robotwin"
robot = "aloha-agilex"
family = "open_laptop"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 60417
diff --git a/tasks/robotwin_2_0_open_microwave-privileged/README.md b/tasks/robotwin_2_0_open_microwave-privileged/README.md
index efb9b47b1b..59720f1df7 100644
--- a/tasks/robotwin_2_0_open_microwave-privileged/README.md
+++ b/tasks/robotwin_2_0_open_microwave-privileged/README.md
@@ -55,3 +55,4 @@ Open the microwave door: pull it open until its hinge joint reaches at least 60
- 2026-10-10 (owner decision, board 0072): the agent's network egress is restricted to the model APIs (`[agent] network_mode = "allowlist"`, the hosts of `scripts/set_agent_egress.py`; since 2026-10-11 these include the package registries, PyPI and the Debian / Ubuntu archives, so that `pip install` and `apt-get install` work); it was open before (Harbor's default), and agents could download upstream sources, reference solutions and robot models from their own shell. The scene, the frozen instance, the tooling and the pass rule are unchanged, so the task keeps its name; results from before this date were obtained with open egress.
- 2026-10-13 (board 0086): the tooling refuses a trajectory the verifier cannot grade (`environment/rcb_traj.py`, `tests/rcb_traj.py`). `Session.save_trajectory` and `Session.replay` raise `ValueError` for more than 200,000 rows (`rcb_traj.MAX_T`, the verifier's `MAX_T` in `tests/grade.py`: 800 s of physics), and `record_stop` warns on stderr. Before, the tool saved and replayed such a file ("success") and the verifier then graded it `bad_trajectory`; the instruction has stated the cap since 2026-10-10. The verifier, the scene, the instance and the pass rule are unchanged: every trajectory is graded as before.
- 2026-10-13 (board 0100): base image `eai-robotwin` 0.1.2. Its scrub of RoboTwin's expert also removes the reference poses that 18 tasks' `load_actors` computed only for the expert, and the expert's comments above the scrubbed method bodies (`images/robotwin/scrub_oracles.py`). This task's `load_actors` set no such value: its scene, instance, tooling, verifier and pass rule are unchanged.
+- 2026-10-14 (board 0113): the tooling only (`environment/rcb_traj.py`, `tests/rcb_traj.py`): while it builds the scene, `glob.glob` returns its matches sorted, so that a task which lists files to draw an object depends on its seed alone, not on the host's directory order (four RoboTwin tasks did; this one lists no files while it builds its scene). Its scene, frozen instance (same initial-state hash), verifier and pass rule are unchanged.
diff --git a/tasks/robotwin_2_0_open_microwave-privileged/environment/rcb_traj.py b/tasks/robotwin_2_0_open_microwave-privileged/environment/rcb_traj.py
index ad0bd95112..bd1cc1c408 100644
--- a/tasks/robotwin_2_0_open_microwave-privileged/environment/rcb_traj.py
+++ b/tasks/robotwin_2_0_open_microwave-privileged/environment/rcb_traj.py
@@ -113,6 +113,26 @@ BASE_ARGS: dict[str, Any] = {
}
+class _SortedGlob:
+ """While a scene is built, glob.glob / glob.iglob return their matches sorted. Four RoboTwin tasks (place_a2b_left,
+ place_a2b_right, place_object_scale, place_object_stand: envs/<task>.py load_actors) list an object's model files
+ with glob.glob and then draw one with the seeded np.random.choice; glob returns the file system's directory order,
+ which differs between hosts (and image layers), so one seed built a different object on another host and the frozen
+ instance's hash check failed there (board 0086 claim 5; never rebuilt off the reference workstation). Sorted, the
+ draw depends on the seed alone. No other task of ours lists files while it builds its scene."""
+
+ def __enter__(self):
+ import glob
+ self._glob, self._orig = glob, (glob.glob, glob.iglob)
+ glob.glob = lambda *a, **k: sorted(self._orig[0](*a, **k))
+ glob.iglob = lambda *a, **k: iter(sorted(self._orig[1](*a, **k)))
+ return self
+
+ def __exit__(self, *exc):
+ self._glob.glob, self._glob.iglob = self._orig
+ return False
+
+
def _chdir_root() -> None:
# RoboTwin resolves assets/config paths relative to the repo root and via ROOT_PATH
if os.getcwd() != ROBOTWIN_ROOT and os.path.isdir(ROBOTWIN_ROOT):
@@ -218,7 +238,8 @@ class Session:
if self.enable_render:
sapien.render.set_camera_shader_dir(self.render_shader)
args = _build_args(self.task, self.embodiment, self.seed)
- self.env.setup_demo(**args)
+ with _SortedGlob(): # the object draw depends on the seed alone, not the host's directory order
+ self.env.setup_demo(**args)
if self.enable_render: # setup_scene() switches the shader to ray tracing; put it back for cheap frames
sapien.render.set_camera_shader_dir(self.render_shader)
self._raw_scene = self.env.scene
diff --git a/tasks/robotwin_2_0_open_microwave-privileged/task.toml b/tasks/robotwin_2_0_open_microwave-privileged/task.toml
index 36beee5025..c50beafd75 100644
--- a/tasks/robotwin_2_0_open_microwave-privileged/task.toml
+++ b/tasks/robotwin_2_0_open_microwave-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "robotwin" # RoboTwin 2.0 = SAPIEN 3 / PhysX + CuRobo
robot = "aloha-agilex"
family = "open_microwave"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 28831 # scene = environment/instance.json (RoboTwin seed + expected initial-state hash); identity = dataset.toml digest
expert_time_estimate_hours = 0.5
diff --git a/tasks/robotwin_2_0_open_microwave-privileged/tests/rcb_traj.py b/tasks/robotwin_2_0_open_microwave-privileged/tests/rcb_traj.py
index ad0bd95112..bd1cc1c408 100644
--- a/tasks/robotwin_2_0_open_microwave-privileged/tests/rcb_traj.py
+++ b/tasks/robotwin_2_0_open_microwave-privileged/tests/rcb_traj.py
@@ -113,6 +113,26 @@ BASE_ARGS: dict[str, Any] = {
}
+class _SortedGlob:
+ """While a scene is built, glob.glob / glob.iglob return their matches sorted. Four RoboTwin tasks (place_a2b_left,
+ place_a2b_right, place_object_scale, place_object_stand: envs/<task>.py load_actors) list an object's model files
+ with glob.glob and then draw one with the seeded np.random.choice; glob returns the file system's directory order,
+ which differs between hosts (and image layers), so one seed built a different object on another host and the frozen
+ instance's hash check failed there (board 0086 claim 5; never rebuilt off the reference workstation). Sorted, the
+ draw depends on the seed alone. No other task of ours lists files while it builds its scene."""
+
+ def __enter__(self):
+ import glob
+ self._glob, self._orig = glob, (glob.glob, glob.iglob)
+ glob.glob = lambda *a, **k: sorted(self._orig[0](*a, **k))
+ glob.iglob = lambda *a, **k: iter(sorted(self._orig[1](*a, **k)))
+ return self
+
+ def __exit__(self, *exc):
+ self._glob.glob, self._glob.iglob = self._orig
+ return False
+
+
def _chdir_root() -> None:
# RoboTwin resolves assets/config paths relative to the repo root and via ROOT_PATH
if os.getcwd() != ROBOTWIN_ROOT and os.path.isdir(ROBOTWIN_ROOT):
@@ -218,7 +238,8 @@ class Session:
if self.enable_render:
sapien.render.set_camera_shader_dir(self.render_shader)
args = _build_args(self.task, self.embodiment, self.seed)
- self.env.setup_demo(**args)
+ with _SortedGlob(): # the object draw depends on the seed alone, not the host's directory order
+ self.env.setup_demo(**args)
if self.enable_render: # setup_scene() switches the shader to ray tracing; put it back for cheap frames
sapien.render.set_camera_shader_dir(self.render_shader)
self._raw_scene = self.env.scene
diff --git a/tasks/robotwin_2_0_open_microwave-standard/README.md b/tasks/robotwin_2_0_open_microwave-standard/README.md
index d0c86449ad..df70aecf4a 100644
--- a/tasks/robotwin_2_0_open_microwave-standard/README.md
+++ b/tasks/robotwin_2_0_open_microwave-standard/README.md
@@ -52,3 +52,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
- 2026-10-10 (owner decision, board 0072): the agent's network egress is restricted to the model APIs (`[agent] network_mode = "allowlist"`, the hosts of `scripts/set_agent_egress.py`; since 2026-10-11 these include the package registries, PyPI and the Debian / Ubuntu archives, so that `pip install` and `apt-get install` work); it was open before (Harbor's default), and agents could download upstream sources, reference solutions and robot models from their own shell. The scene, the frozen instance, the tooling and the pass rule are unchanged, so the task keeps its name; results from before this date were obtained with open egress.
- 2026-10-13 (board 0086): the service's copy of the tooling only (`environment/sim/rcb_traj.py`, one line of `environment/sim/server.py`). The tooling now refuses to save or replay a trajectory of more than 200,000 rows, the privileged verifier's cap; a standard episode has no row cap, so the service lifts it for its own session (`S.max_rows = None`). Nothing an agent sees and nothing the service grades changes.
- 2026-10-13 (board 0100): base image `eai-robotwin` 0.1.2, whose scrub of RoboTwin's expert also removes the reference poses that `load_actors` computed only for the expert. The standard-mode agent never reaches the env, so nothing it observes changes; the scene, the instance and the judge are unchanged.
+- 2026-10-14 (board 0113): the tooling only (`environment/sim/rcb_traj.py`): while it builds the scene, `glob.glob` returns its matches sorted, so that a task which lists files to draw an object depends on its seed alone, not on the host's directory order (four RoboTwin tasks did; this one lists no files while it builds its scene). Its scene, frozen instance (same initial-state hash), verifier and pass rule are unchanged.
diff --git a/tasks/robotwin_2_0_open_microwave-standard/environment/sim/rcb_traj.py b/tasks/robotwin_2_0_open_microwave-standard/environment/sim/rcb_traj.py
index ad0bd95112..bd1cc1c408 100644
--- a/tasks/robotwin_2_0_open_microwave-standard/environment/sim/rcb_traj.py
+++ b/tasks/robotwin_2_0_open_microwave-standard/environment/sim/rcb_traj.py
@@ -113,6 +113,26 @@ BASE_ARGS: dict[str, Any] = {
}
+class _SortedGlob:
+ """While a scene is built, glob.glob / glob.iglob return their matches sorted. Four RoboTwin tasks (place_a2b_left,
+ place_a2b_right, place_object_scale, place_object_stand: envs/<task>.py load_actors) list an object's model files
+ with glob.glob and then draw one with the seeded np.random.choice; glob returns the file system's directory order,
+ which differs between hosts (and image layers), so one seed built a different object on another host and the frozen
+ instance's hash check failed there (board 0086 claim 5; never rebuilt off the reference workstation). Sorted, the
+ draw depends on the seed alone. No other task of ours lists files while it builds its scene."""
+
+ def __enter__(self):
+ import glob
+ self._glob, self._orig = glob, (glob.glob, glob.iglob)
+ glob.glob = lambda *a, **k: sorted(self._orig[0](*a, **k))
+ glob.iglob = lambda *a, **k: iter(sorted(self._orig[1](*a, **k)))
+ return self
+
+ def __exit__(self, *exc):
+ self._glob.glob, self._glob.iglob = self._orig
+ return False
+
+
def _chdir_root() -> None:
# RoboTwin resolves assets/config paths relative to the repo root and via ROOT_PATH
if os.getcwd() != ROBOTWIN_ROOT and os.path.isdir(ROBOTWIN_ROOT):
@@ -218,7 +238,8 @@ class Session:
if self.enable_render:
sapien.render.set_camera_shader_dir(self.render_shader)
args = _build_args(self.task, self.embodiment, self.seed)
- self.env.setup_demo(**args)
+ with _SortedGlob(): # the object draw depends on the seed alone, not the host's directory order
+ self.env.setup_demo(**args)
if self.enable_render: # setup_scene() switches the shader to ray tracing; put it back for cheap frames
sapien.render.set_camera_shader_dir(self.render_shader)
self._raw_scene = self.env.scene
diff --git a/tasks/robotwin_2_0_open_microwave-standard/task.toml b/tasks/robotwin_2_0_open_microwave-standard/task.toml
index 65c8b504ee..b14740fee8 100644
--- a/tasks/robotwin_2_0_open_microwave-standard/task.toml
+++ b/tasks/robotwin_2_0_open_microwave-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "robotwin"
robot = "aloha-agilex"
family = "open_microwave"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 28831
diff --git a/tasks/robotwin_2_0_pick_diverse_bottles-privileged/README.md b/tasks/robotwin_2_0_pick_diverse_bottles-privileged/README.md
index b6635f82e3..89c5cc3e98 100644
--- a/tasks/robotwin_2_0_pick_diverse_bottles-privileged/README.md
+++ b/tasks/robotwin_2_0_pick_diverse_bottles-privileged/README.md
@@ -55,3 +55,4 @@ Pick up one bottle with one arm, and pick up another bottle with the other arm.
- 2026-10-10 (owner decision, board 0072): the agent's network egress is restricted to the model APIs (`[agent] network_mode = "allowlist"`, the hosts of `scripts/set_agent_egress.py`; since 2026-10-11 these include the package registries, PyPI and the Debian / Ubuntu archives, so that `pip install` and `apt-get install` work); it was open before (Harbor's default), and agents could download upstream sources, reference solutions and robot models from their own shell. The scene, the frozen instance, the tooling and the pass rule are unchanged, so the task keeps its name; results from before this date were obtained with open egress.
- 2026-10-13 (board 0086): the tooling refuses a trajectory the verifier cannot grade (`environment/rcb_traj.py`, `tests/rcb_traj.py`). `Session.save_trajectory` and `Session.replay` raise `ValueError` for more than 200,000 rows (`rcb_traj.MAX_T`, the verifier's `MAX_T` in `tests/grade.py`: 800 s of physics), and `record_stop` warns on stderr. Before, the tool saved and replayed such a file ("success") and the verifier then graded it `bad_trajectory`; the instruction has stated the cap since 2026-10-10. The verifier, the scene, the instance and the pass rule are unchanged: every trajectory is graded as before.
- 2026-10-13 (board 0100): base image `eai-robotwin` 0.1.2. Its scrub of RoboTwin's expert also removes the reference poses that 18 tasks' `load_actors` computed only for the expert, and the expert's comments above the scrubbed method bodies (`images/robotwin/scrub_oracles.py`). This task's `load_actors` set no such value: its scene, instance, tooling, verifier and pass rule are unchanged.
+- 2026-10-14 (board 0113): the tooling only (`environment/rcb_traj.py`, `tests/rcb_traj.py`): while it builds the scene, `glob.glob` returns its matches sorted, so that a task which lists files to draw an object depends on its seed alone, not on the host's directory order (four RoboTwin tasks did; this one lists no files while it builds its scene). Its scene, frozen instance (same initial-state hash), verifier and pass rule are unchanged.
diff --git a/tasks/robotwin_2_0_pick_diverse_bottles-privileged/environment/rcb_traj.py b/tasks/robotwin_2_0_pick_diverse_bottles-privileged/environment/rcb_traj.py
index b7a2c212d8..5cc31acf44 100644
--- a/tasks/robotwin_2_0_pick_diverse_bottles-privileged/environment/rcb_traj.py
+++ b/tasks/robotwin_2_0_pick_diverse_bottles-privileged/environment/rcb_traj.py
@@ -113,6 +113,26 @@ BASE_ARGS: dict[str, Any] = {
}
+class _SortedGlob:
+ """While a scene is built, glob.glob / glob.iglob return their matches sorted. Four RoboTwin tasks (place_a2b_left,
+ place_a2b_right, place_object_scale, place_object_stand: envs/<task>.py load_actors) list an object's model files
+ with glob.glob and then draw one with the seeded np.random.choice; glob returns the file system's directory order,
+ which differs between hosts (and image layers), so one seed built a different object on another host and the frozen
+ instance's hash check failed there (board 0086 claim 5; never rebuilt off the reference workstation). Sorted, the
+ draw depends on the seed alone. No other task of ours lists files while it builds its scene."""
+
+ def __enter__(self):
+ import glob
+ self._glob, self._orig = glob, (glob.glob, glob.iglob)
+ glob.glob = lambda *a, **k: sorted(self._orig[0](*a, **k))
+ glob.iglob = lambda *a, **k: iter(sorted(self._orig[1](*a, **k)))
+ return self
+
+ def __exit__(self, *exc):
+ self._glob.glob, self._glob.iglob = self._orig
+ return False
+
+
def _chdir_root() -> None:
# RoboTwin resolves assets/config paths relative to the repo root and via ROOT_PATH
if os.getcwd() != ROBOTWIN_ROOT and os.path.isdir(ROBOTWIN_ROOT):
@@ -218,7 +238,8 @@ class Session:
if self.enable_render:
sapien.render.set_camera_shader_dir(self.render_shader)
args = _build_args(self.task, self.embodiment, self.seed)
- self.env.setup_demo(**args)
+ with _SortedGlob(): # the object draw depends on the seed alone, not the host's directory order
+ self.env.setup_demo(**args)
if self.enable_render: # setup_scene() switches the shader to ray tracing; put it back for cheap frames
sapien.render.set_camera_shader_dir(self.render_shader)
self._raw_scene = self.env.scene
diff --git a/tasks/robotwin_2_0_pick_diverse_bottles-privileged/task.toml b/tasks/robotwin_2_0_pick_diverse_bottles-privileged/task.toml
index e361ed380d..2db9bf4ed8 100644
--- a/tasks/robotwin_2_0_pick_diverse_bottles-privileged/task.toml
+++ b/tasks/robotwin_2_0_pick_diverse_bottles-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "robotwin" # RoboTwin 2.0 = SAPIEN 3 / PhysX + CuRobo
robot = "aloha-agilex"
family = "pick_diverse_bottles"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 77321 # scene = environment/instance.json (RoboTwin seed + expected initial-state hash); identity = dataset.toml digest
expert_time_estimate_hours = 0.5
diff --git a/tasks/robotwin_2_0_pick_diverse_bottles-privileged/tests/rcb_traj.py b/tasks/robotwin_2_0_pick_diverse_bottles-privileged/tests/rcb_traj.py
index b7a2c212d8..5cc31acf44 100644
--- a/tasks/robotwin_2_0_pick_diverse_bottles-privileged/tests/rcb_traj.py
+++ b/tasks/robotwin_2_0_pick_diverse_bottles-privileged/tests/rcb_traj.py
@@ -113,6 +113,26 @@ BASE_ARGS: dict[str, Any] = {
}
+class _SortedGlob:
+ """While a scene is built, glob.glob / glob.iglob return their matches sorted. Four RoboTwin tasks (place_a2b_left,
+ place_a2b_right, place_object_scale, place_object_stand: envs/<task>.py load_actors) list an object's model files
+ with glob.glob and then draw one with the seeded np.random.choice; glob returns the file system's directory order,
+ which differs between hosts (and image layers), so one seed built a different object on another host and the frozen
+ instance's hash check failed there (board 0086 claim 5; never rebuilt off the reference workstation). Sorted, the
+ draw depends on the seed alone. No other task of ours lists files while it builds its scene."""
+
+ def __enter__(self):
+ import glob
+ self._glob, self._orig = glob, (glob.glob, glob.iglob)
+ glob.glob = lambda *a, **k: sorted(self._orig[0](*a, **k))
+ glob.iglob = lambda *a, **k: iter(sorted(self._orig[1](*a, **k)))
+ return self
+
+ def __exit__(self, *exc):
+ self._glob.glob, self._glob.iglob = self._orig
+ return False
+
+
def _chdir_root() -> None:
# RoboTwin resolves assets/config paths relative to the repo root and via ROOT_PATH
if os.getcwd() != ROBOTWIN_ROOT and os.path.isdir(ROBOTWIN_ROOT):
@@ -218,7 +238,8 @@ class Session:
if self.enable_render:
sapien.render.set_camera_shader_dir(self.render_shader)
args = _build_args(self.task, self.embodiment, self.seed)
- self.env.setup_demo(**args)
+ with _SortedGlob(): # the object draw depends on the seed alone, not the host's directory order
+ self.env.setup_demo(**args)
if self.enable_render: # setup_scene() switches the shader to ray tracing; put it back for cheap frames
sapien.render.set_camera_shader_dir(self.render_shader)
self._raw_scene = self.env.scene
diff --git a/tasks/robotwin_2_0_pick_diverse_bottles-standard/README.md b/tasks/robotwin_2_0_pick_diverse_bottles-standard/README.md
index f8c932bb87..864c09e267 100644
--- a/tasks/robotwin_2_0_pick_diverse_bottles-standard/README.md
+++ b/tasks/robotwin_2_0_pick_diverse_bottles-standard/README.md
@@ -53,3 +53,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
- 2026-10-10 (owner decision, board 0072): the agent's network egress is restricted to the model APIs (`[agent] network_mode = "allowlist"`, the hosts of `scripts/set_agent_egress.py`; since 2026-10-11 these include the package registries, PyPI and the Debian / Ubuntu archives, so that `pip install` and `apt-get install` work); it was open before (Harbor's default), and agents could download upstream sources, reference solutions and robot models from their own shell. The scene, the frozen instance, the tooling and the pass rule are unchanged, so the task keeps its name; results from before this date were obtained with open egress.
- 2026-10-13 (board 0086): the service's copy of the tooling only (`environment/sim/rcb_traj.py`, one line of `environment/sim/server.py`). The tooling now refuses to save or replay a trajectory of more than 200,000 rows, the privileged verifier's cap; a standard episode has no row cap, so the service lifts it for its own session (`S.max_rows = None`). Nothing an agent sees and nothing the service grades changes.
- 2026-10-13 (board 0100): base image `eai-robotwin` 0.1.2, whose scrub of RoboTwin's expert also removes the reference poses that `load_actors` computed only for the expert. The standard-mode agent never reaches the env, so nothing it observes changes; the scene, the instance and the judge are unchanged.
+- 2026-10-14 (board 0113): the tooling only (`environment/sim/rcb_traj.py`): while it builds the scene, `glob.glob` returns its matches sorted, so that a task which lists files to draw an object depends on its seed alone, not on the host's directory order (four RoboTwin tasks did; this one lists no files while it builds its scene). Its scene, frozen instance (same initial-state hash), verifier and pass rule are unchanged.
diff --git a/tasks/robotwin_2_0_pick_diverse_bottles-standard/environment/sim/rcb_traj.py b/tasks/robotwin_2_0_pick_diverse_bottles-standard/environment/sim/rcb_traj.py
index b7a2c212d8..5cc31acf44 100644
--- a/tasks/robotwin_2_0_pick_diverse_bottles-standard/environment/sim/rcb_traj.py
+++ b/tasks/robotwin_2_0_pick_diverse_bottles-standard/environment/sim/rcb_traj.py
@@ -113,6 +113,26 @@ BASE_ARGS: dict[str, Any] = {
}
+class _SortedGlob:
+ """While a scene is built, glob.glob / glob.iglob return their matches sorted. Four RoboTwin tasks (place_a2b_left,
+ place_a2b_right, place_object_scale, place_object_stand: envs/<task>.py load_actors) list an object's model files
+ with glob.glob and then draw one with the seeded np.random.choice; glob returns the file system's directory order,
+ which differs between hosts (and image layers), so one seed built a different object on another host and the frozen
+ instance's hash check failed there (board 0086 claim 5; never rebuilt off the reference workstation). Sorted, the
+ draw depends on the seed alone. No other task of ours lists files while it builds its scene."""
+
+ def __enter__(self):
+ import glob
+ self._glob, self._orig = glob, (glob.glob, glob.iglob)
+ glob.glob = lambda *a, **k: sorted(self._orig[0](*a, **k))
+ glob.iglob = lambda *a, **k: iter(sorted(self._orig[1](*a, **k)))
+ return self
+
+ def __exit__(self, *exc):
+ self._glob.glob, self._glob.iglob = self._orig
+ return False
+
+
def _chdir_root() -> None:
# RoboTwin resolves assets/config paths relative to the repo root and via ROOT_PATH
if os.getcwd() != ROBOTWIN_ROOT and os.path.isdir(ROBOTWIN_ROOT):
@@ -218,7 +238,8 @@ class Session:
if self.enable_render:
sapien.render.set_camera_shader_dir(self.render_shader)
args = _build_args(self.task, self.embodiment, self.seed)
- self.env.setup_demo(**args)
+ with _SortedGlob(): # the object draw depends on the seed alone, not the host's directory order
+ self.env.setup_demo(**args)
if self.enable_render: # setup_scene() switches the shader to ray tracing; put it back for cheap frames
sapien.render.set_camera_shader_dir(self.render_shader)
self._raw_scene = self.env.scene
diff --git a/tasks/robotwin_2_0_pick_diverse_bottles-standard/task.toml b/tasks/robotwin_2_0_pick_diverse_bottles-standard/task.toml
index 2799906f9b..82ab3aa915 100644
--- a/tasks/robotwin_2_0_pick_diverse_bottles-standard/task.toml
+++ b/tasks/robotwin_2_0_pick_diverse_bottles-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "robotwin"
robot = "aloha-agilex"
family = "pick_diverse_bottles"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 77321
diff --git a/tasks/robotwin_2_0_pick_dual_bottles-privileged/README.md b/tasks/robotwin_2_0_pick_dual_bottles-privileged/README.md
index 02e4363cf7..bcd38efa41 100644
--- a/tasks/robotwin_2_0_pick_dual_bottles-privileged/README.md
+++ b/tasks/robotwin_2_0_pick_dual_bottles-privileged/README.md
@@ -55,3 +55,4 @@ Pick up one bottle with one arm, and pick up another bottle with the other arm.
- 2026-10-10 (owner decision, board 0072): the agent's network egress is restricted to the model APIs (`[agent] network_mode = "allowlist"`, the hosts of `scripts/set_agent_egress.py`; since 2026-10-11 these include the package registries, PyPI and the Debian / Ubuntu archives, so that `pip install` and `apt-get install` work); it was open before (Harbor's default), and agents could download upstream sources, reference solutions and robot models from their own shell. The scene, the frozen instance, the tooling and the pass rule are unchanged, so the task keeps its name; results from before this date were obtained with open egress.
- 2026-10-13 (board 0086): the tooling refuses a trajectory the verifier cannot grade (`environment/rcb_traj.py`, `tests/rcb_traj.py`). `Session.save_trajectory` and `Session.replay` raise `ValueError` for more than 200,000 rows (`rcb_traj.MAX_T`, the verifier's `MAX_T` in `tests/grade.py`: 800 s of physics), and `record_stop` warns on stderr. Before, the tool saved and replayed such a file ("success") and the verifier then graded it `bad_trajectory`; the instruction has stated the cap since 2026-10-10. The verifier, the scene, the instance and the pass rule are unchanged: every trajectory is graded as before.
- 2026-10-13 (board 0100): base image `eai-robotwin` 0.1.2. Its scrub of RoboTwin's expert also removes the reference poses that 18 tasks' `load_actors` computed only for the expert, and the expert's comments above the scrubbed method bodies (`images/robotwin/scrub_oracles.py`). This task's `load_actors` set no such value: its scene, instance, tooling, verifier and pass rule are unchanged.
+- 2026-10-14 (board 0113): the tooling only (`environment/rcb_traj.py`, `tests/rcb_traj.py`): while it builds the scene, `glob.glob` returns its matches sorted, so that a task which lists files to draw an object depends on its seed alone, not on the host's directory order (four RoboTwin tasks did; this one lists no files while it builds its scene). Its scene, frozen instance (same initial-state hash), verifier and pass rule are unchanged.
diff --git a/tasks/robotwin_2_0_pick_dual_bottles-privileged/environment/rcb_traj.py b/tasks/robotwin_2_0_pick_dual_bottles-privileged/environment/rcb_traj.py
index cc722a1c44..608cb04c96 100644
--- a/tasks/robotwin_2_0_pick_dual_bottles-privileged/environment/rcb_traj.py
+++ b/tasks/robotwin_2_0_pick_dual_bottles-privileged/environment/rcb_traj.py
@@ -113,6 +113,26 @@ BASE_ARGS: dict[str, Any] = {
}
+class _SortedGlob:
+ """While a scene is built, glob.glob / glob.iglob return their matches sorted. Four RoboTwin tasks (place_a2b_left,
+ place_a2b_right, place_object_scale, place_object_stand: envs/<task>.py load_actors) list an object's model files
+ with glob.glob and then draw one with the seeded np.random.choice; glob returns the file system's directory order,
+ which differs between hosts (and image layers), so one seed built a different object on another host and the frozen
+ instance's hash check failed there (board 0086 claim 5; never rebuilt off the reference workstation). Sorted, the
+ draw depends on the seed alone. No other task of ours lists files while it builds its scene."""
+
+ def __enter__(self):
+ import glob
+ self._glob, self._orig = glob, (glob.glob, glob.iglob)
+ glob.glob = lambda *a, **k: sorted(self._orig[0](*a, **k))
+ glob.iglob = lambda *a, **k: iter(sorted(self._orig[1](*a, **k)))
+ return self
+
+ def __exit__(self, *exc):
+ self._glob.glob, self._glob.iglob = self._orig
+ return False
+
+
def _chdir_root() -> None:
# RoboTwin resolves assets/config paths relative to the repo root and via ROOT_PATH
if os.getcwd() != ROBOTWIN_ROOT and os.path.isdir(ROBOTWIN_ROOT):
@@ -218,7 +238,8 @@ class Session:
if self.enable_render:
sapien.render.set_camera_shader_dir(self.render_shader)
args = _build_args(self.task, self.embodiment, self.seed)
- self.env.setup_demo(**args)
+ with _SortedGlob(): # the object draw depends on the seed alone, not the host's directory order
+ self.env.setup_demo(**args)
if self.enable_render: # setup_scene() switches the shader to ray tracing; put it back for cheap frames
sapien.render.set_camera_shader_dir(self.render_shader)
self._raw_scene = self.env.scene
diff --git a/tasks/robotwin_2_0_pick_dual_bottles-privileged/task.toml b/tasks/robotwin_2_0_pick_dual_bottles-privileged/task.toml
index 6085093201..df9b422e62 100644
--- a/tasks/robotwin_2_0_pick_dual_bottles-privileged/task.toml
+++ b/tasks/robotwin_2_0_pick_dual_bottles-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "robotwin" # RoboTwin 2.0 = SAPIEN 3 / PhysX + CuRobo
robot = "aloha-agilex"
family = "pick_dual_bottles"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 60417 # scene = environment/instance.json (RoboTwin seed + expected initial-state hash); identity = dataset.toml digest
expert_time_estimate_hours = 0.5
diff --git a/tasks/robotwin_2_0_pick_dual_bottles-privileged/tests/rcb_traj.py b/tasks/robotwin_2_0_pick_dual_bottles-privileged/tests/rcb_traj.py
index cc722a1c44..608cb04c96 100644
--- a/tasks/robotwin_2_0_pick_dual_bottles-privileged/tests/rcb_traj.py
+++ b/tasks/robotwin_2_0_pick_dual_bottles-privileged/tests/rcb_traj.py
@@ -113,6 +113,26 @@ BASE_ARGS: dict[str, Any] = {
}
+class _SortedGlob:
+ """While a scene is built, glob.glob / glob.iglob return their matches sorted. Four RoboTwin tasks (place_a2b_left,
+ place_a2b_right, place_object_scale, place_object_stand: envs/<task>.py load_actors) list an object's model files
+ with glob.glob and then draw one with the seeded np.random.choice; glob returns the file system's directory order,
+ which differs between hosts (and image layers), so one seed built a different object on another host and the frozen
+ instance's hash check failed there (board 0086 claim 5; never rebuilt off the reference workstation). Sorted, the
+ draw depends on the seed alone. No other task of ours lists files while it builds its scene."""
+
+ def __enter__(self):
+ import glob
+ self._glob, self._orig = glob, (glob.glob, glob.iglob)
+ glob.glob = lambda *a, **k: sorted(self._orig[0](*a, **k))
+ glob.iglob = lambda *a, **k: iter(sorted(self._orig[1](*a, **k)))
+ return self
+
+ def __exit__(self, *exc):
+ self._glob.glob, self._glob.iglob = self._orig
+ return False
+
+
def _chdir_root() -> None:
# RoboTwin resolves assets/config paths relative to the repo root and via ROOT_PATH
if os.getcwd() != ROBOTWIN_ROOT and os.path.isdir(ROBOTWIN_ROOT):
@@ -218,7 +238,8 @@ class Session:
if self.enable_render:
sapien.render.set_camera_shader_dir(self.render_shader)
args = _build_args(self.task, self.embodiment, self.seed)
- self.env.setup_demo(**args)
+ with _SortedGlob(): # the object draw depends on the seed alone, not the host's directory order
+ self.env.setup_demo(**args)
if self.enable_render: # setup_scene() switches the shader to ray tracing; put it back for cheap frames
sapien.render.set_camera_shader_dir(self.render_shader)
self._raw_scene = self.env.scene
diff --git a/tasks/robotwin_2_0_pick_dual_bottles-standard/README.md b/tasks/robotwin_2_0_pick_dual_bottles-standard/README.md
index 93c411b4ff..1160e5bda6 100644
--- a/tasks/robotwin_2_0_pick_dual_bottles-standard/README.md
+++ b/tasks/robotwin_2_0_pick_dual_bottles-standard/README.md
@@ -53,3 +53,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
- 2026-10-10 (owner decision, board 0072): the agent's network egress is restricted to the model APIs (`[agent] network_mode = "allowlist"`, the hosts of `scripts/set_agent_egress.py`; since 2026-10-11 these include the package registries, PyPI and the Debian / Ubuntu archives, so that `pip install` and `apt-get install` work); it was open before (Harbor's default), and agents could download upstream sources, reference solutions and robot models from their own shell. The scene, the frozen instance, the tooling and the pass rule are unchanged, so the task keeps its name; results from before this date were obtained with open egress.
- 2026-10-13 (board 0086): the service's copy of the tooling only (`environment/sim/rcb_traj.py`, one line of `environment/sim/server.py`). The tooling now refuses to save or replay a trajectory of more than 200,000 rows, the privileged verifier's cap; a standard episode has no row cap, so the service lifts it for its own session (`S.max_rows = None`). Nothing an agent sees and nothing the service grades changes.
- 2026-10-13 (board 0100): base image `eai-robotwin` 0.1.2, whose scrub of RoboTwin's expert also removes the reference poses that `load_actors` computed only for the expert. The standard-mode agent never reaches the env, so nothing it observes changes; the scene, the instance and the judge are unchanged.
+- 2026-10-14 (board 0113): the tooling only (`environment/sim/rcb_traj.py`): while it builds the scene, `glob.glob` returns its matches sorted, so that a task which lists files to draw an object depends on its seed alone, not on the host's directory order (four RoboTwin tasks did; this one lists no files while it builds its scene). Its scene, frozen instance (same initial-state hash), verifier and pass rule are unchanged.
diff --git a/tasks/robotwin_2_0_pick_dual_bottles-standard/environment/sim/rcb_traj.py b/tasks/robotwin_2_0_pick_dual_bottles-standard/environment/sim/rcb_traj.py
index cc722a1c44..608cb04c96 100644
--- a/tasks/robotwin_2_0_pick_dual_bottles-standard/environment/sim/rcb_traj.py
+++ b/tasks/robotwin_2_0_pick_dual_bottles-standard/environment/sim/rcb_traj.py
@@ -113,6 +113,26 @@ BASE_ARGS: dict[str, Any] = {
}
+class _SortedGlob:
+ """While a scene is built, glob.glob / glob.iglob return their matches sorted. Four RoboTwin tasks (place_a2b_left,
+ place_a2b_right, place_object_scale, place_object_stand: envs/<task>.py load_actors) list an object's model files
+ with glob.glob and then draw one with the seeded np.random.choice; glob returns the file system's directory order,
+ which differs between hosts (and image layers), so one seed built a different object on another host and the frozen
+ instance's hash check failed there (board 0086 claim 5; never rebuilt off the reference workstation). Sorted, the
+ draw depends on the seed alone. No other task of ours lists files while it builds its scene."""
+
+ def __enter__(self):
+ import glob
+ self._glob, self._orig = glob, (glob.glob, glob.iglob)
+ glob.glob = lambda *a, **k: sorted(self._orig[0](*a, **k))
+ glob.iglob = lambda *a, **k: iter(sorted(self._orig[1](*a, **k)))
+ return self
+
+ def __exit__(self, *exc):
+ self._glob.glob, self._glob.iglob = self._orig
+ return False
+
+
def _chdir_root() -> None:
# RoboTwin resolves assets/config paths relative to the repo root and via ROOT_PATH
if os.getcwd() != ROBOTWIN_ROOT and os.path.isdir(ROBOTWIN_ROOT):
@@ -218,7 +238,8 @@ class Session:
if self.enable_render:
sapien.render.set_camera_shader_dir(self.render_shader)
args = _build_args(self.task, self.embodiment, self.seed)
- self.env.setup_demo(**args)
+ with _SortedGlob(): # the object draw depends on the seed alone, not the host's directory order
+ self.env.setup_demo(**args)
if self.enable_render: # setup_scene() switches the shader to ray tracing; put it back for cheap frames
sapien.render.set_camera_shader_dir(self.render_shader)
self._raw_scene = self.env.scene
diff --git a/tasks/robotwin_2_0_pick_dual_bottles-standard/task.toml b/tasks/robotwin_2_0_pick_dual_bottles-standard/task.toml
index 7f84e8d28c..22c42000a3 100644
--- a/tasks/robotwin_2_0_pick_dual_bottles-standard/task.toml
+++ b/tasks/robotwin_2_0_pick_dual_bottles-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "robotwin"
robot = "aloha-agilex"
family = "pick_dual_bottles"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 60417
diff --git a/tasks/robotwin_2_0_place_a2b_left-privileged/README.md b/tasks/robotwin_2_0_place_a2b_left-privileged/README.md
index 39b7cffaf3..91816e95e5 100644
--- a/tasks/robotwin_2_0_place_a2b_left-privileged/README.md
+++ b/tasks/robotwin_2_0_place_a2b_left-privileged/README.md
@@ -55,3 +55,4 @@ Use appropriate arm to place object A on the left of object B.
- 2026-10-10 (owner decision, board 0072): the agent's network egress is restricted to the model APIs (`[agent] network_mode = "allowlist"`, the hosts of `scripts/set_agent_egress.py`; since 2026-10-11 these include the package registries, PyPI and the Debian / Ubuntu archives, so that `pip install` and `apt-get install` work); it was open before (Harbor's default), and agents could download upstream sources, reference solutions and robot models from their own shell. The scene, the frozen instance, the tooling and the pass rule are unchanged, so the task keeps its name; results from before this date were obtained with open egress.
- 2026-10-13 (board 0086): the tooling refuses a trajectory the verifier cannot grade (`environment/rcb_traj.py`, `tests/rcb_traj.py`). `Session.save_trajectory` and `Session.replay` raise `ValueError` for more than 200,000 rows (`rcb_traj.MAX_T`, the verifier's `MAX_T` in `tests/grade.py`: 800 s of physics), and `record_stop` warns on stderr. Before, the tool saved and replayed such a file ("success") and the verifier then graded it `bad_trajectory`; the instruction has stated the cap since 2026-10-10. The verifier, the scene, the instance and the pass rule are unchanged: every trajectory is graded as before.
- 2026-10-13 (board 0100): base image `eai-robotwin` 0.1.2. Its scrub of RoboTwin's expert also removes the reference poses that 18 tasks' `load_actors` computed only for the expert, and the expert's comments above the scrubbed method bodies (`images/robotwin/scrub_oracles.py`). This task's `load_actors` set no such value: its scene, instance, tooling, verifier and pass rule are unchanged.
+- 2026-10-14 (owner decision C2, board 0113): **the frozen instance was re-frozen** (same seed, initial-state hash 48487255e320 -> 4dace5aa7a4c; object A `057_toycar` (model 0), object B `050_bell` (model 1)). RoboTwin's `load_actors` for this task lists the object's model files with `glob.glob` and draws one with the seeded `np.random.choice`; `glob` returns the file system's directory order, which differs between hosts, so the same seed built a different object on another host and the frozen instance's hash check failed there (board 0086; no run of this task ever got past it on the pool's hosts). The tooling (`environment/rcb_traj.py`, `tests/rcb_traj.py`) now sorts `glob`'s matches while it builds the scene, so the draw depends on the seed alone: the instance rebuilds to the same hash twice in one process, in a fresh container and with the file system's order reversed. The reference solution passes on it (two replays deterministic, verifier reward 1); a trajectory handed in for the old instance is not valid for this one. The task text, the success check and the pass rule are unchanged.
diff --git a/tasks/robotwin_2_0_place_a2b_left-privileged/environment/instance.json b/tasks/robotwin_2_0_place_a2b_left-privileged/environment/instance.json
index 5d1a38843e..aa4eff00c4 100644
--- a/tasks/robotwin_2_0_place_a2b_left-privileged/environment/instance.json
+++ b/tasks/robotwin_2_0_place_a2b_left-privileged/environment/instance.json
@@ -2,7 +2,7 @@
"task": "place_a2b_left",
"seed": 28831,
"embodiment": "aloha-agilex",
- "state_sha": "48487255e320",
+ "state_sha": "4dace5aa7a4c",
"text": "Use appropriate arm to place object A on the left of object B.",
"n_active_joints": 16
}
\ No newline at end of file
diff --git a/tasks/robotwin_2_0_place_a2b_left-privileged/environment/rcb_traj.py b/tasks/robotwin_2_0_place_a2b_left-privileged/environment/rcb_traj.py
index 2008647a8d..6496b53aca 100644
--- a/tasks/robotwin_2_0_place_a2b_left-privileged/environment/rcb_traj.py
+++ b/tasks/robotwin_2_0_place_a2b_left-privileged/environment/rcb_traj.py
@@ -113,6 +113,26 @@ BASE_ARGS: dict[str, Any] = {
}
+class _SortedGlob:
+ """While a scene is built, glob.glob / glob.iglob return their matches sorted. Four RoboTwin tasks (place_a2b_left,
+ place_a2b_right, place_object_scale, place_object_stand: envs/<task>.py load_actors) list an object's model files
+ with glob.glob and then draw one with the seeded np.random.choice; glob returns the file system's directory order,
+ which differs between hosts (and image layers), so one seed built a different object on another host and the frozen
+ instance's hash check failed there (board 0086 claim 5; never rebuilt off the reference workstation). Sorted, the
+ draw depends on the seed alone. No other task of ours lists files while it builds its scene."""
+
+ def __enter__(self):
+ import glob
+ self._glob, self._orig = glob, (glob.glob, glob.iglob)
+ glob.glob = lambda *a, **k: sorted(self._orig[0](*a, **k))
+ glob.iglob = lambda *a, **k: iter(sorted(self._orig[1](*a, **k)))
+ return self
+
+ def __exit__(self, *exc):
+ self._glob.glob, self._glob.iglob = self._orig
+ return False
+
+
def _chdir_root() -> None:
# RoboTwin resolves assets/config paths relative to the repo root and via ROOT_PATH
if os.getcwd() != ROBOTWIN_ROOT and os.path.isdir(ROBOTWIN_ROOT):
@@ -218,7 +238,8 @@ class Session:
if self.enable_render:
sapien.render.set_camera_shader_dir(self.render_shader)
args = _build_args(self.task, self.embodiment, self.seed)
- self.env.setup_demo(**args)
+ with _SortedGlob(): # the object draw depends on the seed alone, not the host's directory order
+ self.env.setup_demo(**args)
if self.enable_render: # setup_scene() switches the shader to ray tracing; put it back for cheap frames
sapien.render.set_camera_shader_dir(self.render_shader)
self._raw_scene = self.env.scene
diff --git a/tasks/robotwin_2_0_place_a2b_left-privileged/task.toml b/tasks/robotwin_2_0_place_a2b_left-privileged/task.toml
index 272cd616cf..edf287d138 100644
--- a/tasks/robotwin_2_0_place_a2b_left-privileged/task.toml
+++ b/tasks/robotwin_2_0_place_a2b_left-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "robotwin" # RoboTwin 2.0 = SAPIEN 3 / PhysX + CuRobo
robot = "aloha-agilex"
family = "place_a2b_left"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 28831 # scene = environment/instance.json (RoboTwin seed + expected initial-state hash); identity = dataset.toml digest
expert_time_estimate_hours = 0.5
diff --git a/tasks/robotwin_2_0_place_a2b_left-privileged/tests/instance.json b/tasks/robotwin_2_0_place_a2b_left-privileged/tests/instance.json
index 5d1a38843e..aa4eff00c4 100644
--- a/tasks/robotwin_2_0_place_a2b_left-privileged/tests/instance.json
+++ b/tasks/robotwin_2_0_place_a2b_left-privileged/tests/instance.json
@@ -2,7 +2,7 @@
"task": "place_a2b_left",
"seed": 28831,
"embodiment": "aloha-agilex",
- "state_sha": "48487255e320",
+ "state_sha": "4dace5aa7a4c",
"text": "Use appropriate arm to place object A on the left of object B.",
"n_active_joints": 16
}
\ No newline at end of file
diff --git a/tasks/robotwin_2_0_place_a2b_left-privileged/tests/rcb_traj.py b/tasks/robotwin_2_0_place_a2b_left-privileged/tests/rcb_traj.py
index 2008647a8d..6496b53aca 100644
--- a/tasks/robotwin_2_0_place_a2b_left-privileged/tests/rcb_traj.py
+++ b/tasks/robotwin_2_0_place_a2b_left-privileged/tests/rcb_traj.py
@@ -113,6 +113,26 @@ BASE_ARGS: dict[str, Any] = {
}
+class _SortedGlob:
+ """While a scene is built, glob.glob / glob.iglob return their matches sorted. Four RoboTwin tasks (place_a2b_left,
+ place_a2b_right, place_object_scale, place_object_stand: envs/<task>.py load_actors) list an object's model files
+ with glob.glob and then draw one with the seeded np.random.choice; glob returns the file system's directory order,
+ which differs between hosts (and image layers), so one seed built a different object on another host and the frozen
+ instance's hash check failed there (board 0086 claim 5; never rebuilt off the reference workstation). Sorted, the
+ draw depends on the seed alone. No other task of ours lists files while it builds its scene."""
+
+ def __enter__(self):
+ import glob
+ self._glob, self._orig = glob, (glob.glob, glob.iglob)
+ glob.glob = lambda *a, **k: sorted(self._orig[0](*a, **k))
+ glob.iglob = lambda *a, **k: iter(sorted(self._orig[1](*a, **k)))
+ return self
+
+ def __exit__(self, *exc):
+ self._glob.glob, self._glob.iglob = self._orig
+ return False
+
+
def _chdir_root() -> None:
# RoboTwin resolves assets/config paths relative to the repo root and via ROOT_PATH
if os.getcwd() != ROBOTWIN_ROOT and os.path.isdir(ROBOTWIN_ROOT):
@@ -218,7 +238,8 @@ class Session:
if self.enable_render:
sapien.render.set_camera_shader_dir(self.render_shader)
args = _build_args(self.task, self.embodiment, self.seed)
- self.env.setup_demo(**args)
+ with _SortedGlob(): # the object draw depends on the seed alone, not the host's directory order
+ self.env.setup_demo(**args)
if self.enable_render: # setup_scene() switches the shader to ray tracing; put it back for cheap frames
sapien.render.set_camera_shader_dir(self.render_shader)
self._raw_scene = self.env.scene
diff --git a/tasks/robotwin_2_0_place_a2b_left-standard/README.md b/tasks/robotwin_2_0_place_a2b_left-standard/README.md
index 3ee7c467c2..0d2fb8be7a 100644
--- a/tasks/robotwin_2_0_place_a2b_left-standard/README.md
+++ b/tasks/robotwin_2_0_place_a2b_left-standard/README.md
@@ -53,3 +53,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
- 2026-10-10 (owner decision, board 0072): the agent's network egress is restricted to the model APIs (`[agent] network_mode = "allowlist"`, the hosts of `scripts/set_agent_egress.py`; since 2026-10-11 these include the package registries, PyPI and the Debian / Ubuntu archives, so that `pip install` and `apt-get install` work); it was open before (Harbor's default), and agents could download upstream sources, reference solutions and robot models from their own shell. The scene, the frozen instance, the tooling and the pass rule are unchanged, so the task keeps its name; results from before this date were obtained with open egress.
- 2026-10-13 (board 0086): the service's copy of the tooling only (`environment/sim/rcb_traj.py`, one line of `environment/sim/server.py`). The tooling now refuses to save or replay a trajectory of more than 200,000 rows, the privileged verifier's cap; a standard episode has no row cap, so the service lifts it for its own session (`S.max_rows = None`). Nothing an agent sees and nothing the service grades changes.
- 2026-10-13 (board 0100): base image `eai-robotwin` 0.1.2, whose scrub of RoboTwin's expert also removes the reference poses that `load_actors` computed only for the expert. The standard-mode agent never reaches the env, so nothing it observes changes; the scene, the instance and the judge are unchanged.
+- 2026-10-14 (owner decision C2, board 0113): **the frozen instance was re-frozen** with the privileged twin (same seed, initial-state hash 48487255e320 -> 4dace5aa7a4c; object A `057_toycar` (model 0), object B `050_bell` (model 1)). RoboTwin's `load_actors` for this task lists the object's model files with `glob.glob` and draws one with the seeded `np.random.choice`; `glob` returns the file system's directory order, which differs between hosts, so the same seed built a different object on another host and the frozen instance's hash check failed there (board 0086; no run of this task ever got past it on the pool's hosts); the service rebuilt the instance and gave up with `instance_mismatch`. The service's copy of the tooling (`environment/sim/rcb_traj.py`) now sorts `glob`'s matches while it builds the scene. The task text, the success check and the pass rule are unchanged.
diff --git a/tasks/robotwin_2_0_place_a2b_left-standard/environment/sim/instance.json b/tasks/robotwin_2_0_place_a2b_left-standard/environment/sim/instance.json
index 5d1a38843e..aa4eff00c4 100644
--- a/tasks/robotwin_2_0_place_a2b_left-standard/environment/sim/instance.json
+++ b/tasks/robotwin_2_0_place_a2b_left-standard/environment/sim/instance.json
@@ -2,7 +2,7 @@
"task": "place_a2b_left",
"seed": 28831,
"embodiment": "aloha-agilex",
- "state_sha": "48487255e320",
+ "state_sha": "4dace5aa7a4c",
"text": "Use appropriate arm to place object A on the left of object B.",
"n_active_joints": 16
}
\ No newline at end of file
diff --git a/tasks/robotwin_2_0_place_a2b_left-standard/environment/sim/rcb_traj.py b/tasks/robotwin_2_0_place_a2b_left-standard/environment/sim/rcb_traj.py
index 2008647a8d..6496b53aca 100644
--- a/tasks/robotwin_2_0_place_a2b_left-standard/environment/sim/rcb_traj.py
+++ b/tasks/robotwin_2_0_place_a2b_left-standard/environment/sim/rcb_traj.py
@@ -113,6 +113,26 @@ BASE_ARGS: dict[str, Any] = {
}
+class _SortedGlob:
+ """While a scene is built, glob.glob / glob.iglob return their matches sorted. Four RoboTwin tasks (place_a2b_left,
+ place_a2b_right, place_object_scale, place_object_stand: envs/<task>.py load_actors) list an object's model files
+ with glob.glob and then draw one with the seeded np.random.choice; glob returns the file system's directory order,
+ which differs between hosts (and image layers), so one seed built a different object on another host and the frozen
+ instance's hash check failed there (board 0086 claim 5; never rebuilt off the reference workstation). Sorted, the
+ draw depends on the seed alone. No other task of ours lists files while it builds its scene."""
+
+ def __enter__(self):
+ import glob
+ self._glob, self._orig = glob, (glob.glob, glob.iglob)
+ glob.glob = lambda *a, **k: sorted(self._orig[0](*a, **k))
+ glob.iglob = lambda *a, **k: iter(sorted(self._orig[1](*a, **k)))
+ return self
+
+ def __exit__(self, *exc):
+ self._glob.glob, self._glob.iglob = self._orig
+ return False
+
+
def _chdir_root() -> None:
# RoboTwin resolves assets/config paths relative to the repo root and via ROOT_PATH
if os.getcwd() != ROBOTWIN_ROOT and os.path.isdir(ROBOTWIN_ROOT):
@@ -218,7 +238,8 @@ class Session:
if self.enable_render:
sapien.render.set_camera_shader_dir(self.render_shader)
args = _build_args(self.task, self.embodiment, self.seed)
- self.env.setup_demo(**args)
+ with _SortedGlob(): # the object draw depends on the seed alone, not the host's directory order
+ self.env.setup_demo(**args)
if self.enable_render: # setup_scene() switches the shader to ray tracing; put it back for cheap frames
sapien.render.set_camera_shader_dir(self.render_shader)
self._raw_scene = self.env.scene
diff --git a/tasks/robotwin_2_0_place_a2b_left-standard/task.toml b/tasks/robotwin_2_0_place_a2b_left-standard/task.toml
index 49123d5954..5bd6af6b67 100644
--- a/tasks/robotwin_2_0_place_a2b_left-standard/task.toml
+++ b/tasks/robotwin_2_0_place_a2b_left-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "robotwin"
robot = "aloha-agilex"
family = "place_a2b_left"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 28831
diff --git a/tasks/robotwin_2_0_place_a2b_right-privileged/README.md b/tasks/robotwin_2_0_place_a2b_right-privileged/README.md
index 5ebf11a221..8c4ae9d961 100644
--- a/tasks/robotwin_2_0_place_a2b_right-privileged/README.md
+++ b/tasks/robotwin_2_0_place_a2b_right-privileged/README.md
@@ -55,3 +55,4 @@ Use appropriate arm to place object A on the right of object B.
- 2026-10-10 (owner decision, board 0072): the agent's network egress is restricted to the model APIs (`[agent] network_mode = "allowlist"`, the hosts of `scripts/set_agent_egress.py`; since 2026-10-11 these include the package registries, PyPI and the Debian / Ubuntu archives, so that `pip install` and `apt-get install` work); it was open before (Harbor's default), and agents could download upstream sources, reference solutions and robot models from their own shell. The scene, the frozen instance, the tooling and the pass rule are unchanged, so the task keeps its name; results from before this date were obtained with open egress.
- 2026-10-13 (board 0086): the tooling refuses a trajectory the verifier cannot grade (`environment/rcb_traj.py`, `tests/rcb_traj.py`). `Session.save_trajectory` and `Session.replay` raise `ValueError` for more than 200,000 rows (`rcb_traj.MAX_T`, the verifier's `MAX_T` in `tests/grade.py`: 800 s of physics), and `record_stop` warns on stderr. Before, the tool saved and replayed such a file ("success") and the verifier then graded it `bad_trajectory`; the instruction has stated the cap since 2026-10-10. The verifier, the scene, the instance and the pass rule are unchanged: every trajectory is graded as before.
- 2026-10-13 (board 0100): base image `eai-robotwin` 0.1.2. Its scrub of RoboTwin's expert also removes the reference poses that 18 tasks' `load_actors` computed only for the expert, and the expert's comments above the scrubbed method bodies (`images/robotwin/scrub_oracles.py`). This task's `load_actors` set no such value: its scene, instance, tooling, verifier and pass rule are unchanged.
+- 2026-10-14 (owner decision C2, board 0113): **the frozen instance was re-frozen** (same seed, initial-state hash 90b641f66006 -> 1ed025a10f0d; object A `081_playingcards` (model 2), object B `107_soap` (model 2)). RoboTwin's `load_actors` for this task lists the object's model files with `glob.glob` and draws one with the seeded `np.random.choice`; `glob` returns the file system's directory order, which differs between hosts, so the same seed built a different object on another host and the frozen instance's hash check failed there (board 0086; no run of this task ever got past it on the pool's hosts). The tooling (`environment/rcb_traj.py`, `tests/rcb_traj.py`) now sorts `glob`'s matches while it builds the scene, so the draw depends on the seed alone: the instance rebuilds to the same hash twice in one process, in a fresh container and with the file system's order reversed. The reference solution passes on it (two replays deterministic, verifier reward 1); a trajectory handed in for the old instance is not valid for this one. The task text, the success check and the pass rule are unchanged.
diff --git a/tasks/robotwin_2_0_place_a2b_right-privileged/environment/instance.json b/tasks/robotwin_2_0_place_a2b_right-privileged/environment/instance.json
index a0d65614be..ee648ce7ef 100644
--- a/tasks/robotwin_2_0_place_a2b_right-privileged/environment/instance.json
+++ b/tasks/robotwin_2_0_place_a2b_right-privileged/environment/instance.json
@@ -2,7 +2,7 @@
"task": "place_a2b_right",
"seed": 28831,
"embodiment": "aloha-agilex",
- "state_sha": "90b641f66006",
+ "state_sha": "1ed025a10f0d",
"text": "Use appropriate arm to place object A on the right of object B.",
"n_active_joints": 16
}
\ No newline at end of file
diff --git a/tasks/robotwin_2_0_place_a2b_right-privileged/environment/rcb_traj.py b/tasks/robotwin_2_0_place_a2b_right-privileged/environment/rcb_traj.py
index 6e6b2d840a..38ea0a477f 100644
--- a/tasks/robotwin_2_0_place_a2b_right-privileged/environment/rcb_traj.py
+++ b/tasks/robotwin_2_0_place_a2b_right-privileged/environment/rcb_traj.py
@@ -113,6 +113,26 @@ BASE_ARGS: dict[str, Any] = {
}
+class _SortedGlob:
+ """While a scene is built, glob.glob / glob.iglob return their matches sorted. Four RoboTwin tasks (place_a2b_left,
+ place_a2b_right, place_object_scale, place_object_stand: envs/<task>.py load_actors) list an object's model files
+ with glob.glob and then draw one with the seeded np.random.choice; glob returns the file system's directory order,
+ which differs between hosts (and image layers), so one seed built a different object on another host and the frozen
+ instance's hash check failed there (board 0086 claim 5; never rebuilt off the reference workstation). Sorted, the
+ draw depends on the seed alone. No other task of ours lists files while it builds its scene."""
+
+ def __enter__(self):
+ import glob
+ self._glob, self._orig = glob, (glob.glob, glob.iglob)
+ glob.glob = lambda *a, **k: sorted(self._orig[0](*a, **k))
+ glob.iglob = lambda *a, **k: iter(sorted(self._orig[1](*a, **k)))
+ return self
+
+ def __exit__(self, *exc):
+ self._glob.glob, self._glob.iglob = self._orig
+ return False
+
+
def _chdir_root() -> None:
# RoboTwin resolves assets/config paths relative to the repo root and via ROOT_PATH
if os.getcwd() != ROBOTWIN_ROOT and os.path.isdir(ROBOTWIN_ROOT):
@@ -218,7 +238,8 @@ class Session:
if self.enable_render:
sapien.render.set_camera_shader_dir(self.render_shader)
args = _build_args(self.task, self.embodiment, self.seed)
- self.env.setup_demo(**args)
+ with _SortedGlob(): # the object draw depends on the seed alone, not the host's directory order
+ self.env.setup_demo(**args)
if self.enable_render: # setup_scene() switches the shader to ray tracing; put it back for cheap frames
sapien.render.set_camera_shader_dir(self.render_shader)
self._raw_scene = self.env.scene
diff --git a/tasks/robotwin_2_0_place_a2b_right-privileged/task.toml b/tasks/robotwin_2_0_place_a2b_right-privileged/task.toml
index a3ffa35d77..fdd488e2a8 100644
--- a/tasks/robotwin_2_0_place_a2b_right-privileged/task.toml
+++ b/tasks/robotwin_2_0_place_a2b_right-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "robotwin" # RoboTwin 2.0 = SAPIEN 3 / PhysX + CuRobo
robot = "aloha-agilex"
family = "place_a2b_right"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 28831 # scene = environment/instance.json (RoboTwin seed + expected initial-state hash); identity = dataset.toml digest
expert_time_estimate_hours = 0.5
diff --git a/tasks/robotwin_2_0_place_a2b_right-privileged/tests/instance.json b/tasks/robotwin_2_0_place_a2b_right-privileged/tests/instance.json
index a0d65614be..ee648ce7ef 100644
--- a/tasks/robotwin_2_0_place_a2b_right-privileged/tests/instance.json
+++ b/tasks/robotwin_2_0_place_a2b_right-privileged/tests/instance.json
@@ -2,7 +2,7 @@
"task": "place_a2b_right",
"seed": 28831,
"embodiment": "aloha-agilex",
- "state_sha": "90b641f66006",
+ "state_sha": "1ed025a10f0d",
"text": "Use appropriate arm to place object A on the right of object B.",
"n_active_joints": 16
}
\ No newline at end of file
diff --git a/tasks/robotwin_2_0_place_a2b_right-privileged/tests/rcb_traj.py b/tasks/robotwin_2_0_place_a2b_right-privileged/tests/rcb_traj.py
index 6e6b2d840a..38ea0a477f 100644
--- a/tasks/robotwin_2_0_place_a2b_right-privileged/tests/rcb_traj.py
+++ b/tasks/robotwin_2_0_place_a2b_right-privileged/tests/rcb_traj.py
@@ -113,6 +113,26 @@ BASE_ARGS: dict[str, Any] = {
}
+class _SortedGlob:
+ """While a scene is built, glob.glob / glob.iglob return their matches sorted. Four RoboTwin tasks (place_a2b_left,
+ place_a2b_right, place_object_scale, place_object_stand: envs/<task>.py load_actors) list an object's model files
+ with glob.glob and then draw one with the seeded np.random.choice; glob returns the file system's directory order,
+ which differs between hosts (and image layers), so one seed built a different object on another host and the frozen
+ instance's hash check failed there (board 0086 claim 5; never rebuilt off the reference workstation). Sorted, the
+ draw depends on the seed alone. No other task of ours lists files while it builds its scene."""
