lerobot-imitation-loop
Use to bring up a new manipulation skill cheaply — teleoperate demos into a LeRobotDataset, train an imitation policy (ACT/diffusion/SmolVLA), evaluate in sim, then deploy, instead of reward-engineeri
SKILL.mdproject-trajectory/skills/lerobot-imitation-loop/SKILL.md
GitHubWhen to use. New manipulation task bringup. Why: imitation from ~50 demos is far cheaper and more reliable than RL reward-shaping for a first working behavior.
Procedure.
- Teleoperate ~50 demonstrations; record to a
LeRobotDataset(standard Parquet+MP4 format). - Train an ACT/diffusion/SmolVLA policy on the dataset.
- Evaluate in sim, then a supervised hardware trial; iterate on failure cases (add targeted demos).
- Done when: policy succeeds ≥ target rate on held-out starts, in sim and on hardware, with the number reported.
Knowledge: KNOWLEDGE-LIBRARY.md §B3. Example: the LeRobot record/train commands there.