What is the task?
Describe actions, tools, objects, initial state, final state, successful completion and permitted recovery cases. Define the unit of a recording.
Human demonstrations can document task-relevant decisions, interactions and outcomes. Before commissioning a dataset, establish which signals your robot-learning or evaluation pipeline can actually use — and which require separate robot demonstrations or sensors.
Human demonstration data records people performing defined tasks under an agreed protocol. Depending on the project, the package may include videos, task identifiers, scene conditions, timestamps, outcome labels, review decisions and documented recording rights. The buyer should define the intended use, data schema and acceptance thresholds before capture.
A dataset of human video is not automatically a dataset of robot actions. Imitation learning, vision-language-action research and other embodied AI approaches have different input and supervision requirements. Human demonstrations may inform a pipeline, but their usefulness must be tested rather than assumed. No model-performance or deployment outcome is guaranteed.
Agree answers with the model team and the people responsible for recording and reviewing the pilot.
Describe actions, tools, objects, initial state, final state, successful completion and permitted recovery cases. Define the unit of a recording.
List visible interactions and any measurements the model requires. Compare first-person, fixed, overhead and multi-view setups for the actual task.
Specify people, object variants, work surfaces, gloves, lighting and environments. Include only safe and authorized variation.
Agree task boundaries, object identities, outcomes, retry tags and any annotation taxonomy, sampling and quality-review process.
Document workplace and participant permissions, permitted uses, licensing, confidentiality limits, privacy handling and delivery access.
Agree sample size, accepted-hour definition, rejection criteria, file compatibility and the buyer's review window before collecting at scale.
An egocentric video may reveal the operator's visual context; an overhead camera may show hand placement more consistently; multiple views may reduce occlusion but increase synchronization and labeling work. If your target pipeline needs calibrated pose, tactile data, force or robot commands, scope those separately.
Ask the model team to load representative files and inspect real frames. Test whether labels, time alignment, field of view, variations and rights are adequate for the intended use. Record where footage is rejected and whether a different setup or data modality is needed.
The purpose of a pilot is a decision: proceed with an agreed collection scope, revise the protocol, or stop. A polished example clip without a clear acceptance test is not enough.
Next, review technical scoping for AI teams, the collection process and the quality and permissions framework.
Describe the capture requirements and intended use. Feasibility is assessed for every project.