+
+ def __enter__(self):
+ import glob
+ self._glob, self._orig = glob, (glob.glob, glob.iglob)
+ glob.glob = lambda *a, **k: sorted(self._orig[0](*a, **k))
+ glob.iglob = lambda *a, **k: iter(sorted(self._orig[1](*a, **k)))
+ return self
+
+ def __exit__(self, *exc):
+ self._glob.glob, self._glob.iglob = self._orig
+ return False
+
+
def _chdir_root() -> None:
# RoboTwin resolves assets/config paths relative to the repo root and via ROOT_PATH
if os.getcwd() != ROBOTWIN_ROOT and os.path.isdir(ROBOTWIN_ROOT):
@@ -218,7 +238,8 @@ class Session:
if self.enable_render:
sapien.render.set_camera_shader_dir(self.render_shader)
args = _build_args(self.task, self.embodiment, self.seed)
- self.env.setup_demo(**args)
+ with _SortedGlob(): # the object draw depends on the seed alone, not the host's directory order
+ self.env.setup_demo(**args)
if self.enable_render: # setup_scene() switches the shader to ray tracing; put it back for cheap frames
sapien.render.set_camera_shader_dir(self.render_shader)
self._raw_scene = self.env.scene
diff --git a/tasks/robotwin_2_0_place_a2b_right-standard/README.md b/tasks/robotwin_2_0_place_a2b_right-standard/README.md
index b998a8cdba..b95aa1f4eb 100644
--- a/tasks/robotwin_2_0_place_a2b_right-standard/README.md
+++ b/tasks/robotwin_2_0_place_a2b_right-standard/README.md
@@ -53,3 +53,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
- 2026-10-10 (owner decision, board 0072): the agent's network egress is restricted to the model APIs (`[agent] network_mode = "allowlist"`, the hosts of `scripts/set_agent_egress.py`; since 2026-10-11 these include the package registries, PyPI and the Debian / Ubuntu archives, so that `pip install` and `apt-get install` work); it was open before (Harbor's default), and agents could download upstream sources, reference solutions and robot models from their own shell. The scene, the frozen instance, the tooling and the pass rule are unchanged, so the task keeps its name; results from before this date were obtained with open egress.
- 2026-10-13 (board 0086): the service's copy of the tooling only (`environment/sim/rcb_traj.py`, one line of `environment/sim/server.py`). The tooling now refuses to save or replay a trajectory of more than 200,000 rows, the privileged verifier's cap; a standard episode has no row cap, so the service lifts it for its own session (`S.max_rows = None`). Nothing an agent sees and nothing the service grades changes.
- 2026-10-13 (board 0100): base image `eai-robotwin` 0.1.2, whose scrub of RoboTwin's expert also removes the reference poses that `load_actors` computed only for the expert. The standard-mode agent never reaches the env, so nothing it observes changes; the scene, the instance and the judge are unchanged.
+- 2026-10-14 (owner decision C2, board 0113): **the frozen instance was re-frozen** with the privileged twin (same seed, initial-state hash 90b641f66006 -> 1ed025a10f0d; object A `081_playingcards` (model 2), object B `107_soap` (model 2)). RoboTwin's `load_actors` for this task lists the object's model files with `glob.glob` and draws one with the seeded `np.random.choice`; `glob` returns the file system's directory order, which differs between hosts, so the same seed built a different object on another host and the frozen instance's hash check failed there (board 0086; no run of this task ever got past it on the pool's hosts); the service rebuilt the instance and gave up with `instance_mismatch`. The service's copy of the tooling (`environment/sim/rcb_traj.py`) now sorts `glob`'s matches while it builds the scene. The task text, the success check and the pass rule are unchanged.
diff --git a/tasks/robotwin_2_0_place_a2b_right-standard/environment/sim/instance.json b/tasks/robotwin_2_0_place_a2b_right-standard/environment/sim/instance.json
index a0d65614be..ee648ce7ef 100644
--- a/tasks/robotwin_2_0_place_a2b_right-standard/environment/sim/instance.json
+++ b/tasks/robotwin_2_0_place_a2b_right-standard/environment/sim/instance.json
@@ -2,7 +2,7 @@
"task": "place_a2b_right",
"seed": 28831,
"embodiment": "aloha-agilex",
- "state_sha": "90b641f66006",
+ "state_sha": "1ed025a10f0d",
"text": "Use appropriate arm to place object A on the right of object B.",
"n_active_joints": 16
}
\ No newline at end of file
diff --git a/tasks/robotwin_2_0_place_a2b_right-standard/environment/sim/rcb_traj.py b/tasks/robotwin_2_0_place_a2b_right-standard/environment/sim/rcb_traj.py
index 6e6b2d840a..38ea0a477f 100644
--- a/tasks/robotwin_2_0_place_a2b_right-standard/environment/sim/rcb_traj.py
+++ b/tasks/robotwin_2_0_place_a2b_right-standard/environment/sim/rcb_traj.py
@@ -113,6 +113,26 @@ BASE_ARGS: dict[str, Any] = {
}
+class _SortedGlob:
+ """While a scene is built, glob.glob / glob.iglob return their matches sorted. Four RoboTwin tasks (place_a2b_left,
+ place_a2b_right, place_object_scale, place_object_stand: envs/<task>.py load_actors) list an object's model files
+ with glob.glob and then draw one with the seeded np.random.choice; glob returns the file system's directory order,
+ which differs between hosts (and image layers), so one seed built a different object on another host and the frozen
+ instance's hash check failed there (board 0086 claim 5; never rebuilt off the reference workstation). Sorted, the
+ draw depends on the seed alone. No other task of ours lists files while it builds its scene."""
+
+ def __enter__(self):
+ import glob
+ self._glob, self._orig = glob, (glob.glob, glob.iglob)
+ glob.glob = lambda *a, **k: sorted(self._orig[0](*a, **k))
+ glob.iglob = lambda *a, **k: iter(sorted(self._orig[1](*a, **k)))
+ return self
+
+ def __exit__(self, *exc):
+ self._glob.glob, self._glob.iglob = self._orig
+ return False
+
+
def _chdir_root() -> None:
# RoboTwin resolves assets/config paths relative to the repo root and via ROOT_PATH
if os.getcwd() != ROBOTWIN_ROOT and os.path.isdir(ROBOTWIN_ROOT):
@@ -218,7 +238,8 @@ class Session:
if self.enable_render:
sapien.render.set_camera_shader_dir(self.render_shader)
args = _build_args(self.task, self.embodiment, self.seed)
- self.env.setup_demo(**args)
+ with _SortedGlob(): # the object draw depends on the seed alone, not the host's directory order
+ self.env.setup_demo(**args)
if self.enable_render: # setup_scene() switches the shader to ray tracing; put it back for cheap frames
sapien.render.set_camera_shader_dir(self.render_shader)
self._raw_scene = self.env.scene
diff --git a/tasks/robotwin_2_0_place_a2b_right-standard/task.toml b/tasks/robotwin_2_0_place_a2b_right-standard/task.toml
index dc3f1b88fe..4aacdda533 100644
--- a/tasks/robotwin_2_0_place_a2b_right-standard/task.toml
+++ b/tasks/robotwin_2_0_place_a2b_right-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "robotwin"
robot = "aloha-agilex"
family = "place_a2b_right"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 28831
diff --git a/tasks/robotwin_2_0_place_bread_basket-privileged/README.md b/tasks/robotwin_2_0_place_bread_basket-privileged/README.md
index d3df179141..8b3576aa54 100644
--- a/tasks/robotwin_2_0_place_bread_basket-privileged/README.md
+++ b/tasks/robotwin_2_0_place_bread_basket-privileged/README.md
@@ -55,3 +55,4 @@ If there is one bread on the table, use one arm to grab the bread and put it in
- 2026-10-10 (owner decision, board 0072): the agent's network egress is restricted to the model APIs (`[agent] network_mode = "allowlist"`, the hosts of `scripts/set_agent_egress.py`; since 2026-10-11 these include the package registries, PyPI and the Debian / Ubuntu archives, so that `pip install` and `apt-get install` work); it was open before (Harbor's default), and agents could download upstream sources, reference solutions and robot models from their own shell. The scene, the frozen instance, the tooling and the pass rule are unchanged, so the task keeps its name; results from before this date were obtained with open egress.
- 2026-10-13 (board 0086): the tooling refuses a trajectory the verifier cannot grade (`environment/rcb_traj.py`, `tests/rcb_traj.py`). `Session.save_trajectory` and `Session.replay` raise `ValueError` for more than 200,000 rows (`rcb_traj.MAX_T`, the verifier's `MAX_T` in `tests/grade.py`: 800 s of physics), and `record_stop` warns on stderr. Before, the tool saved and replayed such a file ("success") and the verifier then graded it `bad_trajectory`; the instruction has stated the cap since 2026-10-10. The verifier, the scene, the instance and the pass rule are unchanged: every trajectory is graded as before.
- 2026-10-13 (board 0100): base image `eai-robotwin` 0.1.2. Its scrub of RoboTwin's expert also removes the reference poses that 18 tasks' `load_actors` computed only for the expert, and the expert's comments above the scrubbed method bodies (`images/robotwin/scrub_oracles.py`). This task's `load_actors` set no such value: its scene, instance, tooling, verifier and pass rule are unchanged.
+- 2026-10-14 (board 0113): the tooling only (`environment/rcb_traj.py`, `tests/rcb_traj.py`): while it builds the scene, `glob.glob` returns its matches sorted, so that a task which lists files to draw an object depends on its seed alone, not on the host's directory order (four RoboTwin tasks did; this one lists no files while it builds its scene). Its scene, frozen instance (same initial-state hash), verifier and pass rule are unchanged.
diff --git a/tasks/robotwin_2_0_place_bread_basket-privileged/environment/rcb_traj.py b/tasks/robotwin_2_0_place_bread_basket-privileged/environment/rcb_traj.py
index 11dce323ac..85634b3439 100644
--- a/tasks/robotwin_2_0_place_bread_basket-privileged/environment/rcb_traj.py
+++ b/tasks/robotwin_2_0_place_bread_basket-privileged/environment/rcb_traj.py
@@ -113,6 +113,26 @@ BASE_ARGS: dict[str, Any] = {
}
+class _SortedGlob:
+ """While a scene is built, glob.glob / glob.iglob return their matches sorted. Four RoboTwin tasks (place_a2b_left,
+ place_a2b_right, place_object_scale, place_object_stand: envs/<task>.py load_actors) list an object's model files
+ with glob.glob and then draw one with the seeded np.random.choice; glob returns the file system's directory order,
+ which differs between hosts (and image layers), so one seed built a different object on another host and the frozen
+ instance's hash check failed there (board 0086 claim 5; never rebuilt off the reference workstation). Sorted, the
+ draw depends on the seed alone. No other task of ours lists files while it builds its scene."""
+
+ def __enter__(self):
+ import glob
+ self._glob, self._orig = glob, (glob.glob, glob.iglob)
+ glob.glob = lambda *a, **k: sorted(self._orig[0](*a, **k))
+ glob.iglob = lambda *a, **k: iter(sorted(self._orig[1](*a, **k)))
+ return self
+
+ def __exit__(self, *exc):
+ self._glob.glob, self._glob.iglob = self._orig
+ return False
+
+
def _chdir_root() -> None:
# RoboTwin resolves assets/config paths relative to the repo root and via ROOT_PATH
if os.getcwd() != ROBOTWIN_ROOT and os.path.isdir(ROBOTWIN_ROOT):
@@ -218,7 +238,8 @@ class Session:
if self.enable_render:
sapien.render.set_camera_shader_dir(self.render_shader)
args = _build_args(self.task, self.embodiment, self.seed)
- self.env.setup_demo(**args)
+ with _SortedGlob(): # the object draw depends on the seed alone, not the host's directory order
+ self.env.setup_demo(**args)
if self.enable_render: # setup_scene() switches the shader to ray tracing; put it back for cheap frames
sapien.render.set_camera_shader_dir(self.render_shader)
self._raw_scene = self.env.scene
diff --git a/tasks/robotwin_2_0_place_bread_basket-privileged/task.toml b/tasks/robotwin_2_0_place_bread_basket-privileged/task.toml
index 6012e0a26a..b5f245ca1b 100644
--- a/tasks/robotwin_2_0_place_bread_basket-privileged/task.toml
+++ b/tasks/robotwin_2_0_place_bread_basket-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "robotwin" # RoboTwin 2.0 = SAPIEN 3 / PhysX + CuRobo
robot = "aloha-agilex"
family = "place_bread_basket"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 60417 # scene = environment/instance.json (RoboTwin seed + expected initial-state hash); identity = dataset.toml digest
expert_time_estimate_hours = 0.5
diff --git a/tasks/robotwin_2_0_place_bread_basket-privileged/tests/rcb_traj.py b/tasks/robotwin_2_0_place_bread_basket-privileged/tests/rcb_traj.py
index 11dce323ac..85634b3439 100644
--- a/tasks/robotwin_2_0_place_bread_basket-privileged/tests/rcb_traj.py
+++ b/tasks/robotwin_2_0_place_bread_basket-privileged/tests/rcb_traj.py
@@ -113,6 +113,26 @@ BASE_ARGS: dict[str, Any] = {
}
+class _SortedGlob:
+ """While a scene is built, glob.glob / glob.iglob return their matches sorted. Four RoboTwin tasks (place_a2b_left,
+ place_a2b_right, place_object_scale, place_object_stand: envs/<task>.py load_actors) list an object's model files
+ with glob.glob and then draw one with the seeded np.random.choice; glob returns the file system's directory order,
+ which differs between hosts (and image layers), so one seed built a different object on another host and the frozen
+ instance's hash check failed there (board 0086 claim 5; never rebuilt off the reference workstation). Sorted, the
+ draw depends on the seed alone. No other task of ours lists files while it builds its scene."""
+
+ def __enter__(self):
+ import glob
+ self._glob, self._orig = glob, (glob.glob, glob.iglob)
+ glob.glob = lambda *a, **k: sorted(self._orig[0](*a, **k))
+ glob.iglob = lambda *a, **k: iter(sorted(self._orig[1](*a, **k)))
+ return self
+
+ def __exit__(self, *exc):
+ self._glob.glob, self._glob.iglob = self._orig
+ return False
+
+
def _chdir_root() -> None:
# RoboTwin resolves assets/config paths relative to the repo root and via ROOT_PATH
if os.getcwd() != ROBOTWIN_ROOT and os.path.isdir(ROBOTWIN_ROOT):
@@ -218,7 +238,8 @@ class Session:
if self.enable_render:
sapien.render.set_camera_shader_dir(self.render_shader)
args = _build_args(self.task, self.embodiment, self.seed)
- self.env.setup_demo(**args)
+ with _SortedGlob(): # the object draw depends on the seed alone, not the host's directory order
+ self.env.setup_demo(**args)
if self.enable_render: # setup_scene() switches the shader to ray tracing; put it back for cheap frames
sapien.render.set_camera_shader_dir(self.render_shader)
self._raw_scene = self.env.scene
diff --git a/tasks/robotwin_2_0_place_bread_basket-standard/README.md b/tasks/robotwin_2_0_place_bread_basket-standard/README.md
index 97e52434b6..ab0aeab1db 100644
--- a/tasks/robotwin_2_0_place_bread_basket-standard/README.md
+++ b/tasks/robotwin_2_0_place_bread_basket-standard/README.md
@@ -53,3 +53,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
- 2026-10-10 (owner decision, board 0072): the agent's network egress is restricted to the model APIs (`[agent] network_mode = "allowlist"`, the hosts of `scripts/set_agent_egress.py`; since 2026-10-11 these include the package registries, PyPI and the Debian / Ubuntu archives, so that `pip install` and `apt-get install` work); it was open before (Harbor's default), and agents could download upstream sources, reference solutions and robot models from their own shell. The scene, the frozen instance, the tooling and the pass rule are unchanged, so the task keeps its name; results from before this date were obtained with open egress.
- 2026-10-13 (board 0086): the service's copy of the tooling only (`environment/sim/rcb_traj.py`, one line of `environment/sim/server.py`). The tooling now refuses to save or replay a trajectory of more than 200,000 rows, the privileged verifier's cap; a standard episode has no row cap, so the service lifts it for its own session (`S.max_rows = None`). Nothing an agent sees and nothing the service grades changes.
- 2026-10-13 (board 0100): base image `eai-robotwin` 0.1.2, whose scrub of RoboTwin's expert also removes the reference poses that `load_actors` computed only for the expert. The standard-mode agent never reaches the env, so nothing it observes changes; the scene, the instance and the judge are unchanged.
+- 2026-10-14 (board 0113): the tooling only (`environment/sim/rcb_traj.py`): while it builds the scene, `glob.glob` returns its matches sorted, so that a task which lists files to draw an object depends on its seed alone, not on the host's directory order (four RoboTwin tasks did; this one lists no files while it builds its scene). Its scene, frozen instance (same initial-state hash), verifier and pass rule are unchanged.
diff --git a/tasks/robotwin_2_0_place_bread_basket-standard/environment/sim/rcb_traj.py b/tasks/robotwin_2_0_place_bread_basket-standard/environment/sim/rcb_traj.py
index 11dce323ac..85634b3439 100644
--- a/tasks/robotwin_2_0_place_bread_basket-standard/environment/sim/rcb_traj.py
+++ b/tasks/robotwin_2_0_place_bread_basket-standard/environment/sim/rcb_traj.py
@@ -113,6 +113,26 @@ BASE_ARGS: dict[str, Any] = {
}
+class _SortedGlob:
+ """While a scene is built, glob.glob / glob.iglob return their matches sorted. Four RoboTwin tasks (place_a2b_left,
+ place_a2b_right, place_object_scale, place_object_stand: envs/<task>.py load_actors) list an object's model files
+ with glob.glob and then draw one with the seeded np.random.choice; glob returns the file system's directory order,
+ which differs between hosts (and image layers), so one seed built a different object on another host and the frozen
+ instance's hash check failed there (board 0086 claim 5; never rebuilt off the reference workstation). Sorted, the
+ draw depends on the seed alone. No other task of ours lists files while it builds its scene."""
+
+ def __enter__(self):
+ import glob
+ self._glob, self._orig = glob, (glob.glob, glob.iglob)
+ glob.glob = lambda *a, **k: sorted(self._orig[0](*a, **k))
+ glob.iglob = lambda *a, **k: iter(sorted(self._orig[1](*a, **k)))
+ return self
+
+ def __exit__(self, *exc):
+ self._glob.glob, self._glob.iglob = self._orig
+ return False
+
+
def _chdir_root() -> None:
# RoboTwin resolves assets/config paths relative to the repo root and via ROOT_PATH
if os.getcwd() != ROBOTWIN_ROOT and os.path.isdir(ROBOTWIN_ROOT):
@@ -218,7 +238,8 @@ class Session:
if self.enable_render:
sapien.render.set_camera_shader_dir(self.render_shader)
args = _build_args(self.task, self.embodiment, self.seed)
- self.env.setup_demo(**args)
+ with _SortedGlob(): # the object draw depends on the seed alone, not the host's directory order
+ self.env.setup_demo(**args)
if self.enable_render: # setup_scene() switches the shader to ray tracing; put it back for cheap frames
sapien.render.set_camera_shader_dir(self.render_shader)
self._raw_scene = self.env.scene
diff --git a/tasks/robotwin_2_0_place_bread_basket-standard/task.toml b/tasks/robotwin_2_0_place_bread_basket-standard/task.toml
index 6cf30a7381..8b2c40c25a 100644
--- a/tasks/robotwin_2_0_place_bread_basket-standard/task.toml
+++ b/tasks/robotwin_2_0_place_bread_basket-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "robotwin"
robot = "aloha-agilex"
family = "place_bread_basket"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 60417
diff --git a/tasks/robotwin_2_0_place_bread_skillet-privileged/README.md b/tasks/robotwin_2_0_place_bread_skillet-privileged/README.md
index 23e719a49d..9c70204b75 100644
--- a/tasks/robotwin_2_0_place_bread_skillet-privileged/README.md
+++ b/tasks/robotwin_2_0_place_bread_skillet-privileged/README.md
@@ -55,3 +55,4 @@ If there is one bread on the table, use one arm to grab the bread and put it int
- 2026-10-10 (owner decision, board 0072): the agent's network egress is restricted to the model APIs (`[agent] network_mode = "allowlist"`, the hosts of `scripts/set_agent_egress.py`; since 2026-10-11 these include the package registries, PyPI and the Debian / Ubuntu archives, so that `pip install` and `apt-get install` work); it was open before (Harbor's default), and agents could download upstream sources, reference solutions and robot models from their own shell. The scene, the frozen instance, the tooling and the pass rule are unchanged, so the task keeps its name; results from before this date were obtained with open egress.
- 2026-10-13 (board 0086): the tooling refuses a trajectory the verifier cannot grade (`environment/rcb_traj.py`, `tests/rcb_traj.py`). `Session.save_trajectory` and `Session.replay` raise `ValueError` for more than 200,000 rows (`rcb_traj.MAX_T`, the verifier's `MAX_T` in `tests/grade.py`: 800 s of physics), and `record_stop` warns on stderr. Before, the tool saved and replayed such a file ("success") and the verifier then graded it `bad_trajectory`; the instruction has stated the cap since 2026-10-10. The verifier, the scene, the instance and the pass rule are unchanged: every trajectory is graded as before.
- 2026-10-13 (board 0100): base image `eai-robotwin` 0.1.2. Its scrub of RoboTwin's expert also removes the reference poses that 18 tasks' `load_actors` computed only for the expert, and the expert's comments above the scrubbed method bodies (`images/robotwin/scrub_oracles.py`). This task's `load_actors` set no such value: its scene, instance, tooling, verifier and pass rule are unchanged.
+- 2026-10-14 (board 0113): the tooling only (`environment/rcb_traj.py`, `tests/rcb_traj.py`): while it builds the scene, `glob.glob` returns its matches sorted, so that a task which lists files to draw an object depends on its seed alone, not on the host's directory order (four RoboTwin tasks did; this one lists no files while it builds its scene). Its scene, frozen instance (same initial-state hash), verifier and pass rule are unchanged.
diff --git a/tasks/robotwin_2_0_place_bread_skillet-privileged/environment/rcb_traj.py b/tasks/robotwin_2_0_place_bread_skillet-privileged/environment/rcb_traj.py
index 0c2db3cadc..8c9ab820e0 100644
--- a/tasks/robotwin_2_0_place_bread_skillet-privileged/environment/rcb_traj.py
+++ b/tasks/robotwin_2_0_place_bread_skillet-privileged/environment/rcb_traj.py
@@ -113,6 +113,26 @@ BASE_ARGS: dict[str, Any] = {
}
+class _SortedGlob:
+ """While a scene is built, glob.glob / glob.iglob return their matches sorted. Four RoboTwin tasks (place_a2b_left,
+ place_a2b_right, place_object_scale, place_object_stand: envs/<task>.py load_actors) list an object's model files
+ with glob.glob and then draw one with the seeded np.random.choice; glob returns the file system's directory order,
+ which differs between hosts (and image layers), so one seed built a different object on another host and the frozen
+ instance's hash check failed there (board 0086 claim 5; never rebuilt off the reference workstation). Sorted, the
+ draw depends on the seed alone. No other task of ours lists files while it builds its scene."""
+
+ def __enter__(self):
+ import glob
+ self._glob, self._orig = glob, (glob.glob, glob.iglob)
+ glob.glob = lambda *a, **k: sorted(self._orig[0](*a, **k))
+ glob.iglob = lambda *a, **k: iter(sorted(self._orig[1](*a, **k)))
+ return self
+
+ def __exit__(self, *exc):
+ self._glob.glob, self._glob.iglob = self._orig
+ return False
+
+
def _chdir_root() -> None:
# RoboTwin resolves assets/config paths relative to the repo root and via ROOT_PATH
if os.getcwd() != ROBOTWIN_ROOT and os.path.isdir(ROBOTWIN_ROOT):
@@ -218,7 +238,8 @@ class Session:
if self.enable_render:
sapien.render.set_camera_shader_dir(self.render_shader)
args = _build_args(self.task, self.embodiment, self.seed)
- self.env.setup_demo(**args)
+ with _SortedGlob(): # the object draw depends on the seed alone, not the host's directory order
+ self.env.setup_demo(**args)
if self.enable_render: # setup_scene() switches the shader to ray tracing; put it back for cheap frames
sapien.render.set_camera_shader_dir(self.render_shader)
self._raw_scene = self.env.scene
diff --git a/tasks/robotwin_2_0_place_bread_skillet-privileged/task.toml b/tasks/robotwin_2_0_place_bread_skillet-privileged/task.toml
index 5a0951959f..ab148a31b6 100644
--- a/tasks/robotwin_2_0_place_bread_skillet-privileged/task.toml
+++ b/tasks/robotwin_2_0_place_bread_skillet-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "robotwin" # RoboTwin 2.0 = SAPIEN 3 / PhysX + CuRobo
robot = "aloha-agilex"
family = "place_bread_skillet"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 28831 # scene = environment/instance.json (RoboTwin seed + expected initial-state hash); identity = dataset.toml digest
expert_time_estimate_hours = 0.5
diff --git a/tasks/robotwin_2_0_place_bread_skillet-privileged/tests/rcb_traj.py b/tasks/robotwin_2_0_place_bread_skillet-privileged/tests/rcb_traj.py
index 0c2db3cadc..8c9ab820e0 100644
--- a/tasks/robotwin_2_0_place_bread_skillet-privileged/tests/rcb_traj.py
+++ b/tasks/robotwin_2_0_place_bread_skillet-privileged/tests/rcb_traj.py
@@ -113,6 +113,26 @@ BASE_ARGS: dict[str, Any] = {
}
+class _SortedGlob:
+ """While a scene is built, glob.glob / glob.iglob return their matches sorted. Four RoboTwin tasks (place_a2b_left,
+ place_a2b_right, place_object_scale, place_object_stand: envs/<task>.py load_actors) list an object's model files
+ with glob.glob and then draw one with the seeded np.random.choice; glob returns the file system's directory order,
+ which differs between hosts (and image layers), so one seed built a different object on another host and the frozen
+ instance's hash check failed there (board 0086 claim 5; never rebuilt off the reference workstation). Sorted, the
+ draw depends on the seed alone. No other task of ours lists files while it builds its scene."""
+
+ def __enter__(self):
+ import glob
+ self._glob, self._orig = glob, (glob.glob, glob.iglob)
+ glob.glob = lambda *a, **k: sorted(self._orig[0](*a, **k))
+ glob.iglob = lambda *a, **k: iter(sorted(self._orig[1](*a, **k)))
+ return self
+
+ def __exit__(self, *exc):
+ self._glob.glob, self._glob.iglob = self._orig
+ return False
+
+
def _chdir_root() -> None:
# RoboTwin resolves assets/config paths relative to the repo root and via ROOT_PATH
if os.getcwd() != ROBOTWIN_ROOT and os.path.isdir(ROBOTWIN_ROOT):
@@ -218,7 +238,8 @@ class Session:
if self.enable_render:
sapien.render.set_camera_shader_dir(self.render_shader)
args = _build_args(self.task, self.embodiment, self.seed)
- self.env.setup_demo(**args)
+ with _SortedGlob(): # the object draw depends on the seed alone, not the host's directory order
+ self.env.setup_demo(**args)
if self.enable_render: # setup_scene() switches the shader to ray tracing; put it back for cheap frames
sapien.render.set_camera_shader_dir(self.render_shader)
self._raw_scene = self.env.scene
diff --git a/tasks/robotwin_2_0_place_bread_skillet-standard/README.md b/tasks/robotwin_2_0_place_bread_skillet-standard/README.md
index 7f25c2413f..34c42d233d 100644
--- a/tasks/robotwin_2_0_place_bread_skillet-standard/README.md
+++ b/tasks/robotwin_2_0_place_bread_skillet-standard/README.md
@@ -53,3 +53,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
- 2026-10-10 (owner decision, board 0072): the agent's network egress is restricted to the model APIs (`[agent] network_mode = "allowlist"`, the hosts of `scripts/set_agent_egress.py`; since 2026-10-11 these include the package registries, PyPI and the Debian / Ubuntu archives, so that `pip install` and `apt-get install` work); it was open before (Harbor's default), and agents could download upstream sources, reference solutions and robot models from their own shell. The scene, the frozen instance, the tooling and the pass rule are unchanged, so the task keeps its name; results from before this date were obtained with open egress.
- 2026-10-13 (board 0086): the service's copy of the tooling only (`environment/sim/rcb_traj.py`, one line of `environment/sim/server.py`). The tooling now refuses to save or replay a trajectory of more than 200,000 rows, the privileged verifier's cap; a standard episode has no row cap, so the service lifts it for its own session (`S.max_rows = None`). Nothing an agent sees and nothing the service grades changes.
- 2026-10-13 (board 0100): base image `eai-robotwin` 0.1.2, whose scrub of RoboTwin's expert also removes the reference poses that `load_actors` computed only for the expert. The standard-mode agent never reaches the env, so nothing it observes changes; the scene, the instance and the judge are unchanged.
+- 2026-10-14 (board 0113): the tooling only (`environment/sim/rcb_traj.py`): while it builds the scene, `glob.glob` returns its matches sorted, so that a task which lists files to draw an object depends on its seed alone, not on the host's directory order (four RoboTwin tasks did; this one lists no files while it builds its scene). Its scene, frozen instance (same initial-state hash), verifier and pass rule are unchanged.
diff --git a/tasks/robotwin_2_0_place_bread_skillet-standard/environment/sim/rcb_traj.py b/tasks/robotwin_2_0_place_bread_skillet-standard/environment/sim/rcb_traj.py
index 0c2db3cadc..8c9ab820e0 100644
--- a/tasks/robotwin_2_0_place_bread_skillet-standard/environment/sim/rcb_traj.py
+++ b/tasks/robotwin_2_0_place_bread_skillet-standard/environment/sim/rcb_traj.py
@@ -113,6 +113,26 @@ BASE_ARGS: dict[str, Any] = {
}
+class _SortedGlob:
+ """While a scene is built, glob.glob / glob.iglob return their matches sorted. Four RoboTwin tasks (place_a2b_left,
+ place_a2b_right, place_object_scale, place_object_stand: envs/<task>.py load_actors) list an object's model files
+ with glob.glob and then draw one with the seeded np.random.choice; glob returns the file system's directory order,
+ which differs between hosts (and image layers), so one seed built a different object on another host and the frozen
+ instance's hash check failed there (board 0086 claim 5; never rebuilt off the reference workstation). Sorted, the
+ draw depends on the seed alone. No other task of ours lists files while it builds its scene."""
+
+ def __enter__(self):
+ import glob
+ self._glob, self._orig = glob, (glob.glob, glob.iglob)
+ glob.glob = lambda *a, **k: sorted(self._orig[0](*a, **k))
+ glob.iglob = lambda *a, **k: iter(sorted(self._orig[1](*a, **k)))
+ return self
+
+ def __exit__(self, *exc):
+ self._glob.glob, self._glob.iglob = self._orig
+ return False
+
+
def _chdir_root() -> None:
# RoboTwin resolves assets/config paths relative to the repo root and via ROOT_PATH
if os.getcwd() != ROBOTWIN_ROOT and os.path.isdir(ROBOTWIN_ROOT):
@@ -218,7 +238,8 @@ class Session:
if self.enable_render:
sapien.render.set_camera_shader_dir(self.render_shader)
args = _build_args(self.task, self.embodiment, self.seed)
- self.env.setup_demo(**args)
+ with _SortedGlob(): # the object draw depends on the seed alone, not the host's directory order
+ self.env.setup_demo(**args)
if self.enable_render: # setup_scene() switches the shader to ray tracing; put it back for cheap frames
sapien.render.set_camera_shader_dir(self.render_shader)
self._raw_scene = self.env.scene
diff --git a/tasks/robotwin_2_0_place_bread_skillet-standard/task.toml b/tasks/robotwin_2_0_place_bread_skillet-standard/task.toml
index 62801f0024..e22e700772 100644
--- a/tasks/robotwin_2_0_place_bread_skillet-standard/task.toml
+++ b/tasks/robotwin_2_0_place_bread_skillet-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "robotwin"
robot = "aloha-agilex"
family = "place_bread_skillet"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 28831
diff --git a/tasks/robotwin_2_0_place_burger_fries-privileged/README.md b/tasks/robotwin_2_0_place_burger_fries-privileged/README.md
index f4bb716f2b..4764bb8162 100644
--- a/tasks/robotwin_2_0_place_burger_fries-privileged/README.md
+++ b/tasks/robotwin_2_0_place_burger_fries-privileged/README.md
@@ -55,3 +55,4 @@ Use dual arm to pick the hamburg and frenchfries and put them onto the tray.
- 2026-10-10 (owner decision, board 0072): the agent's network egress is restricted to the model APIs (`[agent] network_mode = "allowlist"`, the hosts of `scripts/set_agent_egress.py`; since 2026-10-11 these include the package registries, PyPI and the Debian / Ubuntu archives, so that `pip install` and `apt-get install` work); it was open before (Harbor's default), and agents could download upstream sources, reference solutions and robot models from their own shell. The scene, the frozen instance, the tooling and the pass rule are unchanged, so the task keeps its name; results from before this date were obtained with open egress.
- 2026-10-13 (board 0086): the tooling refuses a trajectory the verifier cannot grade (`environment/rcb_traj.py`, `tests/rcb_traj.py`). `Session.save_trajectory` and `Session.replay` raise `ValueError` for more than 200,000 rows (`rcb_traj.MAX_T`, the verifier's `MAX_T` in `tests/grade.py`: 800 s of physics), and `record_stop` warns on stderr. Before, the tool saved and replayed such a file ("success") and the verifier then graded it `bad_trajectory`; the instruction has stated the cap since 2026-10-10. The verifier, the scene, the instance and the pass rule are unchanged: every trajectory is graded as before.
- 2026-10-13 (board 0100): base image `eai-robotwin` 0.1.2. Its scrub of RoboTwin's expert also removes the reference poses that 18 tasks' `load_actors` computed only for the expert, and the expert's comments above the scrubbed method bodies (`images/robotwin/scrub_oracles.py`). This task's `load_actors` set no such value: its scene, instance, tooling, verifier and pass rule are unchanged.
+- 2026-10-14 (board 0113): the tooling only (`environment/rcb_traj.py`, `tests/rcb_traj.py`): while it builds the scene, `glob.glob` returns its matches sorted, so that a task which lists files to draw an object depends on its seed alone, not on the host's directory order (four RoboTwin tasks did; this one lists no files while it builds its scene). Its scene, frozen instance (same initial-state hash), verifier and pass rule are unchanged.
diff --git a/tasks/robotwin_2_0_place_burger_fries-privileged/environment/rcb_traj.py b/tasks/robotwin_2_0_place_burger_fries-privileged/environment/rcb_traj.py
index 0c959c6476..08430cb027 100644
--- a/tasks/robotwin_2_0_place_burger_fries-privileged/environment/rcb_traj.py
+++ b/tasks/robotwin_2_0_place_burger_fries-privileged/environment/rcb_traj.py
@@ -113,6 +113,26 @@ BASE_ARGS: dict[str, Any] = {
}
+class _SortedGlob:
+ """While a scene is built, glob.glob / glob.iglob return their matches sorted. Four RoboTwin tasks (place_a2b_left,
+ place_a2b_right, place_object_scale, place_object_stand: envs/<task>.py load_actors) list an object's model files
+ with glob.glob and then draw one with the seeded np.random.choice; glob returns the file system's directory order,
+ which differs between hosts (and image layers), so one seed built a different object on another host and the frozen
+ instance's hash check failed there (board 0086 claim 5; never rebuilt off the reference workstation). Sorted, the
+ draw depends on the seed alone. No other task of ours lists files while it builds its scene."""
+
+ def __enter__(self):
+ import glob
+ self._glob, self._orig = glob, (glob.glob, glob.iglob)
+ glob.glob = lambda *a, **k: sorted(self._orig[0](*a, **k))
+ glob.iglob = lambda *a, **k: iter(sorted(self._orig[1](*a, **k)))
+ return self
+
+ def __exit__(self, *exc):
+ self._glob.glob, self._glob.iglob = self._orig
+ return False
+
+
def _chdir_root() -> None:
# RoboTwin resolves assets/config paths relative to the repo root and via ROOT_PATH
if os.getcwd() != ROBOTWIN_ROOT and os.path.isdir(ROBOTWIN_ROOT):
@@ -218,7 +238,8 @@ class Session:
if self.enable_render:
sapien.render.set_camera_shader_dir(self.render_shader)
args = _build_args(self.task, self.embodiment, self.seed)
- self.env.setup_demo(**args)
+ with _SortedGlob(): # the object draw depends on the seed alone, not the host's directory order
+ self.env.setup_demo(**args)
if self.enable_render: # setup_scene() switches the shader to ray tracing; put it back for cheap frames
sapien.render.set_camera_shader_dir(self.render_shader)
self._raw_scene = self.env.scene
diff --git a/tasks/robotwin_2_0_place_burger_fries-privileged/task.toml b/tasks/robotwin_2_0_place_burger_fries-privileged/task.toml
index 16b29f30c6..b84d4ba015 100644
--- a/tasks/robotwin_2_0_place_burger_fries-privileged/task.toml
+++ b/tasks/robotwin_2_0_place_burger_fries-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "robotwin" # RoboTwin 2.0 = SAPIEN 3 / PhysX + CuRobo
robot = "aloha-agilex"
family = "place_burger_fries"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 60417 # scene = environment/instance.json (RoboTwin seed + expected initial-state hash); identity = dataset.toml digest
expert_time_estimate_hours = 0.5
diff --git a/tasks/robotwin_2_0_place_burger_fries-privileged/tests/rcb_traj.py b/tasks/robotwin_2_0_place_burger_fries-privileged/tests/rcb_traj.py
index 0c959c6476..08430cb027 100644
--- a/tasks/robotwin_2_0_place_burger_fries-privileged/tests/rcb_traj.py
+++ b/tasks/robotwin_2_0_place_burger_fries-privileged/tests/rcb_traj.py
@@ -113,6 +113,26 @@ BASE_ARGS: dict[str, Any] = {
}
+class _SortedGlob:
+ """While a scene is built, glob.glob / glob.iglob return their matches sorted. Four RoboTwin tasks (place_a2b_left,
+ place_a2b_right, place_object_scale, place_object_stand: envs/<task>.py load_actors) list an object's model files
+ with glob.glob and then draw one with the seeded np.random.choice; glob returns the file system's directory order,
+ which differs between hosts (and image layers), so one seed built a different object on another host and the frozen
+ instance's hash check failed there (board 0086 claim 5; never rebuilt off the reference workstation). Sorted, the
+ draw depends on the seed alone. No other task of ours lists files while it builds its scene."""
+
+ def __enter__(self):
+ import glob
+ self._glob, self._orig = glob, (glob.glob, glob.iglob)
+ glob.glob = lambda *a, **k: sorted(self._orig[0](*a, **k))
+ glob.iglob = lambda *a, **k: iter(sorted(self._orig[1](*a, **k)))
+ return self
+
+ def __exit__(self, *exc):
+ self._glob.glob, self._glob.iglob = self._orig
+ return False
+
+
def _chdir_root() -> None:
# RoboTwin resolves assets/config paths relative to the repo root and via ROOT_PATH
if os.getcwd() != ROBOTWIN_ROOT and os.path.isdir(ROBOTWIN_ROOT):
@@ -218,7 +238,8 @@ class Session:
if self.enable_render:
sapien.render.set_camera_shader_dir(self.render_shader)
args = _build_args(self.task, self.embodiment, self.seed)
- self.env.setup_demo(**args)
+ with _SortedGlob(): # the object draw depends on the seed alone, not the host's directory order
+ self.env.setup_demo(**args)
if self.enable_render: # setup_scene() switches the shader to ray tracing; put it back for cheap frames
sapien.render.set_camera_shader_dir(self.render_shader)
self._raw_scene = self.env.scene
diff --git a/tasks/robotwin_2_0_place_burger_fries-standard/README.md b/tasks/robotwin_2_0_place_burger_fries-standard/README.md
index 97fed29734..68a1578bfe 100644
--- a/tasks/robotwin_2_0_place_burger_fries-standard/README.md
+++ b/tasks/robotwin_2_0_place_burger_fries-standard/README.md
@@ -53,3 +53,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
- 2026-10-10 (owner decision, board 0072): the agent's network egress is restricted to the model APIs (`[agent] network_mode = "allowlist"`, the hosts of `scripts/set_agent_egress.py`; since 2026-10-11 these include the package registries, PyPI and the Debian / Ubuntu archives, so that `pip install` and `apt-get install` work); it was open before (Harbor's default), and agents could download upstream sources, reference solutions and robot models from their own shell. The scene, the frozen instance, the tooling and the pass rule are unchanged, so the task keeps its name; results from before this date were obtained with open egress.
- 2026-10-13 (board 0086): the service's copy of the tooling only (`environment/sim/rcb_traj.py`, one line of `environment/sim/server.py`). The tooling now refuses to save or replay a trajectory of more than 200,000 rows, the privileged verifier's cap; a standard episode has no row cap, so the service lifts it for its own session (`S.max_rows = None`). Nothing an agent sees and nothing the service grades changes.
- 2026-10-13 (board 0100): base image `eai-robotwin` 0.1.2, whose scrub of RoboTwin's expert also removes the reference poses that `load_actors` computed only for the expert. The standard-mode agent never reaches the env, so nothing it observes changes; the scene, the instance and the judge are unchanged.
+- 2026-10-14 (board 0113): the tooling only (`environment/sim/rcb_traj.py`): while it builds the scene, `glob.glob` returns its matches sorted, so that a task which lists files to draw an object depends on its seed alone, not on the host's directory order (four RoboTwin tasks did; this one lists no files while it builds its scene). Its scene, frozen instance (same initial-state hash), verifier and pass rule are unchanged.
diff --git a/tasks/robotwin_2_0_place_burger_fries-standard/environment/sim/rcb_traj.py b/tasks/robotwin_2_0_place_burger_fries-standard/environment/sim/rcb_traj.py
index 0c959c6476..08430cb027 100644
--- a/tasks/robotwin_2_0_place_burger_fries-standard/environment/sim/rcb_traj.py
+++ b/tasks/robotwin_2_0_place_burger_fries-standard/environment/sim/rcb_traj.py
@@ -113,6 +113,26 @@ BASE_ARGS: dict[str, Any] = {
}
+class _SortedGlob:
+ """While a scene is built, glob.glob / glob.iglob return their matches sorted. Four RoboTwin tasks (place_a2b_left,
+ place_a2b_right, place_object_scale, place_object_stand: envs/<task>.py load_actors) list an object's model files
+ with glob.glob and then draw one with the seeded np.random.choice; glob returns the file system's directory order,
+ which differs between hosts (and image layers), so one seed built a different object on another host and the frozen
+ instance's hash check failed there (board 0086 claim 5; never rebuilt off the reference workstation). Sorted, the
+ draw depends on the seed alone. No other task of ours lists files while it builds its scene."""
+
+ def __enter__(self):
+ import glob
+ self._glob, self._orig = glob, (glob.glob, glob.iglob)
+ glob.glob = lambda *a, **k: sorted(self._orig[0](*a, **k))
+ glob.iglob = lambda *a, **k: iter(sorted(self._orig[1](*a, **k)))
+ return self
+
+ def __exit__(self, *exc):
+ self._glob.glob, self._glob.iglob = self._orig
+ return False
+
+
def _chdir_root() -> None:
# RoboTwin resolves assets/config paths relative to the repo root and via ROOT_PATH
if os.getcwd() != ROBOTWIN_ROOT and os.path.isdir(ROBOTWIN_ROOT):
@@ -218,7 +238,8 @@ class Session:
if self.enable_render:
sapien.render.set_camera_shader_dir(self.render_shader)
args = _build_args(self.task, self.embodiment, self.seed)
- self.env.setup_demo(**args)
+ with _SortedGlob(): # the object draw depends on the seed alone, not the host's directory order
+ self.env.setup_demo(**args)
if self.enable_render: # setup_scene() switches the shader to ray tracing; put it back for cheap frames
sapien.render.set_camera_shader_dir(self.render_shader)
self._raw_scene = self.env.scene
diff --git a/tasks/robotwin_2_0_place_burger_fries-standard/task.toml b/tasks/robotwin_2_0_place_burger_fries-standard/task.toml
index 2f02feec88..975784a753 100644
--- a/tasks/robotwin_2_0_place_burger_fries-standard/task.toml
+++ b/tasks/robotwin_2_0_place_burger_fries-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "robotwin"
robot = "aloha-agilex"
family = "place_burger_fries"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 60417
diff --git a/tasks/robotwin_2_0_place_can_basket-privileged/README.md b/tasks/robotwin_2_0_place_can_basket-privileged/README.md
index 220ce5cb2a..ddf0903441 100644
--- a/tasks/robotwin_2_0_place_can_basket-privileged/README.md
+++ b/tasks/robotwin_2_0_place_can_basket-privileged/README.md
@@ -55,3 +55,4 @@ Put the can into the basket, then lift the basket with the other arm and hold it
- 2026-10-10 (owner decision, board 0072): the agent's network egress is restricted to the model APIs (`[agent] network_mode = "allowlist"`, the hosts of `scripts/set_agent_egress.py`; since 2026-10-11 these include the package registries, PyPI and the Debian / Ubuntu archives, so that `pip install` and `apt-get install` work); it was open before (Harbor's default), and agents could download upstream sources, reference solutions and robot models from their own shell. The scene, the frozen instance, the tooling and the pass rule are unchanged, so the task keeps its name; results from before this date were obtained with open egress.
- 2026-10-13 (board 0086): the tooling refuses a trajectory the verifier cannot grade (`environment/rcb_traj.py`, `tests/rcb_traj.py`). `Session.save_trajectory` and `Session.replay` raise `ValueError` for more than 200,000 rows (`rcb_traj.MAX_T`, the verifier's `MAX_T` in `tests/grade.py`: 800 s of physics), and `record_stop` warns on stderr. Before, the tool saved and replayed such a file ("success") and the verifier then graded it `bad_trajectory`; the instruction has stated the cap since 2026-10-10. The verifier, the scene, the instance and the pass rule are unchanged: every trajectory is graded as before.
- 2026-10-13 (board 0100): base image `eai-robotwin` 0.1.2. Its scrub of RoboTwin's expert also removes the reference poses that 18 tasks' `load_actors` computed only for the expert, and the expert's comments above the scrubbed method bodies (`images/robotwin/scrub_oracles.py`). This task's `load_actors` set no such value: its scene, instance, tooling, verifier and pass rule are unchanged.
+- 2026-10-14 (board 0113): the tooling only (`environment/rcb_traj.py`, `tests/rcb_traj.py`): while it builds the scene, `glob.glob` returns its matches sorted, so that a task which lists files to draw an object depends on its seed alone, not on the host's directory order (four RoboTwin tasks did; this one lists no files while it builds its scene). Its scene, frozen instance (same initial-state hash), verifier and pass rule are unchanged.
diff --git a/tasks/robotwin_2_0_place_can_basket-privileged/environment/rcb_traj.py b/tasks/robotwin_2_0_place_can_basket-privileged/environment/rcb_traj.py
index 26a1fec314..758cdbc20d 100644
--- a/tasks/robotwin_2_0_place_can_basket-privileged/environment/rcb_traj.py
+++ b/tasks/robotwin_2_0_place_can_basket-privileged/environment/rcb_traj.py
@@ -113,6 +113,26 @@ BASE_ARGS: dict[str, Any] = {
}
+class _SortedGlob:
+ """While a scene is built, glob.glob / glob.iglob return their matches sorted. Four RoboTwin tasks (place_a2b_left,
+ place_a2b_right, place_object_scale, place_object_stand: envs/<task>.py load_actors) list an object's model files
+ with glob.glob and then draw one with the seeded np.random.choice; glob returns the file system's directory order,
+ which differs between hosts (and image layers), so one seed built a different object on another host and the frozen
+ instance's hash check failed there (board 0086 claim 5; never rebuilt off the reference workstation). Sorted, the
+ draw depends on the seed alone. No other task of ours lists files while it builds its scene."""
+
+ def __enter__(self):
+ import glob
+ self._glob, self._orig = glob, (glob.glob, glob.iglob)
+ glob.glob = lambda *a, **k: sorted(self._orig[0](*a, **k))
+ glob.iglob = lambda *a, **k: iter(sorted(self._orig[1](*a, **k)))
+ return self
+
+ def __exit__(self, *exc):
+ self._glob.glob, self._glob.iglob = self._orig
+ return False
+
+
def _chdir_root() -> None:
# RoboTwin resolves assets/config paths relative to the repo root and via ROOT_PATH
if os.getcwd() != ROBOTWIN_ROOT and os.path.isdir(ROBOTWIN_ROOT):
@@ -218,7 +238,8 @@ class Session:
if self.enable_render:
sapien.render.set_camera_shader_dir(self.render_shader)
args = _build_args(self.task, self.embodiment, self.seed)
- self.env.setup_demo(**args)
+ with _SortedGlob(): # the object draw depends on the seed alone, not the host's directory order
+ self.env.setup_demo(**args)
if self.enable_render: # setup_scene() switches the shader to ray tracing; put it back for cheap frames
sapien.render.set_camera_shader_dir(self.render_shader)
self._raw_scene = self.env.scene
diff --git a/tasks/robotwin_2_0_place_can_basket-privileged/task.toml b/tasks/robotwin_2_0_place_can_basket-privileged/task.toml
index e1aa8d2e09..6298c92029 100644
--- a/tasks/robotwin_2_0_place_can_basket-privileged/task.toml
+++ b/tasks/robotwin_2_0_place_can_basket-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "robotwin" # RoboTwin 2.0 = SAPIEN 3 / PhysX + CuRobo
robot = "aloha-agilex"
family = "place_can_basket"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 60417 # scene = environment/instance.json (RoboTwin seed + expected initial-state hash); identity = dataset.toml digest
expert_time_estimate_hours = 0.5
diff --git a/tasks/robotwin_2_0_place_can_basket-privileged/tests/rcb_traj.py b/tasks/robotwin_2_0_place_can_basket-privileged/tests/rcb_traj.py
index 26a1fec314..758cdbc20d 100644
--- a/tasks/robotwin_2_0_place_can_basket-privileged/tests/rcb_traj.py
+++ b/tasks/robotwin_2_0_place_can_basket-privileged/tests/rcb_traj.py
@@ -113,6 +113,26 @@ BASE_ARGS: dict[str, Any] = {
}
+class _SortedGlob:
+ """While a scene is built, glob.glob / glob.iglob return their matches sorted. Four RoboTwin tasks (place_a2b_left,
+ place_a2b_right, place_object_scale, place_object_stand: envs/<task>.py load_actors) list an object's model files
+ with glob.glob and then draw one with the seeded np.random.choice; glob returns the file system's directory order,
+ which differs between hosts (and image layers), so one seed built a different object on another host and the frozen
+ instance's hash check failed there (board 0086 claim 5; never rebuilt off the reference workstation). Sorted, the
+ draw depends on the seed alone. No other task of ours lists files while it builds its scene."""
+
+ def __enter__(self):
+ import glob
+ self._glob, self._orig = glob, (glob.glob, glob.iglob)
+ glob.glob = lambda *a, **k: sorted(self._orig[0](*a, **k))
+ glob.iglob = lambda *a, **k: iter(sorted(self._orig[1](*a, **k)))
+ return self
+
+ def __exit__(self, *exc):
+ self._glob.glob, self._glob.iglob = self._orig
+ return False
+
+
def _chdir_root() -> None:
# RoboTwin resolves assets/config paths relative to the repo root and via ROOT_PATH
if os.getcwd() != ROBOTWIN_ROOT and os.path.isdir(ROBOTWIN_ROOT):
@@ -218,7 +238,8 @@ class Session:
if self.enable_render:
sapien.render.set_camera_shader_dir(self.render_shader)
args = _build_args(self.task, self.embodiment, self.seed)
- self.env.setup_demo(**args)
+ with _SortedGlob(): # the object draw depends on the seed alone, not the host's directory order
+ self.env.setup_demo(**args)
if self.enable_render: # setup_scene() switches the shader to ray tracing; put it back for cheap frames
sapien.render.set_camera_shader_dir(self.render_shader)
self._raw_scene = self.env.scene
diff --git a/tasks/robotwin_2_0_place_can_basket-standard/README.md b/tasks/robotwin_2_0_place_can_basket-standard/README.md
index 0f0e57959b..ac293d9894 100644
--- a/tasks/robotwin_2_0_place_can_basket-standard/README.md
+++ b/tasks/robotwin_2_0_place_can_basket-standard/README.md
@@ -52,3 +52,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
- 2026-10-10 (owner decision, board 0072): the agent's network egress is restricted to the model APIs (`[agent] network_mode = "allowlist"`, the hosts of `scripts/set_agent_egress.py`; since 2026-10-11 these include the package registries, PyPI and the Debian / Ubuntu archives, so that `pip install` and `apt-get install` work); it was open before (Harbor's default), and agents could download upstream sources, reference solutions and robot models from their own shell. The scene, the frozen instance, the tooling and the pass rule are unchanged, so the task keeps its name; results from before this date were obtained with open egress.
- 2026-10-13 (board 0086): the service's copy of the tooling only (`environment/sim/rcb_traj.py`, one line of `environment/sim/server.py`). The tooling now refuses to save or replay a trajectory of more than 200,000 rows, the privileged verifier's cap; a standard episode has no row cap, so the service lifts it for its own session (`S.max_rows = None`). Nothing an agent sees and nothing the service grades changes.
- 2026-10-13 (board 0100): base image `eai-robotwin` 0.1.2, whose scrub of RoboTwin's expert also removes the reference poses that `load_actors` computed only for the expert. The standard-mode agent never reaches the env, so nothing it observes changes; the scene, the instance and the judge are unchanged.
+- 2026-10-14 (board 0113): the tooling only (`environment/sim/rcb_traj.py`): while it builds the scene, `glob.glob` returns its matches sorted, so that a task which lists files to draw an object depends on its seed alone, not on the host's directory order (four RoboTwin tasks did; this one lists no files while it builds its scene). Its scene, frozen instance (same initial-state hash), verifier and pass rule are unchanged.
diff --git a/tasks/robotwin_2_0_place_can_basket-standard/environment/sim/rcb_traj.py b/tasks/robotwin_2_0_place_can_basket-standard/environment/sim/rcb_traj.py
index 26a1fec314..758cdbc20d 100644
--- a/tasks/robotwin_2_0_place_can_basket-standard/environment/sim/rcb_traj.py
+++ b/tasks/robotwin_2_0_place_can_basket-standard/environment/sim/rcb_traj.py
@@ -113,6 +113,26 @@ BASE_ARGS: dict[str, Any] = {
}
+class _SortedGlob:
+ """While a scene is built, glob.glob / glob.iglob return their matches sorted. Four RoboTwin tasks (place_a2b_left,
+ place_a2b_right, place_object_scale, place_object_stand: envs/<task>.py load_actors) list an object's model files
+ with glob.glob and then draw one with the seeded np.random.choice; glob returns the file system's directory order,
+ which differs between hosts (and image layers), so one seed built a different object on another host and the frozen
+ instance's hash check failed there (board 0086 claim 5; never rebuilt off the reference workstation). Sorted, the
+ draw depends on the seed alone. No other task of ours lists files while it builds its scene."""
+
+ def __enter__(self):
+ import glob
+ self._glob, self._orig = glob, (glob.glob, glob.iglob)
+ glob.glob = lambda *a, **k: sorted(self._orig[0](*a, **k))
+ glob.iglob = lambda *a, **k: iter(sorted(self._orig[1](*a, **k)))
+ return self
+
+ def __exit__(self, *exc):
+ self._glob.glob, self._glob.iglob = self._orig
+ return False
+
+
def _chdir_root() -> None:
# RoboTwin resolves assets/config paths relative to the repo root and via ROOT_PATH
if os.getcwd() != ROBOTWIN_ROOT and os.path.isdir(ROBOTWIN_ROOT):
@@ -218,7 +238,8 @@ class Session:
if self.enable_render:
sapien.render.set_camera_shader_dir(self.render_shader)
args = _build_args(self.task, self.embodiment, self.seed)
- self.env.setup_demo(**args)
+ with _SortedGlob(): # the object draw depends on the seed alone, not the host's directory order
+ self.env.setup_demo(**args)
if self.enable_render: # setup_scene() switches the shader to ray tracing; put it back for cheap frames
sapien.render.set_camera_shader_dir(self.render_shader)
self._raw_scene = self.env.scene
diff --git a/tasks/robotwin_2_0_place_can_basket-standard/task.toml b/tasks/robotwin_2_0_place_can_basket-standard/task.toml
index 67e9688435..a6d9df64d8 100644
--- a/tasks/robotwin_2_0_place_can_basket-standard/task.toml
+++ b/tasks/robotwin_2_0_place_can_basket-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "robotwin"
robot = "aloha-agilex"
family = "place_can_basket"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 60417
diff --git a/tasks/robotwin_2_0_place_cans_plasticbox-privileged/README.md b/tasks/robotwin_2_0_place_cans_plasticbox-privileged/README.md
index e51f2095e6..6d9198216a 100644
--- a/tasks/robotwin_2_0_place_cans_plasticbox-privileged/README.md
+++ b/tasks/robotwin_2_0_place_cans_plasticbox-privileged/README.md
@@ -55,3 +55,4 @@ Use dual arm to pick and place cans into plasticbox.
- 2026-10-10 (owner decision, board 0072): the agent's network egress is restricted to the model APIs (`[agent] network_mode = "allowlist"`, the hosts of `scripts/set_agent_egress.py`; since 2026-10-11 these include the package registries, PyPI and the Debian / Ubuntu archives, so that `pip install` and `apt-get install` work); it was open before (Harbor's default), and agents could download upstream sources, reference solutions and robot models from their own shell. The scene, the frozen instance, the tooling and the pass rule are unchanged, so the task keeps its name; results from before this date were obtained with open egress.
- 2026-10-13 (board 0086): the tooling refuses a trajectory the verifier cannot grade (`environment/rcb_traj.py`, `tests/rcb_traj.py`). `Session.save_trajectory` and `Session.replay` raise `ValueError` for more than 200,000 rows (`rcb_traj.MAX_T`, the verifier's `MAX_T` in `tests/grade.py`: 800 s of physics), and `record_stop` warns on stderr. Before, the tool saved and replayed such a file ("success") and the verifier then graded it `bad_trajectory`; the instruction has stated the cap since 2026-10-10. The verifier, the scene, the instance and the pass rule are unchanged: every trajectory is graded as before.
- 2026-10-13 (board 0100): base image `eai-robotwin` 0.1.2. Its scrub of RoboTwin's expert also removes the reference poses that 18 tasks' `load_actors` computed only for the expert, and the expert's comments above the scrubbed method bodies (`images/robotwin/scrub_oracles.py`). This task's `load_actors` set no such value: its scene, instance, tooling, verifier and pass rule are unchanged.
+- 2026-10-14 (board 0113): the tooling only (`environment/rcb_traj.py`, `tests/rcb_traj.py`): while it builds the scene, `glob.glob` returns its matches sorted, so that a task which lists files to draw an object depends on its seed alone, not on the host's directory order (four RoboTwin tasks did; this one lists no files while it builds its scene). Its scene, frozen instance (same initial-state hash), verifier and pass rule are unchanged.
diff --git a/tasks/robotwin_2_0_place_cans_plasticbox-privileged/environment/rcb_traj.py b/tasks/robotwin_2_0_place_cans_plasticbox-privileged/environment/rcb_traj.py
index 761dec4aa1..868dfd57d3 100644
--- a/tasks/robotwin_2_0_place_cans_plasticbox-privileged/environment/rcb_traj.py
+++ b/tasks/robotwin_2_0_place_cans_plasticbox-privileged/environment/rcb_traj.py
@@ -113,6 +113,26 @@ BASE_ARGS: dict[str, Any] = {
}
+class _SortedGlob:
+ """While a scene is built, glob.glob / glob.iglob return their matches sorted. Four RoboTwin tasks (place_a2b_left,
+ place_a2b_right, place_object_scale, place_object_stand: envs/<task>.py load_actors) list an object's model files
+ with glob.glob and then draw one with the seeded np.random.choice; glob returns the file system's directory order,
+ which differs between hosts (and image layers), so one seed built a different object on another host and the frozen
+ instance's hash check failed there (board 0086 claim 5; never rebuilt off the reference workstation). Sorted, the
+ draw depends on the seed alone. No other task of ours lists files while it builds its scene."""
+
+ def __enter__(self):
+ import glob
+ self._glob, self._orig = glob, (glob.glob, glob.iglob)
+ glob.glob = lambda *a, **k: sorted(self._orig[0](*a, **k))
+ glob.iglob = lambda *a, **k: iter(sorted(self._orig[1](*a, **k)))
+ return self
+
+ def __exit__(self, *exc):
+ self._glob.glob, self._glob.iglob = self._orig
+ return False
+
+
def _chdir_root() -> None:
# RoboTwin resolves assets/config paths relative to the repo root and via ROOT_PATH
if os.getcwd() != ROBOTWIN_ROOT and os.path.isdir(ROBOTWIN_ROOT):
@@ -218,7 +238,8 @@ class Session:
if self.enable_render:
sapien.render.set_camera_shader_dir(self.render_shader)
args = _build_args(self.task, self.embodiment, self.seed)
- self.env.setup_demo(**args)
+ with _SortedGlob(): # the object draw depends on the seed alone, not the host's directory order
+ self.env.setup_demo(**args)
if self.enable_render: # setup_scene() switches the shader to ray tracing; put it back for cheap frames
sapien.render.set_camera_shader_dir(self.render_shader)
self._raw_scene = self.env.scene
diff --git a/tasks/robotwin_2_0_place_cans_plasticbox-privileged/task.toml b/tasks/robotwin_2_0_place_cans_plasticbox-privileged/task.toml
index 7278e4abc9..07471d8226 100644
--- a/tasks/robotwin_2_0_place_cans_plasticbox-privileged/task.toml
+++ b/tasks/robotwin_2_0_place_cans_plasticbox-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "robotwin" # RoboTwin 2.0 = SAPIEN 3 / PhysX + CuRobo
robot = "aloha-agilex"
family = "place_cans_plasticbox"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 60417 # scene = environment/instance.json (RoboTwin seed + expected initial-state hash); identity = dataset.toml digest
expert_time_estimate_hours = 0.5
diff --git a/tasks/robotwin_2_0_place_cans_plasticbox-privileged/tests/rcb_traj.py b/tasks/robotwin_2_0_place_cans_plasticbox-privileged/tests/rcb_traj.py
index 761dec4aa1..868dfd57d3 100644
--- a/tasks/robotwin_2_0_place_cans_plasticbox-privileged/tests/rcb_traj.py
+++ b/tasks/robotwin_2_0_place_cans_plasticbox-privileged/tests/rcb_traj.py
@@ -113,6 +113,26 @@ BASE_ARGS: dict[str, Any] = {
}
+class _SortedGlob:
+ """While a scene is built, glob.glob / glob.iglob return their matches sorted. Four RoboTwin tasks (place_a2b_left,
+ place_a2b_right, place_object_scale, place_object_stand: envs/<task>.py load_actors) list an object's model files
+ with glob.glob and then draw one with the seeded np.random.choice; glob returns the file system's directory order,
+ which differs between hosts (and image layers), so one seed built a different object on another host and the frozen
+ instance's hash check failed there (board 0086 claim 5; never rebuilt off the reference workstation). Sorted, the
+ draw depends on the seed alone. No other task of ours lists files while it builds its scene."""
