Video is the modality with the highest collection overhead — consent, safety, and environment access all have to be arranged before a single clip is recorded — so we scope these projects tightly around a specific action set or environment rather than open-ended "record daily life" briefs, which produce unusable variance.
What a reviewer checks
Action-label accuracy against timestamps, consent documentation for every person and space in frame, and environment diversity against the brief (a robotics buyer asking for "kitchen tasks" needs more than one kitchen layout to be useful).
Worked example. A robotics team needed 5,000 clips of real household activity — folding laundry, loading a dishwasher, food preparation — across a range of home layouts. We scoped a fixed action list, recruited households with consent for identifiable footage, and reviewed each clip against the action-label spec before delivery.