3.1 · Episode Schemas and Collection Policy
Robot data are trajectories generated by a policy interacting with dynamics. Ordering, delay, controller state, calibration, and episode boundaries are part of the signal and must be preserved.
- Design a complete episode schema.
- Explain how collection policy shapes state coverage.
- Record provenance required for adaptation and leakage analysis.
Episodes, not loose samples
Store synchronized observations, actions, rewards or outcomes, termination reasons, timestamps, calibration, and controller state. Preserve episode boundaries so sequence models and rollout analysis reflect the interaction process.
Collection induces a distribution
Kinesthetic teaching, teleoperation, scripted control, autonomy, planning, video inference, and simulation visit different states. High demonstration success can still create a narrow dataset with poor recovery coverage.
Provenance is engineering data
Record robot type, control mode, software revision, operator, site, camera configuration, task taxonomy, intervention reason, license, and success label. Without provenance, compatibility and leakage cannot be audited.
A robot dataset is a sampled interaction process; collection policy is part of the model specification.