+
+ def __enter__(self):
+ import glob
+ self._glob, self._orig = glob, (glob.glob, glob.iglob)
+ glob.glob = lambda *a, **k: sorted(self._orig[0](*a, **k))
+ glob.iglob = lambda *a, **k: iter(sorted(self._orig[1](*a, **k)))
+ return self
+
+ def __exit__(self, *exc):
+ self._glob.glob, self._glob.iglob = self._orig
+ return False
+
+
def _chdir_root() -> None:
# RoboTwin resolves assets/config paths relative to the repo root and via ROOT_PATH
if os.getcwd() != ROBOTWIN_ROOT and os.path.isdir(ROBOTWIN_ROOT):
@@ -218,7 +238,8 @@ class Session:
if self.enable_render:
sapien.render.set_camera_shader_dir(self.render_shader)
args = _build_args(self.task, self.embodiment, self.seed)
- self.env.setup_demo(**args)
+ with _SortedGlob(): # the object draw depends on the seed alone, not the host's directory order
+ self.env.setup_demo(**args)
if self.enable_render: # setup_scene() switches the shader to ray tracing; put it back for cheap frames
sapien.render.set_camera_shader_dir(self.render_shader)
self._raw_scene = self.env.scene
diff --git a/tasks/robotwin_2_0_place_cans_plasticbox-standard/README.md b/tasks/robotwin_2_0_place_cans_plasticbox-standard/README.md
index 04c31cb820..f61343c3f4 100644
--- a/tasks/robotwin_2_0_place_cans_plasticbox-standard/README.md
+++ b/tasks/robotwin_2_0_place_cans_plasticbox-standard/README.md
@@ -53,3 +53,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
- 2026-10-10 (owner decision, board 0072): the agent's network egress is restricted to the model APIs (`[agent] network_mode = "allowlist"`, the hosts of `scripts/set_agent_egress.py`; since 2026-10-11 these include the package registries, PyPI and the Debian / Ubuntu archives, so that `pip install` and `apt-get install` work); it was open before (Harbor's default), and agents could download upstream sources, reference solutions and robot models from their own shell. The scene, the frozen instance, the tooling and the pass rule are unchanged, so the task keeps its name; results from before this date were obtained with open egress.
- 2026-10-13 (board 0086): the service's copy of the tooling only (`environment/sim/rcb_traj.py`, one line of `environment/sim/server.py`). The tooling now refuses to save or replay a trajectory of more than 200,000 rows, the privileged verifier's cap; a standard episode has no row cap, so the service lifts it for its own session (`S.max_rows = None`). Nothing an agent sees and nothing the service grades changes.
- 2026-10-13 (board 0100): base image `eai-robotwin` 0.1.2, whose scrub of RoboTwin's expert also removes the reference poses that `load_actors` computed only for the expert. The standard-mode agent never reaches the env, so nothing it observes changes; the scene, the instance and the judge are unchanged.
+- 2026-10-14 (board 0113): the tooling only (`environment/sim/rcb_traj.py`): while it builds the scene, `glob.glob` returns its matches sorted, so that a task which lists files to draw an object depends on its seed alone, not on the host's directory order (four RoboTwin tasks did; this one lists no files while it builds its scene). Its scene, frozen instance (same initial-state hash), verifier and pass rule are unchanged.
diff --git a/tasks/robotwin_2_0_place_cans_plasticbox-standard/environment/sim/rcb_traj.py b/tasks/robotwin_2_0_place_cans_plasticbox-standard/environment/sim/rcb_traj.py
index 761dec4aa1..868dfd57d3 100644
--- a/tasks/robotwin_2_0_place_cans_plasticbox-standard/environment/sim/rcb_traj.py
+++ b/tasks/robotwin_2_0_place_cans_plasticbox-standard/environment/sim/rcb_traj.py
@@ -113,6 +113,26 @@ BASE_ARGS: dict[str, Any] = {
}
+class _SortedGlob:
+ """While a scene is built, glob.glob / glob.iglob return their matches sorted. Four RoboTwin tasks (place_a2b_left,
+ place_a2b_right, place_object_scale, place_object_stand: envs/<task>.py load_actors) list an object's model files
+ with glob.glob and then draw one with the seeded np.random.choice; glob returns the file system's directory order,
+ which differs between hosts (and image layers), so one seed built a different object on another host and the frozen
+ instance's hash check failed there (board 0086 claim 5; never rebuilt off the reference workstation). Sorted, the
+ draw depends on the seed alone. No other task of ours lists files while it builds its scene."""
+
+ def __enter__(self):
+ import glob
+ self._glob, self._orig = glob, (glob.glob, glob.iglob)
+ glob.glob = lambda *a, **k: sorted(self._orig[0](*a, **k))
+ glob.iglob = lambda *a, **k: iter(sorted(self._orig[1](*a, **k)))
+ return self
+
+ def __exit__(self, *exc):
+ self._glob.glob, self._glob.iglob = self._orig
+ return False
+
+
def _chdir_root() -> None:
# RoboTwin resolves assets/config paths relative to the repo root and via ROOT_PATH
if os.getcwd() != ROBOTWIN_ROOT and os.path.isdir(ROBOTWIN_ROOT):
@@ -218,7 +238,8 @@ class Session:
if self.enable_render:
sapien.render.set_camera_shader_dir(self.render_shader)
args = _build_args(self.task, self.embodiment, self.seed)
- self.env.setup_demo(**args)
+ with _SortedGlob(): # the object draw depends on the seed alone, not the host's directory order
+ self.env.setup_demo(**args)
if self.enable_render: # setup_scene() switches the shader to ray tracing; put it back for cheap frames
sapien.render.set_camera_shader_dir(self.render_shader)
self._raw_scene = self.env.scene
diff --git a/tasks/robotwin_2_0_place_cans_plasticbox-standard/task.toml b/tasks/robotwin_2_0_place_cans_plasticbox-standard/task.toml
index 9b43a6ec5c..43a1977676 100644
--- a/tasks/robotwin_2_0_place_cans_plasticbox-standard/task.toml
+++ b/tasks/robotwin_2_0_place_cans_plasticbox-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "robotwin"
robot = "aloha-agilex"
family = "place_cans_plasticbox"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 60417
diff --git a/tasks/robotwin_2_0_place_container_plate-privileged/README.md b/tasks/robotwin_2_0_place_container_plate-privileged/README.md
index bdca297226..efc5063907 100644
--- a/tasks/robotwin_2_0_place_container_plate-privileged/README.md
+++ b/tasks/robotwin_2_0_place_container_plate-privileged/README.md
@@ -55,3 +55,4 @@ Place the container onto the plate.
- 2026-10-10 (owner decision, board 0072): the agent's network egress is restricted to the model APIs (`[agent] network_mode = "allowlist"`, the hosts of `scripts/set_agent_egress.py`; since 2026-10-11 these include the package registries, PyPI and the Debian / Ubuntu archives, so that `pip install` and `apt-get install` work); it was open before (Harbor's default), and agents could download upstream sources, reference solutions and robot models from their own shell. The scene, the frozen instance, the tooling and the pass rule are unchanged, so the task keeps its name; results from before this date were obtained with open egress.
- 2026-10-13 (board 0086): the tooling refuses a trajectory the verifier cannot grade (`environment/rcb_traj.py`, `tests/rcb_traj.py`). `Session.save_trajectory` and `Session.replay` raise `ValueError` for more than 200,000 rows (`rcb_traj.MAX_T`, the verifier's `MAX_T` in `tests/grade.py`: 800 s of physics), and `record_stop` warns on stderr. Before, the tool saved and replayed such a file ("success") and the verifier then graded it `bad_trajectory`; the instruction has stated the cap since 2026-10-10. The verifier, the scene, the instance and the pass rule are unchanged: every trajectory is graded as before.
- 2026-10-13 (board 0100): base image `eai-robotwin` 0.1.2. Its scrub of RoboTwin's expert also removes the reference poses that 18 tasks' `load_actors` computed only for the expert, and the expert's comments above the scrubbed method bodies (`images/robotwin/scrub_oracles.py`). This task's `load_actors` set no such value: its scene, instance, tooling, verifier and pass rule are unchanged.
+- 2026-10-14 (board 0113): the tooling only (`environment/rcb_traj.py`, `tests/rcb_traj.py`): while it builds the scene, `glob.glob` returns its matches sorted, so that a task which lists files to draw an object depends on its seed alone, not on the host's directory order (four RoboTwin tasks did; this one lists no files while it builds its scene). Its scene, frozen instance (same initial-state hash), verifier and pass rule are unchanged.
diff --git a/tasks/robotwin_2_0_place_container_plate-privileged/environment/rcb_traj.py b/tasks/robotwin_2_0_place_container_plate-privileged/environment/rcb_traj.py
index 48347e3064..3c7d0246ad 100644
--- a/tasks/robotwin_2_0_place_container_plate-privileged/environment/rcb_traj.py
+++ b/tasks/robotwin_2_0_place_container_plate-privileged/environment/rcb_traj.py
@@ -113,6 +113,26 @@ BASE_ARGS: dict[str, Any] = {
}
+class _SortedGlob:
+ """While a scene is built, glob.glob / glob.iglob return their matches sorted. Four RoboTwin tasks (place_a2b_left,
+ place_a2b_right, place_object_scale, place_object_stand: envs/<task>.py load_actors) list an object's model files
+ with glob.glob and then draw one with the seeded np.random.choice; glob returns the file system's directory order,
+ which differs between hosts (and image layers), so one seed built a different object on another host and the frozen
+ instance's hash check failed there (board 0086 claim 5; never rebuilt off the reference workstation). Sorted, the
+ draw depends on the seed alone. No other task of ours lists files while it builds its scene."""
+
+ def __enter__(self):
+ import glob
+ self._glob, self._orig = glob, (glob.glob, glob.iglob)
+ glob.glob = lambda *a, **k: sorted(self._orig[0](*a, **k))
+ glob.iglob = lambda *a, **k: iter(sorted(self._orig[1](*a, **k)))
+ return self
+
+ def __exit__(self, *exc):
+ self._glob.glob, self._glob.iglob = self._orig
+ return False
+
+
def _chdir_root() -> None:
# RoboTwin resolves assets/config paths relative to the repo root and via ROOT_PATH
if os.getcwd() != ROBOTWIN_ROOT and os.path.isdir(ROBOTWIN_ROOT):
@@ -218,7 +238,8 @@ class Session:
if self.enable_render:
sapien.render.set_camera_shader_dir(self.render_shader)
args = _build_args(self.task, self.embodiment, self.seed)
- self.env.setup_demo(**args)
+ with _SortedGlob(): # the object draw depends on the seed alone, not the host's directory order
+ self.env.setup_demo(**args)
if self.enable_render: # setup_scene() switches the shader to ray tracing; put it back for cheap frames
sapien.render.set_camera_shader_dir(self.render_shader)
self._raw_scene = self.env.scene
diff --git a/tasks/robotwin_2_0_place_container_plate-privileged/task.toml b/tasks/robotwin_2_0_place_container_plate-privileged/task.toml
index 8e7a4661b2..55e0133f36 100644
--- a/tasks/robotwin_2_0_place_container_plate-privileged/task.toml
+++ b/tasks/robotwin_2_0_place_container_plate-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "robotwin" # RoboTwin 2.0 = SAPIEN 3 / PhysX + CuRobo
robot = "aloha-agilex"
family = "place_container_plate"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 60417 # scene = environment/instance.json (RoboTwin seed + expected initial-state hash); identity = dataset.toml digest
expert_time_estimate_hours = 0.5
diff --git a/tasks/robotwin_2_0_place_container_plate-privileged/tests/rcb_traj.py b/tasks/robotwin_2_0_place_container_plate-privileged/tests/rcb_traj.py
index 48347e3064..3c7d0246ad 100644
--- a/tasks/robotwin_2_0_place_container_plate-privileged/tests/rcb_traj.py
+++ b/tasks/robotwin_2_0_place_container_plate-privileged/tests/rcb_traj.py
@@ -113,6 +113,26 @@ BASE_ARGS: dict[str, Any] = {
}
+class _SortedGlob:
+ """While a scene is built, glob.glob / glob.iglob return their matches sorted. Four RoboTwin tasks (place_a2b_left,
+ place_a2b_right, place_object_scale, place_object_stand: envs/<task>.py load_actors) list an object's model files
+ with glob.glob and then draw one with the seeded np.random.choice; glob returns the file system's directory order,
+ which differs between hosts (and image layers), so one seed built a different object on another host and the frozen
+ instance's hash check failed there (board 0086 claim 5; never rebuilt off the reference workstation). Sorted, the
+ draw depends on the seed alone. No other task of ours lists files while it builds its scene."""
+
+ def __enter__(self):
+ import glob
+ self._glob, self._orig = glob, (glob.glob, glob.iglob)
+ glob.glob = lambda *a, **k: sorted(self._orig[0](*a, **k))
+ glob.iglob = lambda *a, **k: iter(sorted(self._orig[1](*a, **k)))
+ return self
+
+ def __exit__(self, *exc):
+ self._glob.glob, self._glob.iglob = self._orig
+ return False
+
+
def _chdir_root() -> None:
# RoboTwin resolves assets/config paths relative to the repo root and via ROOT_PATH
if os.getcwd() != ROBOTWIN_ROOT and os.path.isdir(ROBOTWIN_ROOT):
@@ -218,7 +238,8 @@ class Session:
if self.enable_render:
sapien.render.set_camera_shader_dir(self.render_shader)
args = _build_args(self.task, self.embodiment, self.seed)
- self.env.setup_demo(**args)
+ with _SortedGlob(): # the object draw depends on the seed alone, not the host's directory order
+ self.env.setup_demo(**args)
if self.enable_render: # setup_scene() switches the shader to ray tracing; put it back for cheap frames
sapien.render.set_camera_shader_dir(self.render_shader)
self._raw_scene = self.env.scene
diff --git a/tasks/robotwin_2_0_place_container_plate-standard/README.md b/tasks/robotwin_2_0_place_container_plate-standard/README.md
index 954d016281..e383b9cbd2 100644
--- a/tasks/robotwin_2_0_place_container_plate-standard/README.md
+++ b/tasks/robotwin_2_0_place_container_plate-standard/README.md
@@ -53,3 +53,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
- 2026-10-10 (owner decision, board 0072): the agent's network egress is restricted to the model APIs (`[agent] network_mode = "allowlist"`, the hosts of `scripts/set_agent_egress.py`; since 2026-10-11 these include the package registries, PyPI and the Debian / Ubuntu archives, so that `pip install` and `apt-get install` work); it was open before (Harbor's default), and agents could download upstream sources, reference solutions and robot models from their own shell. The scene, the frozen instance, the tooling and the pass rule are unchanged, so the task keeps its name; results from before this date were obtained with open egress.
- 2026-10-13 (board 0086): the service's copy of the tooling only (`environment/sim/rcb_traj.py`, one line of `environment/sim/server.py`). The tooling now refuses to save or replay a trajectory of more than 200,000 rows, the privileged verifier's cap; a standard episode has no row cap, so the service lifts it for its own session (`S.max_rows = None`). Nothing an agent sees and nothing the service grades changes.
- 2026-10-13 (board 0100): base image `eai-robotwin` 0.1.2, whose scrub of RoboTwin's expert also removes the reference poses that `load_actors` computed only for the expert. The standard-mode agent never reaches the env, so nothing it observes changes; the scene, the instance and the judge are unchanged.
+- 2026-10-14 (board 0113): the tooling only (`environment/sim/rcb_traj.py`): while it builds the scene, `glob.glob` returns its matches sorted, so that a task which lists files to draw an object depends on its seed alone, not on the host's directory order (four RoboTwin tasks did; this one lists no files while it builds its scene). Its scene, frozen instance (same initial-state hash), verifier and pass rule are unchanged.
diff --git a/tasks/robotwin_2_0_place_container_plate-standard/environment/sim/rcb_traj.py b/tasks/robotwin_2_0_place_container_plate-standard/environment/sim/rcb_traj.py
index 48347e3064..3c7d0246ad 100644
--- a/tasks/robotwin_2_0_place_container_plate-standard/environment/sim/rcb_traj.py
+++ b/tasks/robotwin_2_0_place_container_plate-standard/environment/sim/rcb_traj.py
@@ -113,6 +113,26 @@ BASE_ARGS: dict[str, Any] = {
}
+class _SortedGlob:
+ """While a scene is built, glob.glob / glob.iglob return their matches sorted. Four RoboTwin tasks (place_a2b_left,
+ place_a2b_right, place_object_scale, place_object_stand: envs/<task>.py load_actors) list an object's model files
+ with glob.glob and then draw one with the seeded np.random.choice; glob returns the file system's directory order,
+ which differs between hosts (and image layers), so one seed built a different object on another host and the frozen
+ instance's hash check failed there (board 0086 claim 5; never rebuilt off the reference workstation). Sorted, the
+ draw depends on the seed alone. No other task of ours lists files while it builds its scene."""
+
+ def __enter__(self):
+ import glob
+ self._glob, self._orig = glob, (glob.glob, glob.iglob)
+ glob.glob = lambda *a, **k: sorted(self._orig[0](*a, **k))
+ glob.iglob = lambda *a, **k: iter(sorted(self._orig[1](*a, **k)))
+ return self
+
+ def __exit__(self, *exc):
+ self._glob.glob, self._glob.iglob = self._orig
+ return False
+
+
def _chdir_root() -> None:
# RoboTwin resolves assets/config paths relative to the repo root and via ROOT_PATH
if os.getcwd() != ROBOTWIN_ROOT and os.path.isdir(ROBOTWIN_ROOT):
@@ -218,7 +238,8 @@ class Session:
if self.enable_render:
sapien.render.set_camera_shader_dir(self.render_shader)
args = _build_args(self.task, self.embodiment, self.seed)
- self.env.setup_demo(**args)
+ with _SortedGlob(): # the object draw depends on the seed alone, not the host's directory order
+ self.env.setup_demo(**args)
if self.enable_render: # setup_scene() switches the shader to ray tracing; put it back for cheap frames
sapien.render.set_camera_shader_dir(self.render_shader)
self._raw_scene = self.env.scene
diff --git a/tasks/robotwin_2_0_place_container_plate-standard/task.toml b/tasks/robotwin_2_0_place_container_plate-standard/task.toml
index c23ccbd090..e56dc722d1 100644
--- a/tasks/robotwin_2_0_place_container_plate-standard/task.toml
+++ b/tasks/robotwin_2_0_place_container_plate-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "robotwin"
robot = "aloha-agilex"
family = "place_container_plate"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 60417
diff --git a/tasks/robotwin_2_0_place_dual_shoes-privileged/README.md b/tasks/robotwin_2_0_place_dual_shoes-privileged/README.md
index d8f215d0ca..56ddc149f6 100644
--- a/tasks/robotwin_2_0_place_dual_shoes-privileged/README.md
+++ b/tasks/robotwin_2_0_place_dual_shoes-privileged/README.md
@@ -56,3 +56,4 @@ Use both arms to pick up the two shoes on the table and put them in the shoebox,
- 2026-10-13 (board 0086): the tooling refuses a trajectory the verifier cannot grade (`environment/rcb_traj.py`, `tests/rcb_traj.py`). `Session.save_trajectory` and `Session.replay` raise `ValueError` for more than 200,000 rows (`rcb_traj.MAX_T`, the verifier's `MAX_T` in `tests/grade.py`: 800 s of physics), and `record_stop` warns on stderr. Before, the tool saved and replayed such a file ("success") and the verifier then graded it `bad_trajectory`; the instruction has stated the cap since 2026-10-10. The verifier, the scene, the instance and the pass rule are unchanged: every trajectory is graded as before.
; the reference solution never used it, and `check_success` does not read it. The scene, the frozen instance (same initial-state hash), the tooling, the verifier and the pass rule are unchanged, so the task keeps its name; privileged results from before this date were obtained with this value reachable.
- 2026-10-13 (board 0100, owner decision): base image `eai-robotwin` 0.1.2. The image's scrub of RoboTwin's expert kept `load_actors` whole, so a value it computes for the expert stayed reachable: `env.right_shoe_middle_pose` (a pose for the expert to move to or place at; unused upstream), and the instruction listed it among the scene objects. It is gone: the env no longer has it, `/opt/RoboTwin/envs/place_dual_shoes.py` no longer shows it, and the scene-object list no longer names it; the reference solution never used it, and `check_success` does not read it. The scene, the frozen instance (same initial-state hash), the tooling, the verifier and the pass rule are unchanged, so the task keeps its name; privileged results from before this date were obtained with this value reachable.
+- 2026-10-14 (board 0113): the tooling only (`environment/rcb_traj.py`, `tests/rcb_traj.py`): while it builds the scene, `glob.glob` returns its matches sorted, so that a task which lists files to draw an object depends on its seed alone, not on the host's directory order (four RoboTwin tasks did; this one lists no files while it builds its scene). Its scene, frozen instance (same initial-state hash), verifier and pass rule are unchanged.
diff --git a/tasks/robotwin_2_0_place_dual_shoes-privileged/environment/rcb_traj.py b/tasks/robotwin_2_0_place_dual_shoes-privileged/environment/rcb_traj.py
index b10319e0ca..981f72e449 100644
--- a/tasks/robotwin_2_0_place_dual_shoes-privileged/environment/rcb_traj.py
+++ b/tasks/robotwin_2_0_place_dual_shoes-privileged/environment/rcb_traj.py
@@ -113,6 +113,26 @@ BASE_ARGS: dict[str, Any] = {
}
+class _SortedGlob:
+ """While a scene is built, glob.glob / glob.iglob return their matches sorted. Four RoboTwin tasks (place_a2b_left,
+ place_a2b_right, place_object_scale, place_object_stand: envs/<task>.py load_actors) list an object's model files
+ with glob.glob and then draw one with the seeded np.random.choice; glob returns the file system's directory order,
+ which differs between hosts (and image layers), so one seed built a different object on another host and the frozen
+ instance's hash check failed there (board 0086 claim 5; never rebuilt off the reference workstation). Sorted, the
+ draw depends on the seed alone. No other task of ours lists files while it builds its scene."""
+
+ def __enter__(self):
+ import glob
+ self._glob, self._orig = glob, (glob.glob, glob.iglob)
+ glob.glob = lambda *a, **k: sorted(self._orig[0](*a, **k))
+ glob.iglob = lambda *a, **k: iter(sorted(self._orig[1](*a, **k)))
+ return self
+
+ def __exit__(self, *exc):
+ self._glob.glob, self._glob.iglob = self._orig
+ return False
+
+
def _chdir_root() -> None:
# RoboTwin resolves assets/config paths relative to the repo root and via ROOT_PATH
if os.getcwd() != ROBOTWIN_ROOT and os.path.isdir(ROBOTWIN_ROOT):
@@ -218,7 +238,8 @@ class Session:
if self.enable_render:
sapien.render.set_camera_shader_dir(self.render_shader)
args = _build_args(self.task, self.embodiment, self.seed)
- self.env.setup_demo(**args)
+ with _SortedGlob(): # the object draw depends on the seed alone, not the host's directory order
+ self.env.setup_demo(**args)
if self.enable_render: # setup_scene() switches the shader to ray tracing; put it back for cheap frames
sapien.render.set_camera_shader_dir(self.render_shader)
self._raw_scene = self.env.scene
diff --git a/tasks/robotwin_2_0_place_dual_shoes-privileged/task.toml b/tasks/robotwin_2_0_place_dual_shoes-privileged/task.toml
index 70c1d171ff..5a8f12b20e 100644
--- a/tasks/robotwin_2_0_place_dual_shoes-privileged/task.toml
+++ b/tasks/robotwin_2_0_place_dual_shoes-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "robotwin" # RoboTwin 2.0 = SAPIEN 3 / PhysX + CuRobo
robot = "aloha-agilex"
family = "place_dual_shoes"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 60417 # scene = environment/instance.json (RoboTwin seed + expected initial-state hash); identity = dataset.toml digest
expert_time_estimate_hours = 0.5
diff --git a/tasks/robotwin_2_0_place_dual_shoes-privileged/tests/rcb_traj.py b/tasks/robotwin_2_0_place_dual_shoes-privileged/tests/rcb_traj.py
index b10319e0ca..981f72e449 100644
--- a/tasks/robotwin_2_0_place_dual_shoes-privileged/tests/rcb_traj.py
+++ b/tasks/robotwin_2_0_place_dual_shoes-privileged/tests/rcb_traj.py
@@ -113,6 +113,26 @@ BASE_ARGS: dict[str, Any] = {
}
+class _SortedGlob:
+ """While a scene is built, glob.glob / glob.iglob return their matches sorted. Four RoboTwin tasks (place_a2b_left,
+ place_a2b_right, place_object_scale, place_object_stand: envs/<task>.py load_actors) list an object's model files
+ with glob.glob and then draw one with the seeded np.random.choice; glob returns the file system's directory order,
+ which differs between hosts (and image layers), so one seed built a different object on another host and the frozen
+ instance's hash check failed there (board 0086 claim 5; never rebuilt off the reference workstation). Sorted, the
+ draw depends on the seed alone. No other task of ours lists files while it builds its scene."""
+
+ def __enter__(self):
+ import glob
+ self._glob, self._orig = glob, (glob.glob, glob.iglob)
+ glob.glob = lambda *a, **k: sorted(self._orig[0](*a, **k))
+ glob.iglob = lambda *a, **k: iter(sorted(self._orig[1](*a, **k)))
+ return self
+
+ def __exit__(self, *exc):
+ self._glob.glob, self._glob.iglob = self._orig
+ return False
+
+
def _chdir_root() -> None:
# RoboTwin resolves assets/config paths relative to the repo root and via ROOT_PATH
if os.getcwd() != ROBOTWIN_ROOT and os.path.isdir(ROBOTWIN_ROOT):
@@ -218,7 +238,8 @@ class Session:
if self.enable_render:
sapien.render.set_camera_shader_dir(self.render_shader)
args = _build_args(self.task, self.embodiment, self.seed)
- self.env.setup_demo(**args)
+ with _SortedGlob(): # the object draw depends on the seed alone, not the host's directory order
+ self.env.setup_demo(**args)
if self.enable_render: # setup_scene() switches the shader to ray tracing; put it back for cheap frames
sapien.render.set_camera_shader_dir(self.render_shader)
self._raw_scene = self.env.scene
diff --git a/tasks/robotwin_2_0_place_dual_shoes-standard/README.md b/tasks/robotwin_2_0_place_dual_shoes-standard/README.md
index 7b529f2387..5aef5ad552 100644
--- a/tasks/robotwin_2_0_place_dual_shoes-standard/README.md
+++ b/tasks/robotwin_2_0_place_dual_shoes-standard/README.md
@@ -53,3 +53,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
- 2026-10-10 (owner decision, board 0072): the agent's network egress is restricted to the model APIs (`[agent] network_mode = "allowlist"`, the hosts of `scripts/set_agent_egress.py`; since 2026-10-11 these include the package registries, PyPI and the Debian / Ubuntu archives, so that `pip install` and `apt-get install` work); it was open before (Harbor's default), and agents could download upstream sources, reference solutions and robot models from their own shell. The scene, the frozen instance, the tooling and the pass rule are unchanged, so the task keeps its name; results from before this date were obtained with open egress.
- 2026-10-13 (board 0086): the service's copy of the tooling only (`environment/sim/rcb_traj.py`, one line of `environment/sim/server.py`). The tooling now refuses to save or replay a trajectory of more than 200,000 rows, the privileged verifier's cap; a standard episode has no row cap, so the service lifts it for its own session (`S.max_rows = None`). Nothing an agent sees and nothing the service grades changes.
- 2026-10-13 (board 0100): base image `eai-robotwin` 0.1.2, whose scrub of RoboTwin's expert also removes the reference poses that `load_actors` computed only for the expert (here `right_shoe_middle_pose`). The standard-mode agent never reaches the env, so nothing it observes changes; the scene, the instance and the judge are unchanged.
+- 2026-10-14 (board 0113): the tooling only (`environment/sim/rcb_traj.py`): while it builds the scene, `glob.glob` returns its matches sorted, so that a task which lists files to draw an object depends on its seed alone, not on the host's directory order (four RoboTwin tasks did; this one lists no files while it builds its scene). Its scene, frozen instance (same initial-state hash), verifier and pass rule are unchanged.
diff --git a/tasks/robotwin_2_0_place_dual_shoes-standard/environment/sim/rcb_traj.py b/tasks/robotwin_2_0_place_dual_shoes-standard/environment/sim/rcb_traj.py
index b10319e0ca..981f72e449 100644
--- a/tasks/robotwin_2_0_place_dual_shoes-standard/environment/sim/rcb_traj.py
+++ b/tasks/robotwin_2_0_place_dual_shoes-standard/environment/sim/rcb_traj.py
@@ -113,6 +113,26 @@ BASE_ARGS: dict[str, Any] = {
}
+class _SortedGlob:
+ """While a scene is built, glob.glob / glob.iglob return their matches sorted. Four RoboTwin tasks (place_a2b_left,
+ place_a2b_right, place_object_scale, place_object_stand: envs/<task>.py load_actors) list an object's model files
+ with glob.glob and then draw one with the seeded np.random.choice; glob returns the file system's directory order,
+ which differs between hosts (and image layers), so one seed built a different object on another host and the frozen
+ instance's hash check failed there (board 0086 claim 5; never rebuilt off the reference workstation). Sorted, the
+ draw depends on the seed alone. No other task of ours lists files while it builds its scene."""
+
+ def __enter__(self):
+ import glob
+ self._glob, self._orig = glob, (glob.glob, glob.iglob)
+ glob.glob = lambda *a, **k: sorted(self._orig[0](*a, **k))
+ glob.iglob = lambda *a, **k: iter(sorted(self._orig[1](*a, **k)))
+ return self
+
+ def __exit__(self, *exc):
+ self._glob.glob, self._glob.iglob = self._orig
+ return False
+
+
def _chdir_root() -> None:
# RoboTwin resolves assets/config paths relative to the repo root and via ROOT_PATH
if os.getcwd() != ROBOTWIN_ROOT and os.path.isdir(ROBOTWIN_ROOT):
@@ -218,7 +238,8 @@ class Session:
if self.enable_render:
sapien.render.set_camera_shader_dir(self.render_shader)
args = _build_args(self.task, self.embodiment, self.seed)
- self.env.setup_demo(**args)
+ with _SortedGlob(): # the object draw depends on the seed alone, not the host's directory order
+ self.env.setup_demo(**args)
if self.enable_render: # setup_scene() switches the shader to ray tracing; put it back for cheap frames
sapien.render.set_camera_shader_dir(self.render_shader)
self._raw_scene = self.env.scene
diff --git a/tasks/robotwin_2_0_place_dual_shoes-standard/task.toml b/tasks/robotwin_2_0_place_dual_shoes-standard/task.toml
index a2dc58b49b..f84209dba7 100644
--- a/tasks/robotwin_2_0_place_dual_shoes-standard/task.toml
+++ b/tasks/robotwin_2_0_place_dual_shoes-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "robotwin"
robot = "aloha-agilex"
family = "place_dual_shoes"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 60417
diff --git a/tasks/robotwin_2_0_place_empty_cup-privileged/README.md b/tasks/robotwin_2_0_place_empty_cup-privileged/README.md
index 1c0d33191d..ecb438fc38 100644
--- a/tasks/robotwin_2_0_place_empty_cup-privileged/README.md
+++ b/tasks/robotwin_2_0_place_empty_cup-privileged/README.md
@@ -55,3 +55,4 @@ Use an arm to place the empty cup on the coaster.
- 2026-10-10 (owner decision, board 0072): the agent's network egress is restricted to the model APIs (`[agent] network_mode = "allowlist"`, the hosts of `scripts/set_agent_egress.py`; since 2026-10-11 these include the package registries, PyPI and the Debian / Ubuntu archives, so that `pip install` and `apt-get install` work); it was open before (Harbor's default), and agents could download upstream sources, reference solutions and robot models from their own shell. The scene, the frozen instance, the tooling and the pass rule are unchanged, so the task keeps its name; results from before this date were obtained with open egress.
- 2026-10-13 (board 0086): the tooling refuses a trajectory the verifier cannot grade (`environment/rcb_traj.py`, `tests/rcb_traj.py`). `Session.save_trajectory` and `Session.replay` raise `ValueError` for more than 200,000 rows (`rcb_traj.MAX_T`, the verifier's `MAX_T` in `tests/grade.py`: 800 s of physics), and `record_stop` warns on stderr. Before, the tool saved and replayed such a file ("success") and the verifier then graded it `bad_trajectory`; the instruction has stated the cap since 2026-10-10. The verifier, the scene, the instance and the pass rule are unchanged: every trajectory is graded as before.
- 2026-10-13 (board 0100): base image `eai-robotwin` 0.1.2. Its scrub of RoboTwin's expert also removes the reference poses that 18 tasks' `load_actors` computed only for the expert, and the expert's comments above the scrubbed method bodies (`images/robotwin/scrub_oracles.py`). This task's `load_actors` set no such value: its scene, instance, tooling, verifier and pass rule are unchanged.
+- 2026-10-14 (board 0113): the tooling only (`environment/rcb_traj.py`, `tests/rcb_traj.py`): while it builds the scene, `glob.glob` returns its matches sorted, so that a task which lists files to draw an object depends on its seed alone, not on the host's directory order (four RoboTwin tasks did; this one lists no files while it builds its scene). Its scene, frozen instance (same initial-state hash), verifier and pass rule are unchanged.
diff --git a/tasks/robotwin_2_0_place_empty_cup-privileged/environment/rcb_traj.py b/tasks/robotwin_2_0_place_empty_cup-privileged/environment/rcb_traj.py
index 8a78b8d6e0..f300f066b5 100644
--- a/tasks/robotwin_2_0_place_empty_cup-privileged/environment/rcb_traj.py
+++ b/tasks/robotwin_2_0_place_empty_cup-privileged/environment/rcb_traj.py
@@ -113,6 +113,26 @@ BASE_ARGS: dict[str, Any] = {
}
+class _SortedGlob:
+ """While a scene is built, glob.glob / glob.iglob return their matches sorted. Four RoboTwin tasks (place_a2b_left,
+ place_a2b_right, place_object_scale, place_object_stand: envs/<task>.py load_actors) list an object's model files
+ with glob.glob and then draw one with the seeded np.random.choice; glob returns the file system's directory order,
+ which differs between hosts (and image layers), so one seed built a different object on another host and the frozen
+ instance's hash check failed there (board 0086 claim 5; never rebuilt off the reference workstation). Sorted, the
+ draw depends on the seed alone. No other task of ours lists files while it builds its scene."""
+
+ def __enter__(self):
+ import glob
+ self._glob, self._orig = glob, (glob.glob, glob.iglob)
+ glob.glob = lambda *a, **k: sorted(self._orig[0](*a, **k))
+ glob.iglob = lambda *a, **k: iter(sorted(self._orig[1](*a, **k)))
+ return self
+
+ def __exit__(self, *exc):
+ self._glob.glob, self._glob.iglob = self._orig
+ return False
+
+
def _chdir_root() -> None:
# RoboTwin resolves assets/config paths relative to the repo root and via ROOT_PATH
if os.getcwd() != ROBOTWIN_ROOT and os.path.isdir(ROBOTWIN_ROOT):
@@ -218,7 +238,8 @@ class Session:
if self.enable_render:
sapien.render.set_camera_shader_dir(self.render_shader)
args = _build_args(self.task, self.embodiment, self.seed)
- self.env.setup_demo(**args)
+ with _SortedGlob(): # the object draw depends on the seed alone, not the host's directory order
+ self.env.setup_demo(**args)
if self.enable_render: # setup_scene() switches the shader to ray tracing; put it back for cheap frames
sapien.render.set_camera_shader_dir(self.render_shader)
self._raw_scene = self.env.scene
diff --git a/tasks/robotwin_2_0_place_empty_cup-privileged/task.toml b/tasks/robotwin_2_0_place_empty_cup-privileged/task.toml
index a4a4ac5f8e..45964dddba 100644
--- a/tasks/robotwin_2_0_place_empty_cup-privileged/task.toml
+++ b/tasks/robotwin_2_0_place_empty_cup-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "robotwin" # RoboTwin 2.0 = SAPIEN 3 / PhysX + CuRobo
robot = "aloha-agilex"
family = "place_empty_cup"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 60417 # scene = environment/instance.json (RoboTwin seed + expected initial-state hash); identity = dataset.toml digest
expert_time_estimate_hours = 0.5
diff --git a/tasks/robotwin_2_0_place_empty_cup-privileged/tests/rcb_traj.py b/tasks/robotwin_2_0_place_empty_cup-privileged/tests/rcb_traj.py
index 8a78b8d6e0..f300f066b5 100644
--- a/tasks/robotwin_2_0_place_empty_cup-privileged/tests/rcb_traj.py
+++ b/tasks/robotwin_2_0_place_empty_cup-privileged/tests/rcb_traj.py
@@ -113,6 +113,26 @@ BASE_ARGS: dict[str, Any] = {
}
+class _SortedGlob:
+ """While a scene is built, glob.glob / glob.iglob return their matches sorted. Four RoboTwin tasks (place_a2b_left,
+ place_a2b_right, place_object_scale, place_object_stand: envs/<task>.py load_actors) list an object's model files
+ with glob.glob and then draw one with the seeded np.random.choice; glob returns the file system's directory order,
+ which differs between hosts (and image layers), so one seed built a different object on another host and the frozen
+ instance's hash check failed there (board 0086 claim 5; never rebuilt off the reference workstation). Sorted, the
+ draw depends on the seed alone. No other task of ours lists files while it builds its scene."""
+
+ def __enter__(self):
+ import glob
+ self._glob, self._orig = glob, (glob.glob, glob.iglob)
+ glob.glob = lambda *a, **k: sorted(self._orig[0](*a, **k))
+ glob.iglob = lambda *a, **k: iter(sorted(self._orig[1](*a, **k)))
+ return self
+
+ def __exit__(self, *exc):
+ self._glob.glob, self._glob.iglob = self._orig
+ return False
+
+
def _chdir_root() -> None:
# RoboTwin resolves assets/config paths relative to the repo root and via ROOT_PATH
if os.getcwd() != ROBOTWIN_ROOT and os.path.isdir(ROBOTWIN_ROOT):
@@ -218,7 +238,8 @@ class Session:
if self.enable_render:
sapien.render.set_camera_shader_dir(self.render_shader)
args = _build_args(self.task, self.embodiment, self.seed)
- self.env.setup_demo(**args)
+ with _SortedGlob(): # the object draw depends on the seed alone, not the host's directory order
+ self.env.setup_demo(**args)
if self.enable_render: # setup_scene() switches the shader to ray tracing; put it back for cheap frames
sapien.render.set_camera_shader_dir(self.render_shader)
self._raw_scene = self.env.scene
diff --git a/tasks/robotwin_2_0_place_empty_cup-standard/README.md b/tasks/robotwin_2_0_place_empty_cup-standard/README.md
index 8947a26830..51e8a964f7 100644
--- a/tasks/robotwin_2_0_place_empty_cup-standard/README.md
+++ b/tasks/robotwin_2_0_place_empty_cup-standard/README.md
@@ -53,3 +53,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
- 2026-10-10 (owner decision, board 0072): the agent's network egress is restricted to the model APIs (`[agent] network_mode = "allowlist"`, the hosts of `scripts/set_agent_egress.py`; since 2026-10-11 these include the package registries, PyPI and the Debian / Ubuntu archives, so that `pip install` and `apt-get install` work); it was open before (Harbor's default), and agents could download upstream sources, reference solutions and robot models from their own shell. The scene, the frozen instance, the tooling and the pass rule are unchanged, so the task keeps its name; results from before this date were obtained with open egress.
- 2026-10-13 (board 0086): the service's copy of the tooling only (`environment/sim/rcb_traj.py`, one line of `environment/sim/server.py`). The tooling now refuses to save or replay a trajectory of more than 200,000 rows, the privileged verifier's cap; a standard episode has no row cap, so the service lifts it for its own session (`S.max_rows = None`). Nothing an agent sees and nothing the service grades changes.
- 2026-10-13 (board 0100): base image `eai-robotwin` 0.1.2, whose scrub of RoboTwin's expert also removes the reference poses that `load_actors` computed only for the expert. The standard-mode agent never reaches the env, so nothing it observes changes; the scene, the instance and the judge are unchanged.
+- 2026-10-14 (board 0113): the tooling only (`environment/sim/rcb_traj.py`): while it builds the scene, `glob.glob` returns its matches sorted, so that a task which lists files to draw an object depends on its seed alone, not on the host's directory order (four RoboTwin tasks did; this one lists no files while it builds its scene). Its scene, frozen instance (same initial-state hash), verifier and pass rule are unchanged.
diff --git a/tasks/robotwin_2_0_place_empty_cup-standard/environment/sim/rcb_traj.py b/tasks/robotwin_2_0_place_empty_cup-standard/environment/sim/rcb_traj.py
index 8a78b8d6e0..f300f066b5 100644
--- a/tasks/robotwin_2_0_place_empty_cup-standard/environment/sim/rcb_traj.py
+++ b/tasks/robotwin_2_0_place_empty_cup-standard/environment/sim/rcb_traj.py
@@ -113,6 +113,26 @@ BASE_ARGS: dict[str, Any] = {
}
+class _SortedGlob:
+ """While a scene is built, glob.glob / glob.iglob return their matches sorted. Four RoboTwin tasks (place_a2b_left,
+ place_a2b_right, place_object_scale, place_object_stand: envs/<task>.py load_actors) list an object's model files
+ with glob.glob and then draw one with the seeded np.random.choice; glob returns the file system's directory order,
+ which differs between hosts (and image layers), so one seed built a different object on another host and the frozen
+ instance's hash check failed there (board 0086 claim 5; never rebuilt off the reference workstation). Sorted, the
+ draw depends on the seed alone. No other task of ours lists files while it builds its scene."""
+
+ def __enter__(self):
+ import glob
+ self._glob, self._orig = glob, (glob.glob, glob.iglob)
+ glob.glob = lambda *a, **k: sorted(self._orig[0](*a, **k))
+ glob.iglob = lambda *a, **k: iter(sorted(self._orig[1](*a, **k)))
+ return self
+
+ def __exit__(self, *exc):
+ self._glob.glob, self._glob.iglob = self._orig
+ return False
+
+
def _chdir_root() -> None:
# RoboTwin resolves assets/config paths relative to the repo root and via ROOT_PATH
if os.getcwd() != ROBOTWIN_ROOT and os.path.isdir(ROBOTWIN_ROOT):
@@ -218,7 +238,8 @@ class Session:
if self.enable_render:
sapien.render.set_camera_shader_dir(self.render_shader)
args = _build_args(self.task, self.embodiment, self.seed)
- self.env.setup_demo(**args)
+ with _SortedGlob(): # the object draw depends on the seed alone, not the host's directory order
+ self.env.setup_demo(**args)
if self.enable_render: # setup_scene() switches the shader to ray tracing; put it back for cheap frames
sapien.render.set_camera_shader_dir(self.render_shader)
self._raw_scene = self.env.scene
diff --git a/tasks/robotwin_2_0_place_empty_cup-standard/task.toml b/tasks/robotwin_2_0_place_empty_cup-standard/task.toml
index ce3c92731e..a22fafbd0f 100644
--- a/tasks/robotwin_2_0_place_empty_cup-standard/task.toml
+++ b/tasks/robotwin_2_0_place_empty_cup-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "robotwin"
robot = "aloha-agilex"
family = "place_empty_cup"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 60417
diff --git a/tasks/robotwin_2_0_place_fan-privileged/README.md b/tasks/robotwin_2_0_place_fan-privileged/README.md
index b81604f038..504da28321 100644
--- a/tasks/robotwin_2_0_place_fan-privileged/README.md
+++ b/tasks/robotwin_2_0_place_fan-privileged/README.md
@@ -55,3 +55,4 @@ Grab the fan and place it on a colored mat, and make sure the fan is facing the
- 2026-10-10 (owner decision, board 0072): the agent's network egress is restricted to the model APIs (`[agent] network_mode = "allowlist"`, the hosts of `scripts/set_agent_egress.py`; since 2026-10-11 these include the package registries, PyPI and the Debian / Ubuntu archives, so that `pip install` and `apt-get install` work); it was open before (Harbor's default), and agents could download upstream sources, reference solutions and robot models from their own shell. The scene, the frozen instance, the tooling and the pass rule are unchanged, so the task keeps its name; results from before this date were obtained with open egress.
- 2026-10-13 (board 0086): the tooling refuses a trajectory the verifier cannot grade (`environment/rcb_traj.py`, `tests/rcb_traj.py`). `Session.save_trajectory` and `Session.replay` raise `ValueError` for more than 200,000 rows (`rcb_traj.MAX_T`, the verifier's `MAX_T` in `tests/grade.py`: 800 s of physics), and `record_stop` warns on stderr. Before, the tool saved and replayed such a file ("success") and the verifier then graded it `bad_trajectory`; the instruction has stated the cap since 2026-10-10. The verifier, the scene, the instance and the pass rule are unchanged: every trajectory is graded as before.
- 2026-10-13 (board 0100): base image `eai-robotwin` 0.1.2. Its scrub of RoboTwin's expert also removes the reference poses that 18 tasks' `load_actors` computed only for the expert, and the expert's comments above the scrubbed method bodies (`images/robotwin/scrub_oracles.py`). This task's `load_actors` set no such value: its scene, instance, tooling, verifier and pass rule are unchanged.
+- 2026-10-14 (board 0113): the tooling only (`environment/rcb_traj.py`, `tests/rcb_traj.py`): while it builds the scene, `glob.glob` returns its matches sorted, so that a task which lists files to draw an object depends on its seed alone, not on the host's directory order (four RoboTwin tasks did; this one lists no files while it builds its scene). Its scene, frozen instance (same initial-state hash), verifier and pass rule are unchanged.
diff --git a/tasks/robotwin_2_0_place_fan-privileged/environment/rcb_traj.py b/tasks/robotwin_2_0_place_fan-privileged/environment/rcb_traj.py
index 735aca2e75..10581a4033 100644
--- a/tasks/robotwin_2_0_place_fan-privileged/environment/rcb_traj.py
+++ b/tasks/robotwin_2_0_place_fan-privileged/environment/rcb_traj.py
@@ -113,6 +113,26 @@ BASE_ARGS: dict[str, Any] = {
}
+class _SortedGlob:
+ """While a scene is built, glob.glob / glob.iglob return their matches sorted. Four RoboTwin tasks (place_a2b_left,
+ place_a2b_right, place_object_scale, place_object_stand: envs/<task>.py load_actors) list an object's model files
+ with glob.glob and then draw one with the seeded np.random.choice; glob returns the file system's directory order,
+ which differs between hosts (and image layers), so one seed built a different object on another host and the frozen
+ instance's hash check failed there (board 0086 claim 5; never rebuilt off the reference workstation). Sorted, the
+ draw depends on the seed alone. No other task of ours lists files while it builds its scene."""
+
+ def __enter__(self):
+ import glob
+ self._glob, self._orig = glob, (glob.glob, glob.iglob)
+ glob.glob = lambda *a, **k: sorted(self._orig[0](*a, **k))
+ glob.iglob = lambda *a, **k: iter(sorted(self._orig[1](*a, **k)))
+ return self
+
+ def __exit__(self, *exc):
+ self._glob.glob, self._glob.iglob = self._orig
+ return False
+
+
def _chdir_root() -> None:
# RoboTwin resolves assets/config paths relative to the repo root and via ROOT_PATH
if os.getcwd() != ROBOTWIN_ROOT and os.path.isdir(ROBOTWIN_ROOT):
@@ -218,7 +238,8 @@ class Session:
if self.enable_render:
sapien.render.set_camera_shader_dir(self.render_shader)
args = _build_args(self.task, self.embodiment, self.seed)
- self.env.setup_demo(**args)
+ with _SortedGlob(): # the object draw depends on the seed alone, not the host's directory order
+ self.env.setup_demo(**args)
if self.enable_render: # setup_scene() switches the shader to ray tracing; put it back for cheap frames
sapien.render.set_camera_shader_dir(self.render_shader)
self._raw_scene = self.env.scene
diff --git a/tasks/robotwin_2_0_place_fan-privileged/task.toml b/tasks/robotwin_2_0_place_fan-privileged/task.toml
index b72355742e..8d3f325933 100644
--- a/tasks/robotwin_2_0_place_fan-privileged/task.toml
+++ b/tasks/robotwin_2_0_place_fan-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "robotwin" # RoboTwin 2.0 = SAPIEN 3 / PhysX + CuRobo
robot = "aloha-agilex"
family = "place_fan"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 60417 # scene = environment/instance.json (RoboTwin seed + expected initial-state hash); identity = dataset.toml digest
expert_time_estimate_hours = 0.5
diff --git a/tasks/robotwin_2_0_place_fan-privileged/tests/rcb_traj.py b/tasks/robotwin_2_0_place_fan-privileged/tests/rcb_traj.py
index 735aca2e75..10581a4033 100644
--- a/tasks/robotwin_2_0_place_fan-privileged/tests/rcb_traj.py
+++ b/tasks/robotwin_2_0_place_fan-privileged/tests/rcb_traj.py
@@ -113,6 +113,26 @@ BASE_ARGS: dict[str, Any] = {
}
+class _SortedGlob:
+ """While a scene is built, glob.glob / glob.iglob return their matches sorted. Four RoboTwin tasks (place_a2b_left,
+ place_a2b_right, place_object_scale, place_object_stand: envs/<task>.py load_actors) list an object's model files
+ with glob.glob and then draw one with the seeded np.random.choice; glob returns the file system's directory order,
+ which differs between hosts (and image layers), so one seed built a different object on another host and the frozen
+ instance's hash check failed there (board 0086 claim 5; never rebuilt off the reference workstation). Sorted, the
+ draw depends on the seed alone. No other task of ours lists files while it builds its scene."""
+
+ def __enter__(self):
+ import glob
+ self._glob, self._orig = glob, (glob.glob, glob.iglob)
+ glob.glob = lambda *a, **k: sorted(self._orig[0](*a, **k))
+ glob.iglob = lambda *a, **k: iter(sorted(self._orig[1](*a, **k)))
+ return self
+
+ def __exit__(self, *exc):
+ self._glob.glob, self._glob.iglob = self._orig
+ return False
+
+
def _chdir_root() -> None:
# RoboTwin resolves assets/config paths relative to the repo root and via ROOT_PATH
if os.getcwd() != ROBOTWIN_ROOT and os.path.isdir(ROBOTWIN_ROOT):
@@ -218,7 +238,8 @@ class Session:
if self.enable_render:
sapien.render.set_camera_shader_dir(self.render_shader)
args = _build_args(self.task, self.embodiment, self.seed)
- self.env.setup_demo(**args)
+ with _SortedGlob(): # the object draw depends on the seed alone, not the host's directory order
+ self.env.setup_demo(**args)
if self.enable_render: # setup_scene() switches the shader to ray tracing; put it back for cheap frames
sapien.render.set_camera_shader_dir(self.render_shader)
self._raw_scene = self.env.scene
diff --git a/tasks/robotwin_2_0_place_fan-standard/README.md b/tasks/robotwin_2_0_place_fan-standard/README.md
index 5f933d550e..e0b8a6e310 100644
--- a/tasks/robotwin_2_0_place_fan-standard/README.md
+++ b/tasks/robotwin_2_0_place_fan-standard/README.md
@@ -53,3 +53,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
- 2026-10-10 (owner decision, board 0072): the agent's network egress is restricted to the model APIs (`[agent] network_mode = "allowlist"`, the hosts of `scripts/set_agent_egress.py`; since 2026-10-11 these include the package registries, PyPI and the Debian / Ubuntu archives, so that `pip install` and `apt-get install` work); it was open before (Harbor's default), and agents could download upstream sources, reference solutions and robot models from their own shell. The scene, the frozen instance, the tooling and the pass rule are unchanged, so the task keeps its name; results from before this date were obtained with open egress.
- 2026-10-13 (board 0086): the service's copy of the tooling only (`environment/sim/rcb_traj.py`, one line of `environment/sim/server.py`). The tooling now refuses to save or replay a trajectory of more than 200,000 rows, the privileged verifier's cap; a standard episode has no row cap, so the service lifts it for its own session (`S.max_rows = None`). Nothing an agent sees and nothing the service grades changes.
- 2026-10-13 (board 0100): base image `eai-robotwin` 0.1.2, whose scrub of RoboTwin's expert also removes the reference poses that `load_actors` computed only for the expert. The standard-mode agent never reaches the env, so nothing it observes changes; the scene, the instance and the judge are unchanged.
+- 2026-10-14 (board 0113): the tooling only (`environment/sim/rcb_traj.py`): while it builds the scene, `glob.glob` returns its matches sorted, so that a task which lists files to draw an object depends on its seed alone, not on the host's directory order (four RoboTwin tasks did; this one lists no files while it builds its scene). Its scene, frozen instance (same initial-state hash), verifier and pass rule are unchanged.
diff --git a/tasks/robotwin_2_0_place_fan-standard/environment/sim/rcb_traj.py b/tasks/robotwin_2_0_place_fan-standard/environment/sim/rcb_traj.py
index 735aca2e75..10581a4033 100644
--- a/tasks/robotwin_2_0_place_fan-standard/environment/sim/rcb_traj.py
+++ b/tasks/robotwin_2_0_place_fan-standard/environment/sim/rcb_traj.py
@@ -113,6 +113,26 @@ BASE_ARGS: dict[str, Any] = {
}
+class _SortedGlob:
+ """While a scene is built, glob.glob / glob.iglob return their matches sorted. Four RoboTwin tasks (place_a2b_left,
+ place_a2b_right, place_object_scale, place_object_stand: envs/<task>.py load_actors) list an object's model files
+ with glob.glob and then draw one with the seeded np.random.choice; glob returns the file system's directory order,
+ which differs between hosts (and image layers), so one seed built a different object on another host and the frozen
+ instance's hash check failed there (board 0086 claim 5; never rebuilt off the reference workstation). Sorted, the
+ draw depends on the seed alone. No other task of ours lists files while it builds its scene."""
+
+ def __enter__(self):
+ import glob
+ self._glob, self._orig = glob, (glob.glob, glob.iglob)
+ glob.glob = lambda *a, **k: sorted(self._orig[0](*a, **k))
+ glob.iglob = lambda *a, **k: iter(sorted(self._orig[1](*a, **k)))
+ return self
+
+ def __exit__(self, *exc):
+ self._glob.glob, self._glob.iglob = self._orig
+ return False
+
+
def _chdir_root() -> None:
# RoboTwin resolves assets/config paths relative to the repo root and via ROOT_PATH
if os.getcwd() != ROBOTWIN_ROOT and os.path.isdir(ROBOTWIN_ROOT):
@@ -218,7 +238,8 @@ class Session:
if self.enable_render:
sapien.render.set_camera_shader_dir(self.render_shader)
args = _build_args(self.task, self.embodiment, self.seed)
- self.env.setup_demo(**args)
+ with _SortedGlob(): # the object draw depends on the seed alone, not the host's directory order
+ self.env.setup_demo(**args)
if self.enable_render: # setup_scene() switches the shader to ray tracing; put it back for cheap frames
sapien.render.set_camera_shader_dir(self.render_shader)
self._raw_scene = self.env.scene
diff --git a/tasks/robotwin_2_0_place_fan-standard/task.toml b/tasks/robotwin_2_0_place_fan-standard/task.toml
index 63788c141c..9b5fad6342 100644
--- a/tasks/robotwin_2_0_place_fan-standard/task.toml
+++ b/tasks/robotwin_2_0_place_fan-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "robotwin"
robot = "aloha-agilex"
family = "place_fan"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 60417
diff --git a/tasks/robotwin_2_0_place_mouse_pad-privileged/README.md b/tasks/robotwin_2_0_place_mouse_pad-privileged/README.md
index 3060bffcee..ff6b3dc094 100644
--- a/tasks/robotwin_2_0_place_mouse_pad-privileged/README.md
+++ b/tasks/robotwin_2_0_place_mouse_pad-privileged/README.md
@@ -56,3 +56,4 @@ Grab the mouse and place it on a colored mat.
- 2026-10-13 (board 0086): the tooling refuses a trajectory the verifier cannot grade (`environment/rcb_traj.py`, `tests/rcb_traj.py`). `Session.save_trajectory` and `Session.replay` raise `ValueError` for more than 200,000 rows (`rcb_traj.MAX_T`, the verifier's `MAX_T` in `tests/grade.py`: 800 s of physics), and `record_stop` warns on stderr. Before, the tool saved and replayed such a file ("success") and the verifier then graded it `bad_trajectory`; the instruction has stated the cap since 2026-10-10. The verifier, the scene, the instance and the pass rule are unchanged: every trajectory is graded as before.
; `solution/` rebuilds the scene with upstream's `load_actors` to get them back for the expert. None of them is read by `check_success`. The scene, the frozen instance (same initial-state hash), the tooling, the verifier and the pass rule are unchanged, so the task keeps its name; privileged results from before this date were obtained with these values reachable.
- 2026-10-13 (board 0100, owner decision): base image `eai-robotwin` 0.1.2. The image's scrub of RoboTwin's expert kept `load_actors` whole, so values it computes only for the removed expert stayed reachable: `env.target_pose` (poses the expert moves to or places at), and the instruction listed them among the scene objects. They are gone: the env no longer has them, `/opt/RoboTwin/envs/place_mouse_pad.py` no longer shows them, and the scene-object list no longer names them; `solution/` rebuilds the scene with upstream's `load_actors` to get them back for the expert. None of them is read by `check_success`. The scene, the frozen instance (same initial-state hash), the tooling, the verifier and the pass rule are unchanged, so the task keeps its name; privileged results from before this date were obtained with these values reachable.
+- 2026-10-14 (board 0113): the tooling only (`environment/rcb_traj.py`, `tests/rcb_traj.py`): while it builds the scene, `glob.glob` returns its matches sorted, so that a task which lists files to draw an object depends on its seed alone, not on the host's directory order (four RoboTwin tasks did; this one lists no files while it builds its scene). Its scene, frozen instance (same initial-state hash), verifier and pass rule are unchanged.
diff --git a/tasks/robotwin_2_0_place_mouse_pad-privileged/environment/rcb_traj.py b/tasks/robotwin_2_0_place_mouse_pad-privileged/environment/rcb_traj.py
index ceb4bc55f6..3cf3c8eab1 100644
--- a/tasks/robotwin_2_0_place_mouse_pad-privileged/environment/rcb_traj.py
+++ b/tasks/robotwin_2_0_place_mouse_pad-privileged/environment/rcb_traj.py
@@ -113,6 +113,26 @@ BASE_ARGS: dict[str, Any] = {
}
+class _SortedGlob:
+ """While a scene is built, glob.glob / glob.iglob return their matches sorted. Four RoboTwin tasks (place_a2b_left,
+ place_a2b_right, place_object_scale, place_object_stand: envs/<task>.py load_actors) list an object's model files
+ with glob.glob and then draw one with the seeded np.random.choice; glob returns the file system's directory order,
+ which differs between hosts (and image layers), so one seed built a different object on another host and the frozen
+ instance's hash check failed there (board 0086 claim 5; never rebuilt off the reference workstation). Sorted, the
+ draw depends on the seed alone. No other task of ours lists files while it builds its scene."""
+
+ def __enter__(self):
+ import glob
+ self._glob, self._orig = glob, (glob.glob, glob.iglob)
+ glob.glob = lambda *a, **k: sorted(self._orig[0](*a, **k))
+ glob.iglob = lambda *a, **k: iter(sorted(self._orig[1](*a, **k)))
+ return self
+
+ def __exit__(self, *exc):
+ self._glob.glob, self._glob.iglob = self._orig
+ return False
+
+
def _chdir_root() -> None:
# RoboTwin resolves assets/config paths relative to the repo root and via ROOT_PATH
if os.getcwd() != ROBOTWIN_ROOT and os.path.isdir(ROBOTWIN_ROOT):
@@ -218,7 +238,8 @@ class Session:
if self.enable_render:
sapien.render.set_camera_shader_dir(self.render_shader)
args = _build_args(self.task, self.embodiment, self.seed)
- self.env.setup_demo(**args)
+ with _SortedGlob(): # the object draw depends on the seed alone, not the host's directory order
+ self.env.setup_demo(**args)
if self.enable_render: # setup_scene() switches the shader to ray tracing; put it back for cheap frames
sapien.render.set_camera_shader_dir(self.render_shader)
self._raw_scene = self.env.scene
diff --git a/tasks/robotwin_2_0_place_mouse_pad-privileged/task.toml b/tasks/robotwin_2_0_place_mouse_pad-privileged/task.toml
index 87e977ed78..473686ad5c 100644
--- a/tasks/robotwin_2_0_place_mouse_pad-privileged/task.toml
+++ b/tasks/robotwin_2_0_place_mouse_pad-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "robotwin" # RoboTwin 2.0 = SAPIEN 3 / PhysX + CuRobo
robot = "aloha-agilex"
family = "place_mouse_pad"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 60417 # scene = environment/instance.json (RoboTwin seed + expected initial-state hash); identity = dataset.toml digest
expert_time_estimate_hours = 0.5
diff --git a/tasks/robotwin_2_0_place_mouse_pad-privileged/tests/rcb_traj.py b/tasks/robotwin_2_0_place_mouse_pad-privileged/tests/rcb_traj.py
index ceb4bc55f6..3cf3c8eab1 100644
--- a/tasks/robotwin_2_0_place_mouse_pad-privileged/tests/rcb_traj.py
+++ b/tasks/robotwin_2_0_place_mouse_pad-privileged/tests/rcb_traj.py
@@ -113,6 +113,26 @@ BASE_ARGS: dict[str, Any] = {
}
+class _SortedGlob:
+ """While a scene is built, glob.glob / glob.iglob return their matches sorted. Four RoboTwin tasks (place_a2b_left,
+ place_a2b_right, place_object_scale, place_object_stand: envs/<task>.py load_actors) list an object's model files
+ with glob.glob and then draw one with the seeded np.random.choice; glob returns the file system's directory order,
+ which differs between hosts (and image layers), so one seed built a different object on another host and the frozen
+ instance's hash check failed there (board 0086 claim 5; never rebuilt off the reference workstation). Sorted, the
+ draw depends on the seed alone. No other task of ours lists files while it builds its scene."""
