Short answer
Depalletizing data for mixed cases is mostly perception data: images and video of stacked cartons of different sizes, faces and wrap states, paired with what was lifted next. Distributors and retailers hold receiving and break-down footage and WMS records. It can train case detection, edge finding and unload ordering, but it contains no robot actions, grip forces or arm trajectories.
The deployment problem
A mixed pallet is a changing puzzle. Cases of different heights are interlocked, shrink wrap and slip sheets cover faces, labels face outward or inward, and the top layer is rarely flat. Before any arm or humanoid hand moves, the system must decide which case is a safe next pick, where its edges are, and whether it will drag its neighbours with it.
That makes depalletizing a perception-led task. A team buying data for it usually needs many views of real stacks first, and motion examples second. Order-picking research frames the stakes: one review estimates that order picking can account for as much as 55% of total warehouse operating expense (de Koster et al., 2007). That figure covers order picking as a whole, not depalletizing alone, so it is used here only as cost context, not as evidence about depalletizing.
This page covers depalletizing. Building mixed pallets is a different decision problem, covered in case picking and mixed palletizing, and emptying trailers is covered in trailer and container unloading.
Task anatomy
| Step | What happens | Failure modes to look for |
|---|---|---|
| Pallet presentation | Pallet arrives at a break-down position; wrap, corner boards and slip sheets are present | Leaning stack, torn wrap, overhang |
| Wrap and strap removal | Wrap cut or pulled away, straps removed | Residual film over faces, loose strap ends |
| Layer and case selection | Operator or system picks the next case | Wrong case lifted, neighbours shifted |
| Lift and clear | Case lifted, rotated if needed, moved off the stack | Corner crush, drop, double pick of two interlocked cases |
| Hand-off | Case placed on a conveyor, cart or tote | Orientation wrong for the scanner |
| Slip sheet and pallet handling | Sheets removed, empty pallet stacked | Sheet catches on the next layer |
| Exception | Damaged or unreadable case set aside | Hold decision not recorded |
What this data can support, and what it cannot
Can support. Segmentation and instance detection of cases in a stack, face and edge detection under wrap, label and barcode visibility checks, estimation of which cases are unobstructed, and pallet-state classification. Video of experienced operators can support step recognition and give a view of unload order. ARMBench, collected in a warehouse pick-and-place workcell, shows the shape such perception data takes: images, videos and metadata for 235K+ pick-and-place activities on 190K+ unique objects, with segmentation, identification and defect-detection tasks (Mitash et al., 2023). It is a workcell for single items, not pallets, so treat it as a model for documentation and task framing.
Cannot support. Existing images and operator video do not contain robot joint states, end-effector commands, suction or grip pressure, or the force a case needs to break free of a stack. A human lifting a carton at waist height tells a humanoid little about its own reach, balance and grasp limits; see the human-to-humanoid embodiment gap. Footage also rarely records why one case was chosen over another. For the broader boundary, see what human demonstration video can and cannot do.
Capture and labels
Images from receiving stations, inbound audit cameras and damage-claim photos are typical existing assets. Useful label fields, which a buyer would need to confirm or request, include:
- case instance masks and visible-face masks, with occlusion flags for wrap and neighbours;
- case dimensions, ideally from a measured source rather than master data (see dimensioning and weighing station records);
- pallet identifier, linked to the logistic unit label where one exists; the GS1 Serial Shipping Container Code is an 18-digit key that uniquely identifies a logistic unit such as a case or pallet (GS1);
- order of removal, with timestamps per case;
- wrap state, slip-sheet positions and pallet condition;
- outcome per pick: clean, damaged, dropped, re-handled.
