Imitation learning¶
The LeRobot integration is an environment, not a robot: python/lerobot_robot_tatbot/
pins the LeRobot install that training, the policy server, the dataset tools
and vision capture run in, and holds the depth-v1 encoder training applies
to recorded D405 depth. Training uses the archived datasets (tatbot data hub): teleoperated recording was removed in 2026-09 and the ManiSkill data
factory in 2026-10, so no source produces new ones. The follower robot class
that ran policies on the blue arm was deleted with that arm’s use in 2026-10;
no robot client consumes the policy server today.
A checkpoint’s camera keys name wrist views of one physical arm, as the vision registry assigns them. A checkpoint trained on two wrist views must be replaced or retrained for a one-view setup; the missing view is never filled with the other arm’s image.
Dataset contract¶
An episode should include task text, action and observation schemas, sampling rate, tool/schema version, source revision, and a clear simulated-versus-real label. Keep personal data, raw recordings, credentials, and private model artifacts out of the public repository.
Reproducibility¶
Pin the code revision and dependency lockfile. Validate shapes and units offline before training. Compare policy results on held-out fixtures and report failed or rejected runs rather than selecting only successful examples.
Training and physical evaluation policies are private acceptance material; this page documents only the public adapter boundary.
Evaluation ladder¶
Offline dataset loss checks tensor and preprocessing compatibility; tatbot rollout bench plausibility checks a policy’s action chunks against a
demonstration-derived contract, offline or against a running policy server,
with no arm. There is no simulated policy screen: the ManiSkill evaluation was
deleted with that simulator, and none replaces it yet.