+
+ def __enter__(self):
+ import glob
+ self._glob, self._orig = glob, (glob.glob, glob.iglob)
+ glob.glob = lambda *a, **k: sorted(self._orig[0](*a, **k))
+ glob.iglob = lambda *a, **k: iter(sorted(self._orig[1](*a, **k)))
+ return self
+
+ def __exit__(self, *exc):
+ self._glob.glob, self._glob.iglob = self._orig
+ return False
+
+
def _chdir_root() -> None:
# RoboTwin resolves assets/config paths relative to the repo root and via ROOT_PATH
if os.getcwd() != ROBOTWIN_ROOT and os.path.isdir(ROBOTWIN_ROOT):
@@ -218,7 +238,8 @@ class Session:
if self.enable_render:
sapien.render.set_camera_shader_dir(self.render_shader)
args = _build_args(self.task, self.embodiment, self.seed)
- self.env.setup_demo(**args)
+ with _SortedGlob(): # the object draw depends on the seed alone, not the host's directory order
+ self.env.setup_demo(**args)
if self.enable_render: # setup_scene() switches the shader to ray tracing; put it back for cheap frames
sapien.render.set_camera_shader_dir(self.render_shader)
self._raw_scene = self.env.scene
diff --git a/tasks/robotwin_2_0_place_mouse_pad-standard/README.md b/tasks/robotwin_2_0_place_mouse_pad-standard/README.md
index 65f572393b..43ceac0560 100644
--- a/tasks/robotwin_2_0_place_mouse_pad-standard/README.md
+++ b/tasks/robotwin_2_0_place_mouse_pad-standard/README.md
@@ -53,3 +53,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
- 2026-10-10 (owner decision, board 0072): the agent's network egress is restricted to the model APIs (`[agent] network_mode = "allowlist"`, the hosts of `scripts/set_agent_egress.py`; since 2026-10-11 these include the package registries, PyPI and the Debian / Ubuntu archives, so that `pip install` and `apt-get install` work); it was open before (Harbor's default), and agents could download upstream sources, reference solutions and robot models from their own shell. The scene, the frozen instance, the tooling and the pass rule are unchanged, so the task keeps its name; results from before this date were obtained with open egress.
- 2026-10-13 (board 0086): the service's copy of the tooling only (`environment/sim/rcb_traj.py`, one line of `environment/sim/server.py`). The tooling now refuses to save or replay a trajectory of more than 200,000 rows, the privileged verifier's cap; a standard episode has no row cap, so the service lifts it for its own session (`S.max_rows = None`). Nothing an agent sees and nothing the service grades changes.
- 2026-10-13 (board 0100): base image `eai-robotwin` 0.1.2, whose scrub of RoboTwin's expert also removes the reference poses that `load_actors` computed only for the expert (here `target_pose`). The standard-mode agent never reaches the env, so nothing it observes changes; the scene, the instance and the judge are unchanged.
+- 2026-10-14 (board 0113): the tooling only (`environment/sim/rcb_traj.py`): while it builds the scene, `glob.glob` returns its matches sorted, so that a task which lists files to draw an object depends on its seed alone, not on the host's directory order (four RoboTwin tasks did; this one lists no files while it builds its scene). Its scene, frozen instance (same initial-state hash), verifier and pass rule are unchanged.
diff --git a/tasks/robotwin_2_0_place_mouse_pad-standard/environment/sim/rcb_traj.py b/tasks/robotwin_2_0_place_mouse_pad-standard/environment/sim/rcb_traj.py
index ceb4bc55f6..3cf3c8eab1 100644
--- a/tasks/robotwin_2_0_place_mouse_pad-standard/environment/sim/rcb_traj.py
+++ b/tasks/robotwin_2_0_place_mouse_pad-standard/environment/sim/rcb_traj.py
@@ -113,6 +113,26 @@ BASE_ARGS: dict[str, Any] = {
}
+class _SortedGlob:
+ """While a scene is built, glob.glob / glob.iglob return their matches sorted. Four RoboTwin tasks (place_a2b_left,
+ place_a2b_right, place_object_scale, place_object_stand: envs/<task>.py load_actors) list an object's model files
+ with glob.glob and then draw one with the seeded np.random.choice; glob returns the file system's directory order,
+ which differs between hosts (and image layers), so one seed built a different object on another host and the frozen
+ instance's hash check failed there (board 0086 claim 5; never rebuilt off the reference workstation). Sorted, the
+ draw depends on the seed alone. No other task of ours lists files while it builds its scene."""
+
+ def __enter__(self):
+ import glob
+ self._glob, self._orig = glob, (glob.glob, glob.iglob)
+ glob.glob = lambda *a, **k: sorted(self._orig[0](*a, **k))
+ glob.iglob = lambda *a, **k: iter(sorted(self._orig[1](*a, **k)))
+ return self
+
+ def __exit__(self, *exc):
+ self._glob.glob, self._glob.iglob = self._orig
+ return False
+
+
def _chdir_root() -> None:
# RoboTwin resolves assets/config paths relative to the repo root and via ROOT_PATH
if os.getcwd() != ROBOTWIN_ROOT and os.path.isdir(ROBOTWIN_ROOT):
@@ -218,7 +238,8 @@ class Session:
if self.enable_render:
sapien.render.set_camera_shader_dir(self.render_shader)
args = _build_args(self.task, self.embodiment, self.seed)
- self.env.setup_demo(**args)
+ with _SortedGlob(): # the object draw depends on the seed alone, not the host's directory order
+ self.env.setup_demo(**args)
if self.enable_render: # setup_scene() switches the shader to ray tracing; put it back for cheap frames
sapien.render.set_camera_shader_dir(self.render_shader)
self._raw_scene = self.env.scene
diff --git a/tasks/robotwin_2_0_place_mouse_pad-standard/task.toml b/tasks/robotwin_2_0_place_mouse_pad-standard/task.toml
index b5f2479099..90f3803cc5 100644
--- a/tasks/robotwin_2_0_place_mouse_pad-standard/task.toml
+++ b/tasks/robotwin_2_0_place_mouse_pad-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "robotwin"
robot = "aloha-agilex"
family = "place_mouse_pad"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 60417
diff --git a/tasks/robotwin_2_0_place_object_basket-privileged/README.md b/tasks/robotwin_2_0_place_object_basket-privileged/README.md
index 5696b7ee9f..76da292bc8 100644
--- a/tasks/robotwin_2_0_place_object_basket-privileged/README.md
+++ b/tasks/robotwin_2_0_place_object_basket-privileged/README.md
@@ -55,3 +55,4 @@ Use one arm to grab the target object and put it in the basket, then use the oth
- 2026-10-10 (owner decision, board 0072): the agent's network egress is restricted to the model APIs (`[agent] network_mode = "allowlist"`, the hosts of `scripts/set_agent_egress.py`; since 2026-10-11 these include the package registries, PyPI and the Debian / Ubuntu archives, so that `pip install` and `apt-get install` work); it was open before (Harbor's default), and agents could download upstream sources, reference solutions and robot models from their own shell. The scene, the frozen instance, the tooling and the pass rule are unchanged, so the task keeps its name; results from before this date were obtained with open egress.
- 2026-10-13 (board 0086): the tooling refuses a trajectory the verifier cannot grade (`environment/rcb_traj.py`, `tests/rcb_traj.py`). `Session.save_trajectory` and `Session.replay` raise `ValueError` for more than 200,000 rows (`rcb_traj.MAX_T`, the verifier's `MAX_T` in `tests/grade.py`: 800 s of physics), and `record_stop` warns on stderr. Before, the tool saved and replayed such a file ("success") and the verifier then graded it `bad_trajectory`; the instruction has stated the cap since 2026-10-10. The verifier, the scene, the instance and the pass rule are unchanged: every trajectory is graded as before.
- 2026-10-13 (board 0100): base image `eai-robotwin` 0.1.2. Its scrub of RoboTwin's expert also removes the reference poses that 18 tasks' `load_actors` computed only for the expert, and the expert's comments above the scrubbed method bodies (`images/robotwin/scrub_oracles.py`). This task's `load_actors` set no such value: its scene, instance, tooling, verifier and pass rule are unchanged.
+- 2026-10-14 (board 0113): the tooling only (`environment/rcb_traj.py`, `tests/rcb_traj.py`): while it builds the scene, `glob.glob` returns its matches sorted, so that a task which lists files to draw an object depends on its seed alone, not on the host's directory order (four RoboTwin tasks did; this one lists no files while it builds its scene). Its scene, frozen instance (same initial-state hash), verifier and pass rule are unchanged.
diff --git a/tasks/robotwin_2_0_place_object_basket-privileged/environment/rcb_traj.py b/tasks/robotwin_2_0_place_object_basket-privileged/environment/rcb_traj.py
index edf11216a4..c01ca13ae7 100644
--- a/tasks/robotwin_2_0_place_object_basket-privileged/environment/rcb_traj.py
+++ b/tasks/robotwin_2_0_place_object_basket-privileged/environment/rcb_traj.py
@@ -113,6 +113,26 @@ BASE_ARGS: dict[str, Any] = {
}
+class _SortedGlob:
+ """While a scene is built, glob.glob / glob.iglob return their matches sorted. Four RoboTwin tasks (place_a2b_left,
+ place_a2b_right, place_object_scale, place_object_stand: envs/<task>.py load_actors) list an object's model files
+ with glob.glob and then draw one with the seeded np.random.choice; glob returns the file system's directory order,
+ which differs between hosts (and image layers), so one seed built a different object on another host and the frozen
+ instance's hash check failed there (board 0086 claim 5; never rebuilt off the reference workstation). Sorted, the
+ draw depends on the seed alone. No other task of ours lists files while it builds its scene."""
+
+ def __enter__(self):
+ import glob
+ self._glob, self._orig = glob, (glob.glob, glob.iglob)
+ glob.glob = lambda *a, **k: sorted(self._orig[0](*a, **k))
+ glob.iglob = lambda *a, **k: iter(sorted(self._orig[1](*a, **k)))
+ return self
+
+ def __exit__(self, *exc):
+ self._glob.glob, self._glob.iglob = self._orig
+ return False
+
+
def _chdir_root() -> None:
# RoboTwin resolves assets/config paths relative to the repo root and via ROOT_PATH
if os.getcwd() != ROBOTWIN_ROOT and os.path.isdir(ROBOTWIN_ROOT):
@@ -218,7 +238,8 @@ class Session:
if self.enable_render:
sapien.render.set_camera_shader_dir(self.render_shader)
args = _build_args(self.task, self.embodiment, self.seed)
- self.env.setup_demo(**args)
+ with _SortedGlob(): # the object draw depends on the seed alone, not the host's directory order
+ self.env.setup_demo(**args)
if self.enable_render: # setup_scene() switches the shader to ray tracing; put it back for cheap frames
sapien.render.set_camera_shader_dir(self.render_shader)
self._raw_scene = self.env.scene
diff --git a/tasks/robotwin_2_0_place_object_basket-privileged/task.toml b/tasks/robotwin_2_0_place_object_basket-privileged/task.toml
index 6dbb0ea15f..778e5a596e 100644
--- a/tasks/robotwin_2_0_place_object_basket-privileged/task.toml
+++ b/tasks/robotwin_2_0_place_object_basket-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "robotwin" # RoboTwin 2.0 = SAPIEN 3 / PhysX + CuRobo
robot = "aloha-agilex"
family = "place_object_basket"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 60417 # scene = environment/instance.json (RoboTwin seed + expected initial-state hash); identity = dataset.toml digest
expert_time_estimate_hours = 0.5
diff --git a/tasks/robotwin_2_0_place_object_basket-privileged/tests/rcb_traj.py b/tasks/robotwin_2_0_place_object_basket-privileged/tests/rcb_traj.py
index edf11216a4..c01ca13ae7 100644
--- a/tasks/robotwin_2_0_place_object_basket-privileged/tests/rcb_traj.py
+++ b/tasks/robotwin_2_0_place_object_basket-privileged/tests/rcb_traj.py
@@ -113,6 +113,26 @@ BASE_ARGS: dict[str, Any] = {
}
+class _SortedGlob:
+ """While a scene is built, glob.glob / glob.iglob return their matches sorted. Four RoboTwin tasks (place_a2b_left,
+ place_a2b_right, place_object_scale, place_object_stand: envs/<task>.py load_actors) list an object's model files
+ with glob.glob and then draw one with the seeded np.random.choice; glob returns the file system's directory order,
+ which differs between hosts (and image layers), so one seed built a different object on another host and the frozen
+ instance's hash check failed there (board 0086 claim 5; never rebuilt off the reference workstation). Sorted, the
+ draw depends on the seed alone. No other task of ours lists files while it builds its scene."""
+
+ def __enter__(self):
+ import glob
+ self._glob, self._orig = glob, (glob.glob, glob.iglob)
+ glob.glob = lambda *a, **k: sorted(self._orig[0](*a, **k))
+ glob.iglob = lambda *a, **k: iter(sorted(self._orig[1](*a, **k)))
+ return self
+
+ def __exit__(self, *exc):
+ self._glob.glob, self._glob.iglob = self._orig
+ return False
+
+
def _chdir_root() -> None:
# RoboTwin resolves assets/config paths relative to the repo root and via ROOT_PATH
if os.getcwd() != ROBOTWIN_ROOT and os.path.isdir(ROBOTWIN_ROOT):
@@ -218,7 +238,8 @@ class Session:
if self.enable_render:
sapien.render.set_camera_shader_dir(self.render_shader)
args = _build_args(self.task, self.embodiment, self.seed)
- self.env.setup_demo(**args)
+ with _SortedGlob(): # the object draw depends on the seed alone, not the host's directory order
+ self.env.setup_demo(**args)
if self.enable_render: # setup_scene() switches the shader to ray tracing; put it back for cheap frames
sapien.render.set_camera_shader_dir(self.render_shader)
self._raw_scene = self.env.scene
diff --git a/tasks/robotwin_2_0_place_object_basket-standard/README.md b/tasks/robotwin_2_0_place_object_basket-standard/README.md
index 9adfebd61f..0b2e04b096 100644
--- a/tasks/robotwin_2_0_place_object_basket-standard/README.md
+++ b/tasks/robotwin_2_0_place_object_basket-standard/README.md
@@ -53,3 +53,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
- 2026-10-10 (owner decision, board 0072): the agent's network egress is restricted to the model APIs (`[agent] network_mode = "allowlist"`, the hosts of `scripts/set_agent_egress.py`; since 2026-10-11 these include the package registries, PyPI and the Debian / Ubuntu archives, so that `pip install` and `apt-get install` work); it was open before (Harbor's default), and agents could download upstream sources, reference solutions and robot models from their own shell. The scene, the frozen instance, the tooling and the pass rule are unchanged, so the task keeps its name; results from before this date were obtained with open egress.
- 2026-10-13 (board 0086): the service's copy of the tooling only (`environment/sim/rcb_traj.py`, one line of `environment/sim/server.py`). The tooling now refuses to save or replay a trajectory of more than 200,000 rows, the privileged verifier's cap; a standard episode has no row cap, so the service lifts it for its own session (`S.max_rows = None`). Nothing an agent sees and nothing the service grades changes.
- 2026-10-13 (board 0100): base image `eai-robotwin` 0.1.2, whose scrub of RoboTwin's expert also removes the reference poses that `load_actors` computed only for the expert. The standard-mode agent never reaches the env, so nothing it observes changes; the scene, the instance and the judge are unchanged.
+- 2026-10-14 (board 0113): the tooling only (`environment/sim/rcb_traj.py`): while it builds the scene, `glob.glob` returns its matches sorted, so that a task which lists files to draw an object depends on its seed alone, not on the host's directory order (four RoboTwin tasks did; this one lists no files while it builds its scene). Its scene, frozen instance (same initial-state hash), verifier and pass rule are unchanged.
diff --git a/tasks/robotwin_2_0_place_object_basket-standard/environment/sim/rcb_traj.py b/tasks/robotwin_2_0_place_object_basket-standard/environment/sim/rcb_traj.py
index edf11216a4..c01ca13ae7 100644
--- a/tasks/robotwin_2_0_place_object_basket-standard/environment/sim/rcb_traj.py
+++ b/tasks/robotwin_2_0_place_object_basket-standard/environment/sim/rcb_traj.py
@@ -113,6 +113,26 @@ BASE_ARGS: dict[str, Any] = {
}
+class _SortedGlob:
+ """While a scene is built, glob.glob / glob.iglob return their matches sorted. Four RoboTwin tasks (place_a2b_left,
+ place_a2b_right, place_object_scale, place_object_stand: envs/<task>.py load_actors) list an object's model files
+ with glob.glob and then draw one with the seeded np.random.choice; glob returns the file system's directory order,
+ which differs between hosts (and image layers), so one seed built a different object on another host and the frozen
+ instance's hash check failed there (board 0086 claim 5; never rebuilt off the reference workstation). Sorted, the
+ draw depends on the seed alone. No other task of ours lists files while it builds its scene."""
+
+ def __enter__(self):
+ import glob
+ self._glob, self._orig = glob, (glob.glob, glob.iglob)
+ glob.glob = lambda *a, **k: sorted(self._orig[0](*a, **k))
+ glob.iglob = lambda *a, **k: iter(sorted(self._orig[1](*a, **k)))
+ return self
+
+ def __exit__(self, *exc):
+ self._glob.glob, self._glob.iglob = self._orig
+ return False
+
+
def _chdir_root() -> None:
# RoboTwin resolves assets/config paths relative to the repo root and via ROOT_PATH
if os.getcwd() != ROBOTWIN_ROOT and os.path.isdir(ROBOTWIN_ROOT):
@@ -218,7 +238,8 @@ class Session:
if self.enable_render:
sapien.render.set_camera_shader_dir(self.render_shader)
args = _build_args(self.task, self.embodiment, self.seed)
- self.env.setup_demo(**args)
+ with _SortedGlob(): # the object draw depends on the seed alone, not the host's directory order
+ self.env.setup_demo(**args)
if self.enable_render: # setup_scene() switches the shader to ray tracing; put it back for cheap frames
sapien.render.set_camera_shader_dir(self.render_shader)
self._raw_scene = self.env.scene
diff --git a/tasks/robotwin_2_0_place_object_basket-standard/task.toml b/tasks/robotwin_2_0_place_object_basket-standard/task.toml
index 7471792ea6..23b9242181 100644
--- a/tasks/robotwin_2_0_place_object_basket-standard/task.toml
+++ b/tasks/robotwin_2_0_place_object_basket-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "robotwin"
robot = "aloha-agilex"
family = "place_object_basket"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 60417
diff --git a/tasks/robotwin_2_0_place_object_scale-privileged/README.md b/tasks/robotwin_2_0_place_object_scale-privileged/README.md
index 8abb458868..ab76b3be2b 100644
--- a/tasks/robotwin_2_0_place_object_scale-privileged/README.md
+++ b/tasks/robotwin_2_0_place_object_scale-privileged/README.md
@@ -55,3 +55,4 @@ Use one arm to grab the object and put it on the scale.
- 2026-10-10 (owner decision, board 0072): the agent's network egress is restricted to the model APIs (`[agent] network_mode = "allowlist"`, the hosts of `scripts/set_agent_egress.py`; since 2026-10-11 these include the package registries, PyPI and the Debian / Ubuntu archives, so that `pip install` and `apt-get install` work); it was open before (Harbor's default), and agents could download upstream sources, reference solutions and robot models from their own shell. The scene, the frozen instance, the tooling and the pass rule are unchanged, so the task keeps its name; results from before this date were obtained with open egress.
- 2026-10-13 (board 0086): the tooling refuses a trajectory the verifier cannot grade (`environment/rcb_traj.py`, `tests/rcb_traj.py`). `Session.save_trajectory` and `Session.replay` raise `ValueError` for more than 200,000 rows (`rcb_traj.MAX_T`, the verifier's `MAX_T` in `tests/grade.py`: 800 s of physics), and `record_stop` warns on stderr. Before, the tool saved and replayed such a file ("success") and the verifier then graded it `bad_trajectory`; the instruction has stated the cap since 2026-10-10. The verifier, the scene, the instance and the pass rule are unchanged: every trajectory is graded as before.
- 2026-10-13 (board 0100): base image `eai-robotwin` 0.1.2. Its scrub of RoboTwin's expert also removes the reference poses that 18 tasks' `load_actors` computed only for the expert, and the expert's comments above the scrubbed method bodies (`images/robotwin/scrub_oracles.py`). This task's `load_actors` set no such value: its scene, instance, tooling, verifier and pass rule are unchanged.
+- 2026-10-14 (owner decision C2, board 0113): **the frozen instance was re-frozen** (same seed, initial-state hash f1442d98b394 -> e21634a59bcf; the object `050_bell` (model 0)). RoboTwin's `load_actors` for this task lists the object's model files with `glob.glob` and draws one with the seeded `np.random.choice`; `glob` returns the file system's directory order, which differs between hosts, so the same seed built a different object on another host and the frozen instance's hash check failed there (board 0086; no run of this task ever got past it on the pool's hosts). The tooling (`environment/rcb_traj.py`, `tests/rcb_traj.py`) now sorts `glob`'s matches while it builds the scene, so the draw depends on the seed alone: the instance rebuilds to the same hash twice in one process, in a fresh container and with the file system's order reversed. The reference solution passes on it (two replays deterministic, verifier reward 1); a trajectory handed in for the old instance is not valid for this one. The task text, the success check and the pass rule are unchanged.
diff --git a/tasks/robotwin_2_0_place_object_scale-privileged/environment/instance.json b/tasks/robotwin_2_0_place_object_scale-privileged/environment/instance.json
index 7df2c25d17..afa39e8d27 100644
--- a/tasks/robotwin_2_0_place_object_scale-privileged/environment/instance.json
+++ b/tasks/robotwin_2_0_place_object_scale-privileged/environment/instance.json
@@ -2,7 +2,7 @@
"task": "place_object_scale",
"seed": 60417,
"embodiment": "aloha-agilex",
- "state_sha": "f1442d98b394",
+ "state_sha": "e21634a59bcf",
"text": "Use one arm to grab the object and put it on the scale.",
"n_active_joints": 16
}
\ No newline at end of file
diff --git a/tasks/robotwin_2_0_place_object_scale-privileged/environment/rcb_traj.py b/tasks/robotwin_2_0_place_object_scale-privileged/environment/rcb_traj.py
index 404d6a7f38..57e5649e23 100644
--- a/tasks/robotwin_2_0_place_object_scale-privileged/environment/rcb_traj.py
+++ b/tasks/robotwin_2_0_place_object_scale-privileged/environment/rcb_traj.py
@@ -113,6 +113,26 @@ BASE_ARGS: dict[str, Any] = {
}
+class _SortedGlob:
+ """While a scene is built, glob.glob / glob.iglob return their matches sorted. Four RoboTwin tasks (place_a2b_left,
+ place_a2b_right, place_object_scale, place_object_stand: envs/<task>.py load_actors) list an object's model files
+ with glob.glob and then draw one with the seeded np.random.choice; glob returns the file system's directory order,
+ which differs between hosts (and image layers), so one seed built a different object on another host and the frozen
+ instance's hash check failed there (board 0086 claim 5; never rebuilt off the reference workstation). Sorted, the
+ draw depends on the seed alone. No other task of ours lists files while it builds its scene."""
+
+ def __enter__(self):
+ import glob
+ self._glob, self._orig = glob, (glob.glob, glob.iglob)
+ glob.glob = lambda *a, **k: sorted(self._orig[0](*a, **k))
+ glob.iglob = lambda *a, **k: iter(sorted(self._orig[1](*a, **k)))
+ return self
+
+ def __exit__(self, *exc):
+ self._glob.glob, self._glob.iglob = self._orig
+ return False
+
+
def _chdir_root() -> None:
# RoboTwin resolves assets/config paths relative to the repo root and via ROOT_PATH
if os.getcwd() != ROBOTWIN_ROOT and os.path.isdir(ROBOTWIN_ROOT):
@@ -218,7 +238,8 @@ class Session:
if self.enable_render:
sapien.render.set_camera_shader_dir(self.render_shader)
args = _build_args(self.task, self.embodiment, self.seed)
- self.env.setup_demo(**args)
+ with _SortedGlob(): # the object draw depends on the seed alone, not the host's directory order
+ self.env.setup_demo(**args)
if self.enable_render: # setup_scene() switches the shader to ray tracing; put it back for cheap frames
sapien.render.set_camera_shader_dir(self.render_shader)
self._raw_scene = self.env.scene
diff --git a/tasks/robotwin_2_0_place_object_scale-privileged/task.toml b/tasks/robotwin_2_0_place_object_scale-privileged/task.toml
index 9bbb485655..2be9324e10 100644
--- a/tasks/robotwin_2_0_place_object_scale-privileged/task.toml
+++ b/tasks/robotwin_2_0_place_object_scale-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "robotwin" # RoboTwin 2.0 = SAPIEN 3 / PhysX + CuRobo
robot = "aloha-agilex"
family = "place_object_scale"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 60417 # scene = environment/instance.json (RoboTwin seed + expected initial-state hash); identity = dataset.toml digest
expert_time_estimate_hours = 0.5
diff --git a/tasks/robotwin_2_0_place_object_scale-privileged/tests/instance.json b/tasks/robotwin_2_0_place_object_scale-privileged/tests/instance.json
index 7df2c25d17..afa39e8d27 100644
--- a/tasks/robotwin_2_0_place_object_scale-privileged/tests/instance.json
+++ b/tasks/robotwin_2_0_place_object_scale-privileged/tests/instance.json
@@ -2,7 +2,7 @@
"task": "place_object_scale",
"seed": 60417,
"embodiment": "aloha-agilex",
- "state_sha": "f1442d98b394",
+ "state_sha": "e21634a59bcf",
"text": "Use one arm to grab the object and put it on the scale.",
"n_active_joints": 16
}
\ No newline at end of file
diff --git a/tasks/robotwin_2_0_place_object_scale-privileged/tests/rcb_traj.py b/tasks/robotwin_2_0_place_object_scale-privileged/tests/rcb_traj.py
index 404d6a7f38..57e5649e23 100644
--- a/tasks/robotwin_2_0_place_object_scale-privileged/tests/rcb_traj.py
+++ b/tasks/robotwin_2_0_place_object_scale-privileged/tests/rcb_traj.py
@@ -113,6 +113,26 @@ BASE_ARGS: dict[str, Any] = {
}
+class _SortedGlob:
+ """While a scene is built, glob.glob / glob.iglob return their matches sorted. Four RoboTwin tasks (place_a2b_left,
+ place_a2b_right, place_object_scale, place_object_stand: envs/<task>.py load_actors) list an object's model files
+ with glob.glob and then draw one with the seeded np.random.choice; glob returns the file system's directory order,
+ which differs between hosts (and image layers), so one seed built a different object on another host and the frozen
+ instance's hash check failed there (board 0086 claim 5; never rebuilt off the reference workstation). Sorted, the
+ draw depends on the seed alone. No other task of ours lists files while it builds its scene."""
+
+ def __enter__(self):
+ import glob
+ self._glob, self._orig = glob, (glob.glob, glob.iglob)
+ glob.glob = lambda *a, **k: sorted(self._orig[0](*a, **k))
+ glob.iglob = lambda *a, **k: iter(sorted(self._orig[1](*a, **k)))
+ return self
+
+ def __exit__(self, *exc):
+ self._glob.glob, self._glob.iglob = self._orig
+ return False
+
+
def _chdir_root() -> None:
# RoboTwin resolves assets/config paths relative to the repo root and via ROOT_PATH
if os.getcwd() != ROBOTWIN_ROOT and os.path.isdir(ROBOTWIN_ROOT):
@@ -218,7 +238,8 @@ class Session:
if self.enable_render:
sapien.render.set_camera_shader_dir(self.render_shader)
args = _build_args(self.task, self.embodiment, self.seed)
- self.env.setup_demo(**args)
+ with _SortedGlob(): # the object draw depends on the seed alone, not the host's directory order
+ self.env.setup_demo(**args)
if self.enable_render: # setup_scene() switches the shader to ray tracing; put it back for cheap frames
sapien.render.set_camera_shader_dir(self.render_shader)
self._raw_scene = self.env.scene
diff --git a/tasks/robotwin_2_0_place_object_scale-standard/README.md b/tasks/robotwin_2_0_place_object_scale-standard/README.md
index 16edac9023..10f6f8d342 100644
--- a/tasks/robotwin_2_0_place_object_scale-standard/README.md
+++ b/tasks/robotwin_2_0_place_object_scale-standard/README.md
@@ -53,3 +53,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
- 2026-10-10 (owner decision, board 0072): the agent's network egress is restricted to the model APIs (`[agent] network_mode = "allowlist"`, the hosts of `scripts/set_agent_egress.py`; since 2026-10-11 these include the package registries, PyPI and the Debian / Ubuntu archives, so that `pip install` and `apt-get install` work); it was open before (Harbor's default), and agents could download upstream sources, reference solutions and robot models from their own shell. The scene, the frozen instance, the tooling and the pass rule are unchanged, so the task keeps its name; results from before this date were obtained with open egress.
- 2026-10-13 (board 0086): the service's copy of the tooling only (`environment/sim/rcb_traj.py`, one line of `environment/sim/server.py`). The tooling now refuses to save or replay a trajectory of more than 200,000 rows, the privileged verifier's cap; a standard episode has no row cap, so the service lifts it for its own session (`S.max_rows = None`). Nothing an agent sees and nothing the service grades changes.
- 2026-10-13 (board 0100): base image `eai-robotwin` 0.1.2, whose scrub of RoboTwin's expert also removes the reference poses that `load_actors` computed only for the expert. The standard-mode agent never reaches the env, so nothing it observes changes; the scene, the instance and the judge are unchanged.
+- 2026-10-14 (owner decision C2, board 0113): **the frozen instance was re-frozen** with the privileged twin (same seed, initial-state hash f1442d98b394 -> e21634a59bcf; the object `050_bell` (model 0)). RoboTwin's `load_actors` for this task lists the object's model files with `glob.glob` and draws one with the seeded `np.random.choice`; `glob` returns the file system's directory order, which differs between hosts, so the same seed built a different object on another host and the frozen instance's hash check failed there (board 0086; no run of this task ever got past it on the pool's hosts); the service rebuilt the instance and gave up with `instance_mismatch`. The service's copy of the tooling (`environment/sim/rcb_traj.py`) now sorts `glob`'s matches while it builds the scene. The task text, the success check and the pass rule are unchanged.
diff --git a/tasks/robotwin_2_0_place_object_scale-standard/environment/sim/instance.json b/tasks/robotwin_2_0_place_object_scale-standard/environment/sim/instance.json
index 7df2c25d17..afa39e8d27 100644
--- a/tasks/robotwin_2_0_place_object_scale-standard/environment/sim/instance.json
+++ b/tasks/robotwin_2_0_place_object_scale-standard/environment/sim/instance.json
@@ -2,7 +2,7 @@
"task": "place_object_scale",
"seed": 60417,
"embodiment": "aloha-agilex",
- "state_sha": "f1442d98b394",
+ "state_sha": "e21634a59bcf",
"text": "Use one arm to grab the object and put it on the scale.",
"n_active_joints": 16
}
\ No newline at end of file
diff --git a/tasks/robotwin_2_0_place_object_scale-standard/environment/sim/rcb_traj.py b/tasks/robotwin_2_0_place_object_scale-standard/environment/sim/rcb_traj.py
index 404d6a7f38..57e5649e23 100644
--- a/tasks/robotwin_2_0_place_object_scale-standard/environment/sim/rcb_traj.py
+++ b/tasks/robotwin_2_0_place_object_scale-standard/environment/sim/rcb_traj.py
@@ -113,6 +113,26 @@ BASE_ARGS: dict[str, Any] = {
}
+class _SortedGlob:
+ """While a scene is built, glob.glob / glob.iglob return their matches sorted. Four RoboTwin tasks (place_a2b_left,
+ place_a2b_right, place_object_scale, place_object_stand: envs/<task>.py load_actors) list an object's model files
+ with glob.glob and then draw one with the seeded np.random.choice; glob returns the file system's directory order,
+ which differs between hosts (and image layers), so one seed built a different object on another host and the frozen
+ instance's hash check failed there (board 0086 claim 5; never rebuilt off the reference workstation). Sorted, the
+ draw depends on the seed alone. No other task of ours lists files while it builds its scene."""
+
+ def __enter__(self):
+ import glob
+ self._glob, self._orig = glob, (glob.glob, glob.iglob)
+ glob.glob = lambda *a, **k: sorted(self._orig[0](*a, **k))
+ glob.iglob = lambda *a, **k: iter(sorted(self._orig[1](*a, **k)))
+ return self
+
+ def __exit__(self, *exc):
+ self._glob.glob, self._glob.iglob = self._orig
+ return False
+
+
def _chdir_root() -> None:
# RoboTwin resolves assets/config paths relative to the repo root and via ROOT_PATH
if os.getcwd() != ROBOTWIN_ROOT and os.path.isdir(ROBOTWIN_ROOT):
@@ -218,7 +238,8 @@ class Session:
if self.enable_render:
sapien.render.set_camera_shader_dir(self.render_shader)
args = _build_args(self.task, self.embodiment, self.seed)
- self.env.setup_demo(**args)
+ with _SortedGlob(): # the object draw depends on the seed alone, not the host's directory order
+ self.env.setup_demo(**args)
if self.enable_render: # setup_scene() switches the shader to ray tracing; put it back for cheap frames
sapien.render.set_camera_shader_dir(self.render_shader)
self._raw_scene = self.env.scene
diff --git a/tasks/robotwin_2_0_place_object_scale-standard/task.toml b/tasks/robotwin_2_0_place_object_scale-standard/task.toml
index 6c46ff5467..4a5a7ef3eb 100644
--- a/tasks/robotwin_2_0_place_object_scale-standard/task.toml
+++ b/tasks/robotwin_2_0_place_object_scale-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "robotwin"
robot = "aloha-agilex"
family = "place_object_scale"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 60417
diff --git a/tasks/robotwin_2_0_place_object_stand-privileged/README.md b/tasks/robotwin_2_0_place_object_stand-privileged/README.md
index 741ff0ca50..37301014cf 100644
--- a/tasks/robotwin_2_0_place_object_stand-privileged/README.md
+++ b/tasks/robotwin_2_0_place_object_stand-privileged/README.md
@@ -55,3 +55,4 @@ Use appropriate arm to place the object on the stand.
- 2026-10-10 (owner decision, board 0072): the agent's network egress is restricted to the model APIs (`[agent] network_mode = "allowlist"`, the hosts of `scripts/set_agent_egress.py`; since 2026-10-11 these include the package registries, PyPI and the Debian / Ubuntu archives, so that `pip install` and `apt-get install` work); it was open before (Harbor's default), and agents could download upstream sources, reference solutions and robot models from their own shell. The scene, the frozen instance, the tooling and the pass rule are unchanged, so the task keeps its name; results from before this date were obtained with open egress.
- 2026-10-13 (board 0086): the tooling refuses a trajectory the verifier cannot grade (`environment/rcb_traj.py`, `tests/rcb_traj.py`). `Session.save_trajectory` and `Session.replay` raise `ValueError` for more than 200,000 rows (`rcb_traj.MAX_T`, the verifier's `MAX_T` in `tests/grade.py`: 800 s of physics), and `record_stop` warns on stderr. Before, the tool saved and replayed such a file ("success") and the verifier then graded it `bad_trajectory`; the instruction has stated the cap since 2026-10-10. The verifier, the scene, the instance and the pass rule are unchanged: every trajectory is graded as before.
- 2026-10-13 (board 0100): base image `eai-robotwin` 0.1.2. Its scrub of RoboTwin's expert also removes the reference poses that 18 tasks' `load_actors` computed only for the expert, and the expert's comments above the scrubbed method bodies (`images/robotwin/scrub_oracles.py`). This task's `load_actors` set no such value: its scene, instance, tooling, verifier and pass rule are unchanged.
+- 2026-10-14 (owner decision C2, board 0113): **the frozen instance was re-frozen** (same seed, initial-state hash 44250d2d7e60 -> 9c68b628c606; the object `050_bell` (model 0)). RoboTwin's `load_actors` for this task lists the object's model files with `glob.glob` and draws one with the seeded `np.random.choice`; `glob` returns the file system's directory order, which differs between hosts, so the same seed built a different object on another host and the frozen instance's hash check failed there (board 0086; no run of this task ever got past it on the pool's hosts). The tooling (`environment/rcb_traj.py`, `tests/rcb_traj.py`) now sorts `glob`'s matches while it builds the scene, so the draw depends on the seed alone: the instance rebuilds to the same hash twice in one process, in a fresh container and with the file system's order reversed. The reference solution passes on it (two replays deterministic, verifier reward 1); a trajectory handed in for the old instance is not valid for this one. The task text, the success check and the pass rule are unchanged.
diff --git a/tasks/robotwin_2_0_place_object_stand-privileged/environment/instance.json b/tasks/robotwin_2_0_place_object_stand-privileged/environment/instance.json
index 97d09ec54c..ce3d96a625 100644
--- a/tasks/robotwin_2_0_place_object_stand-privileged/environment/instance.json
+++ b/tasks/robotwin_2_0_place_object_stand-privileged/environment/instance.json
@@ -2,7 +2,7 @@
"task": "place_object_stand",
"seed": 60417,
"embodiment": "aloha-agilex",
- "state_sha": "44250d2d7e60",
+ "state_sha": "9c68b628c606",
"text": "Use appropriate arm to place the object on the stand.",
"n_active_joints": 16
}
\ No newline at end of file
diff --git a/tasks/robotwin_2_0_place_object_stand-privileged/environment/rcb_traj.py b/tasks/robotwin_2_0_place_object_stand-privileged/environment/rcb_traj.py
index 0d76d65ba8..a1d6b1a71c 100644
--- a/tasks/robotwin_2_0_place_object_stand-privileged/environment/rcb_traj.py
+++ b/tasks/robotwin_2_0_place_object_stand-privileged/environment/rcb_traj.py
@@ -113,6 +113,26 @@ BASE_ARGS: dict[str, Any] = {
}
+class _SortedGlob:
+ """While a scene is built, glob.glob / glob.iglob return their matches sorted. Four RoboTwin tasks (place_a2b_left,
+ place_a2b_right, place_object_scale, place_object_stand: envs/<task>.py load_actors) list an object's model files
+ with glob.glob and then draw one with the seeded np.random.choice; glob returns the file system's directory order,
+ which differs between hosts (and image layers), so one seed built a different object on another host and the frozen
+ instance's hash check failed there (board 0086 claim 5; never rebuilt off the reference workstation). Sorted, the
+ draw depends on the seed alone. No other task of ours lists files while it builds its scene."""
+
+ def __enter__(self):
+ import glob
+ self._glob, self._orig = glob, (glob.glob, glob.iglob)
+ glob.glob = lambda *a, **k: sorted(self._orig[0](*a, **k))
+ glob.iglob = lambda *a, **k: iter(sorted(self._orig[1](*a, **k)))
+ return self
+
+ def __exit__(self, *exc):
+ self._glob.glob, self._glob.iglob = self._orig
+ return False
+
+
def _chdir_root() -> None:
# RoboTwin resolves assets/config paths relative to the repo root and via ROOT_PATH
if os.getcwd() != ROBOTWIN_ROOT and os.path.isdir(ROBOTWIN_ROOT):
@@ -218,7 +238,8 @@ class Session:
if self.enable_render:
sapien.render.set_camera_shader_dir(self.render_shader)
args = _build_args(self.task, self.embodiment, self.seed)
- self.env.setup_demo(**args)
+ with _SortedGlob(): # the object draw depends on the seed alone, not the host's directory order
+ self.env.setup_demo(**args)
if self.enable_render: # setup_scene() switches the shader to ray tracing; put it back for cheap frames
sapien.render.set_camera_shader_dir(self.render_shader)
self._raw_scene = self.env.scene
diff --git a/tasks/robotwin_2_0_place_object_stand-privileged/task.toml b/tasks/robotwin_2_0_place_object_stand-privileged/task.toml
index 3b401689a3..35d927e0ef 100644
--- a/tasks/robotwin_2_0_place_object_stand-privileged/task.toml
+++ b/tasks/robotwin_2_0_place_object_stand-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "robotwin" # RoboTwin 2.0 = SAPIEN 3 / PhysX + CuRobo
robot = "aloha-agilex"
family = "place_object_stand"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 60417 # scene = environment/instance.json (RoboTwin seed + expected initial-state hash); identity = dataset.toml digest
expert_time_estimate_hours = 0.5
diff --git a/tasks/robotwin_2_0_place_object_stand-privileged/tests/instance.json b/tasks/robotwin_2_0_place_object_stand-privileged/tests/instance.json
index 97d09ec54c..ce3d96a625 100644
--- a/tasks/robotwin_2_0_place_object_stand-privileged/tests/instance.json
+++ b/tasks/robotwin_2_0_place_object_stand-privileged/tests/instance.json
@@ -2,7 +2,7 @@
"task": "place_object_stand",
"seed": 60417,
"embodiment": "aloha-agilex",
- "state_sha": "44250d2d7e60",
+ "state_sha": "9c68b628c606",
"text": "Use appropriate arm to place the object on the stand.",
"n_active_joints": 16
}
\ No newline at end of file
diff --git a/tasks/robotwin_2_0_place_object_stand-privileged/tests/rcb_traj.py b/tasks/robotwin_2_0_place_object_stand-privileged/tests/rcb_traj.py
index 0d76d65ba8..a1d6b1a71c 100644
--- a/tasks/robotwin_2_0_place_object_stand-privileged/tests/rcb_traj.py
+++ b/tasks/robotwin_2_0_place_object_stand-privileged/tests/rcb_traj.py
@@ -113,6 +113,26 @@ BASE_ARGS: dict[str, Any] = {
}
+class _SortedGlob:
+ """While a scene is built, glob.glob / glob.iglob return their matches sorted. Four RoboTwin tasks (place_a2b_left,
+ place_a2b_right, place_object_scale, place_object_stand: envs/<task>.py load_actors) list an object's model files
+ with glob.glob and then draw one with the seeded np.random.choice; glob returns the file system's directory order,
+ which differs between hosts (and image layers), so one seed built a different object on another host and the frozen
+ instance's hash check failed there (board 0086 claim 5; never rebuilt off the reference workstation). Sorted, the
+ draw depends on the seed alone. No other task of ours lists files while it builds its scene."""
+
+ def __enter__(self):
+ import glob
+ self._glob, self._orig = glob, (glob.glob, glob.iglob)
+ glob.glob = lambda *a, **k: sorted(self._orig[0](*a, **k))
+ glob.iglob = lambda *a, **k: iter(sorted(self._orig[1](*a, **k)))
+ return self
+
+ def __exit__(self, *exc):
+ self._glob.glob, self._glob.iglob = self._orig
+ return False
+
+
def _chdir_root() -> None:
# RoboTwin resolves assets/config paths relative to the repo root and via ROOT_PATH
if os.getcwd() != ROBOTWIN_ROOT and os.path.isdir(ROBOTWIN_ROOT):
@@ -218,7 +238,8 @@ class Session:
if self.enable_render:
sapien.render.set_camera_shader_dir(self.render_shader)
args = _build_args(self.task, self.embodiment, self.seed)
- self.env.setup_demo(**args)
+ with _SortedGlob(): # the object draw depends on the seed alone, not the host's directory order
+ self.env.setup_demo(**args)
if self.enable_render: # setup_scene() switches the shader to ray tracing; put it back for cheap frames
sapien.render.set_camera_shader_dir(self.render_shader)
self._raw_scene = self.env.scene
diff --git a/tasks/robotwin_2_0_place_object_stand-standard/README.md b/tasks/robotwin_2_0_place_object_stand-standard/README.md
index 516cddfc00..8f42a3aeda 100644
--- a/tasks/robotwin_2_0_place_object_stand-standard/README.md
+++ b/tasks/robotwin_2_0_place_object_stand-standard/README.md
@@ -53,3 +53,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
- 2026-10-10 (owner decision, board 0072): the agent's network egress is restricted to the model APIs (`[agent] network_mode = "allowlist"`, the hosts of `scripts/set_agent_egress.py`; since 2026-10-11 these include the package registries, PyPI and the Debian / Ubuntu archives, so that `pip install` and `apt-get install` work); it was open before (Harbor's default), and agents could download upstream sources, reference solutions and robot models from their own shell. The scene, the frozen instance, the tooling and the pass rule are unchanged, so the task keeps its name; results from before this date were obtained with open egress.
- 2026-10-13 (board 0086): the service's copy of the tooling only (`environment/sim/rcb_traj.py`, one line of `environment/sim/server.py`). The tooling now refuses to save or replay a trajectory of more than 200,000 rows, the privileged verifier's cap; a standard episode has no row cap, so the service lifts it for its own session (`S.max_rows = None`). Nothing an agent sees and nothing the service grades changes.
- 2026-10-13 (board 0100): base image `eai-robotwin` 0.1.2, whose scrub of RoboTwin's expert also removes the reference poses that `load_actors` computed only for the expert. The standard-mode agent never reaches the env, so nothing it observes changes; the scene, the instance and the judge are unchanged.
+- 2026-10-14 (owner decision C2, board 0113): **the frozen instance was re-frozen** with the privileged twin (same seed, initial-state hash 44250d2d7e60 -> 9c68b628c606; the object `050_bell` (model 0)). RoboTwin's `load_actors` for this task lists the object's model files with `glob.glob` and draws one with the seeded `np.random.choice`; `glob` returns the file system's directory order, which differs between hosts, so the same seed built a different object on another host and the frozen instance's hash check failed there (board 0086; no run of this task ever got past it on the pool's hosts); the service rebuilt the instance and gave up with `instance_mismatch`. The service's copy of the tooling (`environment/sim/rcb_traj.py`) now sorts `glob`'s matches while it builds the scene. The task text, the success check and the pass rule are unchanged.
diff --git a/tasks/robotwin_2_0_place_object_stand-standard/environment/sim/instance.json b/tasks/robotwin_2_0_place_object_stand-standard/environment/sim/instance.json
index 97d09ec54c..ce3d96a625 100644
--- a/tasks/robotwin_2_0_place_object_stand-standard/environment/sim/instance.json
+++ b/tasks/robotwin_2_0_place_object_stand-standard/environment/sim/instance.json
@@ -2,7 +2,7 @@
"task": "place_object_stand",
"seed": 60417,
"embodiment": "aloha-agilex",
- "state_sha": "44250d2d7e60",
+ "state_sha": "9c68b628c606",
"text": "Use appropriate arm to place the object on the stand.",
"n_active_joints": 16
}
\ No newline at end of file
diff --git a/tasks/robotwin_2_0_place_object_stand-standard/environment/sim/rcb_traj.py b/tasks/robotwin_2_0_place_object_stand-standard/environment/sim/rcb_traj.py
index 0d76d65ba8..a1d6b1a71c 100644
--- a/tasks/robotwin_2_0_place_object_stand-standard/environment/sim/rcb_traj.py
+++ b/tasks/robotwin_2_0_place_object_stand-standard/environment/sim/rcb_traj.py
@@ -113,6 +113,26 @@ BASE_ARGS: dict[str, Any] = {
}
+class _SortedGlob:
+ """While a scene is built, glob.glob / glob.iglob return their matches sorted. Four RoboTwin tasks (place_a2b_left,
+ place_a2b_right, place_object_scale, place_object_stand: envs/<task>.py load_actors) list an object's model files
+ with glob.glob and then draw one with the seeded np.random.choice; glob returns the file system's directory order,
+ which differs between hosts (and image layers), so one seed built a different object on another host and the frozen
+ instance's hash check failed there (board 0086 claim 5; never rebuilt off the reference workstation). Sorted, the
+ draw depends on the seed alone. No other task of ours lists files while it builds its scene."""
+
+ def __enter__(self):
+ import glob
+ self._glob, self._orig = glob, (glob.glob, glob.iglob)
+ glob.glob = lambda *a, **k: sorted(self._orig[0](*a, **k))
+ glob.iglob = lambda *a, **k: iter(sorted(self._orig[1](*a, **k)))
+ return self
+
+ def __exit__(self, *exc):
+ self._glob.glob, self._glob.iglob = self._orig
+ return False
+
+
def _chdir_root() -> None:
# RoboTwin resolves assets/config paths relative to the repo root and via ROOT_PATH
if os.getcwd() != ROBOTWIN_ROOT and os.path.isdir(ROBOTWIN_ROOT):
@@ -218,7 +238,8 @@ class Session:
if self.enable_render:
sapien.render.set_camera_shader_dir(self.render_shader)
args = _build_args(self.task, self.embodiment, self.seed)
- self.env.setup_demo(**args)
+ with _SortedGlob(): # the object draw depends on the seed alone, not the host's directory order
+ self.env.setup_demo(**args)
if self.enable_render: # setup_scene() switches the shader to ray tracing; put it back for cheap frames
sapien.render.set_camera_shader_dir(self.render_shader)
self._raw_scene = self.env.scene
diff --git a/tasks/robotwin_2_0_place_object_stand-standard/task.toml b/tasks/robotwin_2_0_place_object_stand-standard/task.toml
index f891c573e4..4143656da2 100644
--- a/tasks/robotwin_2_0_place_object_stand-standard/task.toml
+++ b/tasks/robotwin_2_0_place_object_stand-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "robotwin"
robot = "aloha-agilex"
family = "place_object_stand"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 60417
diff --git a/tasks/robotwin_2_0_place_phone_stand-privileged/README.md b/tasks/robotwin_2_0_place_phone_stand-privileged/README.md
index 48bd179520..6576c0feaa 100644
--- a/tasks/robotwin_2_0_place_phone_stand-privileged/README.md
+++ b/tasks/robotwin_2_0_place_phone_stand-privileged/README.md
@@ -55,3 +55,4 @@ Pick up the phone and put it on the phone stand.
- 2026-10-10 (owner decision, board 0072): the agent's network egress is restricted to the model APIs (`[agent] network_mode = "allowlist"`, the hosts of `scripts/set_agent_egress.py`; since 2026-10-11 these include the package registries, PyPI and the Debian / Ubuntu archives, so that `pip install` and `apt-get install` work); it was open before (Harbor's default), and agents could download upstream sources, reference solutions and robot models from their own shell. The scene, the frozen instance, the tooling and the pass rule are unchanged, so the task keeps its name; results from before this date were obtained with open egress.
- 2026-10-13 (board 0086): the tooling refuses a trajectory the verifier cannot grade (`environment/rcb_traj.py`, `tests/rcb_traj.py`). `Session.save_trajectory` and `Session.replay` raise `ValueError` for more than 200,000 rows (`rcb_traj.MAX_T`, the verifier's `MAX_T` in `tests/grade.py`: 800 s of physics), and `record_stop` warns on stderr. Before, the tool saved and replayed such a file ("success") and the verifier then graded it `bad_trajectory`; the instruction has stated the cap since 2026-10-10. The verifier, the scene, the instance and the pass rule are unchanged: every trajectory is graded as before.
- 2026-10-13 (board 0100): base image `eai-robotwin` 0.1.2. Its scrub of RoboTwin's expert also removes the reference poses that 18 tasks' `load_actors` computed only for the expert, and the expert's comments above the scrubbed method bodies (`images/robotwin/scrub_oracles.py`). This task's `load_actors` set no such value: its scene, instance, tooling, verifier and pass rule are unchanged.
+- 2026-10-14 (board 0113): the tooling only (`environment/rcb_traj.py`, `tests/rcb_traj.py`): while it builds the scene, `glob.glob` returns its matches sorted, so that a task which lists files to draw an object depends on its seed alone, not on the host's directory order (four RoboTwin tasks did; this one lists no files while it builds its scene). Its scene, frozen instance (same initial-state hash), verifier and pass rule are unchanged.
diff --git a/tasks/robotwin_2_0_place_phone_stand-privileged/environment/rcb_traj.py b/tasks/robotwin_2_0_place_phone_stand-privileged/environment/rcb_traj.py
index 9445b33944..0d42777fb2 100644
--- a/tasks/robotwin_2_0_place_phone_stand-privileged/environment/rcb_traj.py
+++ b/tasks/robotwin_2_0_place_phone_stand-privileged/environment/rcb_traj.py
@@ -113,6 +113,26 @@ BASE_ARGS: dict[str, Any] = {
}
+class _SortedGlob:
+ """While a scene is built, glob.glob / glob.iglob return their matches sorted. Four RoboTwin tasks (place_a2b_left,
+ place_a2b_right, place_object_scale, place_object_stand: envs/<task>.py load_actors) list an object's model files
+ with glob.glob and then draw one with the seeded np.random.choice; glob returns the file system's directory order,
+ which differs between hosts (and image layers), so one seed built a different object on another host and the frozen
+ instance's hash check failed there (board 0086 claim 5; never rebuilt off the reference workstation). Sorted, the
+ draw depends on the seed alone. No other task of ours lists files while it builds its scene."""
+
+ def __enter__(self):
+ import glob
+ self._glob, self._orig = glob, (glob.glob, glob.iglob)
+ glob.glob = lambda *a, **k: sorted(self._orig[0](*a, **k))
+ glob.iglob = lambda *a, **k: iter(sorted(self._orig[1](*a, **k)))
+ return self
+
+ def __exit__(self, *exc):
+ self._glob.glob, self._glob.iglob = self._orig
+ return False
+
+
def _chdir_root() -> None:
# RoboTwin resolves assets/config paths relative to the repo root and via ROOT_PATH
if os.getcwd() != ROBOTWIN_ROOT and os.path.isdir(ROBOTWIN_ROOT):
@@ -218,7 +238,8 @@ class Session:
if self.enable_render:
sapien.render.set_camera_shader_dir(self.render_shader)
args = _build_args(self.task, self.embodiment, self.seed)
- self.env.setup_demo(**args)
+ with _SortedGlob(): # the object draw depends on the seed alone, not the host's directory order
+ self.env.setup_demo(**args)
if self.enable_render: # setup_scene() switches the shader to ray tracing; put it back for cheap frames
sapien.render.set_camera_shader_dir(self.render_shader)
self._raw_scene = self.env.scene
diff --git a/tasks/robotwin_2_0_place_phone_stand-privileged/task.toml b/tasks/robotwin_2_0_place_phone_stand-privileged/task.toml
index 01bd95a079..9fa90bff71 100644
--- a/tasks/robotwin_2_0_place_phone_stand-privileged/task.toml
+++ b/tasks/robotwin_2_0_place_phone_stand-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "robotwin" # RoboTwin 2.0 = SAPIEN 3 / PhysX + CuRobo
robot = "aloha-agilex"
family = "place_phone_stand"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 60417 # scene = environment/instance.json (RoboTwin seed + expected initial-state hash); identity = dataset.toml digest
expert_time_estimate_hours = 0.5
diff --git a/tasks/robotwin_2_0_place_phone_stand-privileged/tests/rcb_traj.py b/tasks/robotwin_2_0_place_phone_stand-privileged/tests/rcb_traj.py
index 9445b33944..0d42777fb2 100644
--- a/tasks/robotwin_2_0_place_phone_stand-privileged/tests/rcb_traj.py
+++ b/tasks/robotwin_2_0_place_phone_stand-privileged/tests/rcb_traj.py
@@ -113,6 +113,26 @@ BASE_ARGS: dict[str, Any] = {
}
+class _SortedGlob:
+ """While a scene is built, glob.glob / glob.iglob return their matches sorted. Four RoboTwin tasks (place_a2b_left,
+ place_a2b_right, place_object_scale, place_object_stand: envs/<task>.py load_actors) list an object's model files
+ with glob.glob and then draw one with the seeded np.random.choice; glob returns the file system's directory order,
+ which differs between hosts (and image layers), so one seed built a different object on another host and the frozen
+ instance's hash check failed there (board 0086 claim 5; never rebuilt off the reference workstation). Sorted, the
+ draw depends on the seed alone. No other task of ours lists files while it builds its scene."""
+
+ def __enter__(self):
+ import glob
+ self._glob, self._orig = glob, (glob.glob, glob.iglob)
+ glob.glob = lambda *a, **k: sorted(self._orig[0](*a, **k))
+ glob.iglob = lambda *a, **k: iter(sorted(self._orig[1](*a, **k)))
+ return self
+
+ def __exit__(self, *exc):
+ self._glob.glob, self._glob.iglob = self._orig
+ return False
+
+
def _chdir_root() -> None:
# RoboTwin resolves assets/config paths relative to the repo root and via ROOT_PATH
if os.getcwd() != ROBOTWIN_ROOT and os.path.isdir(ROBOTWIN_ROOT):
@@ -218,7 +238,8 @@ class Session:
if self.enable_render:
sapien.render.set_camera_shader_dir(self.render_shader)
args = _build_args(self.task, self.embodiment, self.seed)
- self.env.setup_demo(**args)
+ with _SortedGlob(): # the object draw depends on the seed alone, not the host's directory order
+ self.env.setup_demo(**args)
if self.enable_render: # setup_scene() switches the shader to ray tracing; put it back for cheap frames
sapien.render.set_camera_shader_dir(self.render_shader)
self._raw_scene = self.env.scene
diff --git a/tasks/robotwin_2_0_place_phone_stand-standard/README.md b/tasks/robotwin_2_0_place_phone_stand-standard/README.md
index b063110038..5f99c73eaf 100644
--- a/tasks/robotwin_2_0_place_phone_stand-standard/README.md
+++ b/tasks/robotwin_2_0_place_phone_stand-standard/README.md
@@ -53,3 +53,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
- 2026-10-10 (owner decision, board 0072): the agent's network egress is restricted to the model APIs (`[agent] network_mode = "allowlist"`, the hosts of `scripts/set_agent_egress.py`; since 2026-10-11 these include the package registries, PyPI and the Debian / Ubuntu archives, so that `pip install` and `apt-get install` work); it was open before (Harbor's default), and agents could download upstream sources, reference solutions and robot models from their own shell. The scene, the frozen instance, the tooling and the pass rule are unchanged, so the task keeps its name; results from before this date were obtained with open egress.
- 2026-10-13 (board 0086): the service's copy of the tooling only (`environment/sim/rcb_traj.py`, one line of `environment/sim/server.py`). The tooling now refuses to save or replay a trajectory of more than 200,000 rows, the privileged verifier's cap; a standard episode has no row cap, so the service lifts it for its own session (`S.max_rows = None`). Nothing an agent sees and nothing the service grades changes.
- 2026-10-13 (board 0100): base image `eai-robotwin` 0.1.2, whose scrub of RoboTwin's expert also removes the reference poses that `load_actors` computed only for the expert. The standard-mode agent never reaches the env, so nothing it observes changes; the scene, the instance and the judge are unchanged.
+- 2026-10-14 (board 0113): the tooling only (`environment/sim/rcb_traj.py`): while it builds the scene, `glob.glob` returns its matches sorted, so that a task which lists files to draw an object depends on its seed alone, not on the host's directory order (four RoboTwin tasks did; this one lists no files while it builds its scene). Its scene, frozen instance (same initial-state hash), verifier and pass rule are unchanged.
diff --git a/tasks/robotwin_2_0_place_phone_stand-standard/environment/sim/rcb_traj.py b/tasks/robotwin_2_0_place_phone_stand-standard/environment/sim/rcb_traj.py
index 9445b33944..0d42777fb2 100644
--- a/tasks/robotwin_2_0_place_phone_stand-standard/environment/sim/rcb_traj.py
+++ b/tasks/robotwin_2_0_place_phone_stand-standard/environment/sim/rcb_traj.py
@@ -113,6 +113,26 @@ BASE_ARGS: dict[str, Any] = {
}
+class _SortedGlob:
+ """While a scene is built, glob.glob / glob.iglob return their matches sorted. Four RoboTwin tasks (place_a2b_left,
+ place_a2b_right, place_object_scale, place_object_stand: envs/<task>.py load_actors) list an object's model files
+ with glob.glob and then draw one with the seeded np.random.choice; glob returns the file system's directory order,
+ which differs between hosts (and image layers), so one seed built a different object on another host and the frozen
+ instance's hash check failed there (board 0086 claim 5; never rebuilt off the reference workstation). Sorted, the
+ draw depends on the seed alone. No other task of ours lists files while it builds its scene."""
+
+ def __enter__(self):
+ import glob
+ self._glob, self._orig = glob, (glob.glob, glob.iglob)
+ glob.glob = lambda *a, **k: sorted(self._orig[0](*a, **k))
+ glob.iglob = lambda *a, **k: iter(sorted(self._orig[1](*a, **k)))
+ return self
+
+ def __exit__(self, *exc):
+ self._glob.glob, self._glob.iglob = self._orig
+ return False
+
+
def _chdir_root() -> None:
# RoboTwin resolves assets/config paths relative to the repo root and via ROOT_PATH
if os.getcwd() != ROBOTWIN_ROOT and os.path.isdir(ROBOTWIN_ROOT):
@@ -218,7 +238,8 @@ class Session:
if self.enable_render:
sapien.render.set_camera_shader_dir(self.render_shader)
args = _build_args(self.task, self.embodiment, self.seed)
- self.env.setup_demo(**args)
+ with _SortedGlob(): # the object draw depends on the seed alone, not the host's directory order
+ self.env.setup_demo(**args)
if self.enable_render: # setup_scene() switches the shader to ray tracing; put it back for cheap frames
sapien.render.set_camera_shader_dir(self.render_shader)
self._raw_scene = self.env.scene
diff --git a/tasks/robotwin_2_0_place_phone_stand-standard/task.toml b/tasks/robotwin_2_0_place_phone_stand-standard/task.toml
index b39879e712..8b4eaa887e 100644
--- a/tasks/robotwin_2_0_place_phone_stand-standard/task.toml
+++ b/tasks/robotwin_2_0_place_phone_stand-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "robotwin"
robot = "aloha-agilex"
family = "place_phone_stand"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 60417
diff --git a/tasks/robotwin_2_0_place_shoe-privileged/README.md b/tasks/robotwin_2_0_place_shoe-privileged/README.md
index 5251535860..bb5182061a 100644
--- a/tasks/robotwin_2_0_place_shoe-privileged/README.md
+++ b/tasks/robotwin_2_0_place_shoe-privileged/README.md
@@ -55,3 +55,4 @@ Use one arm to grab the shoe from the table and place it on the mat.
- 2026-10-10 (owner decision, board 0072): the agent's network egress is restricted to the model APIs (`[agent] network_mode = "allowlist"`, the hosts of `scripts/set_agent_egress.py`; since 2026-10-11 these include the package registries, PyPI and the Debian / Ubuntu archives, so that `pip install` and `apt-get install` work); it was open before (Harbor's default), and agents could download upstream sources, reference solutions and robot models from their own shell. The scene, the frozen instance, the tooling and the pass rule are unchanged, so the task keeps its name; results from before this date were obtained with open egress.
