Short answer
Shipping package damage images show crushed corners, punctures, tears, water marks, open flaps and crushed contents on cartons, parcels and pallets, captured at receiving, packing or claims. Their value depends on linking each image to a shipment key, a handoff point and a claim outcome. They show the damage, not how or when it happened, and rarely include handling motion.
Where damage images come from
Damage is recorded at handoff points: when a carrier delivers to a warehouse, when a 3PL receives inbound freight, when an order is packed, when a robot cell rejects a package, and when a customer claims. At each point a different person or system captures the picture, for a different reason. A receiving dock photo is evidence for a carrier claim. A return-processing photo is evidence for a refund. A pick-cell camera frame is a by-product of automation.
The consequences for data are significant. The same damage, a crushed corner, looks different from a dock camera, a phone photo and an overhead packing-station camera. And who took the image decides its metadata: dock photos may be tied to a bill of lading, claim photos to a claim number, cell frames to an order line. Cluster context is at the inspection cluster, within the humanoid data section.
What published work shows about warehouse image data
ARMBench is a public benchmark from an Amazon warehouse pick-and-place workcell, with images, videos and metadata for more than 235,000 pick-and-place activities on more than 190,000 unique objects; its tasks include object segmentation in clutter, object identification and defect detection (Mitash et al., 2023). It illustrates that defect detection can be framed as part of a manipulation workcell rather than a separate inspection stage. It does not tell you whether a given 3PL's images look like those, and a research benchmark should not be mistaken for commercially available data. License terms apply and should be read.
What damage images can and cannot support
They can support classifiers or detectors for damage type and severity, triage models that route damaged packages to a quarantine lane, claims-evidence tools, and evaluation sets for pick-cell vision. When linked to outcomes, as described in inspection images linked to outcomes, they can support models that estimate whether a visible dent implies damaged contents.
They cannot establish cause. Whether a crush occurred in a sort hub, on a truck or at the pick cell is not visible. They cannot show handling technique, and they do not describe the force a gripper applied; see handling-induced damage for what data about robot-caused damage requires. They also cannot show hidden damage: a package may look fine and contain a broken item, which is why linking to claims matters.
Damage categories to specify
| Category | Examples | Notes |
|---|---|---|
| Structural | Crushed corners, collapsed side, bulging | Often confused with design (deliberate recesses) |
| Breach | Punctures, tears, open flaps, torn tape | Small features, tape reflectance |
| Moisture | Water stains, swelling, mold | Color and texture cues |
| Content leakage | Stains, powder, leaked fluid | Needs context of the product |
| Labeling | Torn or covered shipping label | Affects routing; see labels, barcodes and text on packages |
| Pallet and load | Leaning load, torn wrap | Different camera geometry |
| Handling marks | Arrows, fragile marks, symbols | Marks per ISO 780 show handling instructions, they are not damage |
ISO 780:2015 specifies graphical symbols conventionally used on distribution packages to convey handling instructions (ISO, 2015). A model that confuses a "this way up" arrow with a tape strip, or a printed mark with a scuff, is making a labeling error; see handling marks on packaging.
Linking images to shipments
Each logistic unit can be identified by an SSCC, an 18-digit GS1 key identifying a case, pallet or parcel and used with despatch advice data (GS1). Not every carrier or 3PL uses it, but any stable shipment or parcel key serves the purpose. With the key, a damage image can be joined to origin, carrier, lane, handoff scan time, order content and claim result. Without it, images stand alone and the most useful part of the asset is lost.
Illustrative package specification
Illustrative; values are synthetic.
| Element | Example |
|---|---|
| Key | parcel_key (SSCC or carrier tracking hash), order_line_id |
| Image | Photo at receiving, 2 to 4 views, scale reference where available |
| Capture point | dock_receiving, pack_station, claims_photo, robot_cell |
| Damage label | Category, severity grade, location on carton |
| Outcome | content_ok, content_damaged, claim_paid, claim_denied |
| Package info | Box type, dimensions, content class (no personal data) |
| Time | Handoff timestamp and shift |
| Splits | By carrier lane and by week; no shared shipments across splits |
Quality and acceptance criteria
Key coverage. Share of images with a valid shipment key. Low coverage means a pile of unlinked photos.
Severity definition. Who grades severity, and against what? Carrier claim policies and warehouse receiving rules differ.
Capture point consistency. Mixed phone photos and fixed cameras need separate evaluation; see domain shift across lines for the general effect.
Class imbalance. Most packages arrive undamaged. Ask for the undamaged sample and its sampling rule; see rare defects and class imbalance.
Outcome latency. Claims resolve weeks later. A dataset snapshot may lack final outcomes for recent items.
Personal data. Labels show names and addresses; handwritten notes may too. Treat the label area as sensitive, not only the carton.
Supplier types
3PLs hold receiving and packing photos for many clients, and the clients often own the data in contract terms; see 3PL providers as data suppliers. Carriers hold claim photos and hub camera images. E-commerce operators and retailers hold return-processing images; see returned-item condition grading. Insurers and claims adjusters hold annotated damage evidence. Robotics integrators who run pick cells hold frames from customer sites, normally under customer terms.
Rights and provenance
Package images can show recipient names, addresses, tracking numbers, product contents and brand packaging belonging to shippers. A 3PL generally acts for its clients, so ask who can authorize licensing and whether shipper consent is required. Redacting address blocks reduces exposure but does not settle who may license the images. This is not legal advice.
Limitations and extra preparation
Damage images are captured for claims and exceptions, not for training. Photos are taken by different people on different phones under warehouse lighting, often only when damage is already obvious, so mild damage and undamaged packages are under-represented. Labels usually come from claim categories or carrier codes that mix cause, severity and liability, and the claim outcome may be decided weeks later in another system. A fixed dock or conveyor camera sees packages from a different angle than a humanoid's head or wrist camera, so viewpoint mismatch remains.
The images support damage detection, triage and evaluation; they do not show how a damaged package deforms or how a robot should handle it, which needs robot or teleoperation data. There is no general figure for how many images a damage model needs; it depends on the model, the damage classes and the claim being made. Plan for extra preparation: linking images to shipment identifiers and outcomes, removing names, addresses and barcodes that identify recipients, normalizing damage categories, and splitting by site and period.
Questions to ask suppliers
- At which handoff points are images captured, and by what device?
- Do images carry a shipment or parcel key, and can it be mapped to claim outcomes?
- Who assigns damage category and severity, and with what guidance?
- What share of images show undamaged packages, and how were they sampled?
- Are address labels and recipient data visible, and how are they handled?
- Do shipper or client agreements restrict use of images for model training?
- How many carriers, lanes and box types are covered?
- Is there a held-out period or site suitable for evaluation?
Discuss your requirements
If you need shipping package damage images from US logistics businesses, describe the damage types, capture points and permitted uses to SourceX. SourceX reviews the requirement and follows up about potential sources, checks the data and the supplier's licensing permissions, and coordinates the license and delivery. A request does not guarantee a match.
Tell SourceX what your evaluation set needs to cover
Sources
- ARMBench: An Object-centric Benchmark Dataset for Robotic Manipulation — ARMBench: 235K+ pick-and-place activities on 190K+ unique objects from a warehouse workcell, with a defect detection task.
- Serial Shipping Container Code (SSCC) — SSCC: 18-digit key identifying a logistic unit such as a case, pallet or parcel.
- ISO 780:2015 Packaging - Distribution packaging - Graphical symbols for handling and storage of packages — ISO 780 graphical handling symbols marked on distribution packages.