Start with the learning problem
Specify the tasks, target embodiment and intended training or evaluation use. Human video and robot demonstrations answer different questions; neither is a universal substitute for the other.
Inspect a representative sample
Review real episodes, a data dictionary, sensor configuration and annotation examples. Check that the sample includes difficult conditions, not only successful demonstrations.
Separate recorded from accepted volume
Ask how usable duration is calculated. Define incomplete episodes, duplicates, missing channels and excluded footage before comparing price per hour or episode.
Understand timing and coordinates
Robot learning may require synchronized observations, actions and state. Confirm sample rates, timestamp conventions, units, coordinate frames and calibration records.
Agree the label hierarchy
Decide whether labels describe an overall task, an intermediate step or an atomic action. Set boundary tolerance, ambiguity rules and the treatment of failure and recovery.
Set rights and handling requirements
Agree permitted uses, sharing, derivatives, exclusivity and retention. Identify recording permissions and the handling of people, screens and other identifying content.
Validate the release in your stack
Load a sample with the chosen loader and version. Ask for manifests, checksums, known limitations, split definitions and a procedure for rejected batches.
This guide supports project scoping. Actual data fields, performance, rights and delivery obligations depend on the agreed specification and sample.
Back to all resources