Trajectory / Episode Replay

Re-sending a recorded action sequence to a robot or simulator to check whether the data can be reproduced.

Trajectory replay means reading a previously recorded trajectory from a dataset and sending its actions, frame by frame, at the original frequency, to a real robot or a simulator, to see whether the outcome matches what was recorded. It serves three main purposes: quality checking, to confirm that action labels, timestamps, and coordinate frames were not recorded incorrectly — being able to reproduce the original motion means the data is usable for training; checking consistency between robots of the same model; and, in simulation, re-rendering an observation from a different camera viewpoint using the recorded state. Hugging Face's LeRobot provides a lerobot-replay command, which its documentation says is meant to test whether actions are reproducible and whether they transfer between robots of the same model; robomimic's playback_dataset.py can both re-render from state and replay from actions. Note that real-robot replay is open-loop — a small shift in an object's position can make it fail, which does not by itself mean the data is wrong.

ExampleUse lerobot-replay to replay your own recorded demonstration #0 of grasping a block on an SO-101 follower arm; if the motion it produces looks noticeably different from the original recording, that's a sign to check calibration or recording frame rate.

Also called

Episode Replay, Action Replay

Related

Trajectory · Episode · Data Quality Control · Open-loop Control · LeRobot · robomimic

Sources

Imitation Learning on Real-World Robots: Replay an episode (LeRobot docs)
robomimic: Dataset Contents and Visualization