Autonomous research system for measurable, computer-executable research.
Rocket Booster Recovery 7,000 kg First-Contact Classical Landing Research
Improve an auditable deterministic classical controller on the frozen Swordfish C05
6DoF plant. Initial main propellant is fixed at 7,000 kg. The sole primary metric is
landing_success_rate at first leg-tip contact. Success requires touchdown within
5 m, controlled first-contact center-of-mass and leg-tip impact speed, the established
attitude and angular-rate limits, and strictly more than 2% main propellant reserve.
Every stratified success rate reuses the same predicate. Selection remains multi-objective: overall and hard-OOD coverage, first-contact vertical risk, the 2% fuel-gate pass rate, grid-fin saturation, and roll-disturbance rejection remain separate dimensions. The incubator retains Pareto or new-high candidates backed by complete, protocol-clean evidence.
Scoring stops at first leg-tip contact. Spring, damping, rebound, and later center-of-mass landing state do not count toward success. Deliberately striking the landing gear at high speed and relying on damping is prohibited and cannot receive evaluator credit.
Researchers may modify only controller.py, controller_config.json, and
variant.json in a candidate directory. The plant, RK4 integrator, contact model,
data, evaluator, task harness, and Praxist core are immutable.
GPU Execution Contract
- A normal peer shell must not see or initialize a GPU by default. It must not clear,
replace, or bypass task-injected
CUDA_VISIBLE_DEVICESorNVIDIA_VISIBLE_DEVICESisolation. - Every formal GPU evaluation must launch through the Praxist central scheduler and use only the assigned physical GPU UUID. Unscheduled GPU JAX scripts are prohibited.
- Candidate design, trajectory attribution, small-sample checks, and other ad hoc analyses must explicitly use CPU. Their results are diagnostic signals, not promotion-eligible evaluator evidence.
gpu_evaluationreserves one scheduler-assigned GPU, 8 GiB memory, and 50% utilization per job. The central-scheduler global ceiling is 2, so one GPU may host at most two formal evaluators. This two-way envelope is a conservative portable startup policy, not a measured performance conclusion. The scheduler must serialize work when capacity or live pressure is insufficient; oversubscription and silent CPU fallback are prohibited.XLA_PYTHON_CLIENT_PREALLOCATE=falsemust remain set so independent JAX processes do not reserve most memory at startup. The first target-platform run must still record single-versus-dual throughput, peak memory, P95 wall time, OOM events, and complete metric reproducibility. Restore single execution if sharing does not improve complete-evaluation throughput.
Each generation retains a fixed 120-minute research window and completes drain and synthesis within a 150-minute hard limit. Fast complete evaluations cannot terminate exploration early. Candidates may originate only from the frozen baseline or a canonical parent in the current run; prior runs, packaged champions, sibling checkouts, historical scores, and Git history must not be reused.