Short answer
Packaging materials behave differently from the products inside: corrugated board creases and crushes, film stretches, tears and clings, foam compresses and recovers. Packaging suppliers may hold material specifications and product imagery, and sorting or recycling operations may hold in-the-wild images. Attributes and images support recognition and planning priors; they do not measure contact behavior.
Why packaging is not the product
A robot handling a shipping carton rarely handles the item inside. It handles a sheet of fluted paper, tape, a poly bag or a molded insert. Those surfaces determine friction, how much squeeze a gripper can apply before the box buckles, whether a bag can be pinched at a seam, and whether a foam corner springs back. Yet product datasets describe the item and treat packaging as background.
A separate data thread is needed because packaging properties vary within a material family. Two corrugated boxes of the same nominal size can differ in board grade, wall construction, moisture history and age. Two films can differ in thickness, stretch and cling. Recognizing "box" is not the same as knowing how this box will behave, and this page is about the information that sits between the two. For the product-side attribute questions, see material attributes for manipulation.
The cluster overview is the product and object data hub.
Three families and what to ask about each
Corrugated board
Ask for flute profile, board construction (single-wall or multi-wall), liner types, coatings, joint and tape style, and printing. Condition tags matter in practice: new, dented, softened by humidity, re-taped, crushed corner, bulging. Packaging suppliers hold nominal specifications; operating businesses hold how the boxes look after logistics. Neither tells you the force at which a given box deforms unless it was measured.
Plastic film
Film comes as stretch wrap, shrink wrap, bags, mailers and liners. Useful attributes include material family, nominal thickness, finish (clear, tinted, matte), seam type and closure. Transparency and glare create perception problems, and wrinkles create false edges. See transparent and reflective objects and deformable product properties.
Foam and cushioning
Molded pulp, expanded polystyrene, polyethylene foam, air pillows and paper void fill. Attributes include material, density class, form (corner block, sheet, peanut), and visible state (compressed, torn, shedding). Foam often changes shape under the camera's gaze as it is handled, so a single still can mislead.
A published reference for packaging-like scenes
ZeroWaste targets highly deformable, often translucent objects in extreme clutter on a recycling sorting line, and is presented as the first in-the-wild industrial-grade waste detection and segmentation dataset; its project page lists a ZeroWaste-f subset of 4,503 images (Bashkirova et al., 2021).
It is relevant because recycling lines commonly carry cardboard, film and plastics, so it may show how such materials look when crushed and mixed; check its class list rather than assume it. It is a labeling and capture reference only. The setting is recycling, not warehouse packing, and it does not provide material-property measurements. Whether its classes match your packaging mix is something to check in the dataset documentation, not assume.
Illustrative packaging record
This schema is illustrative, with synthetic values. The fields and tags are examples a buyer might adapt; none are standards or minimums.
| Field | Illustrative value | Use |
|---|---|---|
material_family | corrugated | Top-level class |
construction | single_wall, flute_class_x | Nominal supplier specification |
condition_tag | corner_crushed | Real-world state |
moisture_note | humid_dock | Context, not measurement |
closure | taped_h_seam | Affects grasp strategy |
property_source | supplier_nominal / measured_lab / labeler_estimate | Keeps provenance visible |
site_id | site_022 | Domain splits |
The property_source field is the one buyers most often miss. A labeler's guess that a box is "rigid" and a laboratory crush test are different kinds of evidence, and a model should not treat them as interchangeable.
Who holds what
- Packaging manufacturers and distributors. Specification sheets, product photography, 3D dielines and sometimes test data from their own quality programs.
- Contract packers, 3PLs and e-commerce operators. Photos of packing stations, damaged shipments and returns. See 3PL providers as data suppliers.
- Recycling and sorting operations. Large numbers of images of mixed materials, often from fixed cameras. See recyclable material recognition.
- Testing laboratories. Compression, burst or puncture results, if the lab can release them. These are the closest to physical properties, and often confidential to the client who ordered them.
For a buyer who needs measured properties, a small commissioned measurement program on the packaging actually used at your site may beat borrowed specification data. Discuss feasibility with SourceX; scope is case by case.
What this data cannot support
Images and attribute tables describe packaging; they do not capture how it responds to a hand. Material family and a nominal board spec do not give friction coefficients, stiffness at the grasp point, tear force or recovery time, and they say nothing about the interaction between a particular gripper and a particular film. Learning to carry a flexing carton, pinch a bag seam or press foam without crushing the contents needs robot or teleoperation data with contact signals, or instrumented tests. Packaging imagery from suppliers or sorting lines cannot substitute for it.
Additional limits:
- Nominal versus actual. Supplier values describe a specification, not the unit on your dock.
- Environmental effect. Humidity, temperature and age change board and film behavior; metadata rarely records them.
- Perception bias. Sorting-line scenes feature waste, with dirt and crushing that fresh packaging lacks.
- Evaluation. Hold out packaging suppliers or sites, since one supplier's boxes share look and construction.
Property labels that go beyond material family are discussed in physical property annotations, and the container side of the same handling problem is in containers, totes and bins.
Questions to ask suppliers
- Which material families and constructions are covered, and which are not?
- Are property values supplier-nominal, laboratory-measured or labeler-estimated, and is that recorded per row?
- What test methods were used for any measured values, and can the reports be licensed?
- Do images show used, damaged or wet packaging, or only new stock?
- Are closures, tapes and seams labeled?
- What imaging setups were used: fixed camera, studio, handheld, sorting line?
- Do images include product contents, customer labels or addresses?
- Who owns the images and test data, and do customer contracts restrict training use?
- Can a supplier or site be reserved for evaluation?
Discuss your requirements
If you are scoping packaging recognition or property data, describe the materials, conditions and permitted uses to SourceX. SourceX looks for US businesses that hold suitable material, checks the data and the supplier's licensing permissions, and coordinates the license and delivery. A request does not guarantee a match; SourceX follows up about potential sources. The humanoid training data guide maps the other routes, and the catalog entry for supply chain and logistics data is related.
Discuss your humanoid training-data requirements with SourceX
Sources
- ZeroWaste Dataset: Towards Deformable Object Segmentation in Cluttered Scenes — deformable, translucent items in clutter on a sorting line; a labeling reference, not a property dataset.