- 2026-10-13 (board 0086): the tooling refuses a trajectory the verifier cannot grade (`environment/rcb_traj.py`, `tests/rcb_traj.py`). `Session.save_trajectory` and `Session.replay` raise `ValueError` for more than 200,000 rows (`rcb_traj.MAX_T`, the verifier's `MAX_T` in `tests/grade.py`: 800 s of physics), and `record_stop` warns on stderr. Before, the tool saved and replayed such a file ("success") and the verifier then graded it `bad_trajectory`; the instruction has stated the cap since 2026-10-10. The verifier, the scene, the instance and the pass rule are unchanged: every trajectory is graded as before.
- 2026-10-13 (board 0100): base image `eai-robotwin` 0.1.2. Its scrub of RoboTwin's expert also removes the reference poses that 18 tasks' `load_actors` computed only for the expert, and the expert's comments above the scrubbed method bodies (`images/robotwin/scrub_oracles.py`). This task's `load_actors` set no such value: its scene, instance, tooling, verifier and pass rule are unchanged.
+- 2026-10-14 (board 0113): the tooling only (`environment/rcb_traj.py`, `tests/rcb_traj.py`): while it builds the scene, `glob.glob` returns its matches sorted, so that a task which lists files to draw an object depends on its seed alone, not on the host's directory order (four RoboTwin tasks did; this one lists no files while it builds its scene). Its scene, frozen instance (same initial-state hash), verifier and pass rule are unchanged.
diff --git a/tasks/robotwin_2_0_place_shoe-privileged/environment/rcb_traj.py b/tasks/robotwin_2_0_place_shoe-privileged/environment/rcb_traj.py
index 5e957fbe28..3d1dac1ce9 100644
--- a/tasks/robotwin_2_0_place_shoe-privileged/environment/rcb_traj.py
+++ b/tasks/robotwin_2_0_place_shoe-privileged/environment/rcb_traj.py
@@ -113,6 +113,26 @@ BASE_ARGS: dict[str, Any] = {
}
+class _SortedGlob:
+ """While a scene is built, glob.glob / glob.iglob return their matches sorted. Four RoboTwin tasks (place_a2b_left,
+ place_a2b_right, place_object_scale, place_object_stand: envs/<task>.py load_actors) list an object's model files
+ with glob.glob and then draw one with the seeded np.random.choice; glob returns the file system's directory order,
+ which differs between hosts (and image layers), so one seed built a different object on another host and the frozen
+ instance's hash check failed there (board 0086 claim 5; never rebuilt off the reference workstation). Sorted, the
+ draw depends on the seed alone. No other task of ours lists files while it builds its scene."""
+
+ def __enter__(self):
+ import glob
+ self._glob, self._orig = glob, (glob.glob, glob.iglob)
+ glob.glob = lambda *a, **k: sorted(self._orig[0](*a, **k))
+ glob.iglob = lambda *a, **k: iter(sorted(self._orig[1](*a, **k)))
+ return self
+
+ def __exit__(self, *exc):
+ self._glob.glob, self._glob.iglob = self._orig
+ return False
+
+
def _chdir_root() -> None:
# RoboTwin resolves assets/config paths relative to the repo root and via ROOT_PATH
if os.getcwd() != ROBOTWIN_ROOT and os.path.isdir(ROBOTWIN_ROOT):
@@ -218,7 +238,8 @@ class Session:
if self.enable_render:
sapien.render.set_camera_shader_dir(self.render_shader)
args = _build_args(self.task, self.embodiment, self.seed)
- self.env.setup_demo(**args)
+ with _SortedGlob(): # the object draw depends on the seed alone, not the host's directory order
+ self.env.setup_demo(**args)
if self.enable_render: # setup_scene() switches the shader to ray tracing; put it back for cheap frames
sapien.render.set_camera_shader_dir(self.render_shader)
self._raw_scene = self.env.scene
diff --git a/tasks/robotwin_2_0_place_shoe-privileged/task.toml b/tasks/robotwin_2_0_place_shoe-privileged/task.toml
index 447ec51e38..6d2ddbde15 100644
--- a/tasks/robotwin_2_0_place_shoe-privileged/task.toml
+++ b/tasks/robotwin_2_0_place_shoe-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "robotwin" # RoboTwin 2.0 = SAPIEN 3 / PhysX + CuRobo
robot = "aloha-agilex"
family = "place_shoe"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 28831 # scene = environment/instance.json (RoboTwin seed + expected initial-state hash); identity = dataset.toml digest
expert_time_estimate_hours = 0.5
diff --git a/tasks/robotwin_2_0_place_shoe-privileged/tests/rcb_traj.py b/tasks/robotwin_2_0_place_shoe-privileged/tests/rcb_traj.py
index 5e957fbe28..3d1dac1ce9 100644
--- a/tasks/robotwin_2_0_place_shoe-privileged/tests/rcb_traj.py
+++ b/tasks/robotwin_2_0_place_shoe-privileged/tests/rcb_traj.py
@@ -113,6 +113,26 @@ BASE_ARGS: dict[str, Any] = {
}
+class _SortedGlob:
+ """While a scene is built, glob.glob / glob.iglob return their matches sorted. Four RoboTwin tasks (place_a2b_left,
+ place_a2b_right, place_object_scale, place_object_stand: envs/<task>.py load_actors) list an object's model files
+ with glob.glob and then draw one with the seeded np.random.choice; glob returns the file system's directory order,
+ which differs between hosts (and image layers), so one seed built a different object on another host and the frozen
+ instance's hash check failed there (board 0086 claim 5; never rebuilt off the reference workstation). Sorted, the
+ draw depends on the seed alone. No other task of ours lists files while it builds its scene."""
+
+ def __enter__(self):
+ import glob
+ self._glob, self._orig = glob, (glob.glob, glob.iglob)
+ glob.glob = lambda *a, **k: sorted(self._orig[0](*a, **k))
+ glob.iglob = lambda *a, **k: iter(sorted(self._orig[1](*a, **k)))
+ return self
+
+ def __exit__(self, *exc):
+ self._glob.glob, self._glob.iglob = self._orig
+ return False
+
+
def _chdir_root() -> None:
# RoboTwin resolves assets/config paths relative to the repo root and via ROOT_PATH
if os.getcwd() != ROBOTWIN_ROOT and os.path.isdir(ROBOTWIN_ROOT):
@@ -218,7 +238,8 @@ class Session:
if self.enable_render:
sapien.render.set_camera_shader_dir(self.render_shader)
args = _build_args(self.task, self.embodiment, self.seed)
- self.env.setup_demo(**args)
+ with _SortedGlob(): # the object draw depends on the seed alone, not the host's directory order
+ self.env.setup_demo(**args)
if self.enable_render: # setup_scene() switches the shader to ray tracing; put it back for cheap frames
sapien.render.set_camera_shader_dir(self.render_shader)
self._raw_scene = self.env.scene
diff --git a/tasks/robotwin_2_0_place_shoe-standard/README.md b/tasks/robotwin_2_0_place_shoe-standard/README.md
index ff27fbc528..38040adecb 100644
--- a/tasks/robotwin_2_0_place_shoe-standard/README.md
+++ b/tasks/robotwin_2_0_place_shoe-standard/README.md
@@ -53,3 +53,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
- 2026-10-10 (owner decision, board 0072): the agent's network egress is restricted to the model APIs (`[agent] network_mode = "allowlist"`, the hosts of `scripts/set_agent_egress.py`; since 2026-10-11 these include the package registries, PyPI and the Debian / Ubuntu archives, so that `pip install` and `apt-get install` work); it was open before (Harbor's default), and agents could download upstream sources, reference solutions and robot models from their own shell. The scene, the frozen instance, the tooling and the pass rule are unchanged, so the task keeps its name; results from before this date were obtained with open egress.
- 2026-10-13 (board 0086): the service's copy of the tooling only (`environment/sim/rcb_traj.py`, one line of `environment/sim/server.py`). The tooling now refuses to save or replay a trajectory of more than 200,000 rows, the privileged verifier's cap; a standard episode has no row cap, so the service lifts it for its own session (`S.max_rows = None`). Nothing an agent sees and nothing the service grades changes.
- 2026-10-13 (board 0100): base image `eai-robotwin` 0.1.2, whose scrub of RoboTwin's expert also removes the reference poses that `load_actors` computed only for the expert. The standard-mode agent never reaches the env, so nothing it observes changes; the scene, the instance and the judge are unchanged.
+- 2026-10-14 (board 0113): the tooling only (`environment/sim/rcb_traj.py`): while it builds the scene, `glob.glob` returns its matches sorted, so that a task which lists files to draw an object depends on its seed alone, not on the host's directory order (four RoboTwin tasks did; this one lists no files while it builds its scene). Its scene, frozen instance (same initial-state hash), verifier and pass rule are unchanged.
diff --git a/tasks/robotwin_2_0_place_shoe-standard/environment/sim/rcb_traj.py b/tasks/robotwin_2_0_place_shoe-standard/environment/sim/rcb_traj.py
index 5e957fbe28..3d1dac1ce9 100644
--- a/tasks/robotwin_2_0_place_shoe-standard/environment/sim/rcb_traj.py
+++ b/tasks/robotwin_2_0_place_shoe-standard/environment/sim/rcb_traj.py
@@ -113,6 +113,26 @@ BASE_ARGS: dict[str, Any] = {
}
+class _SortedGlob:
+ """While a scene is built, glob.glob / glob.iglob return their matches sorted. Four RoboTwin tasks (place_a2b_left,
+ place_a2b_right, place_object_scale, place_object_stand: envs/<task>.py load_actors) list an object's model files
+ with glob.glob and then draw one with the seeded np.random.choice; glob returns the file system's directory order,
+ which differs between hosts (and image layers), so one seed built a different object on another host and the frozen
+ instance's hash check failed there (board 0086 claim 5; never rebuilt off the reference workstation). Sorted, the
+ draw depends on the seed alone. No other task of ours lists files while it builds its scene."""
+
+ def __enter__(self):
+ import glob
+ self._glob, self._orig = glob, (glob.glob, glob.iglob)
+ glob.glob = lambda *a, **k: sorted(self._orig[0](*a, **k))
+ glob.iglob = lambda *a, **k: iter(sorted(self._orig[1](*a, **k)))
+ return self
+
+ def __exit__(self, *exc):
+ self._glob.glob, self._glob.iglob = self._orig
+ return False
+
+
def _chdir_root() -> None:
# RoboTwin resolves assets/config paths relative to the repo root and via ROOT_PATH
if os.getcwd() != ROBOTWIN_ROOT and os.path.isdir(ROBOTWIN_ROOT):
@@ -218,7 +238,8 @@ class Session:
if self.enable_render:
sapien.render.set_camera_shader_dir(self.render_shader)
args = _build_args(self.task, self.embodiment, self.seed)
- self.env.setup_demo(**args)
+ with _SortedGlob(): # the object draw depends on the seed alone, not the host's directory order
+ self.env.setup_demo(**args)
if self.enable_render: # setup_scene() switches the shader to ray tracing; put it back for cheap frames
sapien.render.set_camera_shader_dir(self.render_shader)
self._raw_scene = self.env.scene
diff --git a/tasks/robotwin_2_0_place_shoe-standard/task.toml b/tasks/robotwin_2_0_place_shoe-standard/task.toml
index ad4a8319a7..858bcd5763 100644
--- a/tasks/robotwin_2_0_place_shoe-standard/task.toml
+++ b/tasks/robotwin_2_0_place_shoe-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "robotwin"
robot = "aloha-agilex"
family = "place_shoe"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 28831
diff --git a/tasks/robotwin_2_0_press_stapler-privileged/README.md b/tasks/robotwin_2_0_press_stapler-privileged/README.md
index a13b717dfb..302dbd482b 100644
--- a/tasks/robotwin_2_0_press_stapler-privileged/README.md
+++ b/tasks/robotwin_2_0_press_stapler-privileged/README.md
@@ -55,3 +55,4 @@ Use one arm to press the stapler.
- 2026-10-10 (owner decision, board 0072): the agent's network egress is restricted to the model APIs (`[agent] network_mode = "allowlist"`, the hosts of `scripts/set_agent_egress.py`; since 2026-10-11 these include the package registries, PyPI and the Debian / Ubuntu archives, so that `pip install` and `apt-get install` work); it was open before (Harbor's default), and agents could download upstream sources, reference solutions and robot models from their own shell. The scene, the frozen instance, the tooling and the pass rule are unchanged, so the task keeps its name; results from before this date were obtained with open egress.
- 2026-10-13 (board 0086): the tooling refuses a trajectory the verifier cannot grade (`environment/rcb_traj.py`, `tests/rcb_traj.py`). `Session.save_trajectory` and `Session.replay` raise `ValueError` for more than 200,000 rows (`rcb_traj.MAX_T`, the verifier's `MAX_T` in `tests/grade.py`: 800 s of physics), and `record_stop` warns on stderr. Before, the tool saved and replayed such a file ("success") and the verifier then graded it `bad_trajectory`; the instruction has stated the cap since 2026-10-10. The verifier, the scene, the instance and the pass rule are unchanged: every trajectory is graded as before.
- 2026-10-13 (board 0100): base image `eai-robotwin` 0.1.2. Its scrub of RoboTwin's expert also removes the reference poses that 18 tasks' `load_actors` computed only for the expert, and the expert's comments above the scrubbed method bodies (`images/robotwin/scrub_oracles.py`). This task's `load_actors` set no such value: its scene, instance, tooling, verifier and pass rule are unchanged.
+- 2026-10-14 (board 0113): the tooling only (`environment/rcb_traj.py`, `tests/rcb_traj.py`): while it builds the scene, `glob.glob` returns its matches sorted, so that a task which lists files to draw an object depends on its seed alone, not on the host's directory order (four RoboTwin tasks did; this one lists no files while it builds its scene). Its scene, frozen instance (same initial-state hash), verifier and pass rule are unchanged.
diff --git a/tasks/robotwin_2_0_press_stapler-privileged/environment/rcb_traj.py b/tasks/robotwin_2_0_press_stapler-privileged/environment/rcb_traj.py
index 99e04ee610..4b1a24de88 100644
--- a/tasks/robotwin_2_0_press_stapler-privileged/environment/rcb_traj.py
+++ b/tasks/robotwin_2_0_press_stapler-privileged/environment/rcb_traj.py
@@ -113,6 +113,26 @@ BASE_ARGS: dict[str, Any] = {
}
+class _SortedGlob:
+ """While a scene is built, glob.glob / glob.iglob return their matches sorted. Four RoboTwin tasks (place_a2b_left,
+ place_a2b_right, place_object_scale, place_object_stand: envs/<task>.py load_actors) list an object's model files
+ with glob.glob and then draw one with the seeded np.random.choice; glob returns the file system's directory order,
+ which differs between hosts (and image layers), so one seed built a different object on another host and the frozen
+ instance's hash check failed there (board 0086 claim 5; never rebuilt off the reference workstation). Sorted, the
+ draw depends on the seed alone. No other task of ours lists files while it builds its scene."""
+
+ def __enter__(self):
+ import glob
+ self._glob, self._orig = glob, (glob.glob, glob.iglob)
+ glob.glob = lambda *a, **k: sorted(self._orig[0](*a, **k))
+ glob.iglob = lambda *a, **k: iter(sorted(self._orig[1](*a, **k)))
+ return self
+
+ def __exit__(self, *exc):
+ self._glob.glob, self._glob.iglob = self._orig
+ return False
+
+
def _chdir_root() -> None:
# RoboTwin resolves assets/config paths relative to the repo root and via ROOT_PATH
if os.getcwd() != ROBOTWIN_ROOT and os.path.isdir(ROBOTWIN_ROOT):
@@ -218,7 +238,8 @@ class Session:
if self.enable_render:
sapien.render.set_camera_shader_dir(self.render_shader)
args = _build_args(self.task, self.embodiment, self.seed)
- self.env.setup_demo(**args)
+ with _SortedGlob(): # the object draw depends on the seed alone, not the host's directory order
+ self.env.setup_demo(**args)
if self.enable_render: # setup_scene() switches the shader to ray tracing; put it back for cheap frames
sapien.render.set_camera_shader_dir(self.render_shader)
self._raw_scene = self.env.scene
diff --git a/tasks/robotwin_2_0_press_stapler-privileged/task.toml b/tasks/robotwin_2_0_press_stapler-privileged/task.toml
index 9009f42267..eac31e6f6e 100644
--- a/tasks/robotwin_2_0_press_stapler-privileged/task.toml
+++ b/tasks/robotwin_2_0_press_stapler-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "robotwin" # RoboTwin 2.0 = SAPIEN 3 / PhysX + CuRobo
robot = "aloha-agilex"
family = "press_stapler"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 60417 # scene = environment/instance.json (RoboTwin seed + expected initial-state hash); identity = dataset.toml digest
expert_time_estimate_hours = 0.5
diff --git a/tasks/robotwin_2_0_press_stapler-privileged/tests/rcb_traj.py b/tasks/robotwin_2_0_press_stapler-privileged/tests/rcb_traj.py
index 99e04ee610..4b1a24de88 100644
--- a/tasks/robotwin_2_0_press_stapler-privileged/tests/rcb_traj.py
+++ b/tasks/robotwin_2_0_press_stapler-privileged/tests/rcb_traj.py
@@ -113,6 +113,26 @@ BASE_ARGS: dict[str, Any] = {
}
+class _SortedGlob:
+ """While a scene is built, glob.glob / glob.iglob return their matches sorted. Four RoboTwin tasks (place_a2b_left,
+ place_a2b_right, place_object_scale, place_object_stand: envs/<task>.py load_actors) list an object's model files
+ with glob.glob and then draw one with the seeded np.random.choice; glob returns the file system's directory order,
+ which differs between hosts (and image layers), so one seed built a different object on another host and the frozen
+ instance's hash check failed there (board 0086 claim 5; never rebuilt off the reference workstation). Sorted, the
+ draw depends on the seed alone. No other task of ours lists files while it builds its scene."""
+
+ def __enter__(self):
+ import glob
+ self._glob, self._orig = glob, (glob.glob, glob.iglob)
+ glob.glob = lambda *a, **k: sorted(self._orig[0](*a, **k))
+ glob.iglob = lambda *a, **k: iter(sorted(self._orig[1](*a, **k)))
+ return self
+
+ def __exit__(self, *exc):
+ self._glob.glob, self._glob.iglob = self._orig
+ return False
+
+
def _chdir_root() -> None:
# RoboTwin resolves assets/config paths relative to the repo root and via ROOT_PATH
if os.getcwd() != ROBOTWIN_ROOT and os.path.isdir(ROBOTWIN_ROOT):
@@ -218,7 +238,8 @@ class Session:
if self.enable_render:
sapien.render.set_camera_shader_dir(self.render_shader)
args = _build_args(self.task, self.embodiment, self.seed)
- self.env.setup_demo(**args)
+ with _SortedGlob(): # the object draw depends on the seed alone, not the host's directory order
+ self.env.setup_demo(**args)
if self.enable_render: # setup_scene() switches the shader to ray tracing; put it back for cheap frames
sapien.render.set_camera_shader_dir(self.render_shader)
self._raw_scene = self.env.scene
diff --git a/tasks/robotwin_2_0_press_stapler-standard/README.md b/tasks/robotwin_2_0_press_stapler-standard/README.md
index abafa5accf..1be0fcd74d 100644
--- a/tasks/robotwin_2_0_press_stapler-standard/README.md
+++ b/tasks/robotwin_2_0_press_stapler-standard/README.md
@@ -53,3 +53,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
- 2026-10-10 (owner decision, board 0072): the agent's network egress is restricted to the model APIs (`[agent] network_mode = "allowlist"`, the hosts of `scripts/set_agent_egress.py`; since 2026-10-11 these include the package registries, PyPI and the Debian / Ubuntu archives, so that `pip install` and `apt-get install` work); it was open before (Harbor's default), and agents could download upstream sources, reference solutions and robot models from their own shell. The scene, the frozen instance, the tooling and the pass rule are unchanged, so the task keeps its name; results from before this date were obtained with open egress.
- 2026-10-13 (board 0086): the service's copy of the tooling only (`environment/sim/rcb_traj.py`, one line of `environment/sim/server.py`). The tooling now refuses to save or replay a trajectory of more than 200,000 rows, the privileged verifier's cap; a standard episode has no row cap, so the service lifts it for its own session (`S.max_rows = None`). Nothing an agent sees and nothing the service grades changes.
- 2026-10-13 (board 0100): base image `eai-robotwin` 0.1.2, whose scrub of RoboTwin's expert also removes the reference poses that `load_actors` computed only for the expert. The standard-mode agent never reaches the env, so nothing it observes changes; the scene, the instance and the judge are unchanged.
+- 2026-10-14 (board 0113): the tooling only (`environment/sim/rcb_traj.py`): while it builds the scene, `glob.glob` returns its matches sorted, so that a task which lists files to draw an object depends on its seed alone, not on the host's directory order (four RoboTwin tasks did; this one lists no files while it builds its scene). Its scene, frozen instance (same initial-state hash), verifier and pass rule are unchanged.
diff --git a/tasks/robotwin_2_0_press_stapler-standard/environment/sim/rcb_traj.py b/tasks/robotwin_2_0_press_stapler-standard/environment/sim/rcb_traj.py
index 99e04ee610..4b1a24de88 100644
--- a/tasks/robotwin_2_0_press_stapler-standard/environment/sim/rcb_traj.py
+++ b/tasks/robotwin_2_0_press_stapler-standard/environment/sim/rcb_traj.py
@@ -113,6 +113,26 @@ BASE_ARGS: dict[str, Any] = {
}
+class _SortedGlob:
+ """While a scene is built, glob.glob / glob.iglob return their matches sorted. Four RoboTwin tasks (place_a2b_left,
+ place_a2b_right, place_object_scale, place_object_stand: envs/<task>.py load_actors) list an object's model files
+ with glob.glob and then draw one with the seeded np.random.choice; glob returns the file system's directory order,
+ which differs between hosts (and image layers), so one seed built a different object on another host and the frozen
+ instance's hash check failed there (board 0086 claim 5; never rebuilt off the reference workstation). Sorted, the
+ draw depends on the seed alone. No other task of ours lists files while it builds its scene."""
+
+ def __enter__(self):
+ import glob
+ self._glob, self._orig = glob, (glob.glob, glob.iglob)
+ glob.glob = lambda *a, **k: sorted(self._orig[0](*a, **k))
+ glob.iglob = lambda *a, **k: iter(sorted(self._orig[1](*a, **k)))
+ return self
+
+ def __exit__(self, *exc):
+ self._glob.glob, self._glob.iglob = self._orig
+ return False
+
+
def _chdir_root() -> None:
# RoboTwin resolves assets/config paths relative to the repo root and via ROOT_PATH
if os.getcwd() != ROBOTWIN_ROOT and os.path.isdir(ROBOTWIN_ROOT):
@@ -218,7 +238,8 @@ class Session:
if self.enable_render:
sapien.render.set_camera_shader_dir(self.render_shader)
args = _build_args(self.task, self.embodiment, self.seed)
- self.env.setup_demo(**args)
+ with _SortedGlob(): # the object draw depends on the seed alone, not the host's directory order
+ self.env.setup_demo(**args)
if self.enable_render: # setup_scene() switches the shader to ray tracing; put it back for cheap frames
sapien.render.set_camera_shader_dir(self.render_shader)
self._raw_scene = self.env.scene
diff --git a/tasks/robotwin_2_0_press_stapler-standard/task.toml b/tasks/robotwin_2_0_press_stapler-standard/task.toml
index bead45df24..5c665c59b1 100644
--- a/tasks/robotwin_2_0_press_stapler-standard/task.toml
+++ b/tasks/robotwin_2_0_press_stapler-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "robotwin"
robot = "aloha-agilex"
family = "press_stapler"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 60417
diff --git a/tasks/robotwin_2_0_put_bottles_dustbin-privileged/README.md b/tasks/robotwin_2_0_put_bottles_dustbin-privileged/README.md
index c0717722f0..fe12c443bd 100644
--- a/tasks/robotwin_2_0_put_bottles_dustbin-privileged/README.md
+++ b/tasks/robotwin_2_0_put_bottles_dustbin-privileged/README.md
@@ -56,3 +56,4 @@ Put all three bottles into the dustbin that stands on the floor at the robot's l
- 2026-10-13 (board 0086): the tooling refuses a trajectory the verifier cannot grade (`environment/rcb_traj.py`, `tests/rcb_traj.py`). `Session.save_trajectory` and `Session.replay` raise `ValueError` for more than 200,000 rows (`rcb_traj.MAX_T`, the verifier's `MAX_T` in `tests/grade.py`: 800 s of physics), and `record_stop` warns on stderr. Before, the tool saved and replayed such a file ("success") and the verifier then graded it `bad_trajectory`; the instruction has stated the cap since 2026-10-10. The verifier, the scene, the instance and the pass rule are unchanged: every trajectory is graded as before.
; `solution/` rebuilds the scene with upstream's `load_actors` to get them back for the expert. None of them is read by `check_success`. The scene, the frozen instance (same initial-state hash), the tooling, the verifier and the pass rule are unchanged, so the task keeps its name; privileged results from before this date were obtained with these values reachable.
- 2026-10-13 (board 0100, owner decision): base image `eai-robotwin` 0.1.2. The image's scrub of RoboTwin's expert kept `load_actors` whole, so values it computes only for the removed expert stayed reachable: `env.right_middle_pose` (poses the expert moves to or places at), and the instruction listed them among the scene objects. They are gone: the env no longer has them, `/opt/RoboTwin/envs/put_bottles_dustbin.py` no longer shows them, and the scene-object list no longer names them; `solution/` rebuilds the scene with upstream's `load_actors` to get them back for the expert. None of them is read by `check_success`. The scene, the frozen instance (same initial-state hash), the tooling, the verifier and the pass rule are unchanged, so the task keeps its name; privileged results from before this date were obtained with these values reachable.
+- 2026-10-14 (board 0113): the tooling only (`environment/rcb_traj.py`, `tests/rcb_traj.py`): while it builds the scene, `glob.glob` returns its matches sorted, so that a task which lists files to draw an object depends on its seed alone, not on the host's directory order (four RoboTwin tasks did; this one lists no files while it builds its scene). Its scene, frozen instance (same initial-state hash), verifier and pass rule are unchanged.
diff --git a/tasks/robotwin_2_0_put_bottles_dustbin-privileged/environment/rcb_traj.py b/tasks/robotwin_2_0_put_bottles_dustbin-privileged/environment/rcb_traj.py
index c45f47ffae..29731c5d19 100644
--- a/tasks/robotwin_2_0_put_bottles_dustbin-privileged/environment/rcb_traj.py
+++ b/tasks/robotwin_2_0_put_bottles_dustbin-privileged/environment/rcb_traj.py
@@ -113,6 +113,26 @@ BASE_ARGS: dict[str, Any] = {
}
+class _SortedGlob:
+ """While a scene is built, glob.glob / glob.iglob return their matches sorted. Four RoboTwin tasks (place_a2b_left,
+ place_a2b_right, place_object_scale, place_object_stand: envs/<task>.py load_actors) list an object's model files
+ with glob.glob and then draw one with the seeded np.random.choice; glob returns the file system's directory order,
+ which differs between hosts (and image layers), so one seed built a different object on another host and the frozen
+ instance's hash check failed there (board 0086 claim 5; never rebuilt off the reference workstation). Sorted, the
+ draw depends on the seed alone. No other task of ours lists files while it builds its scene."""
+
+ def __enter__(self):
+ import glob
+ self._glob, self._orig = glob, (glob.glob, glob.iglob)
+ glob.glob = lambda *a, **k: sorted(self._orig[0](*a, **k))
+ glob.iglob = lambda *a, **k: iter(sorted(self._orig[1](*a, **k)))
+ return self
+
+ def __exit__(self, *exc):
+ self._glob.glob, self._glob.iglob = self._orig
+ return False
+
+
def _chdir_root() -> None:
# RoboTwin resolves assets/config paths relative to the repo root and via ROOT_PATH
if os.getcwd() != ROBOTWIN_ROOT and os.path.isdir(ROBOTWIN_ROOT):
@@ -218,7 +238,8 @@ class Session:
if self.enable_render:
sapien.render.set_camera_shader_dir(self.render_shader)
args = _build_args(self.task, self.embodiment, self.seed)
- self.env.setup_demo(**args)
+ with _SortedGlob(): # the object draw depends on the seed alone, not the host's directory order
+ self.env.setup_demo(**args)
if self.enable_render: # setup_scene() switches the shader to ray tracing; put it back for cheap frames
sapien.render.set_camera_shader_dir(self.render_shader)
self._raw_scene = self.env.scene
diff --git a/tasks/robotwin_2_0_put_bottles_dustbin-privileged/task.toml b/tasks/robotwin_2_0_put_bottles_dustbin-privileged/task.toml
index e839b72d28..5a7d292897 100644
--- a/tasks/robotwin_2_0_put_bottles_dustbin-privileged/task.toml
+++ b/tasks/robotwin_2_0_put_bottles_dustbin-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "robotwin" # RoboTwin 2.0 = SAPIEN 3 / PhysX + CuRobo
robot = "aloha-agilex"
family = "put_bottles_dustbin"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 28831 # scene = environment/instance.json (RoboTwin seed + expected initial-state hash); identity = dataset.toml digest
expert_time_estimate_hours = 0.5
diff --git a/tasks/robotwin_2_0_put_bottles_dustbin-privileged/tests/rcb_traj.py b/tasks/robotwin_2_0_put_bottles_dustbin-privileged/tests/rcb_traj.py
index c45f47ffae..29731c5d19 100644
--- a/tasks/robotwin_2_0_put_bottles_dustbin-privileged/tests/rcb_traj.py
+++ b/tasks/robotwin_2_0_put_bottles_dustbin-privileged/tests/rcb_traj.py
@@ -113,6 +113,26 @@ BASE_ARGS: dict[str, Any] = {
}
+class _SortedGlob:
+ """While a scene is built, glob.glob / glob.iglob return their matches sorted. Four RoboTwin tasks (place_a2b_left,
+ place_a2b_right, place_object_scale, place_object_stand: envs/<task>.py load_actors) list an object's model files
+ with glob.glob and then draw one with the seeded np.random.choice; glob returns the file system's directory order,
+ which differs between hosts (and image layers), so one seed built a different object on another host and the frozen
+ instance's hash check failed there (board 0086 claim 5; never rebuilt off the reference workstation). Sorted, the
+ draw depends on the seed alone. No other task of ours lists files while it builds its scene."""
+
+ def __enter__(self):
+ import glob
+ self._glob, self._orig = glob, (glob.glob, glob.iglob)
+ glob.glob = lambda *a, **k: sorted(self._orig[0](*a, **k))
+ glob.iglob = lambda *a, **k: iter(sorted(self._orig[1](*a, **k)))
+ return self
+
+ def __exit__(self, *exc):
+ self._glob.glob, self._glob.iglob = self._orig
+ return False
+
+
def _chdir_root() -> None:
# RoboTwin resolves assets/config paths relative to the repo root and via ROOT_PATH
if os.getcwd() != ROBOTWIN_ROOT and os.path.isdir(ROBOTWIN_ROOT):
@@ -218,7 +238,8 @@ class Session:
if self.enable_render:
sapien.render.set_camera_shader_dir(self.render_shader)
args = _build_args(self.task, self.embodiment, self.seed)
- self.env.setup_demo(**args)
+ with _SortedGlob(): # the object draw depends on the seed alone, not the host's directory order
+ self.env.setup_demo(**args)
if self.enable_render: # setup_scene() switches the shader to ray tracing; put it back for cheap frames
sapien.render.set_camera_shader_dir(self.render_shader)
self._raw_scene = self.env.scene
diff --git a/tasks/robotwin_2_0_put_bottles_dustbin-standard/README.md b/tasks/robotwin_2_0_put_bottles_dustbin-standard/README.md
index 8630bb50a5..b00b729a7f 100644
--- a/tasks/robotwin_2_0_put_bottles_dustbin-standard/README.md
+++ b/tasks/robotwin_2_0_put_bottles_dustbin-standard/README.md
@@ -52,3 +52,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
- 2026-10-10 (owner decision, board 0072): the agent's network egress is restricted to the model APIs (`[agent] network_mode = "allowlist"`, the hosts of `scripts/set_agent_egress.py`; since 2026-10-11 these include the package registries, PyPI and the Debian / Ubuntu archives, so that `pip install` and `apt-get install` work); it was open before (Harbor's default), and agents could download upstream sources, reference solutions and robot models from their own shell. The scene, the frozen instance, the tooling and the pass rule are unchanged, so the task keeps its name; results from before this date were obtained with open egress.
- 2026-10-13 (board 0086): the service's copy of the tooling only (`environment/sim/rcb_traj.py`, one line of `environment/sim/server.py`). The tooling now refuses to save or replay a trajectory of more than 200,000 rows, the privileged verifier's cap; a standard episode has no row cap, so the service lifts it for its own session (`S.max_rows = None`). Nothing an agent sees and nothing the service grades changes.
- 2026-10-13 (board 0100): base image `eai-robotwin` 0.1.2, whose scrub of RoboTwin's expert also removes the reference poses that `load_actors` computed only for the expert (here `right_middle_pose`). The standard-mode agent never reaches the env, so nothing it observes changes; the scene, the instance and the judge are unchanged.
+- 2026-10-14 (board 0113): the tooling only (`environment/sim/rcb_traj.py`): while it builds the scene, `glob.glob` returns its matches sorted, so that a task which lists files to draw an object depends on its seed alone, not on the host's directory order (four RoboTwin tasks did; this one lists no files while it builds its scene). Its scene, frozen instance (same initial-state hash), verifier and pass rule are unchanged.
diff --git a/tasks/robotwin_2_0_put_bottles_dustbin-standard/environment/sim/rcb_traj.py b/tasks/robotwin_2_0_put_bottles_dustbin-standard/environment/sim/rcb_traj.py
index c45f47ffae..29731c5d19 100644
--- a/tasks/robotwin_2_0_put_bottles_dustbin-standard/environment/sim/rcb_traj.py
+++ b/tasks/robotwin_2_0_put_bottles_dustbin-standard/environment/sim/rcb_traj.py
@@ -113,6 +113,26 @@ BASE_ARGS: dict[str, Any] = {
}
+class _SortedGlob:
+ """While a scene is built, glob.glob / glob.iglob return their matches sorted. Four RoboTwin tasks (place_a2b_left,
+ place_a2b_right, place_object_scale, place_object_stand: envs/<task>.py load_actors) list an object's model files
+ with glob.glob and then draw one with the seeded np.random.choice; glob returns the file system's directory order,
+ which differs between hosts (and image layers), so one seed built a different object on another host and the frozen
+ instance's hash check failed there (board 0086 claim 5; never rebuilt off the reference workstation). Sorted, the
+ draw depends on the seed alone. No other task of ours lists files while it builds its scene."""
+
+ def __enter__(self):
+ import glob
+ self._glob, self._orig = glob, (glob.glob, glob.iglob)
+ glob.glob = lambda *a, **k: sorted(self._orig[0](*a, **k))
+ glob.iglob = lambda *a, **k: iter(sorted(self._orig[1](*a, **k)))
+ return self
+
+ def __exit__(self, *exc):
+ self._glob.glob, self._glob.iglob = self._orig
+ return False
+
+
def _chdir_root() -> None:
# RoboTwin resolves assets/config paths relative to the repo root and via ROOT_PATH
if os.getcwd() != ROBOTWIN_ROOT and os.path.isdir(ROBOTWIN_ROOT):
@@ -218,7 +238,8 @@ class Session:
if self.enable_render:
sapien.render.set_camera_shader_dir(self.render_shader)
args = _build_args(self.task, self.embodiment, self.seed)
- self.env.setup_demo(**args)
+ with _SortedGlob(): # the object draw depends on the seed alone, not the host's directory order
+ self.env.setup_demo(**args)
if self.enable_render: # setup_scene() switches the shader to ray tracing; put it back for cheap frames
sapien.render.set_camera_shader_dir(self.render_shader)
self._raw_scene = self.env.scene
diff --git a/tasks/robotwin_2_0_put_bottles_dustbin-standard/task.toml b/tasks/robotwin_2_0_put_bottles_dustbin-standard/task.toml
index a187d171f8..0d714b8b3e 100644
--- a/tasks/robotwin_2_0_put_bottles_dustbin-standard/task.toml
+++ b/tasks/robotwin_2_0_put_bottles_dustbin-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "robotwin"
robot = "aloha-agilex"
family = "put_bottles_dustbin"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 28831
diff --git a/tasks/robotwin_2_0_rotate_qrcode-privileged/README.md b/tasks/robotwin_2_0_rotate_qrcode-privileged/README.md
index f2ee70dee2..fe3798a8f9 100644
--- a/tasks/robotwin_2_0_rotate_qrcode-privileged/README.md
+++ b/tasks/robotwin_2_0_rotate_qrcode-privileged/README.md
@@ -56,3 +56,4 @@ Use arm to catch the qrcode board on the table, pick it up and rotate to let the
- 2026-10-13 (board 0086): the tooling refuses a trajectory the verifier cannot grade (`environment/rcb_traj.py`, `tests/rcb_traj.py`). `Session.save_trajectory` and `Session.replay` raise `ValueError` for more than 200,000 rows (`rcb_traj.MAX_T`, the verifier's `MAX_T` in `tests/grade.py`: 800 s of physics), and `record_stop` warns on stderr. Before, the tool saved and replayed such a file ("success") and the verifier then graded it `bad_trajectory`; the instruction has stated the cap since 2026-10-10. The verifier, the scene, the instance and the pass rule are unchanged: every trajectory is graded as before.
; `solution/` rebuilds the scene with upstream's `load_actors` to get them back for the expert. None of them is read by `check_success`. The scene, the frozen instance (same initial-state hash), the tooling, the verifier and the pass rule are unchanged, so the task keeps its name; privileged results from before this date were obtained with these values reachable.
- 2026-10-13 (board 0100, owner decision): base image `eai-robotwin` 0.1.2. The image's scrub of RoboTwin's expert kept `load_actors` whole, so values it computes only for the removed expert stayed reachable: `env.target_pose` (poses the expert moves to or places at), and the instruction listed them among the scene objects. They are gone: the env no longer has them, `/opt/RoboTwin/envs/rotate_qrcode.py` no longer shows them, and the scene-object list no longer names them; `solution/` rebuilds the scene with upstream's `load_actors` to get them back for the expert. None of them is read by `check_success`. The scene, the frozen instance (same initial-state hash), the tooling, the verifier and the pass rule are unchanged, so the task keeps its name; privileged results from before this date were obtained with these values reachable.
+- 2026-10-14 (board 0113): the tooling only (`environment/rcb_traj.py`, `tests/rcb_traj.py`): while it builds the scene, `glob.glob` returns its matches sorted, so that a task which lists files to draw an object depends on its seed alone, not on the host's directory order (four RoboTwin tasks did; this one lists no files while it builds its scene). Its scene, frozen instance (same initial-state hash), verifier and pass rule are unchanged.
diff --git a/tasks/robotwin_2_0_rotate_qrcode-privileged/environment/rcb_traj.py b/tasks/robotwin_2_0_rotate_qrcode-privileged/environment/rcb_traj.py
index 940cef8804..e930da8932 100644
--- a/tasks/robotwin_2_0_rotate_qrcode-privileged/environment/rcb_traj.py
+++ b/tasks/robotwin_2_0_rotate_qrcode-privileged/environment/rcb_traj.py
@@ -113,6 +113,26 @@ BASE_ARGS: dict[str, Any] = {
}
+class _SortedGlob:
+ """While a scene is built, glob.glob / glob.iglob return their matches sorted. Four RoboTwin tasks (place_a2b_left,
+ place_a2b_right, place_object_scale, place_object_stand: envs/<task>.py load_actors) list an object's model files
+ with glob.glob and then draw one with the seeded np.random.choice; glob returns the file system's directory order,
+ which differs between hosts (and image layers), so one seed built a different object on another host and the frozen
+ instance's hash check failed there (board 0086 claim 5; never rebuilt off the reference workstation). Sorted, the
+ draw depends on the seed alone. No other task of ours lists files while it builds its scene."""
+
+ def __enter__(self):
+ import glob
+ self._glob, self._orig = glob, (glob.glob, glob.iglob)
+ glob.glob = lambda *a, **k: sorted(self._orig[0](*a, **k))
+ glob.iglob = lambda *a, **k: iter(sorted(self._orig[1](*a, **k)))
+ return self
+
+ def __exit__(self, *exc):
+ self._glob.glob, self._glob.iglob = self._orig
+ return False
+
+
def _chdir_root() -> None:
# RoboTwin resolves assets/config paths relative to the repo root and via ROOT_PATH
if os.getcwd() != ROBOTWIN_ROOT and os.path.isdir(ROBOTWIN_ROOT):
@@ -218,7 +238,8 @@ class Session:
if self.enable_render:
sapien.render.set_camera_shader_dir(self.render_shader)
args = _build_args(self.task, self.embodiment, self.seed)
- self.env.setup_demo(**args)
+ with _SortedGlob(): # the object draw depends on the seed alone, not the host's directory order
+ self.env.setup_demo(**args)
if self.enable_render: # setup_scene() switches the shader to ray tracing; put it back for cheap frames
sapien.render.set_camera_shader_dir(self.render_shader)
self._raw_scene = self.env.scene
diff --git a/tasks/robotwin_2_0_rotate_qrcode-privileged/task.toml b/tasks/robotwin_2_0_rotate_qrcode-privileged/task.toml
index e9c4ff7e3d..89eee63b15 100644
--- a/tasks/robotwin_2_0_rotate_qrcode-privileged/task.toml
+++ b/tasks/robotwin_2_0_rotate_qrcode-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "robotwin" # RoboTwin 2.0 = SAPIEN 3 / PhysX + CuRobo
robot = "aloha-agilex"
family = "rotate_qrcode"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 28831 # scene = environment/instance.json (RoboTwin seed + expected initial-state hash); identity = dataset.toml digest
expert_time_estimate_hours = 0.5
diff --git a/tasks/robotwin_2_0_rotate_qrcode-privileged/tests/rcb_traj.py b/tasks/robotwin_2_0_rotate_qrcode-privileged/tests/rcb_traj.py
index 940cef8804..e930da8932 100644
--- a/tasks/robotwin_2_0_rotate_qrcode-privileged/tests/rcb_traj.py
+++ b/tasks/robotwin_2_0_rotate_qrcode-privileged/tests/rcb_traj.py
@@ -113,6 +113,26 @@ BASE_ARGS: dict[str, Any] = {
}
+class _SortedGlob:
+ """While a scene is built, glob.glob / glob.iglob return their matches sorted. Four RoboTwin tasks (place_a2b_left,
+ place_a2b_right, place_object_scale, place_object_stand: envs/<task>.py load_actors) list an object's model files
+ with glob.glob and then draw one with the seeded np.random.choice; glob returns the file system's directory order,
+ which differs between hosts (and image layers), so one seed built a different object on another host and the frozen
+ instance's hash check failed there (board 0086 claim 5; never rebuilt off the reference workstation). Sorted, the
+ draw depends on the seed alone. No other task of ours lists files while it builds its scene."""
+
+ def __enter__(self):
+ import glob
+ self._glob, self._orig = glob, (glob.glob, glob.iglob)
+ glob.glob = lambda *a, **k: sorted(self._orig[0](*a, **k))
+ glob.iglob = lambda *a, **k: iter(sorted(self._orig[1](*a, **k)))
+ return self
+
+ def __exit__(self, *exc):
+ self._glob.glob, self._glob.iglob = self._orig
+ return False
+
+
def _chdir_root() -> None:
# RoboTwin resolves assets/config paths relative to the repo root and via ROOT_PATH
if os.getcwd() != ROBOTWIN_ROOT and os.path.isdir(ROBOTWIN_ROOT):
@@ -218,7 +238,8 @@ class Session:
if self.enable_render:
sapien.render.set_camera_shader_dir(self.render_shader)
args = _build_args(self.task, self.embodiment, self.seed)
- self.env.setup_demo(**args)
+ with _SortedGlob(): # the object draw depends on the seed alone, not the host's directory order
+ self.env.setup_demo(**args)
if self.enable_render: # setup_scene() switches the shader to ray tracing; put it back for cheap frames
sapien.render.set_camera_shader_dir(self.render_shader)
self._raw_scene = self.env.scene
diff --git a/tasks/robotwin_2_0_rotate_qrcode-standard/README.md b/tasks/robotwin_2_0_rotate_qrcode-standard/README.md
index 1d0e5ae999..83ccb123f1 100644
--- a/tasks/robotwin_2_0_rotate_qrcode-standard/README.md
+++ b/tasks/robotwin_2_0_rotate_qrcode-standard/README.md
@@ -53,3 +53,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
- 2026-10-10 (owner decision, board 0072): the agent's network egress is restricted to the model APIs (`[agent] network_mode = "allowlist"`, the hosts of `scripts/set_agent_egress.py`; since 2026-10-11 these include the package registries, PyPI and the Debian / Ubuntu archives, so that `pip install` and `apt-get install` work); it was open before (Harbor's default), and agents could download upstream sources, reference solutions and robot models from their own shell. The scene, the frozen instance, the tooling and the pass rule are unchanged, so the task keeps its name; results from before this date were obtained with open egress.
- 2026-10-13 (board 0086): the service's copy of the tooling only (`environment/sim/rcb_traj.py`, one line of `environment/sim/server.py`). The tooling now refuses to save or replay a trajectory of more than 200,000 rows, the privileged verifier's cap; a standard episode has no row cap, so the service lifts it for its own session (`S.max_rows = None`). Nothing an agent sees and nothing the service grades changes.
- 2026-10-13 (board 0100): base image `eai-robotwin` 0.1.2, whose scrub of RoboTwin's expert also removes the reference poses that `load_actors` computed only for the expert (here `target_pose`). The standard-mode agent never reaches the env, so nothing it observes changes; the scene, the instance and the judge are unchanged.
+- 2026-10-14 (board 0113): the tooling only (`environment/sim/rcb_traj.py`): while it builds the scene, `glob.glob` returns its matches sorted, so that a task which lists files to draw an object depends on its seed alone, not on the host's directory order (four RoboTwin tasks did; this one lists no files while it builds its scene). Its scene, frozen instance (same initial-state hash), verifier and pass rule are unchanged.
diff --git a/tasks/robotwin_2_0_rotate_qrcode-standard/environment/sim/rcb_traj.py b/tasks/robotwin_2_0_rotate_qrcode-standard/environment/sim/rcb_traj.py
index 940cef8804..e930da8932 100644
--- a/tasks/robotwin_2_0_rotate_qrcode-standard/environment/sim/rcb_traj.py
+++ b/tasks/robotwin_2_0_rotate_qrcode-standard/environment/sim/rcb_traj.py
@@ -113,6 +113,26 @@ BASE_ARGS: dict[str, Any] = {
}
+class _SortedGlob:
+ """While a scene is built, glob.glob / glob.iglob return their matches sorted. Four RoboTwin tasks (place_a2b_left,
+ place_a2b_right, place_object_scale, place_object_stand: envs/<task>.py load_actors) list an object's model files
+ with glob.glob and then draw one with the seeded np.random.choice; glob returns the file system's directory order,
+ which differs between hosts (and image layers), so one seed built a different object on another host and the frozen
+ instance's hash check failed there (board 0086 claim 5; never rebuilt off the reference workstation). Sorted, the
+ draw depends on the seed alone. No other task of ours lists files while it builds its scene."""
+
+ def __enter__(self):
+ import glob
+ self._glob, self._orig = glob, (glob.glob, glob.iglob)
+ glob.glob = lambda *a, **k: sorted(self._orig[0](*a, **k))
+ glob.iglob = lambda *a, **k: iter(sorted(self._orig[1](*a, **k)))
+ return self
+
+ def __exit__(self, *exc):
+ self._glob.glob, self._glob.iglob = self._orig
+ return False
+
+
def _chdir_root() -> None:
# RoboTwin resolves assets/config paths relative to the repo root and via ROOT_PATH
if os.getcwd() != ROBOTWIN_ROOT and os.path.isdir(ROBOTWIN_ROOT):
@@ -218,7 +238,8 @@ class Session:
if self.enable_render:
sapien.render.set_camera_shader_dir(self.render_shader)
args = _build_args(self.task, self.embodiment, self.seed)
- self.env.setup_demo(**args)
+ with _SortedGlob(): # the object draw depends on the seed alone, not the host's directory order
+ self.env.setup_demo(**args)
if self.enable_render: # setup_scene() switches the shader to ray tracing; put it back for cheap frames
sapien.render.set_camera_shader_dir(self.render_shader)
self._raw_scene = self.env.scene
diff --git a/tasks/robotwin_2_0_rotate_qrcode-standard/task.toml b/tasks/robotwin_2_0_rotate_qrcode-standard/task.toml
index 6f74f0f170..4648963b66 100644
--- a/tasks/robotwin_2_0_rotate_qrcode-standard/task.toml
+++ b/tasks/robotwin_2_0_rotate_qrcode-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "robotwin"
robot = "aloha-agilex"
family = "rotate_qrcode"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 28831
diff --git a/tasks/robotwin_2_0_scan_object-privileged/README.md b/tasks/robotwin_2_0_scan_object-privileged/README.md
index 3e31d5abf5..e308b8554c 100644
--- a/tasks/robotwin_2_0_scan_object-privileged/README.md
+++ b/tasks/robotwin_2_0_scan_object-privileged/README.md
@@ -56,3 +56,4 @@ Pick up the scanner with one arm and the tea box with the other, hold the tea bo
- 2026-10-13 (board 0086): the tooling refuses a trajectory the verifier cannot grade (`environment/rcb_traj.py`, `tests/rcb_traj.py`). `Session.save_trajectory` and `Session.replay` raise `ValueError` for more than 200,000 rows (`rcb_traj.MAX_T`, the verifier's `MAX_T` in `tests/grade.py`: 800 s of physics), and `record_stop` warns on stderr. Before, the tool saved and replayed such a file ("success") and the verifier then graded it `bad_trajectory`; the instruction has stated the cap since 2026-10-10. The verifier, the scene, the instance and the pass rule are unchanged: every trajectory is graded as before.
; `solution/` rebuilds the scene with upstream's `load_actors` to get them back for the expert. None of them is read by `check_success`. The scene, the frozen instance (same initial-state hash), the tooling, the verifier and the pass rule are unchanged, so the task keeps its name; privileged results from before this date were obtained with these values reachable.
- 2026-10-13 (board 0100, owner decision): base image `eai-robotwin` 0.1.2. The image's scrub of RoboTwin's expert kept `load_actors` whole, so values it computes only for the removed expert stayed reachable: `env.left_object_target_pose`, `env.right_object_target_pose` (poses the expert moves to or places at), and the instruction listed them among the scene objects. They are gone: the env no longer has them, `/opt/RoboTwin/envs/scan_object.py` no longer shows them, and the scene-object list no longer names them; `solution/` rebuilds the scene with upstream's `load_actors` to get them back for the expert. None of them is read by `check_success`. The scene, the frozen instance (same initial-state hash), the tooling, the verifier and the pass rule are unchanged, so the task keeps its name; privileged results from before this date were obtained with these values reachable.
+- 2026-10-14 (board 0113): the tooling only (`environment/rcb_traj.py`, `tests/rcb_traj.py`): while it builds the scene, `glob.glob` returns its matches sorted, so that a task which lists files to draw an object depends on its seed alone, not on the host's directory order (four RoboTwin tasks did; this one lists no files while it builds its scene). Its scene, frozen instance (same initial-state hash), verifier and pass rule are unchanged.
diff --git a/tasks/robotwin_2_0_scan_object-privileged/environment/rcb_traj.py b/tasks/robotwin_2_0_scan_object-privileged/environment/rcb_traj.py
index 33d10595a3..bf1c187fda 100644
--- a/tasks/robotwin_2_0_scan_object-privileged/environment/rcb_traj.py
+++ b/tasks/robotwin_2_0_scan_object-privileged/environment/rcb_traj.py
@@ -113,6 +113,26 @@ BASE_ARGS: dict[str, Any] = {
}
+class _SortedGlob:
+ """While a scene is built, glob.glob / glob.iglob return their matches sorted. Four RoboTwin tasks (place_a2b_left,
+ place_a2b_right, place_object_scale, place_object_stand: envs/<task>.py load_actors) list an object's model files
+ with glob.glob and then draw one with the seeded np.random.choice; glob returns the file system's directory order,
+ which differs between hosts (and image layers), so one seed built a different object on another host and the frozen
+ instance's hash check failed there (board 0086 claim 5; never rebuilt off the reference workstation). Sorted, the
+ draw depends on the seed alone. No other task of ours lists files while it builds its scene."""
+
+ def __enter__(self):
+ import glob
+ self._glob, self._orig = glob, (glob.glob, glob.iglob)
+ glob.glob = lambda *a, **k: sorted(self._orig[0](*a, **k))
+ glob.iglob = lambda *a, **k: iter(sorted(self._orig[1](*a, **k)))
+ return self
+
+ def __exit__(self, *exc):
+ self._glob.glob, self._glob.iglob = self._orig
+ return False
+
+
def _chdir_root() -> None:
# RoboTwin resolves assets/config paths relative to the repo root and via ROOT_PATH
if os.getcwd() != ROBOTWIN_ROOT and os.path.isdir(ROBOTWIN_ROOT):
@@ -218,7 +238,8 @@ class Session:
if self.enable_render:
sapien.render.set_camera_shader_dir(self.render_shader)
args = _build_args(self.task, self.embodiment, self.seed)
- self.env.setup_demo(**args)
+ with _SortedGlob(): # the object draw depends on the seed alone, not the host's directory order
+ self.env.setup_demo(**args)
if self.enable_render: # setup_scene() switches the shader to ray tracing; put it back for cheap frames
sapien.render.set_camera_shader_dir(self.render_shader)
self._raw_scene = self.env.scene
diff --git a/tasks/robotwin_2_0_scan_object-privileged/task.toml b/tasks/robotwin_2_0_scan_object-privileged/task.toml
index f9930994a3..1ab4f21c74 100644
--- a/tasks/robotwin_2_0_scan_object-privileged/task.toml
+++ b/tasks/robotwin_2_0_scan_object-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "robotwin" # RoboTwin 2.0 = SAPIEN 3 / PhysX + CuRobo
robot = "aloha-agilex"
family = "scan_object"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 60417 # scene = environment/instance.json (RoboTwin seed + expected initial-state hash); identity = dataset.toml digest
expert_time_estimate_hours = 0.5
diff --git a/tasks/robotwin_2_0_scan_object-privileged/tests/rcb_traj.py b/tasks/robotwin_2_0_scan_object-privileged/tests/rcb_traj.py
index 33d10595a3..bf1c187fda 100644
--- a/tasks/robotwin_2_0_scan_object-privileged/tests/rcb_traj.py
+++ b/tasks/robotwin_2_0_scan_object-privileged/tests/rcb_traj.py
@@ -113,6 +113,26 @@ BASE_ARGS: dict[str, Any] = {
}
+class _SortedGlob:
+ """While a scene is built, glob.glob / glob.iglob return their matches sorted. Four RoboTwin tasks (place_a2b_left,
+ place_a2b_right, place_object_scale, place_object_stand: envs/<task>.py load_actors) list an object's model files
+ with glob.glob and then draw one with the seeded np.random.choice; glob returns the file system's directory order,
+ which differs between hosts (and image layers), so one seed built a different object on another host and the frozen
+ instance's hash check failed there (board 0086 claim 5; never rebuilt off the reference workstation). Sorted, the
+ draw depends on the seed alone. No other task of ours lists files while it builds its scene."""
+
+ def __enter__(self):
+ import glob
+ self._glob, self._orig = glob, (glob.glob, glob.iglob)
+ glob.glob = lambda *a, **k: sorted(self._orig[0](*a, **k))
+ glob.iglob = lambda *a, **k: iter(sorted(self._orig[1](*a, **k)))
+ return self
+
+ def __exit__(self, *exc):
+ self._glob.glob, self._glob.iglob = self._orig
+ return False
+
+
def _chdir_root() -> None:
# RoboTwin resolves assets/config paths relative to the repo root and via ROOT_PATH
if os.getcwd() != ROBOTWIN_ROOT and os.path.isdir(ROBOTWIN_ROOT):
@@ -218,7 +238,8 @@ class Session:
if self.enable_render:
sapien.render.set_camera_shader_dir(self.render_shader)
args = _build_args(self.task, self.embodiment, self.seed)
- self.env.setup_demo(**args)
+ with _SortedGlob(): # the object draw depends on the seed alone, not the host's directory order
+ self.env.setup_demo(**args)
if self.enable_render: # setup_scene() switches the shader to ray tracing; put it back for cheap frames
sapien.render.set_camera_shader_dir(self.render_shader)
self._raw_scene = self.env.scene
diff --git a/tasks/robotwin_2_0_scan_object-standard/README.md b/tasks/robotwin_2_0_scan_object-standard/README.md
index ff0c0799e4..26409812e4 100644
--- a/tasks/robotwin_2_0_scan_object-standard/README.md
+++ b/tasks/robotwin_2_0_scan_object-standard/README.md
@@ -52,3 +52,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
- 2026-10-10 (owner decision, board 0072): the agent's network egress is restricted to the model APIs (`[agent] network_mode = "allowlist"`, the hosts of `scripts/set_agent_egress.py`; since 2026-10-11 these include the package registries, PyPI and the Debian / Ubuntu archives, so that `pip install` and `apt-get install` work); it was open before (Harbor's default), and agents could download upstream sources, reference solutions and robot models from their own shell. The scene, the frozen instance, the tooling and the pass rule are unchanged, so the task keeps its name; results from before this date were obtained with open egress.
- 2026-10-13 (board 0086): the service's copy of the tooling only (`environment/sim/rcb_traj.py`, one line of `environment/sim/server.py`). The tooling now refuses to save or replay a trajectory of more than 200,000 rows, the privileged verifier's cap; a standard episode has no row cap, so the service lifts it for its own session (`S.max_rows = None`). Nothing an agent sees and nothing the service grades changes.
- 2026-10-13 (board 0100): base image `eai-robotwin` 0.1.2, whose scrub of RoboTwin's expert also removes the reference poses that `load_actors` computed only for the expert (here `left_object_target_pose`, `right_object_target_pose`). The standard-mode agent never reaches the env, so nothing it observes changes; the scene, the instance and the judge are unchanged.
+- 2026-10-14 (board 0113): the tooling only (`environment/sim/rcb_traj.py`): while it builds the scene, `glob.glob` returns its matches sorted, so that a task which lists files to draw an object depends on its seed alone, not on the host's directory order (four RoboTwin tasks did; this one lists no files while it builds its scene). Its scene, frozen instance (same initial-state hash), verifier and pass rule are unchanged.
diff --git a/tasks/robotwin_2_0_scan_object-standard/environment/sim/rcb_traj.py b/tasks/robotwin_2_0_scan_object-standard/environment/sim/rcb_traj.py
index 33d10595a3..bf1c187fda 100644
--- a/tasks/robotwin_2_0_scan_object-standard/environment/sim/rcb_traj.py
+++ b/tasks/robotwin_2_0_scan_object-standard/environment/sim/rcb_traj.py
@@ -113,6 +113,26 @@ BASE_ARGS: dict[str, Any] = {
}
+class _SortedGlob:
+ """While a scene is built, glob.glob / glob.iglob return their matches sorted. Four RoboTwin tasks (place_a2b_left,
+ place_a2b_right, place_object_scale, place_object_stand: envs/<task>.py load_actors) list an object's model files
+ with glob.glob and then draw one with the seeded np.random.choice; glob returns the file system's directory order,
+ which differs between hosts (and image layers), so one seed built a different object on another host and the frozen
+ instance's hash check failed there (board 0086 claim 5; never rebuilt off the reference workstation). Sorted, the
+ draw depends on the seed alone. No other task of ours lists files while it builds its scene."""
+
+ def __enter__(self):
+ import glob
+ self._glob, self._orig = glob, (glob.glob, glob.iglob)
+ glob.glob = lambda *a, **k: sorted(self._orig[0](*a, **k))
+ glob.iglob = lambda *a, **k: iter(sorted(self._orig[1](*a, **k)))
+ return self
+
+ def __exit__(self, *exc):
+ self._glob.glob, self._glob.iglob = self._orig
+ return False
+
+
def _chdir_root() -> None:
# RoboTwin resolves assets/config paths relative to the repo root and via ROOT_PATH
if os.getcwd() != ROBOTWIN_ROOT and os.path.isdir(ROBOTWIN_ROOT):
@@ -218,7 +238,8 @@ class Session:
if self.enable_render:
sapien.render.set_camera_shader_dir(self.render_shader)
args = _build_args(self.task, self.embodiment, self.seed)
- self.env.setup_demo(**args)
+ with _SortedGlob(): # the object draw depends on the seed alone, not the host's directory order
+ self.env.setup_demo(**args)
if self.enable_render: # setup_scene() switches the shader to ray tracing; put it back for cheap frames
sapien.render.set_camera_shader_dir(self.render_shader)
self._raw_scene = self.env.scene
diff --git a/tasks/robotwin_2_0_scan_object-standard/task.toml b/tasks/robotwin_2_0_scan_object-standard/task.toml
index 14c47ddf12..12f9267e3c 100644
--- a/tasks/robotwin_2_0_scan_object-standard/task.toml
+++ b/tasks/robotwin_2_0_scan_object-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "robotwin"
robot = "aloha-agilex"
family = "scan_object"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 60417
diff --git a/tasks/robotwin_2_0_shake_bottle-privileged/README.md b/tasks/robotwin_2_0_shake_bottle-privileged/README.md
index fce22d7182..4bd6a8160b 100644
--- a/tasks/robotwin_2_0_shake_bottle-privileged/README.md
+++ b/tasks/robotwin_2_0_shake_bottle-privileged/README.md
@@ -55,3 +55,4 @@ Shake the bottle with proper arm.
- 2026-10-10 (owner decision, board 0072): the agent's network egress is restricted to the model APIs (`[agent] network_mode = "allowlist"`, the hosts of `scripts/set_agent_egress.py`; since 2026-10-11 these include the package registries, PyPI and the Debian / Ubuntu archives, so that `pip install` and `apt-get install` work); it was open before (Harbor's default), and agents could download upstream sources, reference solutions and robot models from their own shell. The scene, the frozen instance, the tooling and the pass rule are unchanged, so the task keeps its name; results from before this date were obtained with open egress.
