01
What this environment asks for
Retail aisles put an unusual pressure on object reference. A shelf holds dozens of near-identical items, so a description that names only the product type does not identify which unit the hand actually touched.
Descriptions therefore locate the target as well as naming it, and camera motion is recorded carefully because operators turn and look up and down constantly while working a bay.
02
How the work runs
- Segmented each pick-and-place as a single atomic action rather than batching a run.
- Located the target on the shelf as well as naming the product in each description.
- Marked camera motion as Large wherever the operator turned between bays or shelf levels.
- Separated the hand carrying the case from the hand placing stock.
- Cut end frames before large post-action head motion began.
- Applied blind external review to gate each batch before delivery.
03
What makes it hard
- Dozens of near-identical products on one shelf make the target ambiguous without location.
- Operators work quickly, so actions run close to the half-second floor.
- Large head motion between shelf levels dominates apparent object motion.
- Stock in hand occludes the shelf position it is about to occupy.
- Store lighting varies sharply between aisle and chiller.
- Long repetitive runs make consistency the dominant review risk.
04
Result
Atomic-action coverage of shelf work with location-grounded object reference and honest camera motion attribution, reviewed four times before release.
05
What comes out
Structured records covering pick, place, face and retrieve actions with per-hand language and located object reference.