Short answer
Kit and multi-part data links a whole to its parts: a bill of materials, images of the assembled and disassembled product, and labels saying which part is which. Manufacturers and kitting operations may hold the structure and photos. The data can support contents recognition and completeness checks. It does not teach the motions of assembling or packing a kit.
Why part-whole structure matters
A product is often not one object. A power tool ships with a battery, charger and case. A medical procedure tray holds a dozen instruments. A cabinet arrives as panels, hinges and a bag of fasteners. For a robot that kits, packs, unpacks or assembles, the questions are relational: which parts belong to this kit, which are present, which are missing, and which sit inside which.
Standard recognition datasets label single instances. They have no field for "this is the left hinge of product X, one of four in the kit". Without it, a model can name every item on a table and still not know whether the kit is complete. The part-whole link is a data structure, and it has to come from whoever defines the kit, usually the manufacturer or the kitting operation.
This page covers sourcing for that structure. For completeness checks as an inspection task, read assembly completeness verification. The cluster overview is the product and object data hub.
What a kit dataset contains
- Bill of materials. The list of parts with identifiers and quantities. This is the backbone, and often lives in an ERP or product lifecycle system rather than an image library.
- Images of the whole. The assembled product or the packed kit as shipped.
- Images of the parts. Individually and laid out, from the angles a handler would see them.
- Part-whole mapping. Which image regions correspond to which bill-of-materials entries.
- Variants. Optional parts, regional differences, substitutions and revision levels.
- Packaging layout. Where each part sits in a tray, foam cutout or bag.
Instruction documents may help. Assembly manuals show part order and orientation, and the work-instruction side is covered in what SOPs can and cannot teach a humanoid.
A published reference for part-whole labels
IKEA-Manual pairs 102 IKEA objects with assembly manuals, with annotations including decomposed assembly parts, assembly plans, manual segmentation and 2D-3D correspondence (Wang et al., 2022). It shows what part-level annotation can look like when a manual and a 3D model are aligned.
It is a labeling reference only. Its objects are consumer furniture, and it is about understanding assembly from manuals and shapes, not about a workplace kitting operation or physical assembly by a robot. Whether its annotation conventions suit your kits is a design decision, not something the paper establishes.
Illustrative kit record
The schema below is illustrative, with synthetic values. It is not a standard, and the quantities are examples.
| Field | Illustrative value | Why it matters |
|---|---|---|
kit_id | kit_0041 | The whole |
revision | rev_C | Kits change; labels must be tied to a version |
bom_line | part_A07 x4 | Quantity expected |
part_visible | true | Present in the image |
part_region | [x, y, w, h] or mask ID | Links pixels to the bill of materials |
optional_part | false | Distinguishes missing from not-applicable |
container_slot | tray_slot_3 | Where it sits in packaging |
site_id | site_014 | Held-out-by-site evaluation |
The revision field prevents a quiet error: a picture of rev_B labeled against a rev_C bill of materials will teach the model that a part is missing when the design changed. Ask how the supplier ties images to revisions.
Where the data lives
Manufacturers hold bills of materials, CAD, marketing photos and packaging drawings. They know the structure, but their pictures are usually of finished kits in perfect order.
Kitting and fulfillment operations hold photos taken at packing stations, quality checkpoints and exception handling, including incomplete and wrongly packed kits. That is the interesting material, and the hardest to release because it can show customer orders, labels and staff. See 3PL providers as data suppliers for the warehouse side.
Distributors and service providers hold repair-kit and spare-part sets. Part numbers and naming conventions help, but images are scarce.
If existing sets are incomplete, a purpose-built capture of your kit configurations at your station is a realistic alternative. See existing assets, new collection or specialist partners.
What this data cannot support
Kit data tells a model what a complete kit looks like. It does not show how to build one. It has no grasp sequence, insertion force, orientation correction, snap-fit feel or the order a skilled kitter really follows. Training a humanoid to pick the right part from a bin, orient it and place it in a tray needs robot or teleoperation data, or instrumented human capture with synchronized sensors; images of finished kits, bills of materials and manuals cannot provide it.
Other limits to plan for:
- Occlusion. Nested parts and parts under others are invisible in a single view.
- Near-identical parts. Left and right variants, or two screw lengths, can be separated only by a measurement or a label you must supply. See fasteners and small hardware.
- Revision drift. Kits change faster than images are retaken.
- Confidential structure. A bill of materials can expose supplier names, costs and designs. License terms should say what is delivered and what stays private.
- Evaluation. Hold out kit families or sites, not individual frames of the same tray.
Acceptance checks you can set
- Every image is tied to a kit ID and revision.
- Every labeled part maps to a bill-of-materials line, with optional and substitute parts flagged.
- Incomplete, wrong and reordered kits appear, if completeness checking is the goal, and are labeled as such.
- Part-region labels are audited by someone who knows the product.
- Near-duplicate trays are grouped for splits.
For part-level meaning beyond identity, see affordance and part labels; for fields worth fixing before delivery, the object attribute schema; for geometry per part, manufacturer 3D models and CAD; and for why none of this teaches handling, object recognition versus manipulation data.
Questions to ask suppliers
- Can each image be linked to a kit identifier, revision and bill-of-materials line?
- Do the images include incomplete, mispacked or mixed kits, or only correct ones?
- Who defined the part names, and is there a controlled vocabulary?
- Are part regions labeled, and how were they checked?
- Does the bill of materials reveal supplier or cost data that must be removed?
- Are CAD models or assembly manuals available, and are they covered by the same license?
- What do images show besides the kit: labels, order data, staff?
- Who owns the photographs, and do customer contracts restrict their use?
- Can a kit family be held out for evaluation?
Discuss your requirements
If you need kit or multi-part imagery with a part-whole structure, describe the kit types, label needs 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 CAD and engineering drawings covers a related data type.
Discuss your humanoid training-data requirements with SourceX
Sources
- IKEA-Manual: Seeing Shape Assembly Step by Step — part decomposition and 2D-3D correspondence annotations for furniture; a labeling reference, not a kitting dataset.