- 2026-10-13 (board 0086): the tooling refuses a trajectory the verifier cannot grade (`environment/rcb_traj.py`, `tests/rcb_traj.py`). `Session.save_trajectory` and `Session.replay` raise `ValueError` for more than 200,000 rows (`rcb_traj.MAX_T`, the verifier's `MAX_T` in `tests/grade.py`: 800 s of physics), and `record_stop` warns on stderr. Before, the tool saved and replayed such a file ("success") and the verifier then graded it `bad_trajectory`; the instruction has stated the cap since 2026-10-10. The verifier, the scene, the instance and the pass rule are unchanged: every trajectory is graded as before.
- 2026-10-13 (board 0100): base image `eai-robotwin` 0.1.2. Its scrub of RoboTwin's expert also removes the reference poses that 18 tasks' `load_actors` computed only for the expert, and the expert's comments above the scrubbed method bodies (`images/robotwin/scrub_oracles.py`). This task's `load_actors` set no such value: its scene, instance, tooling, verifier and pass rule are unchanged.
+- 2026-10-14 (board 0113): the tooling only (`environment/rcb_traj.py`, `tests/rcb_traj.py`): while it builds the scene, `glob.glob` returns its matches sorted, so that a task which lists files to draw an object depends on its seed alone, not on the host's directory order (four RoboTwin tasks did; this one lists no files while it builds its scene). Its scene, frozen instance (same initial-state hash), verifier and pass rule are unchanged.
diff --git a/tasks/robotwin_2_0_shake_bottle-privileged/environment/rcb_traj.py b/tasks/robotwin_2_0_shake_bottle-privileged/environment/rcb_traj.py
index 268395ffab..28adde9e02 100644
--- a/tasks/robotwin_2_0_shake_bottle-privileged/environment/rcb_traj.py
+++ b/tasks/robotwin_2_0_shake_bottle-privileged/environment/rcb_traj.py
@@ -113,6 +113,26 @@ BASE_ARGS: dict[str, Any] = {
}
+class _SortedGlob:
+ """While a scene is built, glob.glob / glob.iglob return their matches sorted. Four RoboTwin tasks (place_a2b_left,
+ place_a2b_right, place_object_scale, place_object_stand: envs/<task>.py load_actors) list an object's model files
+ with glob.glob and then draw one with the seeded np.random.choice; glob returns the file system's directory order,
+ which differs between hosts (and image layers), so one seed built a different object on another host and the frozen
+ instance's hash check failed there (board 0086 claim 5; never rebuilt off the reference workstation). Sorted, the
+ draw depends on the seed alone. No other task of ours lists files while it builds its scene."""
+
+ def __enter__(self):
+ import glob
+ self._glob, self._orig = glob, (glob.glob, glob.iglob)
+ glob.glob = lambda *a, **k: sorted(self._orig[0](*a, **k))
+ glob.iglob = lambda *a, **k: iter(sorted(self._orig[1](*a, **k)))
+ return self
+
+ def __exit__(self, *exc):
+ self._glob.glob, self._glob.iglob = self._orig
+ return False
+
+
def _chdir_root() -> None:
# RoboTwin resolves assets/config paths relative to the repo root and via ROOT_PATH
if os.getcwd() != ROBOTWIN_ROOT and os.path.isdir(ROBOTWIN_ROOT):
@@ -218,7 +238,8 @@ class Session:
if self.enable_render:
sapien.render.set_camera_shader_dir(self.render_shader)
args = _build_args(self.task, self.embodiment, self.seed)
- self.env.setup_demo(**args)
+ with _SortedGlob(): # the object draw depends on the seed alone, not the host's directory order
+ self.env.setup_demo(**args)
if self.enable_render: # setup_scene() switches the shader to ray tracing; put it back for cheap frames
sapien.render.set_camera_shader_dir(self.render_shader)
self._raw_scene = self.env.scene
diff --git a/tasks/robotwin_2_0_shake_bottle-privileged/task.toml b/tasks/robotwin_2_0_shake_bottle-privileged/task.toml
index 327fce0a1d..a1604ea827 100644
--- a/tasks/robotwin_2_0_shake_bottle-privileged/task.toml
+++ b/tasks/robotwin_2_0_shake_bottle-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "robotwin" # RoboTwin 2.0 = SAPIEN 3 / PhysX + CuRobo
robot = "aloha-agilex"
family = "shake_bottle"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 60417 # scene = environment/instance.json (RoboTwin seed + expected initial-state hash); identity = dataset.toml digest
expert_time_estimate_hours = 0.5
diff --git a/tasks/robotwin_2_0_shake_bottle-privileged/tests/rcb_traj.py b/tasks/robotwin_2_0_shake_bottle-privileged/tests/rcb_traj.py
index 268395ffab..28adde9e02 100644
--- a/tasks/robotwin_2_0_shake_bottle-privileged/tests/rcb_traj.py
+++ b/tasks/robotwin_2_0_shake_bottle-privileged/tests/rcb_traj.py
@@ -113,6 +113,26 @@ BASE_ARGS: dict[str, Any] = {
}
+class _SortedGlob:
+ """While a scene is built, glob.glob / glob.iglob return their matches sorted. Four RoboTwin tasks (place_a2b_left,
+ place_a2b_right, place_object_scale, place_object_stand: envs/<task>.py load_actors) list an object's model files
+ with glob.glob and then draw one with the seeded np.random.choice; glob returns the file system's directory order,
+ which differs between hosts (and image layers), so one seed built a different object on another host and the frozen
+ instance's hash check failed there (board 0086 claim 5; never rebuilt off the reference workstation). Sorted, the
+ draw depends on the seed alone. No other task of ours lists files while it builds its scene."""
+
+ def __enter__(self):
+ import glob
+ self._glob, self._orig = glob, (glob.glob, glob.iglob)
+ glob.glob = lambda *a, **k: sorted(self._orig[0](*a, **k))
+ glob.iglob = lambda *a, **k: iter(sorted(self._orig[1](*a, **k)))
+ return self
+
+ def __exit__(self, *exc):
+ self._glob.glob, self._glob.iglob = self._orig
+ return False
+
+
def _chdir_root() -> None:
# RoboTwin resolves assets/config paths relative to the repo root and via ROOT_PATH
if os.getcwd() != ROBOTWIN_ROOT and os.path.isdir(ROBOTWIN_ROOT):
@@ -218,7 +238,8 @@ class Session:
if self.enable_render:
sapien.render.set_camera_shader_dir(self.render_shader)
args = _build_args(self.task, self.embodiment, self.seed)
- self.env.setup_demo(**args)
+ with _SortedGlob(): # the object draw depends on the seed alone, not the host's directory order
+ self.env.setup_demo(**args)
if self.enable_render: # setup_scene() switches the shader to ray tracing; put it back for cheap frames
sapien.render.set_camera_shader_dir(self.render_shader)
self._raw_scene = self.env.scene
diff --git a/tasks/robotwin_2_0_shake_bottle-standard/README.md b/tasks/robotwin_2_0_shake_bottle-standard/README.md
index 53ceae8ffb..68bf178754 100644
--- a/tasks/robotwin_2_0_shake_bottle-standard/README.md
+++ b/tasks/robotwin_2_0_shake_bottle-standard/README.md
@@ -53,3 +53,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
- 2026-10-10 (owner decision, board 0072): the agent's network egress is restricted to the model APIs (`[agent] network_mode = "allowlist"`, the hosts of `scripts/set_agent_egress.py`; since 2026-10-11 these include the package registries, PyPI and the Debian / Ubuntu archives, so that `pip install` and `apt-get install` work); it was open before (Harbor's default), and agents could download upstream sources, reference solutions and robot models from their own shell. The scene, the frozen instance, the tooling and the pass rule are unchanged, so the task keeps its name; results from before this date were obtained with open egress.
- 2026-10-13 (board 0086): the service's copy of the tooling only (`environment/sim/rcb_traj.py`, one line of `environment/sim/server.py`). The tooling now refuses to save or replay a trajectory of more than 200,000 rows, the privileged verifier's cap; a standard episode has no row cap, so the service lifts it for its own session (`S.max_rows = None`). Nothing an agent sees and nothing the service grades changes.
- 2026-10-13 (board 0100): base image `eai-robotwin` 0.1.2, whose scrub of RoboTwin's expert also removes the reference poses that `load_actors` computed only for the expert. The standard-mode agent never reaches the env, so nothing it observes changes; the scene, the instance and the judge are unchanged.
+- 2026-10-14 (board 0113): the tooling only (`environment/sim/rcb_traj.py`): while it builds the scene, `glob.glob` returns its matches sorted, so that a task which lists files to draw an object depends on its seed alone, not on the host's directory order (four RoboTwin tasks did; this one lists no files while it builds its scene). Its scene, frozen instance (same initial-state hash), verifier and pass rule are unchanged.
diff --git a/tasks/robotwin_2_0_shake_bottle-standard/environment/sim/rcb_traj.py b/tasks/robotwin_2_0_shake_bottle-standard/environment/sim/rcb_traj.py
index 268395ffab..28adde9e02 100644
--- a/tasks/robotwin_2_0_shake_bottle-standard/environment/sim/rcb_traj.py
+++ b/tasks/robotwin_2_0_shake_bottle-standard/environment/sim/rcb_traj.py
@@ -113,6 +113,26 @@ BASE_ARGS: dict[str, Any] = {
}
+class _SortedGlob:
+ """While a scene is built, glob.glob / glob.iglob return their matches sorted. Four RoboTwin tasks (place_a2b_left,
+ place_a2b_right, place_object_scale, place_object_stand: envs/<task>.py load_actors) list an object's model files
+ with glob.glob and then draw one with the seeded np.random.choice; glob returns the file system's directory order,
+ which differs between hosts (and image layers), so one seed built a different object on another host and the frozen
+ instance's hash check failed there (board 0086 claim 5; never rebuilt off the reference workstation). Sorted, the
+ draw depends on the seed alone. No other task of ours lists files while it builds its scene."""
+
+ def __enter__(self):
+ import glob
+ self._glob, self._orig = glob, (glob.glob, glob.iglob)
+ glob.glob = lambda *a, **k: sorted(self._orig[0](*a, **k))
+ glob.iglob = lambda *a, **k: iter(sorted(self._orig[1](*a, **k)))
+ return self
+
+ def __exit__(self, *exc):
+ self._glob.glob, self._glob.iglob = self._orig
+ return False
+
+
def _chdir_root() -> None:
# RoboTwin resolves assets/config paths relative to the repo root and via ROOT_PATH
if os.getcwd() != ROBOTWIN_ROOT and os.path.isdir(ROBOTWIN_ROOT):
@@ -218,7 +238,8 @@ class Session:
if self.enable_render:
sapien.render.set_camera_shader_dir(self.render_shader)
args = _build_args(self.task, self.embodiment, self.seed)
- self.env.setup_demo(**args)
+ with _SortedGlob(): # the object draw depends on the seed alone, not the host's directory order
+ self.env.setup_demo(**args)
if self.enable_render: # setup_scene() switches the shader to ray tracing; put it back for cheap frames
sapien.render.set_camera_shader_dir(self.render_shader)
self._raw_scene = self.env.scene
diff --git a/tasks/robotwin_2_0_shake_bottle-standard/task.toml b/tasks/robotwin_2_0_shake_bottle-standard/task.toml
index c3869aec83..27e15dd31c 100644
--- a/tasks/robotwin_2_0_shake_bottle-standard/task.toml
+++ b/tasks/robotwin_2_0_shake_bottle-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "robotwin"
robot = "aloha-agilex"
family = "shake_bottle"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 60417
diff --git a/tasks/robotwin_2_0_shake_bottle_horizontally-privileged/README.md b/tasks/robotwin_2_0_shake_bottle_horizontally-privileged/README.md
index 2512c101a7..ad80298489 100644
--- a/tasks/robotwin_2_0_shake_bottle_horizontally-privileged/README.md
+++ b/tasks/robotwin_2_0_shake_bottle_horizontally-privileged/README.md
@@ -55,3 +55,4 @@ Shake the bottle horizontally with proper arm.
- 2026-10-10 (owner decision, board 0072): the agent's network egress is restricted to the model APIs (`[agent] network_mode = "allowlist"`, the hosts of `scripts/set_agent_egress.py`; since 2026-10-11 these include the package registries, PyPI and the Debian / Ubuntu archives, so that `pip install` and `apt-get install` work); it was open before (Harbor's default), and agents could download upstream sources, reference solutions and robot models from their own shell. The scene, the frozen instance, the tooling and the pass rule are unchanged, so the task keeps its name; results from before this date were obtained with open egress.
- 2026-10-13 (board 0086): the tooling refuses a trajectory the verifier cannot grade (`environment/rcb_traj.py`, `tests/rcb_traj.py`). `Session.save_trajectory` and `Session.replay` raise `ValueError` for more than 200,000 rows (`rcb_traj.MAX_T`, the verifier's `MAX_T` in `tests/grade.py`: 800 s of physics), and `record_stop` warns on stderr. Before, the tool saved and replayed such a file ("success") and the verifier then graded it `bad_trajectory`; the instruction has stated the cap since 2026-10-10. The verifier, the scene, the instance and the pass rule are unchanged: every trajectory is graded as before.
- 2026-10-13 (board 0100): base image `eai-robotwin` 0.1.2. Its scrub of RoboTwin's expert also removes the reference poses that 18 tasks' `load_actors` computed only for the expert, and the expert's comments above the scrubbed method bodies (`images/robotwin/scrub_oracles.py`). This task's `load_actors` set no such value: its scene, instance, tooling, verifier and pass rule are unchanged.
+- 2026-10-14 (board 0113): the tooling only (`environment/rcb_traj.py`, `tests/rcb_traj.py`): while it builds the scene, `glob.glob` returns its matches sorted, so that a task which lists files to draw an object depends on its seed alone, not on the host's directory order (four RoboTwin tasks did; this one lists no files while it builds its scene). Its scene, frozen instance (same initial-state hash), verifier and pass rule are unchanged.
diff --git a/tasks/robotwin_2_0_shake_bottle_horizontally-privileged/environment/rcb_traj.py b/tasks/robotwin_2_0_shake_bottle_horizontally-privileged/environment/rcb_traj.py
index c7a054cbaf..e7c9b4b604 100644
--- a/tasks/robotwin_2_0_shake_bottle_horizontally-privileged/environment/rcb_traj.py
+++ b/tasks/robotwin_2_0_shake_bottle_horizontally-privileged/environment/rcb_traj.py
@@ -113,6 +113,26 @@ BASE_ARGS: dict[str, Any] = {
}
+class _SortedGlob:
+ """While a scene is built, glob.glob / glob.iglob return their matches sorted. Four RoboTwin tasks (place_a2b_left,
+ place_a2b_right, place_object_scale, place_object_stand: envs/<task>.py load_actors) list an object's model files
+ with glob.glob and then draw one with the seeded np.random.choice; glob returns the file system's directory order,
+ which differs between hosts (and image layers), so one seed built a different object on another host and the frozen
+ instance's hash check failed there (board 0086 claim 5; never rebuilt off the reference workstation). Sorted, the
+ draw depends on the seed alone. No other task of ours lists files while it builds its scene."""
+
+ def __enter__(self):
+ import glob
+ self._glob, self._orig = glob, (glob.glob, glob.iglob)
+ glob.glob = lambda *a, **k: sorted(self._orig[0](*a, **k))
+ glob.iglob = lambda *a, **k: iter(sorted(self._orig[1](*a, **k)))
+ return self
+
+ def __exit__(self, *exc):
+ self._glob.glob, self._glob.iglob = self._orig
+ return False
+
+
def _chdir_root() -> None:
# RoboTwin resolves assets/config paths relative to the repo root and via ROOT_PATH
if os.getcwd() != ROBOTWIN_ROOT and os.path.isdir(ROBOTWIN_ROOT):
@@ -218,7 +238,8 @@ class Session:
if self.enable_render:
sapien.render.set_camera_shader_dir(self.render_shader)
args = _build_args(self.task, self.embodiment, self.seed)
- self.env.setup_demo(**args)
+ with _SortedGlob(): # the object draw depends on the seed alone, not the host's directory order
+ self.env.setup_demo(**args)
if self.enable_render: # setup_scene() switches the shader to ray tracing; put it back for cheap frames
sapien.render.set_camera_shader_dir(self.render_shader)
self._raw_scene = self.env.scene
diff --git a/tasks/robotwin_2_0_shake_bottle_horizontally-privileged/task.toml b/tasks/robotwin_2_0_shake_bottle_horizontally-privileged/task.toml
index f556eb41e7..9ac2867cd6 100644
--- a/tasks/robotwin_2_0_shake_bottle_horizontally-privileged/task.toml
+++ b/tasks/robotwin_2_0_shake_bottle_horizontally-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "robotwin" # RoboTwin 2.0 = SAPIEN 3 / PhysX + CuRobo
robot = "aloha-agilex"
family = "shake_bottle_horizontally"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 60417 # scene = environment/instance.json (RoboTwin seed + expected initial-state hash); identity = dataset.toml digest
expert_time_estimate_hours = 0.5
diff --git a/tasks/robotwin_2_0_shake_bottle_horizontally-privileged/tests/rcb_traj.py b/tasks/robotwin_2_0_shake_bottle_horizontally-privileged/tests/rcb_traj.py
index c7a054cbaf..e7c9b4b604 100644
--- a/tasks/robotwin_2_0_shake_bottle_horizontally-privileged/tests/rcb_traj.py
+++ b/tasks/robotwin_2_0_shake_bottle_horizontally-privileged/tests/rcb_traj.py
@@ -113,6 +113,26 @@ BASE_ARGS: dict[str, Any] = {
}
+class _SortedGlob:
+ """While a scene is built, glob.glob / glob.iglob return their matches sorted. Four RoboTwin tasks (place_a2b_left,
+ place_a2b_right, place_object_scale, place_object_stand: envs/<task>.py load_actors) list an object's model files
+ with glob.glob and then draw one with the seeded np.random.choice; glob returns the file system's directory order,
+ which differs between hosts (and image layers), so one seed built a different object on another host and the frozen
+ instance's hash check failed there (board 0086 claim 5; never rebuilt off the reference workstation). Sorted, the
+ draw depends on the seed alone. No other task of ours lists files while it builds its scene."""
+
+ def __enter__(self):
+ import glob
+ self._glob, self._orig = glob, (glob.glob, glob.iglob)
+ glob.glob = lambda *a, **k: sorted(self._orig[0](*a, **k))
+ glob.iglob = lambda *a, **k: iter(sorted(self._orig[1](*a, **k)))
+ return self
+
+ def __exit__(self, *exc):
+ self._glob.glob, self._glob.iglob = self._orig
+ return False
+
+
def _chdir_root() -> None:
# RoboTwin resolves assets/config paths relative to the repo root and via ROOT_PATH
if os.getcwd() != ROBOTWIN_ROOT and os.path.isdir(ROBOTWIN_ROOT):
@@ -218,7 +238,8 @@ class Session:
if self.enable_render:
sapien.render.set_camera_shader_dir(self.render_shader)
args = _build_args(self.task, self.embodiment, self.seed)
- self.env.setup_demo(**args)
+ with _SortedGlob(): # the object draw depends on the seed alone, not the host's directory order
+ self.env.setup_demo(**args)
if self.enable_render: # setup_scene() switches the shader to ray tracing; put it back for cheap frames
sapien.render.set_camera_shader_dir(self.render_shader)
self._raw_scene = self.env.scene
diff --git a/tasks/robotwin_2_0_shake_bottle_horizontally-standard/README.md b/tasks/robotwin_2_0_shake_bottle_horizontally-standard/README.md
index 86abed0e43..b24fab61d3 100644
--- a/tasks/robotwin_2_0_shake_bottle_horizontally-standard/README.md
+++ b/tasks/robotwin_2_0_shake_bottle_horizontally-standard/README.md
@@ -53,3 +53,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
- 2026-10-10 (owner decision, board 0072): the agent's network egress is restricted to the model APIs (`[agent] network_mode = "allowlist"`, the hosts of `scripts/set_agent_egress.py`; since 2026-10-11 these include the package registries, PyPI and the Debian / Ubuntu archives, so that `pip install` and `apt-get install` work); it was open before (Harbor's default), and agents could download upstream sources, reference solutions and robot models from their own shell. The scene, the frozen instance, the tooling and the pass rule are unchanged, so the task keeps its name; results from before this date were obtained with open egress.
- 2026-10-13 (board 0086): the service's copy of the tooling only (`environment/sim/rcb_traj.py`, one line of `environment/sim/server.py`). The tooling now refuses to save or replay a trajectory of more than 200,000 rows, the privileged verifier's cap; a standard episode has no row cap, so the service lifts it for its own session (`S.max_rows = None`). Nothing an agent sees and nothing the service grades changes.
- 2026-10-13 (board 0100): base image `eai-robotwin` 0.1.2, whose scrub of RoboTwin's expert also removes the reference poses that `load_actors` computed only for the expert. The standard-mode agent never reaches the env, so nothing it observes changes; the scene, the instance and the judge are unchanged.
+- 2026-10-14 (board 0113): the tooling only (`environment/sim/rcb_traj.py`): while it builds the scene, `glob.glob` returns its matches sorted, so that a task which lists files to draw an object depends on its seed alone, not on the host's directory order (four RoboTwin tasks did; this one lists no files while it builds its scene). Its scene, frozen instance (same initial-state hash), verifier and pass rule are unchanged.
diff --git a/tasks/robotwin_2_0_shake_bottle_horizontally-standard/environment/sim/rcb_traj.py b/tasks/robotwin_2_0_shake_bottle_horizontally-standard/environment/sim/rcb_traj.py
index c7a054cbaf..e7c9b4b604 100644
--- a/tasks/robotwin_2_0_shake_bottle_horizontally-standard/environment/sim/rcb_traj.py
+++ b/tasks/robotwin_2_0_shake_bottle_horizontally-standard/environment/sim/rcb_traj.py
@@ -113,6 +113,26 @@ BASE_ARGS: dict[str, Any] = {
}
+class _SortedGlob:
+ """While a scene is built, glob.glob / glob.iglob return their matches sorted. Four RoboTwin tasks (place_a2b_left,
+ place_a2b_right, place_object_scale, place_object_stand: envs/<task>.py load_actors) list an object's model files
+ with glob.glob and then draw one with the seeded np.random.choice; glob returns the file system's directory order,
+ which differs between hosts (and image layers), so one seed built a different object on another host and the frozen
+ instance's hash check failed there (board 0086 claim 5; never rebuilt off the reference workstation). Sorted, the
+ draw depends on the seed alone. No other task of ours lists files while it builds its scene."""
+
+ def __enter__(self):
+ import glob
+ self._glob, self._orig = glob, (glob.glob, glob.iglob)
+ glob.glob = lambda *a, **k: sorted(self._orig[0](*a, **k))
+ glob.iglob = lambda *a, **k: iter(sorted(self._orig[1](*a, **k)))
+ return self
+
+ def __exit__(self, *exc):
+ self._glob.glob, self._glob.iglob = self._orig
+ return False
+
+
def _chdir_root() -> None:
# RoboTwin resolves assets/config paths relative to the repo root and via ROOT_PATH
if os.getcwd() != ROBOTWIN_ROOT and os.path.isdir(ROBOTWIN_ROOT):
@@ -218,7 +238,8 @@ class Session:
if self.enable_render:
sapien.render.set_camera_shader_dir(self.render_shader)
args = _build_args(self.task, self.embodiment, self.seed)
- self.env.setup_demo(**args)
+ with _SortedGlob(): # the object draw depends on the seed alone, not the host's directory order
+ self.env.setup_demo(**args)
if self.enable_render: # setup_scene() switches the shader to ray tracing; put it back for cheap frames
sapien.render.set_camera_shader_dir(self.render_shader)
self._raw_scene = self.env.scene
diff --git a/tasks/robotwin_2_0_shake_bottle_horizontally-standard/task.toml b/tasks/robotwin_2_0_shake_bottle_horizontally-standard/task.toml
index cc6c8df65c..3e9d1371ca 100644
--- a/tasks/robotwin_2_0_shake_bottle_horizontally-standard/task.toml
+++ b/tasks/robotwin_2_0_shake_bottle_horizontally-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "robotwin"
robot = "aloha-agilex"
family = "shake_bottle_horizontally"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 60417
diff --git a/tasks/robotwin_2_0_stack_blocks_three-privileged/README.md b/tasks/robotwin_2_0_stack_blocks_three-privileged/README.md
index 8e51228459..8bb277175f 100644
--- a/tasks/robotwin_2_0_stack_blocks_three-privileged/README.md
+++ b/tasks/robotwin_2_0_stack_blocks_three-privileged/README.md
@@ -56,3 +56,4 @@ There are three blocks on the table, the color of the blocks is red, green and b
- 2026-10-13 (board 0086): the tooling refuses a trajectory the verifier cannot grade (`environment/rcb_traj.py`, `tests/rcb_traj.py`). `Session.save_trajectory` and `Session.replay` raise `ValueError` for more than 200,000 rows (`rcb_traj.MAX_T`, the verifier's `MAX_T` in `tests/grade.py`: 800 s of physics), and `record_stop` warns on stderr. Before, the tool saved and replayed such a file ("success") and the verifier then graded it `bad_trajectory`; the instruction has stated the cap since 2026-10-10. The verifier, the scene, the instance and the pass rule are unchanged: every trajectory is graded as before.
; the reference solution never used it, and `check_success` does not read it. The scene, the frozen instance (same initial-state hash), the tooling, the verifier and the pass rule are unchanged, so the task keeps its name; privileged results from before this date were obtained with this value reachable.
- 2026-10-13 (board 0100, owner decision): base image `eai-robotwin` 0.1.2. The image's scrub of RoboTwin's expert kept `load_actors` whole, so a value it computes for the expert stayed reachable: `env.block1_target_pose` (a pose for the expert to move to or place at; unused upstream), and the instruction listed it among the scene objects. It is gone: the env no longer has it, `/opt/RoboTwin/envs/stack_blocks_three.py` no longer shows it, and the scene-object list no longer names it; the reference solution never used it, and `check_success` does not read it. The scene, the frozen instance (same initial-state hash), the tooling, the verifier and the pass rule are unchanged, so the task keeps its name; privileged results from before this date were obtained with this value reachable.
+- 2026-10-14 (board 0113): the tooling only (`environment/rcb_traj.py`, `tests/rcb_traj.py`): while it builds the scene, `glob.glob` returns its matches sorted, so that a task which lists files to draw an object depends on its seed alone, not on the host's directory order (four RoboTwin tasks did; this one lists no files while it builds its scene). Its scene, frozen instance (same initial-state hash), verifier and pass rule are unchanged.
diff --git a/tasks/robotwin_2_0_stack_blocks_three-privileged/environment/rcb_traj.py b/tasks/robotwin_2_0_stack_blocks_three-privileged/environment/rcb_traj.py
index 0b564387de..0c6314c152 100644
--- a/tasks/robotwin_2_0_stack_blocks_three-privileged/environment/rcb_traj.py
+++ b/tasks/robotwin_2_0_stack_blocks_three-privileged/environment/rcb_traj.py
@@ -113,6 +113,26 @@ BASE_ARGS: dict[str, Any] = {
}
+class _SortedGlob:
+ """While a scene is built, glob.glob / glob.iglob return their matches sorted. Four RoboTwin tasks (place_a2b_left,
+ place_a2b_right, place_object_scale, place_object_stand: envs/<task>.py load_actors) list an object's model files
+ with glob.glob and then draw one with the seeded np.random.choice; glob returns the file system's directory order,
+ which differs between hosts (and image layers), so one seed built a different object on another host and the frozen
+ instance's hash check failed there (board 0086 claim 5; never rebuilt off the reference workstation). Sorted, the
+ draw depends on the seed alone. No other task of ours lists files while it builds its scene."""
+
+ def __enter__(self):
+ import glob
+ self._glob, self._orig = glob, (glob.glob, glob.iglob)
+ glob.glob = lambda *a, **k: sorted(self._orig[0](*a, **k))
+ glob.iglob = lambda *a, **k: iter(sorted(self._orig[1](*a, **k)))
+ return self
+
+ def __exit__(self, *exc):
+ self._glob.glob, self._glob.iglob = self._orig
+ return False
+
+
def _chdir_root() -> None:
# RoboTwin resolves assets/config paths relative to the repo root and via ROOT_PATH
if os.getcwd() != ROBOTWIN_ROOT and os.path.isdir(ROBOTWIN_ROOT):
@@ -218,7 +238,8 @@ class Session:
if self.enable_render:
sapien.render.set_camera_shader_dir(self.render_shader)
args = _build_args(self.task, self.embodiment, self.seed)
- self.env.setup_demo(**args)
+ with _SortedGlob(): # the object draw depends on the seed alone, not the host's directory order
+ self.env.setup_demo(**args)
if self.enable_render: # setup_scene() switches the shader to ray tracing; put it back for cheap frames
sapien.render.set_camera_shader_dir(self.render_shader)
self._raw_scene = self.env.scene
diff --git a/tasks/robotwin_2_0_stack_blocks_three-privileged/task.toml b/tasks/robotwin_2_0_stack_blocks_three-privileged/task.toml
index 393819e4b3..b1da68bddf 100644
--- a/tasks/robotwin_2_0_stack_blocks_three-privileged/task.toml
+++ b/tasks/robotwin_2_0_stack_blocks_three-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "robotwin" # RoboTwin 2.0 = SAPIEN 3 / PhysX + CuRobo
robot = "aloha-agilex"
family = "stack_blocks_three"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 60417 # scene = environment/instance.json (RoboTwin seed + expected initial-state hash); identity = dataset.toml digest
expert_time_estimate_hours = 0.5
diff --git a/tasks/robotwin_2_0_stack_blocks_three-privileged/tests/rcb_traj.py b/tasks/robotwin_2_0_stack_blocks_three-privileged/tests/rcb_traj.py
index 0b564387de..0c6314c152 100644
--- a/tasks/robotwin_2_0_stack_blocks_three-privileged/tests/rcb_traj.py
+++ b/tasks/robotwin_2_0_stack_blocks_three-privileged/tests/rcb_traj.py
@@ -113,6 +113,26 @@ BASE_ARGS: dict[str, Any] = {
}
+class _SortedGlob:
+ """While a scene is built, glob.glob / glob.iglob return their matches sorted. Four RoboTwin tasks (place_a2b_left,
+ place_a2b_right, place_object_scale, place_object_stand: envs/<task>.py load_actors) list an object's model files
+ with glob.glob and then draw one with the seeded np.random.choice; glob returns the file system's directory order,
+ which differs between hosts (and image layers), so one seed built a different object on another host and the frozen
+ instance's hash check failed there (board 0086 claim 5; never rebuilt off the reference workstation). Sorted, the
+ draw depends on the seed alone. No other task of ours lists files while it builds its scene."""
+
+ def __enter__(self):
+ import glob
+ self._glob, self._orig = glob, (glob.glob, glob.iglob)
+ glob.glob = lambda *a, **k: sorted(self._orig[0](*a, **k))
+ glob.iglob = lambda *a, **k: iter(sorted(self._orig[1](*a, **k)))
+ return self
+
+ def __exit__(self, *exc):
+ self._glob.glob, self._glob.iglob = self._orig
+ return False
+
+
def _chdir_root() -> None:
# RoboTwin resolves assets/config paths relative to the repo root and via ROOT_PATH
if os.getcwd() != ROBOTWIN_ROOT and os.path.isdir(ROBOTWIN_ROOT):
@@ -218,7 +238,8 @@ class Session:
if self.enable_render:
sapien.render.set_camera_shader_dir(self.render_shader)
args = _build_args(self.task, self.embodiment, self.seed)
- self.env.setup_demo(**args)
+ with _SortedGlob(): # the object draw depends on the seed alone, not the host's directory order
+ self.env.setup_demo(**args)
if self.enable_render: # setup_scene() switches the shader to ray tracing; put it back for cheap frames
sapien.render.set_camera_shader_dir(self.render_shader)
self._raw_scene = self.env.scene
diff --git a/tasks/robotwin_2_0_stack_blocks_three-standard/README.md b/tasks/robotwin_2_0_stack_blocks_three-standard/README.md
index c3f1d4dfad..16950e7261 100644
--- a/tasks/robotwin_2_0_stack_blocks_three-standard/README.md
+++ b/tasks/robotwin_2_0_stack_blocks_three-standard/README.md
@@ -53,3 +53,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
- 2026-10-10 (owner decision, board 0072): the agent's network egress is restricted to the model APIs (`[agent] network_mode = "allowlist"`, the hosts of `scripts/set_agent_egress.py`; since 2026-10-11 these include the package registries, PyPI and the Debian / Ubuntu archives, so that `pip install` and `apt-get install` work); it was open before (Harbor's default), and agents could download upstream sources, reference solutions and robot models from their own shell. The scene, the frozen instance, the tooling and the pass rule are unchanged, so the task keeps its name; results from before this date were obtained with open egress.
- 2026-10-13 (board 0086): the service's copy of the tooling only (`environment/sim/rcb_traj.py`, one line of `environment/sim/server.py`). The tooling now refuses to save or replay a trajectory of more than 200,000 rows, the privileged verifier's cap; a standard episode has no row cap, so the service lifts it for its own session (`S.max_rows = None`). Nothing an agent sees and nothing the service grades changes.
- 2026-10-13 (board 0100): base image `eai-robotwin` 0.1.2, whose scrub of RoboTwin's expert also removes the reference poses that `load_actors` computed only for the expert (here `block1_target_pose`). The standard-mode agent never reaches the env, so nothing it observes changes; the scene, the instance and the judge are unchanged.
+- 2026-10-14 (board 0113): the tooling only (`environment/sim/rcb_traj.py`): while it builds the scene, `glob.glob` returns its matches sorted, so that a task which lists files to draw an object depends on its seed alone, not on the host's directory order (four RoboTwin tasks did; this one lists no files while it builds its scene). Its scene, frozen instance (same initial-state hash), verifier and pass rule are unchanged.
diff --git a/tasks/robotwin_2_0_stack_blocks_three-standard/environment/sim/rcb_traj.py b/tasks/robotwin_2_0_stack_blocks_three-standard/environment/sim/rcb_traj.py
index 0b564387de..0c6314c152 100644
--- a/tasks/robotwin_2_0_stack_blocks_three-standard/environment/sim/rcb_traj.py
+++ b/tasks/robotwin_2_0_stack_blocks_three-standard/environment/sim/rcb_traj.py
@@ -113,6 +113,26 @@ BASE_ARGS: dict[str, Any] = {
}
+class _SortedGlob:
+ """While a scene is built, glob.glob / glob.iglob return their matches sorted. Four RoboTwin tasks (place_a2b_left,
+ place_a2b_right, place_object_scale, place_object_stand: envs/<task>.py load_actors) list an object's model files
+ with glob.glob and then draw one with the seeded np.random.choice; glob returns the file system's directory order,
+ which differs between hosts (and image layers), so one seed built a different object on another host and the frozen
+ instance's hash check failed there (board 0086 claim 5; never rebuilt off the reference workstation). Sorted, the
+ draw depends on the seed alone. No other task of ours lists files while it builds its scene."""
+
+ def __enter__(self):
+ import glob
+ self._glob, self._orig = glob, (glob.glob, glob.iglob)
+ glob.glob = lambda *a, **k: sorted(self._orig[0](*a, **k))
+ glob.iglob = lambda *a, **k: iter(sorted(self._orig[1](*a, **k)))
+ return self
+
+ def __exit__(self, *exc):
+ self._glob.glob, self._glob.iglob = self._orig
+ return False
+
+
def _chdir_root() -> None:
# RoboTwin resolves assets/config paths relative to the repo root and via ROOT_PATH
if os.getcwd() != ROBOTWIN_ROOT and os.path.isdir(ROBOTWIN_ROOT):
@@ -218,7 +238,8 @@ class Session:
if self.enable_render:
sapien.render.set_camera_shader_dir(self.render_shader)
args = _build_args(self.task, self.embodiment, self.seed)
- self.env.setup_demo(**args)
+ with _SortedGlob(): # the object draw depends on the seed alone, not the host's directory order
+ self.env.setup_demo(**args)
if self.enable_render: # setup_scene() switches the shader to ray tracing; put it back for cheap frames
sapien.render.set_camera_shader_dir(self.render_shader)
self._raw_scene = self.env.scene
diff --git a/tasks/robotwin_2_0_stack_blocks_three-standard/task.toml b/tasks/robotwin_2_0_stack_blocks_three-standard/task.toml
index b9cc60134e..2450fd142e 100644
--- a/tasks/robotwin_2_0_stack_blocks_three-standard/task.toml
+++ b/tasks/robotwin_2_0_stack_blocks_three-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "robotwin"
robot = "aloha-agilex"
family = "stack_blocks_three"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 60417
diff --git a/tasks/robotwin_2_0_stack_blocks_two-privileged/README.md b/tasks/robotwin_2_0_stack_blocks_two-privileged/README.md
index 04111dd3bd..afa212b3ed 100644
--- a/tasks/robotwin_2_0_stack_blocks_two-privileged/README.md
+++ b/tasks/robotwin_2_0_stack_blocks_two-privileged/README.md
@@ -55,3 +55,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
- 2026-10-13 (board 0086): the tooling refuses a trajectory the verifier cannot grade (`environment/rcb_traj.py`, `tests/rcb_traj.py`). `Session.save_trajectory` and `Session.replay` raise `ValueError` for more than 200,000 rows (`rcb_traj.MAX_T`, the verifier's `MAX_T` in `tests/grade.py`: 800 s of physics), and `record_stop` warns on stderr. Before, the tool saved and replayed such a file ("success") and the verifier then graded it `bad_trajectory`; the instruction has stated the cap since 2026-10-10. The verifier, the scene, the instance and the pass rule are unchanged: every trajectory is graded as before.
; the reference solution never used it, and `check_success` does not read it. The scene, the frozen instance (same initial-state hash), the tooling, the verifier and the pass rule are unchanged, so the task keeps its name; privileged results from before this date were obtained with this value reachable.
- 2026-10-13 (board 0100, owner decision): base image `eai-robotwin` 0.1.2. The image's scrub of RoboTwin's expert kept `load_actors` whole, so a value it computes for the expert stayed reachable: `env.block1_target_pose` (a pose for the expert to move to or place at; unused upstream). It is gone: the env no longer has it, `/opt/RoboTwin/envs/stack_blocks_two.py` no longer shows it; the reference solution never used it, and `check_success` does not read it. The scene, the frozen instance (same initial-state hash), the tooling, the verifier and the pass rule are unchanged, so the task keeps its name; privileged results from before this date were obtained with this value reachable.
+- 2026-10-14 (board 0113): the tooling only (`environment/rcb_traj.py`, `tests/rcb_traj.py`): while it builds the scene, `glob.glob` returns its matches sorted, so that a task which lists files to draw an object depends on its seed alone, not on the host's directory order (four RoboTwin tasks did; this one lists no files while it builds its scene). Its scene, frozen instance (same initial-state hash), verifier and pass rule are unchanged.
diff --git a/tasks/robotwin_2_0_stack_blocks_two-privileged/environment/rcb_traj.py b/tasks/robotwin_2_0_stack_blocks_two-privileged/environment/rcb_traj.py
index 27fcbfa20f..7ea0ead342 100644
--- a/tasks/robotwin_2_0_stack_blocks_two-privileged/environment/rcb_traj.py
+++ b/tasks/robotwin_2_0_stack_blocks_two-privileged/environment/rcb_traj.py
@@ -112,6 +112,26 @@ BASE_ARGS: dict[str, Any] = {
}
+class _SortedGlob:
+ """While a scene is built, glob.glob / glob.iglob return their matches sorted. Four RoboTwin tasks (place_a2b_left,
+ place_a2b_right, place_object_scale, place_object_stand: envs/<task>.py load_actors) list an object's model files
+ with glob.glob and then draw one with the seeded np.random.choice; glob returns the file system's directory order,
+ which differs between hosts (and image layers), so one seed built a different object on another host and the frozen
+ instance's hash check failed there (board 0086 claim 5; never rebuilt off the reference workstation). Sorted, the
+ draw depends on the seed alone. No other task of ours lists files while it builds its scene."""
+
+ def __enter__(self):
+ import glob
+ self._glob, self._orig = glob, (glob.glob, glob.iglob)
+ glob.glob = lambda *a, **k: sorted(self._orig[0](*a, **k))
+ glob.iglob = lambda *a, **k: iter(sorted(self._orig[1](*a, **k)))
+ return self
+
+ def __exit__(self, *exc):
+ self._glob.glob, self._glob.iglob = self._orig
+ return False
+
+
def _chdir_root() -> None:
# RoboTwin resolves assets/config paths relative to the repo root and via ROOT_PATH
if os.getcwd() != ROBOTWIN_ROOT and os.path.isdir(ROBOTWIN_ROOT):
@@ -217,7 +237,8 @@ class Session:
if self.enable_render:
sapien.render.set_camera_shader_dir(self.render_shader)
args = _build_args(self.task, self.embodiment, self.seed)
- self.env.setup_demo(**args)
+ with _SortedGlob(): # the object draw depends on the seed alone, not the host's directory order
+ self.env.setup_demo(**args)
if self.enable_render: # setup_scene() switches the shader to ray tracing; put it back for cheap frames
sapien.render.set_camera_shader_dir(self.render_shader)
self._raw_scene = self.env.scene
diff --git a/tasks/robotwin_2_0_stack_blocks_two-privileged/task.toml b/tasks/robotwin_2_0_stack_blocks_two-privileged/task.toml
index 07d0902ca8..c5d2be5bc4 100644
--- a/tasks/robotwin_2_0_stack_blocks_two-privileged/task.toml
+++ b/tasks/robotwin_2_0_stack_blocks_two-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "robotwin" # RoboTwin 2.0 = SAPIEN 3 / PhysX + CuRobo
robot = "aloha-agilex"
family = "stack_blocks_two"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 60417 # scene = environment/instance.json (RoboTwin seed + expected initial-state hash); identity = dataset.toml digest
expert_time_estimate_hours = 0.5
diff --git a/tasks/robotwin_2_0_stack_blocks_two-privileged/tests/rcb_traj.py b/tasks/robotwin_2_0_stack_blocks_two-privileged/tests/rcb_traj.py
index 27fcbfa20f..7ea0ead342 100644
--- a/tasks/robotwin_2_0_stack_blocks_two-privileged/tests/rcb_traj.py
+++ b/tasks/robotwin_2_0_stack_blocks_two-privileged/tests/rcb_traj.py
@@ -112,6 +112,26 @@ BASE_ARGS: dict[str, Any] = {
}
+class _SortedGlob:
+ """While a scene is built, glob.glob / glob.iglob return their matches sorted. Four RoboTwin tasks (place_a2b_left,
+ place_a2b_right, place_object_scale, place_object_stand: envs/<task>.py load_actors) list an object's model files
+ with glob.glob and then draw one with the seeded np.random.choice; glob returns the file system's directory order,
+ which differs between hosts (and image layers), so one seed built a different object on another host and the frozen
+ instance's hash check failed there (board 0086 claim 5; never rebuilt off the reference workstation). Sorted, the
+ draw depends on the seed alone. No other task of ours lists files while it builds its scene."""
+
+ def __enter__(self):
+ import glob
+ self._glob, self._orig = glob, (glob.glob, glob.iglob)
+ glob.glob = lambda *a, **k: sorted(self._orig[0](*a, **k))
+ glob.iglob = lambda *a, **k: iter(sorted(self._orig[1](*a, **k)))
+ return self
+
+ def __exit__(self, *exc):
+ self._glob.glob, self._glob.iglob = self._orig
+ return False
+
+
def _chdir_root() -> None:
# RoboTwin resolves assets/config paths relative to the repo root and via ROOT_PATH
if os.getcwd() != ROBOTWIN_ROOT and os.path.isdir(ROBOTWIN_ROOT):
@@ -217,7 +237,8 @@ class Session:
if self.enable_render:
sapien.render.set_camera_shader_dir(self.render_shader)
args = _build_args(self.task, self.embodiment, self.seed)
- self.env.setup_demo(**args)
+ with _SortedGlob(): # the object draw depends on the seed alone, not the host's directory order
+ self.env.setup_demo(**args)
if self.enable_render: # setup_scene() switches the shader to ray tracing; put it back for cheap frames
sapien.render.set_camera_shader_dir(self.render_shader)
self._raw_scene = self.env.scene
diff --git a/tasks/robotwin_2_0_stack_blocks_two-standard/README.md b/tasks/robotwin_2_0_stack_blocks_two-standard/README.md
index 8fb9d19045..29e8681f4b 100644
--- a/tasks/robotwin_2_0_stack_blocks_two-standard/README.md
+++ b/tasks/robotwin_2_0_stack_blocks_two-standard/README.md
@@ -52,3 +52,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
- 2026-10-10 (owner decision, board 0072): the agent's network egress is restricted to the model APIs (`[agent] network_mode = "allowlist"`, the hosts of `scripts/set_agent_egress.py`; since 2026-10-11 these include the package registries, PyPI and the Debian / Ubuntu archives, so that `pip install` and `apt-get install` work); it was open before (Harbor's default), and agents could download upstream sources, reference solutions and robot models from their own shell. The scene, the frozen instance, the tooling and the pass rule are unchanged, so the task keeps its name; results from before this date were obtained with open egress.
- 2026-10-13 (board 0086): the service's copy of the tooling only (`environment/sim/rcb_traj.py`, one line of `environment/sim/server.py`). The tooling now refuses to save or replay a trajectory of more than 200,000 rows, the privileged verifier's cap; a standard episode has no row cap, so the service lifts it for its own session (`S.max_rows = None`). Nothing an agent sees and nothing the service grades changes.
- 2026-10-13 (board 0100): base image `eai-robotwin` 0.1.2, whose scrub of RoboTwin's expert also removes the reference poses that `load_actors` computed only for the expert (here `block1_target_pose`). The standard-mode agent never reaches the env, so nothing it observes changes; the scene, the instance and the judge are unchanged.
+- 2026-10-14 (board 0113): the tooling only (`environment/sim/rcb_traj.py`): while it builds the scene, `glob.glob` returns its matches sorted, so that a task which lists files to draw an object depends on its seed alone, not on the host's directory order (four RoboTwin tasks did; this one lists no files while it builds its scene). Its scene, frozen instance (same initial-state hash), verifier and pass rule are unchanged.
diff --git a/tasks/robotwin_2_0_stack_blocks_two-standard/environment/sim/rcb_traj.py b/tasks/robotwin_2_0_stack_blocks_two-standard/environment/sim/rcb_traj.py
index 27fcbfa20f..7ea0ead342 100644
--- a/tasks/robotwin_2_0_stack_blocks_two-standard/environment/sim/rcb_traj.py
+++ b/tasks/robotwin_2_0_stack_blocks_two-standard/environment/sim/rcb_traj.py
@@ -112,6 +112,26 @@ BASE_ARGS: dict[str, Any] = {
}
+class _SortedGlob:
+ """While a scene is built, glob.glob / glob.iglob return their matches sorted. Four RoboTwin tasks (place_a2b_left,
+ place_a2b_right, place_object_scale, place_object_stand: envs/<task>.py load_actors) list an object's model files
+ with glob.glob and then draw one with the seeded np.random.choice; glob returns the file system's directory order,
+ which differs between hosts (and image layers), so one seed built a different object on another host and the frozen
+ instance's hash check failed there (board 0086 claim 5; never rebuilt off the reference workstation). Sorted, the
+ draw depends on the seed alone. No other task of ours lists files while it builds its scene."""
+
+ def __enter__(self):
+ import glob
+ self._glob, self._orig = glob, (glob.glob, glob.iglob)
+ glob.glob = lambda *a, **k: sorted(self._orig[0](*a, **k))
+ glob.iglob = lambda *a, **k: iter(sorted(self._orig[1](*a, **k)))
+ return self
+
+ def __exit__(self, *exc):
+ self._glob.glob, self._glob.iglob = self._orig
+ return False
+
+
def _chdir_root() -> None:
# RoboTwin resolves assets/config paths relative to the repo root and via ROOT_PATH
if os.getcwd() != ROBOTWIN_ROOT and os.path.isdir(ROBOTWIN_ROOT):
@@ -217,7 +237,8 @@ class Session:
if self.enable_render:
sapien.render.set_camera_shader_dir(self.render_shader)
args = _build_args(self.task, self.embodiment, self.seed)
- self.env.setup_demo(**args)
+ with _SortedGlob(): # the object draw depends on the seed alone, not the host's directory order
+ self.env.setup_demo(**args)
if self.enable_render: # setup_scene() switches the shader to ray tracing; put it back for cheap frames
sapien.render.set_camera_shader_dir(self.render_shader)
self._raw_scene = self.env.scene
diff --git a/tasks/robotwin_2_0_stack_blocks_two-standard/task.toml b/tasks/robotwin_2_0_stack_blocks_two-standard/task.toml
index 848c83905f..935b640692 100644
--- a/tasks/robotwin_2_0_stack_blocks_two-standard/task.toml
+++ b/tasks/robotwin_2_0_stack_blocks_two-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "robotwin"
robot = "aloha-agilex"
family = "stack_blocks_two"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 60417
diff --git a/tasks/robotwin_2_0_stack_bowls_three-privileged/README.md b/tasks/robotwin_2_0_stack_bowls_three-privileged/README.md
index 457c45c1fa..9936fb93c2 100644
--- a/tasks/robotwin_2_0_stack_bowls_three-privileged/README.md
+++ b/tasks/robotwin_2_0_stack_bowls_three-privileged/README.md
@@ -56,3 +56,4 @@ Stack the three bowls on top of each other.
- 2026-10-13 (board 0086): the tooling refuses a trajectory the verifier cannot grade (`environment/rcb_traj.py`, `tests/rcb_traj.py`). `Session.save_trajectory` and `Session.replay` raise `ValueError` for more than 200,000 rows (`rcb_traj.MAX_T`, the verifier's `MAX_T` in `tests/grade.py`: 800 s of physics), and `record_stop` warns on stderr. Before, the tool saved and replayed such a file ("success") and the verifier then graded it `bad_trajectory`; the instruction has stated the cap since 2026-10-10. The verifier, the scene, the instance and the pass rule are unchanged: every trajectory is graded as before.
; `solution/` rebuilds the scene with upstream's `load_actors` to get them back for the expert. None of them is read by `check_success`. The scene, the frozen instance (same initial-state hash), the tooling, the verifier and the pass rule are unchanged, so the task keeps its name; privileged results from before this date were obtained with these values reachable.
- 2026-10-13 (board 0100, owner decision): base image `eai-robotwin` 0.1.2. The image's scrub of RoboTwin's expert kept `load_actors` whole, so values it computes only for the removed expert stayed reachable: `env.bowl1_target_pose`, `env.quat_of_target_pose` (poses the expert moves to or places at), and the instruction listed them among the scene objects. They are gone: the env no longer has them, `/opt/RoboTwin/envs/stack_bowls_three.py` no longer shows them, and the scene-object list no longer names them; `solution/` rebuilds the scene with upstream's `load_actors` to get them back for the expert. None of them is read by `check_success`. The scene, the frozen instance (same initial-state hash), the tooling, the verifier and the pass rule are unchanged, so the task keeps its name; privileged results from before this date were obtained with these values reachable.
+- 2026-10-14 (board 0113): the tooling only (`environment/rcb_traj.py`, `tests/rcb_traj.py`): while it builds the scene, `glob.glob` returns its matches sorted, so that a task which lists files to draw an object depends on its seed alone, not on the host's directory order (four RoboTwin tasks did; this one lists no files while it builds its scene). Its scene, frozen instance (same initial-state hash), verifier and pass rule are unchanged.
diff --git a/tasks/robotwin_2_0_stack_bowls_three-privileged/environment/rcb_traj.py b/tasks/robotwin_2_0_stack_bowls_three-privileged/environment/rcb_traj.py
index 46b36ace8b..92a79e7c5e 100644
--- a/tasks/robotwin_2_0_stack_bowls_three-privileged/environment/rcb_traj.py
+++ b/tasks/robotwin_2_0_stack_bowls_three-privileged/environment/rcb_traj.py
@@ -113,6 +113,26 @@ BASE_ARGS: dict[str, Any] = {
}
+class _SortedGlob:
+ """While a scene is built, glob.glob / glob.iglob return their matches sorted. Four RoboTwin tasks (place_a2b_left,
+ place_a2b_right, place_object_scale, place_object_stand: envs/<task>.py load_actors) list an object's model files
+ with glob.glob and then draw one with the seeded np.random.choice; glob returns the file system's directory order,
+ which differs between hosts (and image layers), so one seed built a different object on another host and the frozen
+ instance's hash check failed there (board 0086 claim 5; never rebuilt off the reference workstation). Sorted, the
+ draw depends on the seed alone. No other task of ours lists files while it builds its scene."""
+
+ def __enter__(self):
+ import glob
+ self._glob, self._orig = glob, (glob.glob, glob.iglob)
+ glob.glob = lambda *a, **k: sorted(self._orig[0](*a, **k))
+ glob.iglob = lambda *a, **k: iter(sorted(self._orig[1](*a, **k)))
+ return self
+
+ def __exit__(self, *exc):
+ self._glob.glob, self._glob.iglob = self._orig
+ return False
+
+
def _chdir_root() -> None:
# RoboTwin resolves assets/config paths relative to the repo root and via ROOT_PATH
if os.getcwd() != ROBOTWIN_ROOT and os.path.isdir(ROBOTWIN_ROOT):
@@ -218,7 +238,8 @@ class Session:
if self.enable_render:
sapien.render.set_camera_shader_dir(self.render_shader)
args = _build_args(self.task, self.embodiment, self.seed)
- self.env.setup_demo(**args)
+ with _SortedGlob(): # the object draw depends on the seed alone, not the host's directory order
+ self.env.setup_demo(**args)
if self.enable_render: # setup_scene() switches the shader to ray tracing; put it back for cheap frames
sapien.render.set_camera_shader_dir(self.render_shader)
self._raw_scene = self.env.scene
diff --git a/tasks/robotwin_2_0_stack_bowls_three-privileged/task.toml b/tasks/robotwin_2_0_stack_bowls_three-privileged/task.toml
index 5a4f55c1ca..3849720028 100644
--- a/tasks/robotwin_2_0_stack_bowls_three-privileged/task.toml
+++ b/tasks/robotwin_2_0_stack_bowls_three-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "robotwin" # RoboTwin 2.0 = SAPIEN 3 / PhysX + CuRobo
robot = "aloha-agilex"
family = "stack_bowls_three"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 60417 # scene = environment/instance.json (RoboTwin seed + expected initial-state hash); identity = dataset.toml digest
expert_time_estimate_hours = 0.5
diff --git a/tasks/robotwin_2_0_stack_bowls_three-privileged/tests/rcb_traj.py b/tasks/robotwin_2_0_stack_bowls_three-privileged/tests/rcb_traj.py
index 46b36ace8b..92a79e7c5e 100644
--- a/tasks/robotwin_2_0_stack_bowls_three-privileged/tests/rcb_traj.py
+++ b/tasks/robotwin_2_0_stack_bowls_three-privileged/tests/rcb_traj.py
@@ -113,6 +113,26 @@ BASE_ARGS: dict[str, Any] = {
}
+class _SortedGlob:
+ """While a scene is built, glob.glob / glob.iglob return their matches sorted. Four RoboTwin tasks (place_a2b_left,
+ place_a2b_right, place_object_scale, place_object_stand: envs/<task>.py load_actors) list an object's model files
+ with glob.glob and then draw one with the seeded np.random.choice; glob returns the file system's directory order,
+ which differs between hosts (and image layers), so one seed built a different object on another host and the frozen
+ instance's hash check failed there (board 0086 claim 5; never rebuilt off the reference workstation). Sorted, the
+ draw depends on the seed alone. No other task of ours lists files while it builds its scene."""
+
+ def __enter__(self):
+ import glob
+ self._glob, self._orig = glob, (glob.glob, glob.iglob)
+ glob.glob = lambda *a, **k: sorted(self._orig[0](*a, **k))
+ glob.iglob = lambda *a, **k: iter(sorted(self._orig[1](*a, **k)))
+ return self
+
+ def __exit__(self, *exc):
+ self._glob.glob, self._glob.iglob = self._orig
+ return False
+
+
def _chdir_root() -> None:
# RoboTwin resolves assets/config paths relative to the repo root and via ROOT_PATH
if os.getcwd() != ROBOTWIN_ROOT and os.path.isdir(ROBOTWIN_ROOT):
@@ -218,7 +238,8 @@ class Session:
if self.enable_render:
sapien.render.set_camera_shader_dir(self.render_shader)
args = _build_args(self.task, self.embodiment, self.seed)
- self.env.setup_demo(**args)
+ with _SortedGlob(): # the object draw depends on the seed alone, not the host's directory order
+ self.env.setup_demo(**args)
if self.enable_render: # setup_scene() switches the shader to ray tracing; put it back for cheap frames
sapien.render.set_camera_shader_dir(self.render_shader)
self._raw_scene = self.env.scene
diff --git a/tasks/robotwin_2_0_stack_bowls_three-standard/README.md b/tasks/robotwin_2_0_stack_bowls_three-standard/README.md
index 70fe63be72..242640c859 100644
--- a/tasks/robotwin_2_0_stack_bowls_three-standard/README.md
+++ b/tasks/robotwin_2_0_stack_bowls_three-standard/README.md
@@ -53,3 +53,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
- 2026-10-10 (owner decision, board 0072): the agent's network egress is restricted to the model APIs (`[agent] network_mode = "allowlist"`, the hosts of `scripts/set_agent_egress.py`; since 2026-10-11 these include the package registries, PyPI and the Debian / Ubuntu archives, so that `pip install` and `apt-get install` work); it was open before (Harbor's default), and agents could download upstream sources, reference solutions and robot models from their own shell. The scene, the frozen instance, the tooling and the pass rule are unchanged, so the task keeps its name; results from before this date were obtained with open egress.
- 2026-10-13 (board 0086): the service's copy of the tooling only (`environment/sim/rcb_traj.py`, one line of `environment/sim/server.py`). The tooling now refuses to save or replay a trajectory of more than 200,000 rows, the privileged verifier's cap; a standard episode has no row cap, so the service lifts it for its own session (`S.max_rows = None`). Nothing an agent sees and nothing the service grades changes.
- 2026-10-13 (board 0100): base image `eai-robotwin` 0.1.2, whose scrub of RoboTwin's expert also removes the reference poses that `load_actors` computed only for the expert (here `bowl1_target_pose`, `quat_of_target_pose`). The standard-mode agent never reaches the env, so nothing it observes changes; the scene, the instance and the judge are unchanged.
+- 2026-10-14 (board 0113): the tooling only (`environment/sim/rcb_traj.py`): while it builds the scene, `glob.glob` returns its matches sorted, so that a task which lists files to draw an object depends on its seed alone, not on the host's directory order (four RoboTwin tasks did; this one lists no files while it builds its scene). Its scene, frozen instance (same initial-state hash), verifier and pass rule are unchanged.
diff --git a/tasks/robotwin_2_0_stack_bowls_three-standard/environment/sim/rcb_traj.py b/tasks/robotwin_2_0_stack_bowls_three-standard/environment/sim/rcb_traj.py
index 46b36ace8b..92a79e7c5e 100644
--- a/tasks/robotwin_2_0_stack_bowls_three-standard/environment/sim/rcb_traj.py
+++ b/tasks/robotwin_2_0_stack_bowls_three-standard/environment/sim/rcb_traj.py
@@ -113,6 +113,26 @@ BASE_ARGS: dict[str, Any] = {
}
+class _SortedGlob:
+ """While a scene is built, glob.glob / glob.iglob return their matches sorted. Four RoboTwin tasks (place_a2b_left,
+ place_a2b_right, place_object_scale, place_object_stand: envs/<task>.py load_actors) list an object's model files
+ with glob.glob and then draw one with the seeded np.random.choice; glob returns the file system's directory order,
+ which differs between hosts (and image layers), so one seed built a different object on another host and the frozen
+ instance's hash check failed there (board 0086 claim 5; never rebuilt off the reference workstation). Sorted, the
+ draw depends on the seed alone. No other task of ours lists files while it builds its scene."""
+
+ def __enter__(self):
+ import glob
+ self._glob, self._orig = glob, (glob.glob, glob.iglob)
+ glob.glob = lambda *a, **k: sorted(self._orig[0](*a, **k))
+ glob.iglob = lambda *a, **k: iter(sorted(self._orig[1](*a, **k)))
+ return self
+
+ def __exit__(self, *exc):
+ self._glob.glob, self._glob.iglob = self._orig
+ return False
+
+
def _chdir_root() -> None:
# RoboTwin resolves assets/config paths relative to the repo root and via ROOT_PATH
if os.getcwd() != ROBOTWIN_ROOT and os.path.isdir(ROBOTWIN_ROOT):
@@ -218,7 +238,8 @@ class Session:
if self.enable_render:
sapien.render.set_camera_shader_dir(self.render_shader)
args = _build_args(self.task, self.embodiment, self.seed)
- self.env.setup_demo(**args)
+ with _SortedGlob(): # the object draw depends on the seed alone, not the host's directory order
+ self.env.setup_demo(**args)
if self.enable_render: # setup_scene() switches the shader to ray tracing; put it back for cheap frames
sapien.render.set_camera_shader_dir(self.render_shader)
self._raw_scene = self.env.scene
diff --git a/tasks/robotwin_2_0_stack_bowls_three-standard/task.toml b/tasks/robotwin_2_0_stack_bowls_three-standard/task.toml
index 29a11333a8..a45f4723fc 100644
--- a/tasks/robotwin_2_0_stack_bowls_three-standard/task.toml
+++ b/tasks/robotwin_2_0_stack_bowls_three-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "robotwin"
robot = "aloha-agilex"
family = "stack_bowls_three"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 60417
diff --git a/tasks/robotwin_2_0_stack_bowls_two-privileged/README.md b/tasks/robotwin_2_0_stack_bowls_two-privileged/README.md
index dd8df3ee31..d73aaff7c6 100644
--- a/tasks/robotwin_2_0_stack_bowls_two-privileged/README.md
+++ b/tasks/robotwin_2_0_stack_bowls_two-privileged/README.md
@@ -56,3 +56,4 @@ Stack the two bowls on top of each other.
- 2026-10-13 (board 0086): the tooling refuses a trajectory the verifier cannot grade (`environment/rcb_traj.py`, `tests/rcb_traj.py`). `Session.save_trajectory` and `Session.replay` raise `ValueError` for more than 200,000 rows (`rcb_traj.MAX_T`, the verifier's `MAX_T` in `tests/grade.py`: 800 s of physics), and `record_stop` warns on stderr. Before, the tool saved and replayed such a file ("success") and the verifier then graded it `bad_trajectory`; the instruction has stated the cap since 2026-10-10. The verifier, the scene, the instance and the pass rule are unchanged: every trajectory is graded as before.