Illustrative capture specification
# Illustrative only - synthetic values for a buyer to adapt
task: depalletize_mixed_cases
site_id: site_017
views:
- overhead_stack: 1 camera, full pallet in frame, fixed
- side_oblique: 2 cameras, one per long face
- operator_station: 1 camera on the hand-off point
frame_rate: constant, stated per camera
per_pallet_record:
pallet_id: pal_000412
layers_visible: integer
wrap_state: intact | cut | removed
case_events: [{case_id, t_lift_start, t_clear, outcome}]
labels: [case_mask, visible_face_mask, occlusion_flag, removal_order]
excluded: [robot_state, grip_force, joint_angles]
holdout: split by site_id and by product family, never by frame
Supplier types
Distributors, retailers' regional distribution centers and 3PLs receive mixed and single-SKU pallets daily and often photograph inbound loads for claims. A 3PL's footage is usually covered by client contracts; see 3PL providers as warehouse data suppliers and client-owned inventory rights. Existing material may be enough for perception pretraining. If you need a fixed rig, a defined viewpoint or robot-matched views, discuss whether a commissioned recording at a cooperating site is feasible, and read existing footage vs commissioned recording.
Limitations and additional preparation
- Camera placement was chosen for security or claims, not for case segmentation. Check whether faces are visible and whether the angle changes between pallets.
- Product identity in the images may belong to a client of the facility, not the facility.
- Wrap glare, dark corrugate and label peel are common failure sources for vision models; ask for examples, not only clean pallets.
- Records of the pallet usually exist as ASN or WMS rows. They need time alignment and join keys before they label images.
- Evaluation sets must be held out by site and by product mix to test generalization, not by random frame. See warehouse pick evaluation protocol.
- None of this replaces robot or teleoperated data for the lift itself.
How to sample a depalletizing offer
Before licensing, ask for a stratified sample rather than a showcase set. Check the following on 30 or so pallets chosen by you, not the supplier:
- Layer coverage. Do images show the top layer only, or each layer as it is exposed? Mid-stack views, where cases lean on neighbours, are the hardest and the most useful.
- Wrap condition. Count pallets shown before and after wrap removal. A set that only shows unwrapped pallets will not prepare a model for film glare.
- Case variety. Count distinct case sizes, print styles and colors per pallet. Uniform pallets are easy and teach little about mixed stacks.
- Label side. Check how often a case label or barcode faces the camera, since downstream scanning depends on it.
- Sequence. If video exists, verify that the order of removal can be recovered from timestamps or event rows.
Write these checks into the acceptance terms, with thresholds you choose. If a supplier cannot provide a random sample, treat that as information about the archive.
Questions to ask suppliers
- Which cameras cover the break-down position, and at what height and angle?
- Are pallets recorded from the first wrap cut through the last case, or only part of the sequence?
- Is wrap present in most images, or is footage taken after removal?
- What share of pallets are mixed-SKU, and how many distinct case sizes appear?
- Do case events link to a pallet identifier or logistic unit label?
- Are measured case dimensions available, or only master-data dimensions?
- How are damaged, dropped and re-handled cases recorded?
- Who owns the goods and labels visible in the images, and has the owner agreed to licensing?
- Are workers' faces or names visible, and what notice applied?
Permissions and provenance
Packaging, product labels and client names can appear on every image. Ask who may grant rights over them before assuming the facility can. Workers may be visible; see worker consent and identifiable people in workplace recordings. Blurring faces or labels does not resolve a permission question and can reduce perception value. This is not legal advice.
Discuss your requirements
If you can describe the pallet types, views and label fields you need, describe the requirement to SourceX. SourceX reviews it and follows up about potential sources among US businesses, checks the data and the supplier's licensing permissions, and coordinates the license and delivery. A request does not guarantee a match. For the wider picture, see the humanoid training data guide and the warehouse and fulfillment data hub. Catalog context for logistics records is on the supply chain and logistics data page.
Explore sourcing and licensing options with SourceX
Sources
- Design and control of warehouse order picking: A literature review — order-picking cost share estimate.
- ARMBench: An Object-centric Benchmark Dataset for Robotic Manipulation — example of warehouse perception dataset scope.
- Serial Shipping Container Code (SSCC) — logistic unit identifier.