Imported from robotics-team-1213/rosy-platform (
learning/AGENTS.md). Install upstream withnpx skills add robotics-team-1213/rosy-platform --skill learning. Copyright stays with the author.
learning
Parent context: ../AGENTS.md
Generated: 2026-10-07 · Updated: 2026-10-12
Purpose
D-427 learning part. Perception curation, replay evaluation and offline training, shared artifact/provenance and policy registries, and Isaac Sim. Only the ROS package isaac_sim is a colcon_roots entry and only that package is allowed onto a device payload (D-427 Q8). The other folders are host-side tools. Local development outputs stay on X:, not in this repo. Existing remote data and store paths retain their declared lineage.
Subdirectories
| Directory | Purpose |
|---|---|
curation/ |
Perception recording/review/dataset preparation, OMX LeRobot export, and Pinky Episode conversion (see curation/AGENTS.md) |
evaluation/ |
Replay evaluation holder. The implementation is perception/ (see evaluation/AGENTS.md) |
envs/ |
Simulation environments. isaac/ is ROS package isaac_sim (see envs/AGENTS.md) |
registry/ |
Shared perception artifact/provenance store and model lifecycle, plus offline policy metadata (see registry/AGENTS.md) |
training/ |
Offline jobs for perception, OMX ACT, and Pinky speed models (see training/AGENTS.md) |
For AI Agents
Working In This Directory
- Do not install torch, LeRobot, or these job CLIs onto the robot image or the contract wheels.
- A training result is not promotion and not a Fleet dispatch. Registry promotion is metadata only.
- New run outputs go to a new directory on X:. Do not overwrite a previous run.
Testing Requirements
Each child names its suite. Run pytest only for committed HEAD through tools/remote/remote_pytest.py --log-dir X:/DevTemp/<session> -- <suite>, then compare the returned log with test/known_failures.py (D-584). The unified perception suite is test/learning/perception/; host results do not replace device or field acceptance.
Common Patterns
Korean README titles, English module names. Jobs take explicit train and eval sessions and refuse overlapping dataset names or MCAP hashes.
Dependencies
Internal
contracts/learning/src/rosy/contracts/learning/for Episode and PolicyArtifact shapes; standalone loaders addcontracts/learning/srcto the import path.contracts/foundation/for existing protocol validators where required.tools/perception/for the cross-ownerrosy_ml.pyCLI and pinned operator SSH options.deploy/robot/omx/for the LeRobot export requirements file.
External
- Separate Python environments and CPU/GPU dependencies named by each job README; perception can use a model PC or Colab. Keep them separate from the robot venv.