; `solution/` rebuilds the scene with upstream's `load_actors` to get them back for the expert. None of them is read by `check_success`. The scene, the frozen instance (same initial-state hash), the tooling, the verifier and the pass rule are unchanged, so the task keeps its name; privileged results from before this date were obtained with these values reachable.
- 2026-10-13 (board 0100, owner decision): base image `eai-robotwin` 0.1.2. The image's scrub of RoboTwin's expert kept `load_actors` whole, so values it computes only for the removed expert stayed reachable: `env.bowl1_target_pose`, `env.quat_of_target_pose` (poses the expert moves to or places at), and the instruction listed them among the scene objects. They are gone: the env no longer has them, `/opt/RoboTwin/envs/stack_bowls_two.py` no longer shows them, and the scene-object list no longer names them; `solution/` rebuilds the scene with upstream's `load_actors` to get them back for the expert. None of them is read by `check_success`. The scene, the frozen instance (same initial-state hash), the tooling, the verifier and the pass rule are unchanged, so the task keeps its name; privileged results from before this date were obtained with these values reachable.
+- 2026-10-14 (board 0113): the tooling only (`environment/rcb_traj.py`, `tests/rcb_traj.py`): while it builds the scene, `glob.glob` returns its matches sorted, so that a task which lists files to draw an object depends on its seed alone, not on the host's directory order (four RoboTwin tasks did; this one lists no files while it builds its scene). Its scene, frozen instance (same initial-state hash), verifier and pass rule are unchanged.
diff --git a/tasks/robotwin_2_0_stack_bowls_two-privileged/environment/rcb_traj.py b/tasks/robotwin_2_0_stack_bowls_two-privileged/environment/rcb_traj.py
index 6d5e412be6..87e9f7b426 100644
--- a/tasks/robotwin_2_0_stack_bowls_two-privileged/environment/rcb_traj.py
+++ b/tasks/robotwin_2_0_stack_bowls_two-privileged/environment/rcb_traj.py
@@ -113,6 +113,26 @@ BASE_ARGS: dict[str, Any] = {
}
+class _SortedGlob:
+ """While a scene is built, glob.glob / glob.iglob return their matches sorted. Four RoboTwin tasks (place_a2b_left,
+ place_a2b_right, place_object_scale, place_object_stand: envs/<task>.py load_actors) list an object's model files
+ with glob.glob and then draw one with the seeded np.random.choice; glob returns the file system's directory order,
+ which differs between hosts (and image layers), so one seed built a different object on another host and the frozen
+ instance's hash check failed there (board 0086 claim 5; never rebuilt off the reference workstation). Sorted, the
+ draw depends on the seed alone. No other task of ours lists files while it builds its scene."""
+
+ def __enter__(self):
+ import glob
+ self._glob, self._orig = glob, (glob.glob, glob.iglob)
+ glob.glob = lambda *a, **k: sorted(self._orig[0](*a, **k))
+ glob.iglob = lambda *a, **k: iter(sorted(self._orig[1](*a, **k)))
+ return self
+
+ def __exit__(self, *exc):
+ self._glob.glob, self._glob.iglob = self._orig
+ return False
+
+
def _chdir_root() -> None:
# RoboTwin resolves assets/config paths relative to the repo root and via ROOT_PATH
if os.getcwd() != ROBOTWIN_ROOT and os.path.isdir(ROBOTWIN_ROOT):
@@ -218,7 +238,8 @@ class Session:
if self.enable_render:
sapien.render.set_camera_shader_dir(self.render_shader)
args = _build_args(self.task, self.embodiment, self.seed)
- self.env.setup_demo(**args)
+ with _SortedGlob(): # the object draw depends on the seed alone, not the host's directory order
+ self.env.setup_demo(**args)
if self.enable_render: # setup_scene() switches the shader to ray tracing; put it back for cheap frames
sapien.render.set_camera_shader_dir(self.render_shader)
self._raw_scene = self.env.scene
diff --git a/tasks/robotwin_2_0_stack_bowls_two-privileged/task.toml b/tasks/robotwin_2_0_stack_bowls_two-privileged/task.toml
index 2ffe064be5..4dba6f3e86 100644
--- a/tasks/robotwin_2_0_stack_bowls_two-privileged/task.toml
+++ b/tasks/robotwin_2_0_stack_bowls_two-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "robotwin" # RoboTwin 2.0 = SAPIEN 3 / PhysX + CuRobo
robot = "aloha-agilex"
family = "stack_bowls_two"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 60417 # scene = environment/instance.json (RoboTwin seed + expected initial-state hash); identity = dataset.toml digest
expert_time_estimate_hours = 0.5
diff --git a/tasks/robotwin_2_0_stack_bowls_two-privileged/tests/rcb_traj.py b/tasks/robotwin_2_0_stack_bowls_two-privileged/tests/rcb_traj.py
index 6d5e412be6..87e9f7b426 100644
--- a/tasks/robotwin_2_0_stack_bowls_two-privileged/tests/rcb_traj.py
+++ b/tasks/robotwin_2_0_stack_bowls_two-privileged/tests/rcb_traj.py
@@ -113,6 +113,26 @@ BASE_ARGS: dict[str, Any] = {
}
+class _SortedGlob:
+ """While a scene is built, glob.glob / glob.iglob return their matches sorted. Four RoboTwin tasks (place_a2b_left,
+ place_a2b_right, place_object_scale, place_object_stand: envs/<task>.py load_actors) list an object's model files
+ with glob.glob and then draw one with the seeded np.random.choice; glob returns the file system's directory order,
+ which differs between hosts (and image layers), so one seed built a different object on another host and the frozen
+ instance's hash check failed there (board 0086 claim 5; never rebuilt off the reference workstation). Sorted, the
+ draw depends on the seed alone. No other task of ours lists files while it builds its scene."""
+
+ def __enter__(self):
+ import glob
+ self._glob, self._orig = glob, (glob.glob, glob.iglob)
+ glob.glob = lambda *a, **k: sorted(self._orig[0](*a, **k))
+ glob.iglob = lambda *a, **k: iter(sorted(self._orig[1](*a, **k)))
+ return self
+
+ def __exit__(self, *exc):
+ self._glob.glob, self._glob.iglob = self._orig
+ return False
+
+
def _chdir_root() -> None:
# RoboTwin resolves assets/config paths relative to the repo root and via ROOT_PATH
if os.getcwd() != ROBOTWIN_ROOT and os.path.isdir(ROBOTWIN_ROOT):
@@ -218,7 +238,8 @@ class Session:
if self.enable_render:
sapien.render.set_camera_shader_dir(self.render_shader)
args = _build_args(self.task, self.embodiment, self.seed)
- self.env.setup_demo(**args)
+ with _SortedGlob(): # the object draw depends on the seed alone, not the host's directory order
+ self.env.setup_demo(**args)
if self.enable_render: # setup_scene() switches the shader to ray tracing; put it back for cheap frames
sapien.render.set_camera_shader_dir(self.render_shader)
self._raw_scene = self.env.scene
diff --git a/tasks/robotwin_2_0_stack_bowls_two-standard/README.md b/tasks/robotwin_2_0_stack_bowls_two-standard/README.md
index 3cad683480..85a16d3b16 100644
--- a/tasks/robotwin_2_0_stack_bowls_two-standard/README.md
+++ b/tasks/robotwin_2_0_stack_bowls_two-standard/README.md
@@ -53,3 +53,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
- 2026-10-10 (owner decision, board 0072): the agent's network egress is restricted to the model APIs (`[agent] network_mode = "allowlist"`, the hosts of `scripts/set_agent_egress.py`; since 2026-10-11 these include the package registries, PyPI and the Debian / Ubuntu archives, so that `pip install` and `apt-get install` work); it was open before (Harbor's default), and agents could download upstream sources, reference solutions and robot models from their own shell. The scene, the frozen instance, the tooling and the pass rule are unchanged, so the task keeps its name; results from before this date were obtained with open egress.
- 2026-10-13 (board 0086): the service's copy of the tooling only (`environment/sim/rcb_traj.py`, one line of `environment/sim/server.py`). The tooling now refuses to save or replay a trajectory of more than 200,000 rows, the privileged verifier's cap; a standard episode has no row cap, so the service lifts it for its own session (`S.max_rows = None`). Nothing an agent sees and nothing the service grades changes.
- 2026-10-13 (board 0100): base image `eai-robotwin` 0.1.2, whose scrub of RoboTwin's expert also removes the reference poses that `load_actors` computed only for the expert (here `bowl1_target_pose`, `quat_of_target_pose`). The standard-mode agent never reaches the env, so nothing it observes changes; the scene, the instance and the judge are unchanged.
+- 2026-10-14 (board 0113): the tooling only (`environment/sim/rcb_traj.py`): while it builds the scene, `glob.glob` returns its matches sorted, so that a task which lists files to draw an object depends on its seed alone, not on the host's directory order (four RoboTwin tasks did; this one lists no files while it builds its scene). Its scene, frozen instance (same initial-state hash), verifier and pass rule are unchanged.
diff --git a/tasks/robotwin_2_0_stack_bowls_two-standard/environment/sim/rcb_traj.py b/tasks/robotwin_2_0_stack_bowls_two-standard/environment/sim/rcb_traj.py
index 6d5e412be6..87e9f7b426 100644
--- a/tasks/robotwin_2_0_stack_bowls_two-standard/environment/sim/rcb_traj.py
+++ b/tasks/robotwin_2_0_stack_bowls_two-standard/environment/sim/rcb_traj.py
@@ -113,6 +113,26 @@ BASE_ARGS: dict[str, Any] = {
}
+class _SortedGlob:
+ """While a scene is built, glob.glob / glob.iglob return their matches sorted. Four RoboTwin tasks (place_a2b_left,
+ place_a2b_right, place_object_scale, place_object_stand: envs/<task>.py load_actors) list an object's model files
+ with glob.glob and then draw one with the seeded np.random.choice; glob returns the file system's directory order,
+ which differs between hosts (and image layers), so one seed built a different object on another host and the frozen
+ instance's hash check failed there (board 0086 claim 5; never rebuilt off the reference workstation). Sorted, the
+ draw depends on the seed alone. No other task of ours lists files while it builds its scene."""
+
+ def __enter__(self):
+ import glob
+ self._glob, self._orig = glob, (glob.glob, glob.iglob)
+ glob.glob = lambda *a, **k: sorted(self._orig[0](*a, **k))
+ glob.iglob = lambda *a, **k: iter(sorted(self._orig[1](*a, **k)))
+ return self
+
+ def __exit__(self, *exc):
+ self._glob.glob, self._glob.iglob = self._orig
+ return False
+
+
def _chdir_root() -> None:
# RoboTwin resolves assets/config paths relative to the repo root and via ROOT_PATH
if os.getcwd() != ROBOTWIN_ROOT and os.path.isdir(ROBOTWIN_ROOT):
@@ -218,7 +238,8 @@ class Session:
if self.enable_render:
sapien.render.set_camera_shader_dir(self.render_shader)
args = _build_args(self.task, self.embodiment, self.seed)
- self.env.setup_demo(**args)
+ with _SortedGlob(): # the object draw depends on the seed alone, not the host's directory order
+ self.env.setup_demo(**args)
if self.enable_render: # setup_scene() switches the shader to ray tracing; put it back for cheap frames
sapien.render.set_camera_shader_dir(self.render_shader)
self._raw_scene = self.env.scene
diff --git a/tasks/robotwin_2_0_stack_bowls_two-standard/task.toml b/tasks/robotwin_2_0_stack_bowls_two-standard/task.toml
index de2bb3f5f9..7d2235317b 100644
--- a/tasks/robotwin_2_0_stack_bowls_two-standard/task.toml
+++ b/tasks/robotwin_2_0_stack_bowls_two-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "robotwin"
robot = "aloha-agilex"
family = "stack_bowls_two"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 60417
diff --git a/tasks/robotwin_2_0_stamp_seal-privileged/README.md b/tasks/robotwin_2_0_stamp_seal-privileged/README.md
index 9de60a451b..774c50d7a0 100644
--- a/tasks/robotwin_2_0_stamp_seal-privileged/README.md
+++ b/tasks/robotwin_2_0_stamp_seal-privileged/README.md
@@ -55,3 +55,4 @@ Grab the stamp and stamp onto the specific color mat.
- 2026-10-10 (owner decision, board 0072): the agent's network egress is restricted to the model APIs (`[agent] network_mode = "allowlist"`, the hosts of `scripts/set_agent_egress.py`; since 2026-10-11 these include the package registries, PyPI and the Debian / Ubuntu archives, so that `pip install` and `apt-get install` work); it was open before (Harbor's default), and agents could download upstream sources, reference solutions and robot models from their own shell. The scene, the frozen instance, the tooling and the pass rule are unchanged, so the task keeps its name; results from before this date were obtained with open egress.
- 2026-10-13 (board 0086): the tooling refuses a trajectory the verifier cannot grade (`environment/rcb_traj.py`, `tests/rcb_traj.py`). `Session.save_trajectory` and `Session.replay` raise `ValueError` for more than 200,000 rows (`rcb_traj.MAX_T`, the verifier's `MAX_T` in `tests/grade.py`: 800 s of physics), and `record_stop` warns on stderr. Before, the tool saved and replayed such a file ("success") and the verifier then graded it `bad_trajectory`; the instruction has stated the cap since 2026-10-10. The verifier, the scene, the instance and the pass rule are unchanged: every trajectory is graded as before.
- 2026-10-13 (board 0100): base image `eai-robotwin` 0.1.2. Its scrub of RoboTwin's expert also removes the reference poses that 18 tasks' `load_actors` computed only for the expert, and the expert's comments above the scrubbed method bodies (`images/robotwin/scrub_oracles.py`). This task's `load_actors` set no such value: its scene, instance, tooling, verifier and pass rule are unchanged.
+- 2026-10-14 (board 0113): the tooling only (`environment/rcb_traj.py`, `tests/rcb_traj.py`): while it builds the scene, `glob.glob` returns its matches sorted, so that a task which lists files to draw an object depends on its seed alone, not on the host's directory order (four RoboTwin tasks did; this one lists no files while it builds its scene). Its scene, frozen instance (same initial-state hash), verifier and pass rule are unchanged.
diff --git a/tasks/robotwin_2_0_stamp_seal-privileged/environment/rcb_traj.py b/tasks/robotwin_2_0_stamp_seal-privileged/environment/rcb_traj.py
index c6853cc9c9..278cac21d3 100644
--- a/tasks/robotwin_2_0_stamp_seal-privileged/environment/rcb_traj.py
+++ b/tasks/robotwin_2_0_stamp_seal-privileged/environment/rcb_traj.py
@@ -113,6 +113,26 @@ BASE_ARGS: dict[str, Any] = {
}
+class _SortedGlob:
+ """While a scene is built, glob.glob / glob.iglob return their matches sorted. Four RoboTwin tasks (place_a2b_left,
+ place_a2b_right, place_object_scale, place_object_stand: envs/<task>.py load_actors) list an object's model files
+ with glob.glob and then draw one with the seeded np.random.choice; glob returns the file system's directory order,
+ which differs between hosts (and image layers), so one seed built a different object on another host and the frozen
+ instance's hash check failed there (board 0086 claim 5; never rebuilt off the reference workstation). Sorted, the
+ draw depends on the seed alone. No other task of ours lists files while it builds its scene."""
+
+ def __enter__(self):
+ import glob
+ self._glob, self._orig = glob, (glob.glob, glob.iglob)
+ glob.glob = lambda *a, **k: sorted(self._orig[0](*a, **k))
+ glob.iglob = lambda *a, **k: iter(sorted(self._orig[1](*a, **k)))
+ return self
+
+ def __exit__(self, *exc):
+ self._glob.glob, self._glob.iglob = self._orig
+ return False
+
+
def _chdir_root() -> None:
# RoboTwin resolves assets/config paths relative to the repo root and via ROOT_PATH
if os.getcwd() != ROBOTWIN_ROOT and os.path.isdir(ROBOTWIN_ROOT):
@@ -218,7 +238,8 @@ class Session:
if self.enable_render:
sapien.render.set_camera_shader_dir(self.render_shader)
args = _build_args(self.task, self.embodiment, self.seed)
- self.env.setup_demo(**args)
+ with _SortedGlob(): # the object draw depends on the seed alone, not the host's directory order
+ self.env.setup_demo(**args)
if self.enable_render: # setup_scene() switches the shader to ray tracing; put it back for cheap frames
sapien.render.set_camera_shader_dir(self.render_shader)
self._raw_scene = self.env.scene
diff --git a/tasks/robotwin_2_0_stamp_seal-privileged/task.toml b/tasks/robotwin_2_0_stamp_seal-privileged/task.toml
index 2a02676d2e..78bf27739f 100644
--- a/tasks/robotwin_2_0_stamp_seal-privileged/task.toml
+++ b/tasks/robotwin_2_0_stamp_seal-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "robotwin" # RoboTwin 2.0 = SAPIEN 3 / PhysX + CuRobo
robot = "aloha-agilex"
family = "stamp_seal"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 91573 # scene = environment/instance.json (RoboTwin seed + expected initial-state hash); identity = dataset.toml digest
expert_time_estimate_hours = 0.5
diff --git a/tasks/robotwin_2_0_stamp_seal-privileged/tests/rcb_traj.py b/tasks/robotwin_2_0_stamp_seal-privileged/tests/rcb_traj.py
index c6853cc9c9..278cac21d3 100644
--- a/tasks/robotwin_2_0_stamp_seal-privileged/tests/rcb_traj.py
+++ b/tasks/robotwin_2_0_stamp_seal-privileged/tests/rcb_traj.py
@@ -113,6 +113,26 @@ BASE_ARGS: dict[str, Any] = {
}
+class _SortedGlob:
+ """While a scene is built, glob.glob / glob.iglob return their matches sorted. Four RoboTwin tasks (place_a2b_left,
+ place_a2b_right, place_object_scale, place_object_stand: envs/<task>.py load_actors) list an object's model files
+ with glob.glob and then draw one with the seeded np.random.choice; glob returns the file system's directory order,
+ which differs between hosts (and image layers), so one seed built a different object on another host and the frozen
+ instance's hash check failed there (board 0086 claim 5; never rebuilt off the reference workstation). Sorted, the
+ draw depends on the seed alone. No other task of ours lists files while it builds its scene."""
+
+ def __enter__(self):
+ import glob
+ self._glob, self._orig = glob, (glob.glob, glob.iglob)
+ glob.glob = lambda *a, **k: sorted(self._orig[0](*a, **k))
+ glob.iglob = lambda *a, **k: iter(sorted(self._orig[1](*a, **k)))
+ return self
+
+ def __exit__(self, *exc):
+ self._glob.glob, self._glob.iglob = self._orig
+ return False
+
+
def _chdir_root() -> None:
# RoboTwin resolves assets/config paths relative to the repo root and via ROOT_PATH
if os.getcwd() != ROBOTWIN_ROOT and os.path.isdir(ROBOTWIN_ROOT):
@@ -218,7 +238,8 @@ class Session:
if self.enable_render:
sapien.render.set_camera_shader_dir(self.render_shader)
args = _build_args(self.task, self.embodiment, self.seed)
- self.env.setup_demo(**args)
+ with _SortedGlob(): # the object draw depends on the seed alone, not the host's directory order
+ self.env.setup_demo(**args)
if self.enable_render: # setup_scene() switches the shader to ray tracing; put it back for cheap frames
sapien.render.set_camera_shader_dir(self.render_shader)
self._raw_scene = self.env.scene
diff --git a/tasks/robotwin_2_0_stamp_seal-standard/README.md b/tasks/robotwin_2_0_stamp_seal-standard/README.md
index 5ba854aac4..75ef859099 100644
--- a/tasks/robotwin_2_0_stamp_seal-standard/README.md
+++ b/tasks/robotwin_2_0_stamp_seal-standard/README.md
@@ -53,3 +53,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
- 2026-10-10 (owner decision, board 0072): the agent's network egress is restricted to the model APIs (`[agent] network_mode = "allowlist"`, the hosts of `scripts/set_agent_egress.py`; since 2026-10-11 these include the package registries, PyPI and the Debian / Ubuntu archives, so that `pip install` and `apt-get install` work); it was open before (Harbor's default), and agents could download upstream sources, reference solutions and robot models from their own shell. The scene, the frozen instance, the tooling and the pass rule are unchanged, so the task keeps its name; results from before this date were obtained with open egress.
- 2026-10-13 (board 0086): the service's copy of the tooling only (`environment/sim/rcb_traj.py`, one line of `environment/sim/server.py`). The tooling now refuses to save or replay a trajectory of more than 200,000 rows, the privileged verifier's cap; a standard episode has no row cap, so the service lifts it for its own session (`S.max_rows = None`). Nothing an agent sees and nothing the service grades changes.
- 2026-10-13 (board 0100): base image `eai-robotwin` 0.1.2, whose scrub of RoboTwin's expert also removes the reference poses that `load_actors` computed only for the expert. The standard-mode agent never reaches the env, so nothing it observes changes; the scene, the instance and the judge are unchanged.
+- 2026-10-14 (board 0113): the tooling only (`environment/sim/rcb_traj.py`): while it builds the scene, `glob.glob` returns its matches sorted, so that a task which lists files to draw an object depends on its seed alone, not on the host's directory order (four RoboTwin tasks did; this one lists no files while it builds its scene). Its scene, frozen instance (same initial-state hash), verifier and pass rule are unchanged.
diff --git a/tasks/robotwin_2_0_stamp_seal-standard/environment/sim/rcb_traj.py b/tasks/robotwin_2_0_stamp_seal-standard/environment/sim/rcb_traj.py
index c6853cc9c9..278cac21d3 100644
--- a/tasks/robotwin_2_0_stamp_seal-standard/environment/sim/rcb_traj.py
+++ b/tasks/robotwin_2_0_stamp_seal-standard/environment/sim/rcb_traj.py
@@ -113,6 +113,26 @@ BASE_ARGS: dict[str, Any] = {
}
+class _SortedGlob:
+ """While a scene is built, glob.glob / glob.iglob return their matches sorted. Four RoboTwin tasks (place_a2b_left,
+ place_a2b_right, place_object_scale, place_object_stand: envs/<task>.py load_actors) list an object's model files
+ with glob.glob and then draw one with the seeded np.random.choice; glob returns the file system's directory order,
+ which differs between hosts (and image layers), so one seed built a different object on another host and the frozen
+ instance's hash check failed there (board 0086 claim 5; never rebuilt off the reference workstation). Sorted, the
+ draw depends on the seed alone. No other task of ours lists files while it builds its scene."""
+
+ def __enter__(self):
+ import glob
+ self._glob, self._orig = glob, (glob.glob, glob.iglob)
+ glob.glob = lambda *a, **k: sorted(self._orig[0](*a, **k))
+ glob.iglob = lambda *a, **k: iter(sorted(self._orig[1](*a, **k)))
+ return self
+
+ def __exit__(self, *exc):
+ self._glob.glob, self._glob.iglob = self._orig
+ return False
+
+
def _chdir_root() -> None:
# RoboTwin resolves assets/config paths relative to the repo root and via ROOT_PATH
if os.getcwd() != ROBOTWIN_ROOT and os.path.isdir(ROBOTWIN_ROOT):
@@ -218,7 +238,8 @@ class Session:
if self.enable_render:
sapien.render.set_camera_shader_dir(self.render_shader)
args = _build_args(self.task, self.embodiment, self.seed)
- self.env.setup_demo(**args)
+ with _SortedGlob(): # the object draw depends on the seed alone, not the host's directory order
+ self.env.setup_demo(**args)
if self.enable_render: # setup_scene() switches the shader to ray tracing; put it back for cheap frames
sapien.render.set_camera_shader_dir(self.render_shader)
self._raw_scene = self.env.scene
diff --git a/tasks/robotwin_2_0_stamp_seal-standard/task.toml b/tasks/robotwin_2_0_stamp_seal-standard/task.toml
index 34f52064c7..6b80ff6d49 100644
--- a/tasks/robotwin_2_0_stamp_seal-standard/task.toml
+++ b/tasks/robotwin_2_0_stamp_seal-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "robotwin"
robot = "aloha-agilex"
family = "stamp_seal"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 91573
diff --git a/tasks/robotwin_2_0_turn_switch-privileged/README.md b/tasks/robotwin_2_0_turn_switch-privileged/README.md
index 88f64439c8..819d859085 100644
--- a/tasks/robotwin_2_0_turn_switch-privileged/README.md
+++ b/tasks/robotwin_2_0_turn_switch-privileged/README.md
@@ -55,3 +55,4 @@ Use the robotic arm to click the switch.
- 2026-10-10 (owner decision, board 0072): the agent's network egress is restricted to the model APIs (`[agent] network_mode = "allowlist"`, the hosts of `scripts/set_agent_egress.py`; since 2026-10-11 these include the package registries, PyPI and the Debian / Ubuntu archives, so that `pip install` and `apt-get install` work); it was open before (Harbor's default), and agents could download upstream sources, reference solutions and robot models from their own shell. The scene, the frozen instance, the tooling and the pass rule are unchanged, so the task keeps its name; results from before this date were obtained with open egress.
- 2026-10-13 (board 0086): the tooling refuses a trajectory the verifier cannot grade (`environment/rcb_traj.py`, `tests/rcb_traj.py`). `Session.save_trajectory` and `Session.replay` raise `ValueError` for more than 200,000 rows (`rcb_traj.MAX_T`, the verifier's `MAX_T` in `tests/grade.py`: 800 s of physics), and `record_stop` warns on stderr. Before, the tool saved and replayed such a file ("success") and the verifier then graded it `bad_trajectory`; the instruction has stated the cap since 2026-10-10. The verifier, the scene, the instance and the pass rule are unchanged: every trajectory is graded as before.
- 2026-10-13 (board 0100): base image `eai-robotwin` 0.1.2. Its scrub of RoboTwin's expert also removes the reference poses that 18 tasks' `load_actors` computed only for the expert, and the expert's comments above the scrubbed method bodies (`images/robotwin/scrub_oracles.py`). This task's `load_actors` set no such value: its scene, instance, tooling, verifier and pass rule are unchanged.
+- 2026-10-14 (board 0113): the tooling only (`environment/rcb_traj.py`, `tests/rcb_traj.py`): while it builds the scene, `glob.glob` returns its matches sorted, so that a task which lists files to draw an object depends on its seed alone, not on the host's directory order (four RoboTwin tasks did; this one lists no files while it builds its scene). Its scene, frozen instance (same initial-state hash), verifier and pass rule are unchanged.
diff --git a/tasks/robotwin_2_0_turn_switch-privileged/environment/rcb_traj.py b/tasks/robotwin_2_0_turn_switch-privileged/environment/rcb_traj.py
index 013426cf20..25d160b3ab 100644
--- a/tasks/robotwin_2_0_turn_switch-privileged/environment/rcb_traj.py
+++ b/tasks/robotwin_2_0_turn_switch-privileged/environment/rcb_traj.py
@@ -113,6 +113,26 @@ BASE_ARGS: dict[str, Any] = {
}
+class _SortedGlob:
+ """While a scene is built, glob.glob / glob.iglob return their matches sorted. Four RoboTwin tasks (place_a2b_left,
+ place_a2b_right, place_object_scale, place_object_stand: envs/<task>.py load_actors) list an object's model files
+ with glob.glob and then draw one with the seeded np.random.choice; glob returns the file system's directory order,
+ which differs between hosts (and image layers), so one seed built a different object on another host and the frozen
+ instance's hash check failed there (board 0086 claim 5; never rebuilt off the reference workstation). Sorted, the
+ draw depends on the seed alone. No other task of ours lists files while it builds its scene."""
+
+ def __enter__(self):
+ import glob
+ self._glob, self._orig = glob, (glob.glob, glob.iglob)
+ glob.glob = lambda *a, **k: sorted(self._orig[0](*a, **k))
+ glob.iglob = lambda *a, **k: iter(sorted(self._orig[1](*a, **k)))
+ return self
+
+ def __exit__(self, *exc):
+ self._glob.glob, self._glob.iglob = self._orig
+ return False
+
+
def _chdir_root() -> None:
# RoboTwin resolves assets/config paths relative to the repo root and via ROOT_PATH
if os.getcwd() != ROBOTWIN_ROOT and os.path.isdir(ROBOTWIN_ROOT):
@@ -218,7 +238,8 @@ class Session:
if self.enable_render:
sapien.render.set_camera_shader_dir(self.render_shader)
args = _build_args(self.task, self.embodiment, self.seed)
- self.env.setup_demo(**args)
+ with _SortedGlob(): # the object draw depends on the seed alone, not the host's directory order
+ self.env.setup_demo(**args)
if self.enable_render: # setup_scene() switches the shader to ray tracing; put it back for cheap frames
sapien.render.set_camera_shader_dir(self.render_shader)
self._raw_scene = self.env.scene
diff --git a/tasks/robotwin_2_0_turn_switch-privileged/task.toml b/tasks/robotwin_2_0_turn_switch-privileged/task.toml
index ebea607917..c47d524d55 100644
--- a/tasks/robotwin_2_0_turn_switch-privileged/task.toml
+++ b/tasks/robotwin_2_0_turn_switch-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "robotwin" # RoboTwin 2.0 = SAPIEN 3 / PhysX + CuRobo
robot = "aloha-agilex"
family = "turn_switch"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 91573 # scene = environment/instance.json (RoboTwin seed + expected initial-state hash); identity = dataset.toml digest
expert_time_estimate_hours = 0.5
diff --git a/tasks/robotwin_2_0_turn_switch-privileged/tests/rcb_traj.py b/tasks/robotwin_2_0_turn_switch-privileged/tests/rcb_traj.py
index 013426cf20..25d160b3ab 100644
--- a/tasks/robotwin_2_0_turn_switch-privileged/tests/rcb_traj.py
+++ b/tasks/robotwin_2_0_turn_switch-privileged/tests/rcb_traj.py
@@ -113,6 +113,26 @@ BASE_ARGS: dict[str, Any] = {
}
+class _SortedGlob:
+ """While a scene is built, glob.glob / glob.iglob return their matches sorted. Four RoboTwin tasks (place_a2b_left,
+ place_a2b_right, place_object_scale, place_object_stand: envs/<task>.py load_actors) list an object's model files
+ with glob.glob and then draw one with the seeded np.random.choice; glob returns the file system's directory order,
+ which differs between hosts (and image layers), so one seed built a different object on another host and the frozen
+ instance's hash check failed there (board 0086 claim 5; never rebuilt off the reference workstation). Sorted, the
+ draw depends on the seed alone. No other task of ours lists files while it builds its scene."""
+
+ def __enter__(self):
+ import glob
+ self._glob, self._orig = glob, (glob.glob, glob.iglob)
+ glob.glob = lambda *a, **k: sorted(self._orig[0](*a, **k))
+ glob.iglob = lambda *a, **k: iter(sorted(self._orig[1](*a, **k)))
+ return self
+
+ def __exit__(self, *exc):
+ self._glob.glob, self._glob.iglob = self._orig
+ return False
+
+
def _chdir_root() -> None:
# RoboTwin resolves assets/config paths relative to the repo root and via ROOT_PATH
if os.getcwd() != ROBOTWIN_ROOT and os.path.isdir(ROBOTWIN_ROOT):
@@ -218,7 +238,8 @@ class Session:
if self.enable_render:
sapien.render.set_camera_shader_dir(self.render_shader)
args = _build_args(self.task, self.embodiment, self.seed)
- self.env.setup_demo(**args)
+ with _SortedGlob(): # the object draw depends on the seed alone, not the host's directory order
+ self.env.setup_demo(**args)
if self.enable_render: # setup_scene() switches the shader to ray tracing; put it back for cheap frames
sapien.render.set_camera_shader_dir(self.render_shader)
self._raw_scene = self.env.scene
diff --git a/tasks/robotwin_2_0_turn_switch-standard/README.md b/tasks/robotwin_2_0_turn_switch-standard/README.md
index d9e6745a7f..7f741d73a5 100644
--- a/tasks/robotwin_2_0_turn_switch-standard/README.md
+++ b/tasks/robotwin_2_0_turn_switch-standard/README.md
@@ -53,3 +53,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
- 2026-10-10 (owner decision, board 0072): the agent's network egress is restricted to the model APIs (`[agent] network_mode = "allowlist"`, the hosts of `scripts/set_agent_egress.py`; since 2026-10-11 these include the package registries, PyPI and the Debian / Ubuntu archives, so that `pip install` and `apt-get install` work); it was open before (Harbor's default), and agents could download upstream sources, reference solutions and robot models from their own shell. The scene, the frozen instance, the tooling and the pass rule are unchanged, so the task keeps its name; results from before this date were obtained with open egress.
- 2026-10-13 (board 0086): the service's copy of the tooling only (`environment/sim/rcb_traj.py`, one line of `environment/sim/server.py`). The tooling now refuses to save or replay a trajectory of more than 200,000 rows, the privileged verifier's cap; a standard episode has no row cap, so the service lifts it for its own session (`S.max_rows = None`). Nothing an agent sees and nothing the service grades changes.
- 2026-10-13 (board 0100): base image `eai-robotwin` 0.1.2, whose scrub of RoboTwin's expert also removes the reference poses that `load_actors` computed only for the expert. The standard-mode agent never reaches the env, so nothing it observes changes; the scene, the instance and the judge are unchanged.
+- 2026-10-14 (board 0113): the tooling only (`environment/sim/rcb_traj.py`): while it builds the scene, `glob.glob` returns its matches sorted, so that a task which lists files to draw an object depends on its seed alone, not on the host's directory order (four RoboTwin tasks did; this one lists no files while it builds its scene). Its scene, frozen instance (same initial-state hash), verifier and pass rule are unchanged.
diff --git a/tasks/robotwin_2_0_turn_switch-standard/environment/sim/rcb_traj.py b/tasks/robotwin_2_0_turn_switch-standard/environment/sim/rcb_traj.py
index 013426cf20..25d160b3ab 100644
--- a/tasks/robotwin_2_0_turn_switch-standard/environment/sim/rcb_traj.py
+++ b/tasks/robotwin_2_0_turn_switch-standard/environment/sim/rcb_traj.py
@@ -113,6 +113,26 @@ BASE_ARGS: dict[str, Any] = {
}
+class _SortedGlob:
+ """While a scene is built, glob.glob / glob.iglob return their matches sorted. Four RoboTwin tasks (place_a2b_left,
+ place_a2b_right, place_object_scale, place_object_stand: envs/<task>.py load_actors) list an object's model files
+ with glob.glob and then draw one with the seeded np.random.choice; glob returns the file system's directory order,
+ which differs between hosts (and image layers), so one seed built a different object on another host and the frozen
+ instance's hash check failed there (board 0086 claim 5; never rebuilt off the reference workstation). Sorted, the
+ draw depends on the seed alone. No other task of ours lists files while it builds its scene."""
+
+ def __enter__(self):
+ import glob
+ self._glob, self._orig = glob, (glob.glob, glob.iglob)
+ glob.glob = lambda *a, **k: sorted(self._orig[0](*a, **k))
+ glob.iglob = lambda *a, **k: iter(sorted(self._orig[1](*a, **k)))
+ return self
+
+ def __exit__(self, *exc):
+ self._glob.glob, self._glob.iglob = self._orig
+ return False
+
+
def _chdir_root() -> None:
# RoboTwin resolves assets/config paths relative to the repo root and via ROOT_PATH
if os.getcwd() != ROBOTWIN_ROOT and os.path.isdir(ROBOTWIN_ROOT):
@@ -218,7 +238,8 @@ class Session:
if self.enable_render:
sapien.render.set_camera_shader_dir(self.render_shader)
args = _build_args(self.task, self.embodiment, self.seed)
- self.env.setup_demo(**args)
+ with _SortedGlob(): # the object draw depends on the seed alone, not the host's directory order
+ self.env.setup_demo(**args)
if self.enable_render: # setup_scene() switches the shader to ray tracing; put it back for cheap frames
sapien.render.set_camera_shader_dir(self.render_shader)
self._raw_scene = self.env.scene
diff --git a/tasks/robotwin_2_0_turn_switch-standard/task.toml b/tasks/robotwin_2_0_turn_switch-standard/task.toml
index efd6cf4738..7d4edbcf90 100644
--- a/tasks/robotwin_2_0_turn_switch-standard/task.toml
+++ b/tasks/robotwin_2_0_turn_switch-standard/task.toml
@@ -27,6 +27,7 @@ simulator = "robotwin"
robot = "aloha-agilex"
family = "turn_switch"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 91573
diff --git a/tasks/robowits-align-blocks-i00-privileged/task.toml b/tasks/robowits-align-blocks-i00-privileged/task.toml
index 1e135c803b..c1c1d4cc9b 100644
--- a/tasks/robowits-align-blocks-i00-privileged/task.toml
+++ b/tasks/robowits-align-blocks-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "robowits" # RoboWits @ 9cc30ae on Genesis 0.4.7 (CPU phy
robot = "marvin-bimanual"
family = "align_blocks"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution: RoboWits ships no expert or planner for its seed tasks. Solvable: model trial robowits-align-blocks-i00-limite__bbbgbmS (robowits-lunaxh-limited-0929, standard) graded success=1, deterministic=1 by the separate verifier; nop scored 0. These records are from earlier versions of the task (before protocol v1.0, on the same frozen instance and predicate); not re-run on this version (static validation only, 2026-10-01). Human review per family (protocol §2.7)."
instance_seed = 0 # scene = environment/instance.json (entry 0 of RoboWits' eval_dataset_50/01.json + expected state hash); identity = dataset.toml digest
diff --git a/tasks/robowits-align-blocks-i00-standard/task.toml b/tasks/robowits-align-blocks-i00-standard/task.toml
index 11af4c40a5..faa1ee269b 100644
--- a/tasks/robowits-align-blocks-i00-standard/task.toml
+++ b/tasks/robowits-align-blocks-i00-standard/task.toml
@@ -24,6 +24,7 @@ simulator = "robowits"
robot = "marvin-bimanual"
family = "align_blocks"
mode = "standard" # RGB + proprioception only; no object poses, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution: RoboWits ships no expert or planner for its seed tasks. Solvable: model trial robowits-align-blocks-i00-limite__bbbgbmS (robowits-lunaxh-limited-0929, standard) graded success=1, deterministic=1 by the separate verifier; nop scored 0. These records are from earlier versions of the task (before protocol v1.0, on the same frozen instance and predicate); not re-run on this version (static validation only, 2026-10-01). Human review per family (protocol §2.7)."
diff --git a/tasks/robowits-ball-into-bottle-i00-privileged/README.md b/tasks/robowits-ball-into-bottle-i00-privileged/README.md
index ec19bebc1a..63fbb05841 100644
--- a/tasks/robowits-ball-into-bottle-i00-privileged/README.md
+++ b/tasks/robowits-ball-into-bottle-i00-privileged/README.md
@@ -64,3 +64,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
- 2026-10-10 (review of board 0067; findings RW-1, RW-5, RW-6, RW-7): the verifier reports an infrastructure error (`infra_error = 1`) instead of the agent's 0 when it cannot build the frozen scene or the scene it rebuilt does not match the frozen instance on its host (`instance_ok = 0`: another CPU vendor, docs/benchmarks/robowits.md, Hosts); a `trajectory.npz` that cannot be read or lacks `task` / `actions` is `bad_trajectory`, no longer `grader_error`; `reward.json` always carries `infra_error`; the verifier runs without network (`network_mode = "no-network"`); the instruction states the tooling's cap of 200000 control steps per trajectory instead of leaving it unsaid. The pass rule, the frozen instance, the tooling and the predicate are unchanged; validated with synthetic tests of the grader, not re-run in the simulator.
- 2026-10-10 (review round 2, board 0067; minor 2): `tests/grade.py` time-boxes the replay video, which it renders after the verdict is on disk: it is skipped or stopped (`video_timed_out = 1`) 180 s before the verifier's `timeout_sec` (`tests/test.sh` passes it as `--budget`). Before, the video replay of a very long trajectory could run into Harbor's verifier timeout, which discards the verdict already written. The verdict, the pass rule, the frozen instance and the tooling are unchanged; validated with a synthetic test of the grader, not re-run in the simulator.
- 2026-10-10 (owner decision, board 0072): the agent's network egress is restricted to the model APIs (`[agent] network_mode = "allowlist"`, the hosts of `scripts/set_agent_egress.py`; since 2026-10-11 these include the package registries, PyPI and the Debian / Ubuntu archives, so that `pip install` and `apt-get install` work); it was open before (Harbor's default), and agents could download upstream sources, reference solutions and robot models from their own shell. The scene, the frozen instance, the tooling and the pass rule are unchanged, so the task keeps its name; results from before this date were obtained with open egress.
+- 2026-10-14 (board 0113, from board 0086): the instruction says that the bottle's base is thin in the simulator and a fast-falling ball can pass through it (board 0086: a privileged trial measured 9 of 16 slightly perturbed replays of one pour losing the ball through the base; every standard run failed). Text only: the scene, the frozen instance, the tooling, the predicate and the verifier are unchanged, so the task keeps its name. A thicker base would change the frozen scene (not done).
diff --git a/tasks/robowits-ball-into-bottle-i00-privileged/instruction.md b/tasks/robowits-ball-into-bottle-i00-privileged/instruction.md
index cfa81e7259..ca6c43aad4 100644
--- a/tasks/robowits-ball-into-bottle-i00-privileged/instruction.md
+++ b/tasks/robowits-ball-into-bottle-i00-privileged/instruction.md
@@ -3,6 +3,8 @@
You control a simulated **bimanual Marvin** robot (two 7-DoF arms with Pika parallel grippers; RoboWits on Genesis
0.4.7 physics) via Python. **Task:** Transfer the ball from the blue container into the red bottle.
+The red bottle's base is thin in the simulator (its collision shape is about 3.5 mm thick, and the bottle is fixed to the table): a ball that falls fast down the bottle can pass through the base onto the table, which fails the task. In one measured case about half of the pours did (16 replays of one pouring trajectory with tiny noise, 9 balls through the base): let the ball in slowly.
+
The scene objects are the env's entities (`s.objects()`: `red bottle`, `funnel`, `ball container`, `small ball`; see
`/opt/robowits/gs_gym/envs/robowits/12_ball_into_bottle.py`). RoboWits' success predicate is `_check_success` in that file:
```python
diff --git a/tasks/robowits-ball-into-bottle-i00-privileged/task.toml b/tasks/robowits-ball-into-bottle-i00-privileged/task.toml
index 1f863103f2..d471136a36 100644
--- a/tasks/robowits-ball-into-bottle-i00-privileged/task.toml
+++ b/tasks/robowits-ball-into-bottle-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "robowits" # RoboWits @ 9cc30ae on Genesis 0.4.7 (CPU phy
robot = "marvin-bimanual"
family = "ball_into_bottle"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution: RoboWits ships no expert or planner for its seed tasks. Solvable: model trial robowits-ball-into-bottle-i00__aUXXC4E (robowits-lunaxh-0925-2042, privileged) graded success=1, deterministic=1 by the separate verifier; nop scored 0. These records are from earlier versions of the task (before protocol v1.0, on the same frozen instance and predicate); not re-run on this version (static validation only, 2026-10-01). Human review per family (protocol §2.7)."
instance_seed = 0 # scene = environment/instance.json (entry 0 of RoboWits' eval_dataset_50/12.json + expected state hash); identity = dataset.toml digest
diff --git a/tasks/robowits-ball-into-bottle-i00-standard/README.md b/tasks/robowits-ball-into-bottle-i00-standard/README.md
index 34ac3d7e13..c6d913111f 100644
--- a/tasks/robowits-ball-into-bottle-i00-standard/README.md
+++ b/tasks/robowits-ball-into-bottle-i00-standard/README.md
@@ -66,3 +66,4 @@ Minimum allocation for one trial (docs/PROTOCOL.md §2.9; `[metadata.resources]`
- 2026-10-10 (review round 3, board 0067): the shared service layer, the client and a per-call image limit in the service (`environment/sim/rcb_limited.py`, `environment/rcb_client.py`, `environment/sim/server.py`). A camera named more than once in `cameras` is rendered and returned once (N copies of one name made the service build N x N images within the frame budget, which could exhaust its memory). A service that has given up on the episode (it could not start it, or could not resume it after its crashes) answers every request with 409 and the reason, and `Robot()` / `wait_ready()` raise `EpisodeOver` at once; before, nothing answered and they waited out their timeout (up to an hour). One call returns at most 1500 images (`/spec` `max_frames_per_call`: cameras x frames, a frame with per-frame depth counts 4): a request that would exceed it is a 400 before anything runs, and `ee_pose` rows that take longer than planned stop adding frames there (`meta["frames_skipped"]`). Before, a call had no such bound: one long observation at the full frame rate could hold more frames than the sidecar's memory and get the service killed, so a trial that ended `sim_crashed*` or with `resource_service_oom` right after such a request before this date was that. The collect hook, the grader, the scene and the pass rule are unchanged.
- 2026-10-10 (owner decision, board 0072): the agent's network egress is restricted to the model APIs (`[agent] network_mode = "allowlist"`, the hosts of `scripts/set_agent_egress.py`; since 2026-10-11 these include the package registries, PyPI and the Debian / Ubuntu archives, so that `pip install` and `apt-get install` work); it was open before (Harbor's default), and agents could download upstream sources, reference solutions and robot models from their own shell. The scene, the frozen instance, the tooling and the pass rule are unchanged, so the task keeps its name; results from before this date were obtained with open egress.
- 2026-10-13 (owner, boards 0085 and 0087): the sim service's clock (`environment/sim/rcb_limited.py`, `environment/sim/server.py`, the docstring of `environment/rcb_client.py`). `meta["time_s"]` and `meta["time_left_s"]` count from the agent's first request to the service (any endpoint), and a restarted service keeps that moment; before, they counted from the service's own start of the episode (when its scene was ready). Where the service is ready before the agent's phase begins (the harness is installed in between: a median of 79 s over the recorded standard trials), that clock ran ahead of the agent's and `time_left_s` reached 0 while the agent still had that time. `/spec` says what the clock counts from (`budget.clock`), and the instruction's `## Time limit` tells the agent to plan by `time_left_s`. The service never ended an episode on this clock: the scene, the frozen instance, the pass rule and the grader are unchanged, so the task keeps its name and earlier results stand.
+- 2026-10-14 (board 0113, from board 0086): the instruction says that the bottle's base is thin in the simulator and a fast-falling ball can pass through it (board 0086: a privileged trial measured 9 of 16 slightly perturbed replays of one pour losing the ball through the base; every standard run failed). Text only: the scene, the frozen instance, the tooling, the predicate and the verifier are unchanged, so the task keeps its name. A thicker base would change the frozen scene (not done).
diff --git a/tasks/robowits-ball-into-bottle-i00-standard/instruction.md b/tasks/robowits-ball-into-bottle-i00-standard/instruction.md
index 7ef0a86d29..9293dec875 100644
--- a/tasks/robowits-ball-into-bottle-i00-standard/instruction.md
+++ b/tasks/robowits-ball-into-bottle-i00-standard/instruction.md
@@ -4,6 +4,8 @@ You control a bimanual **Marvin** robot (two 7-DoF arms with Pika parallel gripp
the simulator runs in a service you cannot open. You get one client and one episode. **Task:** Transfer the ball from
the blue container into the red bottle.
+The red bottle's base is thin in the simulator (its collision shape is about 3.5 mm thick, and the bottle is fixed to the table): a ball that falls fast down the bottle can pass through the base onto the table, which fails the task. In one measured case about half of the pours did (16 replays of one pouring trajectory with tiny noise, 9 balls through the base): let the ball in slowly.
+
## Success criterion (as graded)
This is exactly what the grader checks (RoboWits' success predicate, stated in words). Where the task text says
something else, this section is what counts. The service evaluates it after every `submit` and after every `get_obs`
diff --git a/tasks/robowits-ball-into-bottle-i00-standard/task.toml b/tasks/robowits-ball-into-bottle-i00-standard/task.toml
index 3db1fb1a9b..7beb296572 100644
--- a/tasks/robowits-ball-into-bottle-i00-standard/task.toml
+++ b/tasks/robowits-ball-into-bottle-i00-standard/task.toml
@@ -24,6 +24,7 @@ simulator = "robowits"
robot = "marvin-bimanual"
family = "ball_into_bottle"
mode = "standard" # RGB + proprioception only; no object poses, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution: RoboWits ships no expert or planner for its seed tasks. Solvable: model trial robowits-ball-into-bottle-i00__aUXXC4E (robowits-lunaxh-0925-2042, privileged) graded success=1, deterministic=1 by the separate verifier; nop scored 0. These records are from earlier versions of the task (before protocol v1.0, on the same frozen instance and predicate); not re-run on this version (static validation only, 2026-10-01). Human review per family (protocol §2.7)."
diff --git a/tasks/robowits-ball-onto-tower-i00-privileged/task.toml b/tasks/robowits-ball-onto-tower-i00-privileged/task.toml
index 7c669b3932..6a66805a58 100644
--- a/tasks/robowits-ball-onto-tower-i00-privileged/task.toml
+++ b/tasks/robowits-ball-onto-tower-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "robowits" # RoboWits @ 9cc30ae on Genesis 0.4.7 (CPU phy
robot = "marvin-bimanual"
family = "ball_onto_tower"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution: RoboWits ships no expert or planner for its seed tasks. Solvable: model trial robowits-ball-onto-tower-i00-lim__LQbxymJ (robowits-lunaxh-limited-0929, standard) graded success=1, deterministic=1 by the separate verifier; nop scored 0. These records are from earlier versions of the task (before protocol v1.0, on the same frozen instance and predicate); not re-run on this version (static validation only, 2026-10-01). Human review per family (protocol §2.7)."
instance_seed = 0 # scene = environment/instance.json (entry 0 of RoboWits' eval_dataset_50/17.json + expected state hash); identity = dataset.toml digest
diff --git a/tasks/robowits-ball-onto-tower-i00-standard/task.toml b/tasks/robowits-ball-onto-tower-i00-standard/task.toml
index bb04815ecf..f6bee2f5b2 100644
--- a/tasks/robowits-ball-onto-tower-i00-standard/task.toml
+++ b/tasks/robowits-ball-onto-tower-i00-standard/task.toml
@@ -24,6 +24,7 @@ simulator = "robowits"
robot = "marvin-bimanual"
family = "ball_onto_tower"
mode = "standard" # RGB + proprioception only; no object poses, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution: RoboWits ships no expert or planner for its seed tasks. Solvable: model trial robowits-ball-onto-tower-i00-lim__LQbxymJ (robowits-lunaxh-limited-0929, standard) graded success=1, deterministic=1 by the separate verifier; nop scored 0. These records are from earlier versions of the task (before protocol v1.0, on the same frozen instance and predicate); not re-run on this version (static validation only, 2026-10-01). Human review per family (protocol §2.7)."
diff --git a/tasks/robowits-collect-screws-i00-privileged/task.toml b/tasks/robowits-collect-screws-i00-privileged/task.toml
index fb4536673c..abe306d5af 100644
--- a/tasks/robowits-collect-screws-i00-privileged/task.toml
+++ b/tasks/robowits-collect-screws-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "robowits" # RoboWits @ 9cc30ae on Genesis 0.4.7 (CPU phy
robot = "marvin-bimanual"
family = "collect_screws"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution: RoboWits ships no expert or planner for its seed tasks. Solvable: model trial robowits-collect-screws-i00-limi__xxcMFF8 (robowits-lunaxh-limited-0929, standard) graded success=1, deterministic=1 by the separate verifier; nop scored 0. These records are from earlier versions of the task (before protocol v1.0, on the same frozen instance and predicate); not re-run on this version (static validation only, 2026-10-01). Human review per family (protocol §2.7)."
instance_seed = 0 # scene = environment/instance.json (entry 0 of RoboWits' eval_dataset_50/10.json + expected state hash); identity = dataset.toml digest
diff --git a/tasks/robowits-collect-screws-i00-standard/task.toml b/tasks/robowits-collect-screws-i00-standard/task.toml
index e9de5190ae..d022074655 100644
--- a/tasks/robowits-collect-screws-i00-standard/task.toml
+++ b/tasks/robowits-collect-screws-i00-standard/task.toml
@@ -24,6 +24,7 @@ simulator = "robowits"
robot = "marvin-bimanual"
family = "collect_screws"
mode = "standard" # RGB + proprioception only; no object poses, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution: RoboWits ships no expert or planner for its seed tasks. Solvable: model trial robowits-collect-screws-i00-limi__xxcMFF8 (robowits-lunaxh-limited-0929, standard) graded success=1, deterministic=1 by the separate verifier; nop scored 0. These records are from earlier versions of the task (before protocol v1.0, on the same frozen instance and predicate); not re-run on this version (static validation only, 2026-10-01). Human review per family (protocol §2.7)."
diff --git a/tasks/robowits-cover-with-lid-i00-privileged/task.toml b/tasks/robowits-cover-with-lid-i00-privileged/task.toml
index e161f2d079..8b737d028c 100644
--- a/tasks/robowits-cover-with-lid-i00-privileged/task.toml
+++ b/tasks/robowits-cover-with-lid-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "robowits" # RoboWits @ 9cc30ae on Genesis 0.4.7 (CPU phy
robot = "marvin-bimanual"
family = "cover_with_lid"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution: RoboWits ships no expert or planner for its seed tasks. Solvable: model trial robowits-cover-with-lid-i00__Jb7Zzve (robowits-lunaxh-0925-2042, privileged) graded success=1, deterministic=1 by the separate verifier; nop scored 0. These records are from earlier versions of the task (before protocol v1.0, on the same frozen instance and predicate); not re-run on this version (static validation only, 2026-10-01). Human review per family (protocol §2.7)."
instance_seed = 0 # scene = environment/instance.json (entry 0 of RoboWits' eval_dataset_50/13.json + expected state hash); identity = dataset.toml digest
diff --git a/tasks/robowits-cover-with-lid-i00-standard/task.toml b/tasks/robowits-cover-with-lid-i00-standard/task.toml
index 923e843101..ed58789819 100644
--- a/tasks/robowits-cover-with-lid-i00-standard/task.toml
+++ b/tasks/robowits-cover-with-lid-i00-standard/task.toml
@@ -24,6 +24,7 @@ simulator = "robowits"
robot = "marvin-bimanual"
family = "cover_with_lid"
mode = "standard" # RGB + proprioception only; no object poses, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution: RoboWits ships no expert or planner for its seed tasks. Solvable: model trial robowits-cover-with-lid-i00__Jb7Zzve (robowits-lunaxh-0925-2042, privileged) graded success=1, deterministic=1 by the separate verifier; nop scored 0. These records are from earlier versions of the task (before protocol v1.0, on the same frozen instance and predicate); not re-run on this version (static validation only, 2026-10-01). Human review per family (protocol §2.7)."
diff --git a/tasks/robowits-dominos-i00-privileged/task.toml b/tasks/robowits-dominos-i00-privileged/task.toml
index 79a1da98e2..6e2f6a36b6 100644
--- a/tasks/robowits-dominos-i00-privileged/task.toml
+++ b/tasks/robowits-dominos-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "robowits" # RoboWits @ 9cc30ae on Genesis 0.4.7 (CPU phy
robot = "marvin-bimanual"
family = "dominos"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "partial"
oracle_evidence = "Reference solution solution/oracle_dominos.py (written for this instance; RoboWits ships no expert). It scored 1 under Harbor -a oracle on an earlier version of this task: job robowits-dominos-i00-oracle-thin-0925-000706, trial robowits-dominos-i00__qL9Rvsb, 2026-09-25 (laptop, Intel; success=1, deterministic=1, instance_exact=1; an uncommitted tree just before dev/qineng f15dbdb7). Same since then: the frozen instance (state hash 4f933a76ce2ce34b) and the oracle script (byte-identical to that run's /app/output/gen.py). Changed since then, so the label is partial: the grader tests/grade.py (dev/qineng 82d76c61 added the GPU-backend exception; this PR writes the verdict before the video) and the tooling rcb_robowits.py (dev/qineng 5aa848bd, bddfc520, adce3fef, 8f0ebd55, 82d76c61, 6c0b439e, fe5fb9e6, 1515ea4e). Not re-run on this version (static validation only, 2026-10-01)."
instance_seed = 0 # scene = environment/instance.json (entry 0 of RoboWits' eval_dataset_50/06.json + expected state hash); identity = dataset.toml digest
diff --git a/tasks/robowits-dominos-i00-standard/task.toml b/tasks/robowits-dominos-i00-standard/task.toml
index c1d7f928ea..3f2b5e3512 100644
--- a/tasks/robowits-dominos-i00-standard/task.toml
+++ b/tasks/robowits-dominos-i00-standard/task.toml
@@ -24,6 +24,7 @@ simulator = "robowits"
robot = "marvin-bimanual"
family = "dominos"
mode = "standard" # RGB + proprioception only; no object poses, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "partial"
oracle_evidence = "Reference solution solution/oracle_replay.py, which streams the privileged oracle's recorded joint rows (solution/oracle_qpos.npy) through the client. It scored 1 under Harbor -a oracle on an earlier version of this task: job robowits-dominos-i00-limited-oracle-v2b-0924-2125, trial robowits-dominos-i00-limited__pnfVz4w, 2026-09-25 (laptop, Intel; the protocol's 2.0 service, then named rcb-limited; success=1, live 1, replay 1, deterministic 1; an uncommitted tree just before dev/qineng f15dbdb7). Same since then: the frozen instance (state hash 4f933a76ce2ce34b) and the joint rows (equal to that run's recorded trajectory). Changed since then, so the label is partial: the grader (dev/qineng 82d76c61; this PR: the crash-recovery fields and infra_error), the sim adapter server.py (dev/qineng 6c0b439e: protocol 2.2; this PR: crash recovery), the shared layer rcb_limited.py / rcb_client.py (dev/qineng 42813008; this PR: the eai-standard/2.2 layer with crash recovery), the tooling rcb_robowits.py (as in the privileged twin), the collect hook, the client image and pins, and the oracle player (this PR: it recovers from a crash). Not re-run on this version (static validation only, 2026-10-01)."
diff --git a/tasks/robowits-gap-retrieve-i00-privileged/task.toml b/tasks/robowits-gap-retrieve-i00-privileged/task.toml
index faa5eb4d9e..e907af6dd2 100644
--- a/tasks/robowits-gap-retrieve-i00-privileged/task.toml
+++ b/tasks/robowits-gap-retrieve-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "robowits" # RoboWits @ 9cc30ae on Genesis 0.4.7 (CPU phy
robot = "marvin-bimanual"
family = "gap_retrieve"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution: RoboWits ships no expert or planner for its seed tasks. Solvable: model trial robowits-gap-retrieve-i00-limite__hTYhCVz (robowits-lunaxh-limited-0929, standard) graded success=1, deterministic=1 by the separate verifier; nop scored 0. These records are from earlier versions of the task (before protocol v1.0, on the same frozen instance and predicate); not re-run on this version (static validation only, 2026-10-01). Human review per family (protocol §2.7)."
instance_seed = 0 # scene = environment/instance.json (entry 0 of RoboWits' eval_dataset_50/03.json + expected state hash); identity = dataset.toml digest
diff --git a/tasks/robowits-gap-retrieve-i00-standard/task.toml b/tasks/robowits-gap-retrieve-i00-standard/task.toml
index 90a97fe551..b8a005dc47 100644
--- a/tasks/robowits-gap-retrieve-i00-standard/task.toml
+++ b/tasks/robowits-gap-retrieve-i00-standard/task.toml
@@ -24,6 +24,7 @@ simulator = "robowits"
robot = "marvin-bimanual"
family = "gap_retrieve"
mode = "standard" # RGB + proprioception only; no object poses, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution: RoboWits ships no expert or planner for its seed tasks. Solvable: model trial robowits-gap-retrieve-i00-limite__hTYhCVz (robowits-lunaxh-limited-0929, standard) graded success=1, deterministic=1 by the separate verifier; nop scored 0. These records are from earlier versions of the task (before protocol v1.0, on the same frozen instance and predicate); not re-run on this version (static validation only, 2026-10-01). Human review per family (protocol §2.7)."
diff --git a/tasks/robowits-hold-cup-i00-privileged/task.toml b/tasks/robowits-hold-cup-i00-privileged/task.toml
index b236cad234..382ca1dbc4 100644
--- a/tasks/robowits-hold-cup-i00-privileged/task.toml
+++ b/tasks/robowits-hold-cup-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "robowits" # RoboWits @ 9cc30ae on Genesis 0.4.7 (CPU phy
robot = "marvin-bimanual"
family = "hold_cup"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution: RoboWits ships no expert or planner for its seed tasks. Solvable: model trial robowits-hold-cup-i00-limited__BiMATyh (robowits-lunaxh-limited-0929, standard) graded success=1, deterministic=1 by the separate verifier; nop scored 0. These records are from earlier versions of the task (before protocol v1.0, on the same frozen instance and predicate); not re-run on this version (static validation only, 2026-10-01). Human review per family (protocol §2.7)."
instance_seed = 0 # scene = environment/instance.json (entry 0 of RoboWits' eval_dataset_50/09.json + expected state hash); identity = dataset.toml digest
diff --git a/tasks/robowits-hold-cup-i00-standard/task.toml b/tasks/robowits-hold-cup-i00-standard/task.toml
index 375678d2bf..0e7c154cd9 100644
--- a/tasks/robowits-hold-cup-i00-standard/task.toml
+++ b/tasks/robowits-hold-cup-i00-standard/task.toml
@@ -24,6 +24,7 @@ simulator = "robowits"
robot = "marvin-bimanual"
family = "hold_cup"
mode = "standard" # RGB + proprioception only; no object poses, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution: RoboWits ships no expert or planner for its seed tasks. Solvable: model trial robowits-hold-cup-i00-limited__BiMATyh (robowits-lunaxh-limited-0929, standard) graded success=1, deterministic=1 by the separate verifier; nop scored 0. These records are from earlier versions of the task (before protocol v1.0, on the same frozen instance and predicate); not re-run on this version (static validation only, 2026-10-01). Human review per family (protocol §2.7)."
