Name the embodiment and controller
A task label such as “pick and place” does not describe the robot or its action space. Include the platform, end effector, control mode, and any limitations on acceptable configurations. If your project can tolerate several embodiments, explain how they will be represented and compared.
Define every signal
List the camera streams, state variables, and action fields you need. Specify units and coordinate frames and distinguish observed states from issued commands. Ask for an example episode schema that makes those meanings explicit before discussing a larger collection.
Align observations with actions
Describe timestamp conventions, sensor rates, and acceptable gaps. Establish how frames, commands, and outcomes are linked. A collection with all the requested file types can still be unsuitable if the temporal relationship between them is not documented.
Make acceptance task-specific
Define episode boundaries, reset procedures, task instructions, and success criteria. Decide how failures and interrupted attempts are labeled. Request a representative pilot and review it against the actual training pipeline before treating collection feasibility as settled.
Review checklist
- Task and target embodiment
- Required observation and action signals
- Episode counts and success criteria
- Sensor timing and annotation requirements
- Source permissions and intended use