Choose representative episodes
Include ordinary task execution and the situations most likely to expose mismatches: contact transitions, gripper changes, interruptions, and missing optional streams. Review complete episodes rather than only rendered highlights. The pilot should test the proposed source and configuration against the buyer’s requirements, not merely demonstrate that one clean recording can be played.
Review semantics before volume
Confirm the robot setup, action representation, units, frames, and controller context. Ask reviewers to interpret a simple motion from the records and compare it with the observation. Resolve ambiguity before aggregating trajectories. A large set of structurally valid arrays cannot compensate for uncertainty about what an action value means or which observation belongs to it.
Test the actual pipeline
Load the pilot through the planned preprocessing and training input path. Inspect image transformations, timing associations, missing-data handling, episode boundaries, and derived targets. Preserve original references so mismatches can be traced. A separate visualization tool may apply conventions differently, so successful playback should be considered one piece of evidence rather than the entire acceptance result.
Record decisions and outstanding issues
For each issue, identify affected examples, likely cause, required correction, and the next review step. Keep technical acceptance, task coverage, source documentation, and commercial terms distinguishable. Pilot acceptance applies to the reviewed version and conditions. If the configuration or export changes, assess which earlier conclusions remain valid before assuming the expanded collection will behave identically. Distinguish a correction to documentation from a change to the underlying records.
Review checklist
- Include ordinary and difficult episodes
- Confirm action and embodiment semantics
- Inspect the intended preprocessing path
- Version decisions and unresolved issues