Short answer
Manufacturer CAD and 3D product models give exact geometry, part identity and a starting point for simulation assets and pose estimation. They do not show real-world appearance, wear or how an object behaves when handled, and physics properties exist only if someone recorded them. Rights are the main obstacle: designs are often proprietary or owned by the manufacturer's customers.
The buyer's problem
A perception or simulation team wants the true shape of the objects a humanoid will meet: parts, products, fixtures, tools. Scans are noisy and artist-made models are approximations. The manufacturer's own CAD is exact, carries part numbers, and can be rendered, voxelized or loaded into a physics engine.
The difficulty is that this is the most sensitive asset a manufacturer owns. A buyer needs to know which geometry can leave the company, in what form, and under whose permission.
What the data can support - and what it cannot
| Use | Supported by manufacturer geometry? | Notes |
|---|---|---|
| Instance recognition by shape | Yes | Render views with controlled variation |
| 6D pose estimation for known parts | Yes | Needs real images to close the sim-to-real gap |
| Grasp candidate generation | Partly | Requires gripper model and friction assumptions |
| Simulation assets | Partly | Needs mass, inertia, friction, compliance |
| Real-world appearance | No | Finish, wear, lighting and packaging are absent |
| Behavior when handled | No | Deformation, sliding and slip are not in geometry |
| Assembly relationships | If assembly CAD is provided | Constraints may be stripped on export |
Geometry supports recognition and planning; it does not show handling. That is the same line drawn in object recognition data vs manipulation training data. Pair CAD-derived renders with real images (see synthetic renders vs real images), use the geometry as the reference for 6D pose annotation, take part identifiers from PLM and ERP part master data, and expect to add physics yourself or buy it separately; see simulation-ready object assets.
What published 3D collections show about the options
Public collections illustrate the trade-offs between scanned, artist-made and engineering sources. Amazon Berkeley Objects pairs catalog images and metadata with artist-designed 3D meshes that have physically based materials, and benchmarks reconstruction, material estimation and multi-view retrieval (Collins et al., 2021). Google Scanned Objects provides scanned household items preprocessed for physics simulators under a Creative Commons license (Downs et al., 2022). Objaverse gathers a very large set of 3D models from a public source under Creative Commons licenses, with captions and tags (Deitke et al., 2022).
None of these is manufacturer engineering data, and their licenses do not transfer to proprietary CAD. They set expectations for metadata and format, and they show that license terms are as much a part of the asset as the mesh. Check any public collection's terms for commercial use before relying on it; see public dataset commercial license checks.
Supplier types
- Product manufacturers hold design CAD, often in native formats from their design system, plus released drawings and bills of materials.
- Contract manufacturers hold customer-owned CAD supplied for production. They usually cannot license it.
- Component vendors publish neutral-format models of catalog parts for customers to design in. Terms often restrict use to design-in, which is not the same as training use.
- Industrial distributors and retailers hold catalog images and metadata more often than geometry; see product catalog data for robot perception.
- 3D scanning and modeling studios produce models on commission, which gives clearer rights but not engineering accuracy.
Existing assets vs commissioned models
Existing CAD is exact and already exists, but is rights-heavy and may carry features the manufacturer will not release. Commissioned scans or models of a defined object set can be designed for your needs, such as including mass measurements and multiple states, but cost time and add a rights chain of their own. For a general route decision, see existing assets, new collection or specialist providers.
Illustrative package: geometry export for a part family
# Illustrative only - synthetic values, not an available dataset
package: part_family_geometry_v0
parts: 120 released part numbers (synthetic)
geometry:
format: STEP (AP214) plus tessellated mesh (OBJ or glTF)
units: millimeters
detail: external surfaces; internal features removed
tolerance_note: nominal dimensions only; no tolerance stack-ups
metadata:
part_number: pseudonymized_id
material_class: [steel, ABS, aluminum]
revision: rev_C
physics_fields:
mass_g: measured on 10 sample parts, others unknown
friction: not provided
appearance:
finish_note: text description; no textures
excluded: tooling, fixtures, supplier names, customer-owned designs
rights: training and internal evaluation; no redistribution; derived meshes allowed for internal use
Change the fields to match your pipeline. A pose-estimation team may need textures and several part revisions. A simulation team will care more about mass, inertia and friction than about surface detail.
Evaluating an offer
- Is the geometry exact CAD, a mesh derived from it, or a scan? Ask for a sample part in the delivery format.
- Are internal features, tolerances and proprietary details removed, and is the removal documented?
- Do part identifiers map to real catalog or ERP numbers, and are revisions tracked?
- Are mass, center of mass or material properties recorded, measured or estimated?
- Do textures or finishes match production parts, or are they generic?
- Do the models load in your simulator or renderer without repair? Test five before accepting the rest.
Questions to ask suppliers
- Does the supplier own the design, or are some models customer-owned or licensed in?
- Does any supplier contract restrict disclosure, derived works or use in machine learning?
- Can the supplier strip proprietary features and still preserve the geometry you need for grasping?
- Which neutral formats can they export, and who performs and checks the export?
- Are there export-control, defense or safety classifications that affect release?
- Can physical samples be measured for mass and friction, and who does it?
- Can models be licensed with the right to render synthetic images for training?
- What happens when a part is revised or discontinued?
Treat unresolved design-IP and customer-ownership questions as blockers, not details. A clean geometry file with an unclear chain of title is a risk for as long as you hold it.
Discuss your requirements
If your perception or simulation work depends on exact object geometry, describe the parts, formats and permitted uses to SourceX. SourceX looks for US businesses that may hold suitable material, checks the data and the supplier's licensing permissions, and coordinates the license and delivery; nothing is contracted until a supplier agrees, and a request does not guarantee a match. The humanoid training data guide and the product and object data hub show the surrounding options, and the catalog entry for CAD and PCB engineering files covers the related document type. For license terms, see licensing terms for robotics training data.
Explore sourcing and licensing options with SourceX
Sources
- ABO: Dataset and Benchmarks for Real-World 3D Object Understanding — catalog images and metadata paired with artist-made 3D models and benchmarks.
- Google Scanned Objects: A High-Quality Dataset of 3D Scanned Household Items — scanned items prepared for physics simulators under a Creative Commons license.
- Objaverse: A Universe of Annotated 3D Objects — large 3D model collection under Creative Commons licenses with captions and tags.