Skip to content

armgit / moss

Public Imported from GitHub
0

MOSS: a small, mostly 3D-printed litter-picking rover. Cleaner streets, one maker at a time.

Python

Data: no recording format named before it is decided; recording goes through dimOS

metrox3bb2034017d9
diff --git a/README.md b/README.md
index d79a7f0..2515a97 100644
--- a/README.md
+++ b/README.md
@@ -35,7 +35,7 @@ The Jetson configuration is the one on the bench. Every line is in the [BOM](BOM
- An SO-101 arm (open source) with a NormaCore parallel-jaw gripper, widened to a 90 mm grip.
- Two cameras: a depth camera at the nose, a colour camera in the hand.
- A 2D lidar for navigation.
-- Software: dimOS for teleop and navigation, LeRobot for data and policies.
+- Software: dimOS for teleop, recording and navigation.
I pick up litter by hand on my walks. MOSS is the robot I want next to me doing the same. One less can on the ground.
diff --git a/Software.md b/Software.md
index 0347d4a..58efcc3 100644
--- a/Software.md
+++ b/Software.md
@@ -44,7 +44,7 @@ Nothing moves until you arm with the deadman released. Hold the deadman to drive
## Data and policies
-Demonstrations are recorded in the LeRobot format with the same camera names on every rover, so datasets pool without rework. The rules are in [docs/data.md](docs/data.md). The simulator (MuJoCo body, three recorded pickup missions) is in [moss-jev](https://github.com/metrox-eth/moss-jev).
+Demonstrations are recorded with the same camera names on every rover, so datasets pool without rework. The rules are in [docs/data.md](docs/data.md). The simulator (MuJoCo body, three recorded pickup missions) is in [moss-jev](https://github.com/metrox-eth/moss-jev).
## Tests
diff --git a/docs/data.md b/docs/data.md
index 0418a3f..cda4635 100644
--- a/docs/data.md
+++ b/docs/data.md
@@ -4,9 +4,9 @@ What we want is a robot that works. Data is how it gets there.
Proposed rules for MOSS builders:
-1. Teleop demonstrations are recorded in the LeRobot dataset format, with the same camera names on every robot, so datasets can be pooled without rework.
-2. Pooled datasets are published under an open license on Hugging Face. Anyone can train on them.
+1. Teleop demonstrations are recorded with the same camera names on every robot, so datasets can be pooled without rework.
+2. Pooled datasets are published under an open license. Anyone can train on them.
3. Each builder keeps the right to do what they want with their own recordings, including selling them. The pooled copy stays open.
4. The first policy trained on pooled data is evaluated on real litter, on a real roadside, and the numbers are published whether they are good or not.
-The camera layout and the recording tooling will be fixed in this repository before the first pooled dataset exists.
+The camera layout and the recording format are fixed in this repository before the first pooled dataset exists.