Cooking video data for physical AI
Define cooking video tasks around visible ingredients, utensils and sequences, with practical framing and clear limits on inferred outcomes.
Cooking combines repeated actions with substantial variation in ingredients, utensils and workspace layout. For AI research, the valuable signal may be a transfer between containers, the order of preparation or the way a person recovers from an awkward placement. Name that signal before deciding how much footage to collect.
A kitchen recording should make the relevant process observable without turning ordinary cooking into a performance for the camera. Capture should fit established safe routines and the permissions of the location. Visual descriptions can explain what appears to happen; they cannot establish food safety, temperature or taste.
Tasks to include in the brief
Gathering ingredients
Show selecting approved ingredients and placing them at the preparation area. Include the source and destination so the sequence distinguishes searching, picking up, carrying and putting down.
Transferring between containers
Record an ordinary pour or scoop with both containers visible. A remaining amount, a pause or a second attempt can be useful context without deliberately spilling or wasting ingredients.
Mixing and combining
Specify whether the target is utensil use, ingredient order or the visible mixing movement. Keep the bowl and working hand in view rather than recording only the participant’s upper body.
Packing a prepared item
Capture arranging an item in a container, positioning a lid and checking the closure. Distinguish a visibly completed closure from an ending hidden by the hands or camera angle.
Capture requirements
- Choose framing that preserves the preparation surface and the relevant hands. Reflective bowls, steam and tall containers can obscure interactions even when the overall image looks clear.
- Document meaningful variation such as container shape, ingredient texture, utensil type and workspace arrangement. Avoid changing every variable at once when comparing a specific action.
- Preserve the natural order of steps and ordinary pauses. If a recipe or task instruction provides context, record its identifier separately instead of filming a private screen.
- Include successful and naturally incomplete outcomes when permitted by the brief. An incorrectly seated lid is observable; a claim that a dish is correctly cooked needs additional evidence.
Useful annotation layers
- Action intervals can identify selecting, pouring, stirring, transferring and closing, paired with agreed ingredient or utensil names. Boundary conventions should explain when each movement starts and ends.
- Contributor context can supply recipe or ingredient information that the image alone cannot resolve. Keep those declarations distinguishable from generated observations and uncertain classifications.
- Text analysis does not measure heat, weight, nutritional properties, contact forces or hygiene. Fine contact frames, masks and calibrated geometry require dedicated annotations or sensors.
Permissions, boundaries and exclusions
- Never change knife handling, heat use or established kitchen procedures for a better view. Exclude tasks the participant is not already authorized and equipped to perform.
- Avoid customers, household members, delivery labels, payment details and confidential orders. A commercial kitchen brief must identify approved recording areas and participating staff.
- Do not stage burns, spills, contamination or other hazardous failures. If an important interaction is blocked, report the limitation and revise the capture plan.