Short answer
A pick face is the physical presentation of items a picker reaches: a static shelf, a gravity flow rack, a tote or bin, or a pallet. Each changes reach, line of sight and grasp approach, so a picking dataset recorded at one face type may not represent another. Specify the face types you need, record them as metadata, and hold them out when evaluating.
Why "picking data" is underspecified without the face
A buyer asking for picking video usually has an item in mind and forgets the surface it sits on. The surface decides almost everything the camera sees and the hand does. Picking a small item from a shelf, from a tote in a goods-to-person port, from the front of a flow rack and from a layer on a pallet are four different problems that share the word "pick".
As general background, a review of the warehouse literature organizes order picking around layout, storage assignment, routing, batching and zoning, and describes picking as the most labor-intensive and costly warehouse activity (de Koster et al., 2007). It does not catalog pick-face geometry or its effect on robot perception, so the geometry points below are reasoning for a buyer to test, not published findings. The warehouse task taxonomy places these faces within the wider task landscape.
The four face families
| Face type | What it looks like | Geometry consequences to check |
|---|---|---|
| Static shelf or rack bay | Items sit on fixed shelves at several heights, usually facing an aisle | Reach height and depth range; shelf lip and upper shelf occlusion; items behind others |
| Flow rack | Gravity or roller lanes feed items to a front pick position | A small, repeated pick position; lane fronts can jam or tilt; items arrive at the front rather than being reached into |
| Tote or bin | Open-top container, from shallow to deep, often presented at a station | Top-down view; wall interference; clutter inside; many items overlapping |
| Pallet | Stacked cases or layers on a base | Large flat faces, layer patterns, tall stacks, partial layers, shrink-wrap |
These are broad categories, and sites mix and modify them: a shelf with bins on it, a tote on a flow rack, a half pallet on a rack level. Record what the face actually is, not just its nominal class.
How face type changes the data
Reach. Shelf height and depth determine whether a picker bends, stretches or leans. A humanoid's workspace differs from a person's, so pick-face height distribution in the data should be known, not assumed. A flow rack concentrates picks at a repeated height; a pallet spans a range as layers are removed.
View. A camera on a chest or head sees a shelf from the front, a tote from above and a pallet from a standoff distance. Occlusion differs: items hide behind the lip of a tote, behind each other in a flow lane or behind shrink-wrap. The rack and shelf location images and bin picking image datasets pages cover the image side of those faces.
Grasp approach. Top-down access to a tote differs from lateral access to a shelf, and case picks from a pallet involve larger, heavier units than eaches from a bin. See each-picking demonstrations and case picking and mixed palletizing for those task variants.
Container handling. Totes themselves are objects to lift, tilt and move, which belongs to tote and container handling rather than picking from within them.
Illustrative face-coverage matrix
Use something like the following to state what you need before approaching suppliers. The cells are examples of what a buyer might mark, not a recommended mix.
| Item class (buyer-defined) | Shelf | Flow rack | Tote | Pallet |
|---|---|---|---|---|
| Small rigid items | required | required | required | not needed |
| Soft or bagged items | required | optional | required | not needed |
| Cases | optional | not needed | not needed | required |
| Irregular items | optional | optional | required | not needed |
# Illustrative only - synthetic values, not an available dataset
clip_id: clip_004471
site_id: site_017
pick_face:
type: flow_rack # shelf | flow_rack | tote | pallet | mixed
height_band: mid # low | mid | high | above_shoulder
depth: front_pick_position
occlusion: lip_and_neighbors
item_class: class_4
camera_view: chest_mounted
excluded: robot actions, grasp forces
If the matrix is used for sizing, note that the grid above has sixteen cells; that count is an arbitrary input and not a minimum. The volume per cell, and which cells matter, depend on the architecture, task and embodiment.
What to buy, by decision
- If you plan a single station type (for example a tote station), concentrate on that face and make sure the video is from the station's real geometry, including tote walls and lighting.
- If you plan to roam aisles, you need shelf and flow-rack faces at multiple heights, plus the transitions between them.
- If you plan case work, pallet faces and layer patterns matter more than item variety.
- If you plan to test generalization across faces, treat face type as a split dimension, not a nuisance variable.
Face type interacts with the supply side. A site that runs mostly goods-to-person stations is likely to have more tote footage than shelf footage, while a retail-style distribution site may be dominated by shelves. Check what a supplier's footage covers before you assume it matches your face needs. The warehouse layout and facility data page covers the layout context that predicts which faces a site has.
What this data cannot support
- Robot control. Whatever face type is covered, learning to reach and grasp on it needs robot or teleoperation data. Footage, location images and recognition labels alone do not contain actions or contact information.
- Reach feasibility for a given robot. Face height and depth recorded for people say nothing about a humanoid's workspace; test with the robot's own kinematics.
- Face-type transfer. A dataset on shelves does not demonstrate tote performance.
- Layout optimization. Pick-face data records how work was done, not how it should be designed.
Keep evaluation faces separate from training faces when the claim concerns generalization, and keep any evaluation of perception separate from monitoring of the people who appear in the footage.
Permissions and provenance
Pick-face images and video reveal client products, quantities and storage practices. A third-party logistics provider may not be able to license images of client goods, so ask early. See client-owned inventory rights. Footage that includes staff raises a separate consent question. This is not legal advice.
Questions to ask suppliers
- Which pick-face types does the site run, and in what proportions?
- At what heights and depths are picks made on each face?
- Do cameras capture the face from a view close to a robot's, and can that be checked in samples?
- Are face type and location recorded as metadata per clip?
- Which faces are mixed or modified, such as bins on shelves or totes in flow lanes?
- Are exceptions recorded, such as jammed lanes or partial layers?
- Can footage be separated by face for held-out evaluation?
- Which client goods are visible, and what restrictions apply?
- Can new recording target a specified face mix?
Discuss your requirements
If you are specifying pick-face coverage, describe the face types, item classes and permitted uses to SourceX. SourceX reviews the requirement, looks for US businesses that may hold suitable material, checks the data and the supplier's licensing permissions, and follows up about potential sources. If no existing asset fits, SourceX can scope a commissioned collection of new recordings. A request does not guarantee a match. The humanoid training data guide and the warehouse data hub give the wider map; the catalog page on images and inspection photos is related.
Discuss your humanoid training-data requirements with SourceX
Sources
- Design and control of warehouse order picking: A literature review — picking design context; it does not cover pick-face geometry.