diff --git a/tasks/robowits-move-cube-i00-privileged/task.toml b/tasks/robowits-move-cube-i00-privileged/task.toml
index 00d96a1198..65006d7707 100644
--- a/tasks/robowits-move-cube-i00-privileged/task.toml
+++ b/tasks/robowits-move-cube-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "robowits" # RoboWits @ 9cc30ae on Genesis 0.4.7 (CPU phy
robot = "marvin-bimanual"
family = "move_cube"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution: RoboWits ships no expert or planner for its seed tasks. Solvable: model trial robowits-move-cube-i00-limited__9Vd2dpc (robowits-lunaxh-limited-0929, standard) graded success=1, deterministic=1 by the separate verifier; nop scored 0. These records are from earlier versions of the task (before protocol v1.0, on the same frozen instance and predicate); not re-run on this version (static validation only, 2026-10-01). Human review per family (protocol §2.7)."
instance_seed = 0 # scene = environment/instance.json (entry 0 of RoboWits' eval_dataset_50/28.json + expected state hash); identity = dataset.toml digest
diff --git a/tasks/robowits-move-cube-i00-standard/task.toml b/tasks/robowits-move-cube-i00-standard/task.toml
index 2adaaec50c..38ff5e80be 100644
--- a/tasks/robowits-move-cube-i00-standard/task.toml
+++ b/tasks/robowits-move-cube-i00-standard/task.toml
@@ -24,6 +24,7 @@ simulator = "robowits"
robot = "marvin-bimanual"
family = "move_cube"
mode = "standard" # RGB + proprioception only; no object poses, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution: RoboWits ships no expert or planner for its seed tasks. Solvable: model trial robowits-move-cube-i00-limited__9Vd2dpc (robowits-lunaxh-limited-0929, standard) graded success=1, deterministic=1 by the separate verifier; nop scored 0. These records are from earlier versions of the task (before protocol v1.0, on the same frozen instance and predicate); not re-run on this version (static validation only, 2026-10-01). Human review per family (protocol §2.7)."
diff --git a/tasks/robowits-raise-platform-i00-privileged/task.toml b/tasks/robowits-raise-platform-i00-privileged/task.toml
index cabb59eec7..5529db81fc 100644
--- a/tasks/robowits-raise-platform-i00-privileged/task.toml
+++ b/tasks/robowits-raise-platform-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "robowits" # RoboWits @ 9cc30ae on Genesis 0.4.7 (CPU phy
robot = "marvin-bimanual"
family = "raise_platform"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution: RoboWits ships no expert or planner for its seed tasks. Solvable: model trial robowits-raise-platform-i00__uqturjE (robowits-lunaxh-0925-2042, privileged) graded success=1, deterministic=1 by the separate verifier; nop scored 0. These records are from earlier versions of the task (before protocol v1.0, on the same frozen instance and predicate); not re-run on this version (static validation only, 2026-10-01). Human review per family (protocol §2.7)."
instance_seed = 0 # scene = environment/instance.json (entry 0 of RoboWits' eval_dataset_50/24.json + expected state hash); identity = dataset.toml digest
diff --git a/tasks/robowits-raise-platform-i00-standard/task.toml b/tasks/robowits-raise-platform-i00-standard/task.toml
index 7e9ba64c39..dc6dc3000b 100644
--- a/tasks/robowits-raise-platform-i00-standard/task.toml
+++ b/tasks/robowits-raise-platform-i00-standard/task.toml
@@ -24,6 +24,7 @@ simulator = "robowits"
robot = "marvin-bimanual"
family = "raise_platform"
mode = "standard" # RGB + proprioception only; no object poses, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution: RoboWits ships no expert or planner for its seed tasks. Solvable: model trial robowits-raise-platform-i00__uqturjE (robowits-lunaxh-0925-2042, privileged) graded success=1, deterministic=1 by the separate verifier; nop scored 0. These records are from earlier versions of the task (before protocol v1.0, on the same frozen instance and predicate); not re-run on this version (static validation only, 2026-10-01). Human review per family (protocol §2.7)."
diff --git a/tasks/robowits-retrieve-cube-i00-privileged/task.toml b/tasks/robowits-retrieve-cube-i00-privileged/task.toml
index f633fa07c3..0af810f724 100644
--- a/tasks/robowits-retrieve-cube-i00-privileged/task.toml
+++ b/tasks/robowits-retrieve-cube-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "robowits" # RoboWits @ 9cc30ae on Genesis 0.4.7 (CPU phy
robot = "marvin-bimanual"
family = "retrieve_cube"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution: RoboWits ships no expert or planner for its seed tasks. Solvable: model trial robowits-retrieve-cube-i00__eRfLscE (robowits-lunaxh-0925-2042, privileged) graded success=1, deterministic=1 by the separate verifier; nop scored 0. These records are from earlier versions of the task (before protocol v1.0, on the same frozen instance and predicate); not re-run on this version (static validation only, 2026-10-01). Human review per family (protocol §2.7)."
instance_seed = 0 # scene = environment/instance.json (entry 0 of RoboWits' eval_dataset_50/02.json + expected state hash); identity = dataset.toml digest
diff --git a/tasks/robowits-retrieve-cube-i00-standard/task.toml b/tasks/robowits-retrieve-cube-i00-standard/task.toml
index 06df829430..bb949ec8f3 100644
--- a/tasks/robowits-retrieve-cube-i00-standard/task.toml
+++ b/tasks/robowits-retrieve-cube-i00-standard/task.toml
@@ -24,6 +24,7 @@ simulator = "robowits"
robot = "marvin-bimanual"
family = "retrieve_cube"
mode = "standard" # RGB + proprioception only; no object poses, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution: RoboWits ships no expert or planner for its seed tasks. Solvable: model trial robowits-retrieve-cube-i00__eRfLscE (robowits-lunaxh-0925-2042, privileged) graded success=1, deterministic=1 by the separate verifier; nop scored 0. These records are from earlier versions of the task (before protocol v1.0, on the same frozen instance and predicate); not re-run on this version (static validation only, 2026-10-01). Human review per family (protocol §2.7)."
diff --git a/tasks/robowits-stack-bowls-i00-privileged/task.toml b/tasks/robowits-stack-bowls-i00-privileged/task.toml
index 4b6f9ea4fb..4d134f325b 100644
--- a/tasks/robowits-stack-bowls-i00-privileged/task.toml
+++ b/tasks/robowits-stack-bowls-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "robowits" # RoboWits @ 9cc30ae on Genesis 0.4.7 (CPU phy
robot = "marvin-bimanual"
family = "stack_bowls"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution: RoboWits ships no expert or planner for its seed tasks. Solvable: model trial robowits-stack-bowls-i00__6pB7rjT (robowits-lunaxh-0925-2042, privileged) graded success=1, deterministic=1 by the separate verifier; nop scored 0. These records are from earlier versions of the task (before protocol v1.0, on the same frozen instance and predicate); not re-run on this version (static validation only, 2026-10-01). Human review per family (protocol §2.7)."
instance_seed = 0 # scene = environment/instance.json (entry 0 of RoboWits' eval_dataset_50/19.json + expected state hash); identity = dataset.toml digest
diff --git a/tasks/robowits-stack-bowls-i00-standard/task.toml b/tasks/robowits-stack-bowls-i00-standard/task.toml
index 94de662e2f..87e105705a 100644
--- a/tasks/robowits-stack-bowls-i00-standard/task.toml
+++ b/tasks/robowits-stack-bowls-i00-standard/task.toml
@@ -24,6 +24,7 @@ simulator = "robowits"
robot = "marvin-bimanual"
family = "stack_bowls"
mode = "standard" # RGB + proprioception only; no object poses, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution: RoboWits ships no expert or planner for its seed tasks. Solvable: model trial robowits-stack-bowls-i00__6pB7rjT (robowits-lunaxh-0925-2042, privileged) graded success=1, deterministic=1 by the separate verifier; nop scored 0. These records are from earlier versions of the task (before protocol v1.0, on the same frozen instance and predicate); not re-run on this version (static validation only, 2026-10-01). Human review per family (protocol §2.7)."
diff --git a/tasks/robowits-stack-cubes-i00-privileged/task.toml b/tasks/robowits-stack-cubes-i00-privileged/task.toml
index 8f93305e0d..0d0eb5f21c 100644
--- a/tasks/robowits-stack-cubes-i00-privileged/task.toml
+++ b/tasks/robowits-stack-cubes-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "robowits" # RoboWits @ 9cc30ae on Genesis 0.4.7 (CPU phy
robot = "marvin-bimanual"
family = "stack_cubes"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "partial"
oracle_evidence = "Reference solution solution/oracle_stack_cubes.py with its skills in solution/ref_primitives.py (written for this instance; RoboWits ships no expert). It scored 1 under Harbor -a oracle on an earlier version of this task: job robowits-stack-cubes-i00-oracle-thin-0925-000706, trial robowits-stack-cubes-i00__2rVdsVe, 2026-09-25 (laptop, Intel; success=1, deterministic=1, instance_exact=1; an uncommitted tree just before dev/qineng f15dbdb7). Same since then: the frozen instance (state hash 41c4d7e7d7193b0d) and the oracle script (byte-identical to that run's /app/output/gen.py; ref_primitives.py unchanged since f15dbdb7). Changed since then, so the label is partial: the grader tests/grade.py (dev/qineng 82d76c61 added the GPU-backend exception; this PR writes the verdict before the video) and the tooling rcb_robowits.py (dev/qineng 5aa848bd, bddfc520, adce3fef, 8f0ebd55, 82d76c61, 6c0b439e, fe5fb9e6, 1515ea4e). Not re-run on this version (static validation only, 2026-10-01)."
instance_seed = 0 # scene = environment/instance.json (entry 0 of RoboWits' eval_dataset_50/14.json + expected state hash); identity = dataset.toml digest
diff --git a/tasks/robowits-stack-cubes-i00-standard/task.toml b/tasks/robowits-stack-cubes-i00-standard/task.toml
index 70b0cb22f7..974ec51095 100644
--- a/tasks/robowits-stack-cubes-i00-standard/task.toml
+++ b/tasks/robowits-stack-cubes-i00-standard/task.toml
@@ -24,6 +24,7 @@ simulator = "robowits"
robot = "marvin-bimanual"
family = "stack_cubes"
mode = "standard" # RGB + proprioception only; no object poses, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "partial"
oracle_evidence = "Reference solution solution/oracle_replay.py, which streams the privileged oracle's recorded joint rows (solution/oracle_qpos.npy) through the client. It scored 1 under Harbor -a oracle on the dev/qineng 9eb18edf version of this task (protocol 2.2, then named rcb-limited): job v22-laptop-oracle-stack-cubes-0413, trial robowits-stack-cubes-i00-limited__Yw2oaio, 2026-09-29 (laptop, Intel; success=1, live 1, replay 1, deterministic 1). Same since then: the frozen instance (state hash 41c4d7e7d7193b0d), the joint rows (equal to that run's recorded trajectory) and the tooling rcb_robowits.py. Changed since then, so the label is partial: the grader (this PR: the crash-recovery fields and infra_error), the sim adapter server.py and the shared layer rcb_limited.py / rcb_client.py (this PR: crash recovery), the collect hook, the client image and pins, and the oracle player (this PR: it recovers from a crash). Not re-run on this version (static validation only, 2026-10-01)."
diff --git a/tasks/robowits-stand-bulb-i00-privileged/task.toml b/tasks/robowits-stand-bulb-i00-privileged/task.toml
index 76c0008644..352656a170 100644
--- a/tasks/robowits-stand-bulb-i00-privileged/task.toml
+++ b/tasks/robowits-stand-bulb-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "robowits" # RoboWits @ 9cc30ae on Genesis 0.4.7 (CPU phy
robot = "marvin-bimanual"
family = "stand_bulb"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution: RoboWits ships no expert or planner for its seed tasks. Solvable: model trial robowits-stand-bulb-i00__BrTmH6k (robowits-lunaxh-0925-2042, privileged) graded success=1, deterministic=1 by the separate verifier; nop scored 0. These records are from earlier versions of the task (before protocol v1.0, on the same frozen instance and predicate); not re-run on this version (static validation only, 2026-10-01). Human review per family (protocol §2.7)."
instance_seed = 0 # scene = environment/instance.json (entry 0 of RoboWits' eval_dataset_50/16.json + expected state hash); identity = dataset.toml digest
diff --git a/tasks/robowits-stand-bulb-i00-standard/task.toml b/tasks/robowits-stand-bulb-i00-standard/task.toml
index 70b309375e..d179cf1b43 100644
--- a/tasks/robowits-stand-bulb-i00-standard/task.toml
+++ b/tasks/robowits-stand-bulb-i00-standard/task.toml
@@ -24,6 +24,7 @@ simulator = "robowits"
robot = "marvin-bimanual"
family = "stand_bulb"
mode = "standard" # RGB + proprioception only; no object poses, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution: RoboWits ships no expert or planner for its seed tasks. Solvable: model trial robowits-stand-bulb-i00__BrTmH6k (robowits-lunaxh-0925-2042, privileged) graded success=1, deterministic=1 by the separate verifier; nop scored 0. These records are from earlier versions of the task (before protocol v1.0, on the same frozen instance and predicate); not re-run on this version (static validation only, 2026-10-01). Human review per family (protocol §2.7)."
diff --git a/tasks/robowits-stand-pages-i00-privileged/task.toml b/tasks/robowits-stand-pages-i00-privileged/task.toml
index 1358122217..12e0ddd94f 100644
--- a/tasks/robowits-stand-pages-i00-privileged/task.toml
+++ b/tasks/robowits-stand-pages-i00-privileged/task.toml
@@ -14,6 +14,7 @@ simulator = "robowits" # RoboWits @ 9cc30ae on Genesis 0.4.7 (CPU phy
robot = "marvin-bimanual"
family = "stand_pages"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "none"
oracle_evidence = "No reference solution: RoboWits ships no expert or planner for its seed tasks. Solvable: model trial robowits-stand-pages-i00__npAKeQv (calib-high-1944, privileged) graded success=1, deterministic=1 by the separate verifier; nop scored 0. These records are from earlier versions of the task (before protocol v1.0, on the same frozen instance and predicate); not re-run on this version (static validation only, 2026-10-01). Human review per family (protocol §2.7)."
instance_seed = 0 # scene = environment/instance.json (entry 0 of RoboWits' eval_dataset_50/07.json + expected state hash); identity = dataset.toml digest
diff --git a/tasks/robowits-stand-pages-i00-standard/task.toml b/tasks/robowits-stand-pages-i00-standard/task.toml
index 35c377599c..c1cc3b8a62 100644
--- a/tasks/robowits-stand-pages-i00-standard/task.toml
+++ b/tasks/robowits-stand-pages-i00-standard/task.toml
@@ -24,6 +24,7 @@ simulator = "robowits"
robot = "marvin-bimanual"
family = "stand_pages"
mode = "standard" # RGB + proprioception only; no object poses, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "none"
oracle_evidence = "No reference solution: RoboWits ships no expert or planner for its seed tasks. Solvable: model trial robowits-stand-pages-i00__npAKeQv (calib-high-1944, privileged) graded success=1, deterministic=1 by the separate verifier; nop scored 0. These records are from earlier versions of the task (before protocol v1.0, on the same frozen instance and predicate); not re-run on this version (static validation only, 2026-10-01). Human review per family (protocol §2.7)."
diff --git a/tasks/vlabench-add-condiment-i00-privileged/task.toml b/tasks/vlabench-add-condiment-i00-privileged/task.toml
index dd77cc7d82..dd4463a85d 100644
--- a/tasks/vlabench-add-condiment-i00-privileged/task.toml
+++ b/tasks/vlabench-add-condiment-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "vlabench"
robot = "franka-panda"
family = "vlabench/add_condiment"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 0 # VLABench's episode configuration for this task, seeded build
expert_time_estimate_hours = 1.0
diff --git a/tasks/vlabench-add-condiment-i00-standard/task.toml b/tasks/vlabench-add-condiment-i00-standard/task.toml
index d74d584862..db54fe0a8f 100644
--- a/tasks/vlabench-add-condiment-i00-standard/task.toml
+++ b/tasks/vlabench-add-condiment-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "vlabench"
robot = "franka-panda"
family = "vlabench/add_condiment"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 0
diff --git a/tasks/vlabench-book-rearrange-i00-privileged/task.toml b/tasks/vlabench-book-rearrange-i00-privileged/task.toml
index d633971e2a..adf42a8b6f 100644
--- a/tasks/vlabench-book-rearrange-i00-privileged/task.toml
+++ b/tasks/vlabench-book-rearrange-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "vlabench"
robot = "franka-panda"
family = "vlabench/book_rearrange"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 0 # VLABench's episode configuration for this task, seeded build
expert_time_estimate_hours = 1.0
diff --git a/tasks/vlabench-book-rearrange-i00-standard/task.toml b/tasks/vlabench-book-rearrange-i00-standard/task.toml
index b51a5594a2..054a482c2b 100644
--- a/tasks/vlabench-book-rearrange-i00-standard/task.toml
+++ b/tasks/vlabench-book-rearrange-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "vlabench"
robot = "franka-panda"
family = "vlabench/book_rearrange"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 0
diff --git a/tasks/vlabench-find-unseen-object-i00-privileged/task.toml b/tasks/vlabench-find-unseen-object-i00-privileged/task.toml
index f5d0336e8d..b4c17d60d0 100644
--- a/tasks/vlabench-find-unseen-object-i00-privileged/task.toml
+++ b/tasks/vlabench-find-unseen-object-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "vlabench"
robot = "franka-panda"
family = "vlabench/find_unseen_object"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 0 # VLABench's episode configuration for this task, seeded build
expert_time_estimate_hours = 1.0
diff --git a/tasks/vlabench-find-unseen-object-i00-standard/task.toml b/tasks/vlabench-find-unseen-object-i00-standard/task.toml
index 44d3f715f4..30efa2dbb2 100644
--- a/tasks/vlabench-find-unseen-object-i00-standard/task.toml
+++ b/tasks/vlabench-find-unseen-object-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "vlabench"
robot = "franka-panda"
family = "vlabench/find_unseen_object"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 0
diff --git a/tasks/vlabench-friction-qa-i00-privileged/task.toml b/tasks/vlabench-friction-qa-i00-privileged/task.toml
index c23c4b8595..83a3aa5280 100644
--- a/tasks/vlabench-friction-qa-i00-privileged/task.toml
+++ b/tasks/vlabench-friction-qa-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "vlabench"
robot = "franka-panda"
family = "vlabench/friction_qa"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 0 # VLABench's episode configuration for this task, seeded build
expert_time_estimate_hours = 1.0
diff --git a/tasks/vlabench-friction-qa-i00-standard/task.toml b/tasks/vlabench-friction-qa-i00-standard/task.toml
index 571164b9cb..e994624324 100644
--- a/tasks/vlabench-friction-qa-i00-standard/task.toml
+++ b/tasks/vlabench-friction-qa-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "vlabench"
robot = "franka-panda"
family = "vlabench/friction_qa"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 0
diff --git a/tasks/vlabench-hammer-nail-and-hang-picture-i00-privileged/task.toml b/tasks/vlabench-hammer-nail-and-hang-picture-i00-privileged/task.toml
index 0d2fe495fd..979412c929 100644
--- a/tasks/vlabench-hammer-nail-and-hang-picture-i00-privileged/task.toml
+++ b/tasks/vlabench-hammer-nail-and-hang-picture-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "vlabench"
robot = "franka-panda"
family = "vlabench/hammer_nail_and_hang_picture"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 0 # VLABench's episode configuration for this task, seeded build
expert_time_estimate_hours = 1.0
diff --git a/tasks/vlabench-hammer-nail-and-hang-picture-i00-standard/task.toml b/tasks/vlabench-hammer-nail-and-hang-picture-i00-standard/task.toml
index 1d50ea4bfe..b931828150 100644
--- a/tasks/vlabench-hammer-nail-and-hang-picture-i00-standard/task.toml
+++ b/tasks/vlabench-hammer-nail-and-hang-picture-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "vlabench"
robot = "franka-panda"
family = "vlabench/hammer_nail_and_hang_picture"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 0
diff --git a/tasks/vlabench-heat-food-i00-privileged/task.toml b/tasks/vlabench-heat-food-i00-privileged/task.toml
index aadbe3d4fc..a5c5a52d61 100644
--- a/tasks/vlabench-heat-food-i00-privileged/task.toml
+++ b/tasks/vlabench-heat-food-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "vlabench"
robot = "franka-panda"
family = "vlabench/heat_food"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 0 # VLABench's episode configuration for this task, seeded build
expert_time_estimate_hours = 1.0
diff --git a/tasks/vlabench-heat-food-i00-standard/task.toml b/tasks/vlabench-heat-food-i00-standard/task.toml
index 35a79d1f93..49f7ae9cba 100644
--- a/tasks/vlabench-heat-food-i00-standard/task.toml
+++ b/tasks/vlabench-heat-food-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "vlabench"
robot = "franka-panda"
family = "vlabench/heat_food"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 0
diff --git a/tasks/vlabench-insert-bloom-flower-i00-privileged/task.toml b/tasks/vlabench-insert-bloom-flower-i00-privileged/task.toml
index 53b4ddcb17..336acd7535 100644
--- a/tasks/vlabench-insert-bloom-flower-i00-privileged/task.toml
+++ b/tasks/vlabench-insert-bloom-flower-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "vlabench"
robot = "franka-panda"
family = "vlabench/insert_bloom_flower"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 0 # VLABench's episode configuration for this task, seeded build
expert_time_estimate_hours = 1.0
diff --git a/tasks/vlabench-insert-bloom-flower-i00-standard/task.toml b/tasks/vlabench-insert-bloom-flower-i00-standard/task.toml
index 4a0d1f9343..c5b16da96f 100644
--- a/tasks/vlabench-insert-bloom-flower-i00-standard/task.toml
+++ b/tasks/vlabench-insert-bloom-flower-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "vlabench"
robot = "franka-panda"
family = "vlabench/insert_bloom_flower"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 0
diff --git a/tasks/vlabench-insert-flower-i00-privileged/task.toml b/tasks/vlabench-insert-flower-i00-privileged/task.toml
index af7786ba11..ee337de942 100644
--- a/tasks/vlabench-insert-flower-i00-privileged/task.toml
+++ b/tasks/vlabench-insert-flower-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "vlabench"
robot = "franka-panda"
family = "vlabench/insert_flower"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 4 # VLABench's episode configuration for this task, seeded build
expert_time_estimate_hours = 1.0
diff --git a/tasks/vlabench-insert-flower-i00-standard/task.toml b/tasks/vlabench-insert-flower-i00-standard/task.toml
index 752f5b39a2..74a8fac804 100644
--- a/tasks/vlabench-insert-flower-i00-standard/task.toml
+++ b/tasks/vlabench-insert-flower-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "vlabench"
robot = "franka-panda"
family = "vlabench/insert_flower"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 4
diff --git a/tasks/vlabench-magnetism-qa-i00-privileged/task.toml b/tasks/vlabench-magnetism-qa-i00-privileged/task.toml
index 425c9196fa..b682d452c2 100644
--- a/tasks/vlabench-magnetism-qa-i00-privileged/task.toml
+++ b/tasks/vlabench-magnetism-qa-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "vlabench"
robot = "franka-panda"
family = "vlabench/magnetism_qa"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 0 # VLABench's episode configuration for this task, seeded build
expert_time_estimate_hours = 1.0
diff --git a/tasks/vlabench-magnetism-qa-i00-standard/task.toml b/tasks/vlabench-magnetism-qa-i00-standard/task.toml
index 5c3ca6f79f..10fc8b4f74 100644
--- a/tasks/vlabench-magnetism-qa-i00-standard/task.toml
+++ b/tasks/vlabench-magnetism-qa-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "vlabench"
robot = "franka-panda"
family = "vlabench/magnetism_qa"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 0
diff --git a/tasks/vlabench-play-math-game-i00-privileged/task.toml b/tasks/vlabench-play-math-game-i00-privileged/task.toml
index db355ba611..72b688e6e1 100644
--- a/tasks/vlabench-play-math-game-i00-privileged/task.toml
+++ b/tasks/vlabench-play-math-game-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "vlabench"
robot = "franka-panda"
family = "vlabench/play_math_game"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 0 # VLABench's episode configuration for this task, seeded build
expert_time_estimate_hours = 1.0
diff --git a/tasks/vlabench-play-math-game-i00-standard/task.toml b/tasks/vlabench-play-math-game-i00-standard/task.toml
index f2c54f5c61..684dbaf5a8 100644
--- a/tasks/vlabench-play-math-game-i00-standard/task.toml
+++ b/tasks/vlabench-play-math-game-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "vlabench"
robot = "franka-panda"
family = "vlabench/play_math_game"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 0
diff --git a/tasks/vlabench-put-box-on-painting-i00-privileged/task.toml b/tasks/vlabench-put-box-on-painting-i00-privileged/task.toml
index 6bd1e4a0fd..ae3af25aec 100644
--- a/tasks/vlabench-put-box-on-painting-i00-privileged/task.toml
+++ b/tasks/vlabench-put-box-on-painting-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "vlabench"
robot = "franka-panda"
family = "vlabench/put_box_on_painting"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 0 # VLABench's episode configuration for this task, seeded build
expert_time_estimate_hours = 1.0
diff --git a/tasks/vlabench-put-box-on-painting-i00-standard/task.toml b/tasks/vlabench-put-box-on-painting-i00-standard/task.toml
index ca6c10499f..8bdfa20db0 100644
--- a/tasks/vlabench-put-box-on-painting-i00-standard/task.toml
+++ b/tasks/vlabench-put-box-on-painting-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "vlabench"
robot = "franka-panda"
family = "vlabench/put_box_on_painting"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 0
diff --git a/tasks/vlabench-reflection-qa-i00-privileged/task.toml b/tasks/vlabench-reflection-qa-i00-privileged/task.toml
index 938093d36d..03d917aa3e 100644
--- a/tasks/vlabench-reflection-qa-i00-privileged/task.toml
+++ b/tasks/vlabench-reflection-qa-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "vlabench"
robot = "franka-panda"
family = "vlabench/reflection_qa"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 0 # VLABench's episode configuration for this task, seeded build
expert_time_estimate_hours = 1.0
diff --git a/tasks/vlabench-reflection-qa-i00-standard/task.toml b/tasks/vlabench-reflection-qa-i00-standard/task.toml
index ab34f32fe6..0f02afc51b 100644
--- a/tasks/vlabench-reflection-qa-i00-standard/task.toml
+++ b/tasks/vlabench-reflection-qa-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "vlabench"
robot = "franka-panda"
family = "vlabench/reflection_qa"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 0
diff --git a/tasks/vlabench-select-billiards-common-sense-i00-privileged/task.toml b/tasks/vlabench-select-billiards-common-sense-i00-privileged/task.toml
index 9840c2ac36..bc3b871ba1 100644
--- a/tasks/vlabench-select-billiards-common-sense-i00-privileged/task.toml
+++ b/tasks/vlabench-select-billiards-common-sense-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "vlabench"
robot = "franka-panda"
family = "vlabench/select_billiards_common_sense"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 0 # VLABench's episode configuration for this task, seeded build
expert_time_estimate_hours = 1.0
diff --git a/tasks/vlabench-select-billiards-common-sense-i00-standard/task.toml b/tasks/vlabench-select-billiards-common-sense-i00-standard/task.toml
index 1c84634e9b..5f86ea3af8 100644
--- a/tasks/vlabench-select-billiards-common-sense-i00-standard/task.toml
+++ b/tasks/vlabench-select-billiards-common-sense-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "vlabench"
robot = "franka-panda"
family = "vlabench/select_billiards_common_sense"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 0
diff --git a/tasks/vlabench-select-billiards-i00-privileged/task.toml b/tasks/vlabench-select-billiards-i00-privileged/task.toml
index ab0a816413..1044fe3221 100644
--- a/tasks/vlabench-select-billiards-i00-privileged/task.toml
+++ b/tasks/vlabench-select-billiards-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "vlabench"
robot = "franka-panda"
family = "vlabench/select_billiards"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 0 # VLABench's episode configuration for this task, seeded build
expert_time_estimate_hours = 1.0
diff --git a/tasks/vlabench-select-billiards-i00-standard/task.toml b/tasks/vlabench-select-billiards-i00-standard/task.toml
index 0b71c68dcc..d573c05051 100644
--- a/tasks/vlabench-select-billiards-i00-standard/task.toml
+++ b/tasks/vlabench-select-billiards-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "vlabench"
robot = "franka-panda"
family = "vlabench/select_billiards"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 0
diff --git a/tasks/vlabench-select-billiards-spatial-i00-privileged/task.toml b/tasks/vlabench-select-billiards-spatial-i00-privileged/task.toml
index 18b798cbdc..0e9dacee37 100644
--- a/tasks/vlabench-select-billiards-spatial-i00-privileged/task.toml
+++ b/tasks/vlabench-select-billiards-spatial-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "vlabench"
robot = "franka-panda"
family = "vlabench/select_billiards_spatial"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 0 # VLABench's episode configuration for this task, seeded build
expert_time_estimate_hours = 1.0
diff --git a/tasks/vlabench-select-billiards-spatial-i00-standard/task.toml b/tasks/vlabench-select-billiards-spatial-i00-standard/task.toml
index 57130d94f3..a088897ba0 100644
--- a/tasks/vlabench-select-billiards-spatial-i00-standard/task.toml
+++ b/tasks/vlabench-select-billiards-spatial-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "vlabench"
robot = "franka-panda"
family = "vlabench/select_billiards_spatial"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 0
diff --git a/tasks/vlabench-select-book-spatial-i00-privileged/task.toml b/tasks/vlabench-select-book-spatial-i00-privileged/task.toml
index 8e0072b5d6..fa845bfb5e 100644
--- a/tasks/vlabench-select-book-spatial-i00-privileged/task.toml
+++ b/tasks/vlabench-select-book-spatial-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "vlabench"
robot = "franka-panda"
family = "vlabench/select_book_spatial"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 0 # VLABench's episode configuration for this task, seeded build
expert_time_estimate_hours = 1.0
diff --git a/tasks/vlabench-select-book-spatial-i00-standard/task.toml b/tasks/vlabench-select-book-spatial-i00-standard/task.toml
index acdd0f287d..9e882d699b 100644
--- a/tasks/vlabench-select-book-spatial-i00-standard/task.toml
+++ b/tasks/vlabench-select-book-spatial-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "vlabench"
robot = "franka-panda"
family = "vlabench/select_book_spatial"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 0
diff --git a/tasks/vlabench-select-chemistry-tube-common-sense-i00-privileged/task.toml b/tasks/vlabench-select-chemistry-tube-common-sense-i00-privileged/task.toml
index 3702b94f54..af309db120 100644
--- a/tasks/vlabench-select-chemistry-tube-common-sense-i00-privileged/task.toml
+++ b/tasks/vlabench-select-chemistry-tube-common-sense-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "vlabench"
robot = "franka-panda"
family = "vlabench/select_chemistry_tube_common_sense"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 0 # VLABench's episode configuration for this task, seeded build
expert_time_estimate_hours = 1.0
diff --git a/tasks/vlabench-select-chemistry-tube-common-sense-i00-standard/task.toml b/tasks/vlabench-select-chemistry-tube-common-sense-i00-standard/task.toml
index 3b15411b6b..4a57e55615 100644
--- a/tasks/vlabench-select-chemistry-tube-common-sense-i00-standard/task.toml
+++ b/tasks/vlabench-select-chemistry-tube-common-sense-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "vlabench"
robot = "franka-panda"
family = "vlabench/select_chemistry_tube_common_sense"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 0
diff --git a/tasks/vlabench-select-chemistry-tube-i00-privileged/task.toml b/tasks/vlabench-select-chemistry-tube-i00-privileged/task.toml
index 9e3a2e3147..4a93b2eb4b 100644
--- a/tasks/vlabench-select-chemistry-tube-i00-privileged/task.toml
+++ b/tasks/vlabench-select-chemistry-tube-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "vlabench"
robot = "franka-panda"
family = "vlabench/select_chemistry_tube"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 0 # VLABench's episode configuration for this task, seeded build
expert_time_estimate_hours = 1.0
diff --git a/tasks/vlabench-select-chemistry-tube-i00-standard/task.toml b/tasks/vlabench-select-chemistry-tube-i00-standard/task.toml
index 3c86169ea1..af20441e19 100644
--- a/tasks/vlabench-select-chemistry-tube-i00-standard/task.toml
+++ b/tasks/vlabench-select-chemistry-tube-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "vlabench"
robot = "franka-panda"
family = "vlabench/select_chemistry_tube"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 0
diff --git a/tasks/vlabench-select-chemistry-tube-semantic-i00-privileged/task.toml b/tasks/vlabench-select-chemistry-tube-semantic-i00-privileged/task.toml
index b80247606e..d640f80f89 100644
--- a/tasks/vlabench-select-chemistry-tube-semantic-i00-privileged/task.toml
+++ b/tasks/vlabench-select-chemistry-tube-semantic-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "vlabench"
robot = "franka-panda"
family = "vlabench/select_chemistry_tube_semantic"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 0 # VLABench's episode configuration for this task, seeded build
expert_time_estimate_hours = 1.0
diff --git a/tasks/vlabench-select-chemistry-tube-semantic-i00-standard/task.toml b/tasks/vlabench-select-chemistry-tube-semantic-i00-standard/task.toml
index 4588ddd707..3362199aba 100644
--- a/tasks/vlabench-select-chemistry-tube-semantic-i00-standard/task.toml
+++ b/tasks/vlabench-select-chemistry-tube-semantic-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "vlabench"
robot = "franka-panda"
family = "vlabench/select_chemistry_tube_semantic"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 0
diff --git a/tasks/vlabench-select-chemistry-tube-spatial-i00-privileged/task.toml b/tasks/vlabench-select-chemistry-tube-spatial-i00-privileged/task.toml
index 479ed9326c..8f8ce73dae 100644
--- a/tasks/vlabench-select-chemistry-tube-spatial-i00-privileged/task.toml
+++ b/tasks/vlabench-select-chemistry-tube-spatial-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "vlabench"
robot = "franka-panda"
family = "vlabench/select_chemistry_tube_spatial"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 0 # VLABench's episode configuration for this task, seeded build
expert_time_estimate_hours = 1.0
diff --git a/tasks/vlabench-select-chemistry-tube-spatial-i00-standard/task.toml b/tasks/vlabench-select-chemistry-tube-spatial-i00-standard/task.toml
index 7a23c8d77f..d855cf394a 100644
--- a/tasks/vlabench-select-chemistry-tube-spatial-i00-standard/task.toml
+++ b/tasks/vlabench-select-chemistry-tube-spatial-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "vlabench"
robot = "franka-panda"
family = "vlabench/select_chemistry_tube_spatial"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 0
diff --git a/tasks/vlabench-select-drink-i00-privileged/task.toml b/tasks/vlabench-select-drink-i00-privileged/task.toml
index cf5a988cea..d9e1e8c094 100644
--- a/tasks/vlabench-select-drink-i00-privileged/task.toml
+++ b/tasks/vlabench-select-drink-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "vlabench"
robot = "franka-panda"
family = "vlabench/select_drink"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 0 # VLABench's episode configuration for this task, seeded build
expert_time_estimate_hours = 1.0
diff --git a/tasks/vlabench-select-drink-i00-standard/task.toml b/tasks/vlabench-select-drink-i00-standard/task.toml
index 143bf9d87c..88cd31ef43 100644
--- a/tasks/vlabench-select-drink-i00-standard/task.toml
+++ b/tasks/vlabench-select-drink-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "vlabench"
robot = "franka-panda"
family = "vlabench/select_drink"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 0
diff --git a/tasks/vlabench-select-fruit-common-sense-i00-privileged/task.toml b/tasks/vlabench-select-fruit-common-sense-i00-privileged/task.toml
index 1e650565c3..dac4ebe89d 100644
--- a/tasks/vlabench-select-fruit-common-sense-i00-privileged/task.toml
+++ b/tasks/vlabench-select-fruit-common-sense-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "vlabench"
robot = "franka-panda"
family = "vlabench/select_fruit_common_sense"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 0 # VLABench's episode configuration for this task, seeded build
expert_time_estimate_hours = 1.0
diff --git a/tasks/vlabench-select-fruit-common-sense-i00-standard/task.toml b/tasks/vlabench-select-fruit-common-sense-i00-standard/task.toml
index 3b6655c792..de943c1472 100644
--- a/tasks/vlabench-select-fruit-common-sense-i00-standard/task.toml
+++ b/tasks/vlabench-select-fruit-common-sense-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "vlabench"
robot = "franka-panda"
family = "vlabench/select_fruit_common_sense"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 0
diff --git a/tasks/vlabench-select-fruit-i00-privileged/task.toml b/tasks/vlabench-select-fruit-i00-privileged/task.toml
index 0483c55624..42e8773e4a 100644
--- a/tasks/vlabench-select-fruit-i00-privileged/task.toml
+++ b/tasks/vlabench-select-fruit-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "vlabench"
robot = "franka-panda"
family = "vlabench/select_fruit"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 0 # VLABench's episode configuration for this task, seeded build
expert_time_estimate_hours = 1.0
diff --git a/tasks/vlabench-select-fruit-i00-standard/task.toml b/tasks/vlabench-select-fruit-i00-standard/task.toml
index d91b1ff697..f6096d1533 100644
--- a/tasks/vlabench-select-fruit-i00-standard/task.toml
+++ b/tasks/vlabench-select-fruit-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "vlabench"
robot = "franka-panda"
family = "vlabench/select_fruit"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 0
diff --git a/tasks/vlabench-select-mahjong-i00-privileged/task.toml b/tasks/vlabench-select-mahjong-i00-privileged/task.toml
index 5c4c39715e..62263a01e0 100644
--- a/tasks/vlabench-select-mahjong-i00-privileged/task.toml
+++ b/tasks/vlabench-select-mahjong-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "vlabench"
robot = "franka-panda"
family = "vlabench/select_mahjong"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 0 # VLABench's episode configuration for this task, seeded build
expert_time_estimate_hours = 1.0
diff --git a/tasks/vlabench-select-mahjong-i00-standard/task.toml b/tasks/vlabench-select-mahjong-i00-standard/task.toml
index e42284a69a..9d3ff52682 100644
--- a/tasks/vlabench-select-mahjong-i00-standard/task.toml
+++ b/tasks/vlabench-select-mahjong-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "vlabench"
robot = "franka-panda"
family = "vlabench/select_mahjong"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 0
diff --git a/tasks/vlabench-select-mahjong-semantic-i00-privileged/task.toml b/tasks/vlabench-select-mahjong-semantic-i00-privileged/task.toml
index 545c280017..11a464a12e 100644
--- a/tasks/vlabench-select-mahjong-semantic-i00-privileged/task.toml
+++ b/tasks/vlabench-select-mahjong-semantic-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "vlabench"
robot = "franka-panda"
family = "vlabench/select_mahjong_semantic"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 0 # VLABench's episode configuration for this task, seeded build
expert_time_estimate_hours = 1.0
diff --git a/tasks/vlabench-select-mahjong-semantic-i00-standard/task.toml b/tasks/vlabench-select-mahjong-semantic-i00-standard/task.toml
index f0c55dee38..3a331afbc2 100644
--- a/tasks/vlabench-select-mahjong-semantic-i00-standard/task.toml
+++ b/tasks/vlabench-select-mahjong-semantic-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "vlabench"
robot = "franka-panda"
family = "vlabench/select_mahjong_semantic"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 0
diff --git a/tasks/vlabench-select-mahjong-spatial-i00-privileged/task.toml b/tasks/vlabench-select-mahjong-spatial-i00-privileged/task.toml
index 95f6b12bd6..5ea34fcf50 100644
--- a/tasks/vlabench-select-mahjong-spatial-i00-privileged/task.toml
+++ b/tasks/vlabench-select-mahjong-spatial-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "vlabench"
robot = "franka-panda"
family = "vlabench/select_mahjong_spatial"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 0 # VLABench's episode configuration for this task, seeded build
expert_time_estimate_hours = 1.0
diff --git a/tasks/vlabench-select-mahjong-spatial-i00-standard/task.toml b/tasks/vlabench-select-mahjong-spatial-i00-standard/task.toml
index 13977fc2d9..e9507c927b 100644
--- a/tasks/vlabench-select-mahjong-spatial-i00-standard/task.toml
+++ b/tasks/vlabench-select-mahjong-spatial-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "vlabench"
robot = "franka-panda"
family = "vlabench/select_mahjong_spatial"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 0
diff --git a/tasks/vlabench-select-nth-largest-poker-i00-privileged/task.toml b/tasks/vlabench-select-nth-largest-poker-i00-privileged/task.toml
index 4ddd2945f2..fe6724e740 100644
--- a/tasks/vlabench-select-nth-largest-poker-i00-privileged/task.toml
+++ b/tasks/vlabench-select-nth-largest-poker-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "vlabench"
robot = "franka-panda"
family = "vlabench/select_nth_largest_poker"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 0 # VLABench's episode configuration for this task, seeded build
expert_time_estimate_hours = 1.0
diff --git a/tasks/vlabench-select-nth-largest-poker-i00-standard/task.toml b/tasks/vlabench-select-nth-largest-poker-i00-standard/task.toml
index 29140e657d..278e0b888b 100644
--- a/tasks/vlabench-select-nth-largest-poker-i00-standard/task.toml
+++ b/tasks/vlabench-select-nth-largest-poker-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "vlabench"
robot = "franka-panda"
family = "vlabench/select_nth_largest_poker"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 0
diff --git a/tasks/vlabench-select-painting-i00-privileged/task.toml b/tasks/vlabench-select-painting-i00-privileged/task.toml
index dcd09f9a05..d34832cc6c 100644
--- a/tasks/vlabench-select-painting-i00-privileged/task.toml
+++ b/tasks/vlabench-select-painting-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "vlabench"
robot = "franka-panda"
family = "vlabench/select_painting"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 0 # VLABench's episode configuration for this task, seeded build
expert_time_estimate_hours = 1.0
diff --git a/tasks/vlabench-select-painting-i00-standard/task.toml b/tasks/vlabench-select-painting-i00-standard/task.toml
index eafaef3725..560fbc2fa7 100644
--- a/tasks/vlabench-select-painting-i00-standard/task.toml
+++ b/tasks/vlabench-select-painting-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "vlabench"
robot = "franka-panda"
family = "vlabench/select_painting"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 0
diff --git a/tasks/vlabench-select-poker-i00-privileged/task.toml b/tasks/vlabench-select-poker-i00-privileged/task.toml
index 2a0d29d2e9..22124a26c7 100644
--- a/tasks/vlabench-select-poker-i00-privileged/task.toml
+++ b/tasks/vlabench-select-poker-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "vlabench"
robot = "franka-panda"
family = "vlabench/select_poker"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 0 # VLABench's episode configuration for this task, seeded build
expert_time_estimate_hours = 1.0
diff --git a/tasks/vlabench-select-poker-i00-standard/task.toml b/tasks/vlabench-select-poker-i00-standard/task.toml
index 0e51ef285c..7082d49776 100644
--- a/tasks/vlabench-select-poker-i00-standard/task.toml
+++ b/tasks/vlabench-select-poker-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "vlabench"
robot = "franka-panda"
family = "vlabench/select_poker"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 0
diff --git a/tasks/vlabench-select-poker-semantic-i00-privileged/task.toml b/tasks/vlabench-select-poker-semantic-i00-privileged/task.toml
index ef3aa7a1ab..8ac7d3df4f 100644
--- a/tasks/vlabench-select-poker-semantic-i00-privileged/task.toml
+++ b/tasks/vlabench-select-poker-semantic-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "vlabench"
robot = "franka-panda"
family = "vlabench/select_poker_semantic"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 0 # VLABench's episode configuration for this task, seeded build
expert_time_estimate_hours = 1.0
diff --git a/tasks/vlabench-select-poker-semantic-i00-standard/task.toml b/tasks/vlabench-select-poker-semantic-i00-standard/task.toml
index f48b82b8e8..05c01c056d 100644
--- a/tasks/vlabench-select-poker-semantic-i00-standard/task.toml
+++ b/tasks/vlabench-select-poker-semantic-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "vlabench"
robot = "franka-panda"
family = "vlabench/select_poker_semantic"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 0
diff --git a/tasks/vlabench-select-poker-spatial-i00-privileged/task.toml b/tasks/vlabench-select-poker-spatial-i00-privileged/task.toml
index 953e73b908..6686ee426c 100644
--- a/tasks/vlabench-select-poker-spatial-i00-privileged/task.toml
+++ b/tasks/vlabench-select-poker-spatial-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "vlabench"
robot = "franka-panda"
family = "vlabench/select_poker_spatial"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 0 # VLABench's episode configuration for this task, seeded build
expert_time_estimate_hours = 1.0
diff --git a/tasks/vlabench-select-poker-spatial-i00-standard/task.toml b/tasks/vlabench-select-poker-spatial-i00-standard/task.toml
index 10679d152e..1ebecce8f5 100644
--- a/tasks/vlabench-select-poker-spatial-i00-standard/task.toml
+++ b/tasks/vlabench-select-poker-spatial-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "vlabench"
robot = "franka-panda"
family = "vlabench/select_poker_spatial"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 0
diff --git a/tasks/vlabench-select-specific-type-book-i00-privileged/task.toml b/tasks/vlabench-select-specific-type-book-i00-privileged/task.toml
index 5da808242d..48132787aa 100644
--- a/tasks/vlabench-select-specific-type-book-i00-privileged/task.toml
+++ b/tasks/vlabench-select-specific-type-book-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "vlabench"
robot = "franka-panda"
family = "vlabench/select_specific_type_book"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 0 # VLABench's episode configuration for this task, seeded build
expert_time_estimate_hours = 1.0
diff --git a/tasks/vlabench-select-specific-type-book-i00-standard/task.toml b/tasks/vlabench-select-specific-type-book-i00-standard/task.toml
index de964ad8d1..c6585daa5b 100644
--- a/tasks/vlabench-select-specific-type-book-i00-standard/task.toml
+++ b/tasks/vlabench-select-specific-type-book-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "vlabench"
robot = "franka-panda"
family = "vlabench/select_specific_type_book"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 0
diff --git a/tasks/vlabench-select-toy-i00-privileged/task.toml b/tasks/vlabench-select-toy-i00-privileged/task.toml
index d776e9d396..e3fb6bfba1 100644
--- a/tasks/vlabench-select-toy-i00-privileged/task.toml
+++ b/tasks/vlabench-select-toy-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "vlabench"
robot = "franka-panda"
family = "vlabench/select_toy"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 0 # VLABench's episode configuration for this task, seeded build
expert_time_estimate_hours = 1.0
diff --git a/tasks/vlabench-select-toy-i00-standard/task.toml b/tasks/vlabench-select-toy-i00-standard/task.toml
index b03069c10f..9db33e8c21 100644
--- a/tasks/vlabench-select-toy-i00-standard/task.toml
+++ b/tasks/vlabench-select-toy-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "vlabench"
robot = "franka-panda"
family = "vlabench/select_toy"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 0
diff --git a/tasks/vlabench-select-unique-type-mahjong-i00-privileged/task.toml b/tasks/vlabench-select-unique-type-mahjong-i00-privileged/task.toml
index e11b70f3c1..bd67459ba5 100644
--- a/tasks/vlabench-select-unique-type-mahjong-i00-privileged/task.toml
+++ b/tasks/vlabench-select-unique-type-mahjong-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "vlabench"
robot = "franka-panda"
family = "vlabench/select_unique_type_mahjong"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 0 # VLABench's episode configuration for this task, seeded build
expert_time_estimate_hours = 1.0
diff --git a/tasks/vlabench-select-unique-type-mahjong-i00-standard/task.toml b/tasks/vlabench-select-unique-type-mahjong-i00-standard/task.toml
index 7529572702..6a9b2d249c 100644
--- a/tasks/vlabench-select-unique-type-mahjong-i00-standard/task.toml
+++ b/tasks/vlabench-select-unique-type-mahjong-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "vlabench"
robot = "franka-panda"
family = "vlabench/select_unique_type_mahjong"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 0
diff --git a/tasks/vlabench-take-out-cool-drink-i00-privileged/task.toml b/tasks/vlabench-take-out-cool-drink-i00-privileged/task.toml
index ed3409f919..dff256645c 100644
--- a/tasks/vlabench-take-out-cool-drink-i00-privileged/task.toml
+++ b/tasks/vlabench-take-out-cool-drink-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "vlabench"
robot = "franka-panda"
family = "vlabench/take_out_cool_drink"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 0 # VLABench's episode configuration for this task, seeded build
expert_time_estimate_hours = 1.0
diff --git a/tasks/vlabench-take-out-cool-drink-i00-standard/task.toml b/tasks/vlabench-take-out-cool-drink-i00-standard/task.toml
index 0f8471d33c..643fe08012 100644
--- a/tasks/vlabench-take-out-cool-drink-i00-standard/task.toml
+++ b/tasks/vlabench-take-out-cool-drink-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "vlabench"
robot = "franka-panda"
family = "vlabench/take_out_cool_drink"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 0
diff --git a/tasks/vlabench-texas-holdem-i00-privileged/task.toml b/tasks/vlabench-texas-holdem-i00-privileged/task.toml
index b79fe70088..f3d3c8ab85 100644
--- a/tasks/vlabench-texas-holdem-i00-privileged/task.toml
+++ b/tasks/vlabench-texas-holdem-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "vlabench"
robot = "franka-panda"
family = "vlabench/texas_holdem"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
instance_seed = 0 # VLABench's episode configuration for this task, seeded build
expert_time_estimate_hours = 1.0
diff --git a/tasks/vlabench-texas-holdem-i00-standard/task.toml b/tasks/vlabench-texas-holdem-i00-standard/task.toml
index d53c3fc9cc..01e88516be 100644
--- a/tasks/vlabench-texas-holdem-i00-standard/task.toml
+++ b/tasks/vlabench-texas-holdem-i00-standard/task.toml
@@ -26,6 +26,7 @@ simulator = "vlabench"
robot = "franka-panda"
family = "vlabench/texas_holdem"
mode = "standard"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.1"
oracle = "full"
instance_seed = 0
diff --git a/tasks/yamlab-hang-mug-on-tree-i10-privileged/task.toml b/tasks/yamlab-hang-mug-on-tree-i10-privileged/task.toml
index d55c931db8..09881c3700 100644
--- a/tasks/yamlab-hang-mug-on-tree-i10-privileged/task.toml
+++ b/tasks/yamlab-hang-mug-on-tree-i10-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "yamlab" # YAMLab @ ec0455d on Isaac Lab @ f4aa17f / Is
robot = "i2rt-yam-bimanual"
family = "hang_mug_on_tree"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
expert_time_estimate_hours = 1.5
diff --git a/tasks/yamlab-hang-mug-on-tree-i10-standard/task.toml b/tasks/yamlab-hang-mug-on-tree-i10-standard/task.toml
index 8eabf7bee7..c775a7e361 100644
--- a/tasks/yamlab-hang-mug-on-tree-i10-standard/task.toml
+++ b/tasks/yamlab-hang-mug-on-tree-i10-standard/task.toml
@@ -28,6 +28,7 @@ simulator = "yamlab"
robot = "i2rt-yam-bimanual"
family = "hang_mug_on_tree"
mode = "standard" # robot as a service: RGB, depth + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "full"
expert_time_estimate_hours = 3.0
diff --git a/tasks/yamlab-hang-mug-on-tree-i13-privileged/task.toml b/tasks/yamlab-hang-mug-on-tree-i13-privileged/task.toml
index 2d2aefa61b..3b1694a1f1 100644
--- a/tasks/yamlab-hang-mug-on-tree-i13-privileged/task.toml
+++ b/tasks/yamlab-hang-mug-on-tree-i13-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "yamlab" # YAMLab @ ec0455d on Isaac Lab @ f4aa17f / Is
robot = "i2rt-yam-bimanual"
family = "hang_mug_on_tree"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
expert_time_estimate_hours = 1.5
diff --git a/tasks/yamlab-hang-mug-on-tree-i13-standard/task.toml b/tasks/yamlab-hang-mug-on-tree-i13-standard/task.toml
index f75216e1cf..7c7efc194c 100644
--- a/tasks/yamlab-hang-mug-on-tree-i13-standard/task.toml
+++ b/tasks/yamlab-hang-mug-on-tree-i13-standard/task.toml
@@ -28,6 +28,7 @@ simulator = "yamlab"
robot = "i2rt-yam-bimanual"
family = "hang_mug_on_tree"
mode = "standard" # robot as a service: RGB, depth + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "full"
expert_time_estimate_hours = 3.0
diff --git a/tasks/yamlab-hang-mug-on-tree-i15-privileged/task.toml b/tasks/yamlab-hang-mug-on-tree-i15-privileged/task.toml
index 28f4d9e54c..73ba831183 100644
--- a/tasks/yamlab-hang-mug-on-tree-i15-privileged/task.toml
+++ b/tasks/yamlab-hang-mug-on-tree-i15-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "yamlab" # YAMLab @ ec0455d on Isaac Lab @ f4aa17f / Is
robot = "i2rt-yam-bimanual"
family = "hang_mug_on_tree"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
expert_time_estimate_hours = 1.5
diff --git a/tasks/yamlab-hang-mug-on-tree-i15-standard/task.toml b/tasks/yamlab-hang-mug-on-tree-i15-standard/task.toml
index 04cac838a7..470d1f3f1f 100644
--- a/tasks/yamlab-hang-mug-on-tree-i15-standard/task.toml
+++ b/tasks/yamlab-hang-mug-on-tree-i15-standard/task.toml
@@ -28,6 +28,7 @@ simulator = "yamlab"
robot = "i2rt-yam-bimanual"
family = "hang_mug_on_tree"
mode = "standard" # robot as a service: RGB, depth + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "full"
expert_time_estimate_hours = 3.0
diff --git a/tasks/yamlab-hang-mug-on-tree-i20-privileged/task.toml b/tasks/yamlab-hang-mug-on-tree-i20-privileged/task.toml
index 8788819f1f..c8fe5bc1de 100644
--- a/tasks/yamlab-hang-mug-on-tree-i20-privileged/task.toml
+++ b/tasks/yamlab-hang-mug-on-tree-i20-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "yamlab" # YAMLab @ ec0455d on Isaac Lab @ f4aa17f / Is
robot = "i2rt-yam-bimanual"
family = "hang_mug_on_tree"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
expert_time_estimate_hours = 1.5
diff --git a/tasks/yamlab-hang-mug-on-tree-i20-standard/task.toml b/tasks/yamlab-hang-mug-on-tree-i20-standard/task.toml
index ff17b1624d..7062756c0e 100644
--- a/tasks/yamlab-hang-mug-on-tree-i20-standard/task.toml
+++ b/tasks/yamlab-hang-mug-on-tree-i20-standard/task.toml
@@ -28,6 +28,7 @@ simulator = "yamlab"
robot = "i2rt-yam-bimanual"
family = "hang_mug_on_tree"
mode = "standard" # robot as a service: RGB, depth + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "full"
expert_time_estimate_hours = 3.0
diff --git a/tasks/yamlab-hang-mug-on-tree-i21-privileged/task.toml b/tasks/yamlab-hang-mug-on-tree-i21-privileged/task.toml
index 02f7eddbc6..2085f116d7 100644
--- a/tasks/yamlab-hang-mug-on-tree-i21-privileged/task.toml
+++ b/tasks/yamlab-hang-mug-on-tree-i21-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "yamlab" # YAMLab @ ec0455d on Isaac Lab @ f4aa17f / Is
robot = "i2rt-yam-bimanual"
family = "hang_mug_on_tree"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
expert_time_estimate_hours = 1.5
diff --git a/tasks/yamlab-hang-mug-on-tree-i21-standard/task.toml b/tasks/yamlab-hang-mug-on-tree-i21-standard/task.toml
index 4e07914751..379d3413f4 100644
--- a/tasks/yamlab-hang-mug-on-tree-i21-standard/task.toml
+++ b/tasks/yamlab-hang-mug-on-tree-i21-standard/task.toml
@@ -28,6 +28,7 @@ simulator = "yamlab"
robot = "i2rt-yam-bimanual"
family = "hang_mug_on_tree"
mode = "standard" # robot as a service: RGB, depth + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "full"
expert_time_estimate_hours = 3.0
diff --git a/tasks/yamlab-put-pot-on-cooktop-i00-privileged/task.toml b/tasks/yamlab-put-pot-on-cooktop-i00-privileged/task.toml
index d924cbc3a9..4171d3e379 100644
--- a/tasks/yamlab-put-pot-on-cooktop-i00-privileged/task.toml
+++ b/tasks/yamlab-put-pot-on-cooktop-i00-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "yamlab" # YAMLab @ ec0455d on Isaac Lab @ f4aa17f / Is
robot = "i2rt-yam-bimanual"
family = "put_pot_on_cooktop"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
expert_time_estimate_hours = 1.5
diff --git a/tasks/yamlab-put-pot-on-cooktop-i00-standard/task.toml b/tasks/yamlab-put-pot-on-cooktop-i00-standard/task.toml
index eef3e0b6bd..1df26ba284 100644
--- a/tasks/yamlab-put-pot-on-cooktop-i00-standard/task.toml
+++ b/tasks/yamlab-put-pot-on-cooktop-i00-standard/task.toml
@@ -28,6 +28,7 @@ simulator = "yamlab"
robot = "i2rt-yam-bimanual"
family = "put_pot_on_cooktop"
mode = "standard" # robot as a service: RGB, depth + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "full"
expert_time_estimate_hours = 3.0
diff --git a/tasks/yamlab-put-pot-on-cooktop-i03-privileged/task.toml b/tasks/yamlab-put-pot-on-cooktop-i03-privileged/task.toml
index 9b8811dde6..97678f9572 100644
--- a/tasks/yamlab-put-pot-on-cooktop-i03-privileged/task.toml
+++ b/tasks/yamlab-put-pot-on-cooktop-i03-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "yamlab" # YAMLab @ ec0455d on Isaac Lab @ f4aa17f / Is
robot = "i2rt-yam-bimanual"
family = "put_pot_on_cooktop"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
expert_time_estimate_hours = 1.5
diff --git a/tasks/yamlab-put-pot-on-cooktop-i03-standard/task.toml b/tasks/yamlab-put-pot-on-cooktop-i03-standard/task.toml
index 8cd320ce02..93108a8110 100644
--- a/tasks/yamlab-put-pot-on-cooktop-i03-standard/task.toml
+++ b/tasks/yamlab-put-pot-on-cooktop-i03-standard/task.toml
@@ -28,6 +28,7 @@ simulator = "yamlab"
robot = "i2rt-yam-bimanual"
family = "put_pot_on_cooktop"
mode = "standard" # robot as a service: RGB, depth + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "full"
expert_time_estimate_hours = 3.0
diff --git a/tasks/yamlab-put-pot-on-cooktop-i04-privileged/task.toml b/tasks/yamlab-put-pot-on-cooktop-i04-privileged/task.toml
index 25e4782ad5..466a880e99 100644
--- a/tasks/yamlab-put-pot-on-cooktop-i04-privileged/task.toml
+++ b/tasks/yamlab-put-pot-on-cooktop-i04-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "yamlab" # YAMLab @ ec0455d on Isaac Lab @ f4aa17f / Is
robot = "i2rt-yam-bimanual"
family = "put_pot_on_cooktop"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
expert_time_estimate_hours = 1.5
diff --git a/tasks/yamlab-put-pot-on-cooktop-i04-standard/task.toml b/tasks/yamlab-put-pot-on-cooktop-i04-standard/task.toml
index a1b0de66e9..e104dde1a6 100644
--- a/tasks/yamlab-put-pot-on-cooktop-i04-standard/task.toml
+++ b/tasks/yamlab-put-pot-on-cooktop-i04-standard/task.toml
@@ -28,6 +28,7 @@ simulator = "yamlab"
robot = "i2rt-yam-bimanual"
family = "put_pot_on_cooktop"
mode = "standard" # robot as a service: RGB, depth + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "full"
expert_time_estimate_hours = 3.0
diff --git a/tasks/yamlab-put-pot-on-cooktop-i07-privileged/task.toml b/tasks/yamlab-put-pot-on-cooktop-i07-privileged/task.toml
index 2e87afffea..f38581bd28 100644
--- a/tasks/yamlab-put-pot-on-cooktop-i07-privileged/task.toml
+++ b/tasks/yamlab-put-pot-on-cooktop-i07-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "yamlab" # YAMLab @ ec0455d on Isaac Lab @ f4aa17f / Is
robot = "i2rt-yam-bimanual"
family = "put_pot_on_cooktop"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
expert_time_estimate_hours = 1.5
diff --git a/tasks/yamlab-put-pot-on-cooktop-i07-standard/task.toml b/tasks/yamlab-put-pot-on-cooktop-i07-standard/task.toml
index fc9d1be7d8..b367341d06 100644
--- a/tasks/yamlab-put-pot-on-cooktop-i07-standard/task.toml
+++ b/tasks/yamlab-put-pot-on-cooktop-i07-standard/task.toml
@@ -28,6 +28,7 @@ simulator = "yamlab"
robot = "i2rt-yam-bimanual"
family = "put_pot_on_cooktop"
mode = "standard" # robot as a service: RGB, depth + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "full"
expert_time_estimate_hours = 3.0
diff --git a/tasks/yamlab-put-pot-on-cooktop-i08-privileged/task.toml b/tasks/yamlab-put-pot-on-cooktop-i08-privileged/task.toml
index b855fd94b6..5bbca04586 100644
--- a/tasks/yamlab-put-pot-on-cooktop-i08-privileged/task.toml
+++ b/tasks/yamlab-put-pot-on-cooktop-i08-privileged/task.toml
@@ -16,6 +16,7 @@ simulator = "yamlab" # YAMLab @ ec0455d on Isaac Lab @ f4aa17f / Is
robot = "i2rt-yam-bimanual"
family = "put_pot_on_cooktop"
mode = "privileged"
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
oracle = "full"
expert_time_estimate_hours = 1.5
diff --git a/tasks/yamlab-put-pot-on-cooktop-i08-standard/task.toml b/tasks/yamlab-put-pot-on-cooktop-i08-standard/task.toml
index 8edea4b0b0..5fbd229831 100644
--- a/tasks/yamlab-put-pot-on-cooktop-i08-standard/task.toml
+++ b/tasks/yamlab-put-pot-on-cooktop-i08-standard/task.toml
@@ -28,6 +28,7 @@ simulator = "yamlab"
robot = "i2rt-yam-bimanual"
family = "put_pot_on_cooktop"
mode = "standard" # robot as a service: RGB, depth + proprioception only; no object poses, no success predicate, no reset
+dynamics = "static" # scene dynamics (docs/PROTOCOL.md §2.14): scripts/task_dynamics.toml
protocol = "eai-standard/2.2"
oracle = "full"
expert_time_estimate_hours = 3.0