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.md

GitHub

When 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.

  1. Teleoperate ~50 demonstrations; record to a LeRobotDataset (standard Parquet+MP4 format).
  2. Train an ACT/diffusion/SmolVLA policy on the dataset.
  3. Evaluate in sim, then a supervised hardware trial; iterate on failure cases (add targeted demos).
  4. 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.