Human video shows a task unfolding
A person’s recording can reveal object interactions, activity order, and visible outcomes. It may provide context for visual recognition or task understanding. Its usefulness depends on the viewpoint, completeness, and labels that accompany it.
Robot data describes a particular system
An observation-action episode can include signals tied to hardware and a controller. The meaning of an action depends on its representation and units. Moving between systems therefore requires a clear account of what each recorded field describes; the word “robotics” alone does not establish compatibility.
Specify the missing information
When reviewing a proposed source, identify what is directly recorded, what is annotated later, and what would need to be inferred. Do not silently equate a visual hand movement with a known robot command. Any proposed conversion should have its own evaluation criteria.
Choose against the research objective
Begin with the model objective and select the data accordingly. Some projects need visual task diversity; others require tightly synchronized actions. A mixed sourcing plan can be discussed, but each component should have a defined role and a separate acceptance specification.
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