Short answer
Point clouds and depth scans record the measured shape of real product samples, including wear, tolerances and sensor noise that CAD files omit. They support pose estimation, shape completion and simulation assets. They do not carry mass, friction or compliance unless someone records them, and rights over scans of branded or proprietary designs need checking before any license is agreed.
What a scan gives you that a CAD model does not
A CAD model describes design intent: exact edges, nominal dimensions, clean surfaces. A scan describes one physical object as a sensor saw it. That difference is the reason to buy scans at all. A robot rarely meets the nominal part; it meets a worn, slightly warped, label-peeled sample, observed through a depth sensor with its own noise.
Scans therefore help most where perception must cope with real appearance: matching a partial depth view against a known shape, completing the hidden side of an object, or building a simulation asset that behaves like the real item's geometry. For the design-file route, see manufacturer 3D models and CAD, and for assets prepared for physics engines, simulation-ready object assets.
What published scan collections include
Public collections show what a well-specified scan package contains. Google Scanned Objects offers over one thousand 3D-scanned household items captured with a structured-light rig, released under a Creative Commons license and preprocessed for the Ignition Gazebo and Bullet simulators (Downs et al., 2022). OmniObject3D scanned about 6,000 real objects across 190 everyday categories and provides textured meshes, point clouds, multi-view rendered images and real-captured videos with masks and COLMAP camera poses (Wu et al., 2023).
The YCB object set took a manipulation-first approach: everyday objects of varied shape, size, texture, weight and rigidity, released with high-resolution RGB-D scans, physical properties and geometric models (Calli et al., 2015). It is the clearest example of scans being paired with the physical measurements that geometry alone lacks.
These are research collections of household items. They show a format, not the availability of any product class you need, such as industrial fittings or a specific retailer's assortment.
Where product scans come from
| Supplier type | What they may hold | Typical gap |
|---|---|---|
| Manufacturers with metrology or quality labs | Scans of parts for dimensional checks, often as meshes or point clouds | Built for measurement, not appearance; may lack color |
| Scanning vendors and 3D-asset studios | Textured scans made for e-commerce or visual effects | Usually cosmetic detail rather than accurate scale or physics |
| Retailers and marketplaces with 3D programs | 3D viewers' assets for selected products | Often artist-built meshes, not scans |
| Reverse-engineering and repair shops | Scans of legacy parts | Narrow, irregular coverage; unclear ownership of the design |
| Commissioned collection | Scans of the exact samples you specify | Cost, time, and still needs sample access and permission |
Because many scanned products are someone else's designs, the practical route is usually either licensing existing scan archives from the business that made them or commissioning a new scan of specified samples. Commissioned work can be designed to your spec but costs time and money; see choosing between existing assets, new collection and specialist providers.
Illustrative package: scanned product samples
# Illustrative only - field names and values are examples, not an available dataset
package: product_scan_pack_v0
objects: 150 sample items, 3 samples per SKU (new, used, damaged-packaging)
per_object:
point_cloud: PLY, metric units, per-point normals, 6+ registered views
mesh: OBJ/GLB, watertight where possible, texture 2048 px
depth_frames: raw RGB-D with intrinsics and sensor model recorded
pose_ground_truth: marker-board or turntable transform per frame
scale_check: caliper measurements of 3 axes, instrument ID logged
physical: mass_g, measured_by, center_of_mass_estimate (blank if unmeasured)
flags: transparent_regions, specular_regions, filled_holes (yes/no)
metadata: SKU, GTIN, scan_date, scanner_model, operator_A12, lighting_note
rights: AI training and internal evaluation; no redistribution; design-IP warranty stated
excluded: friction, compliance, handling behavior (not captured by scanning)
A buyer doing 6D pose work needs the per-frame pose truth. A buyer building simulation assets cares more about watertight meshes and recorded mass. Decide which use comes first before choosing a supplier.
Acceptance checks for scan quality
- Scale. Compare scan dimensions to caliper measurements on a sample; ask for the residuals, not a statement of accuracy.
- Completeness. Count holes and filled regions, especially on top and underside faces that scanners often miss.
- Registration. Check whether views overlap cleanly or show ghost edges and doubled surfaces.
- Material problems. Check how transparent, shiny and dark surfaces were handled; scanners often fail there, and the failure may be silently interpolated. See transparent and reflective objects.
- Identity. Confirm each scan links to a SKU or part number, so it can join catalog attributes.
- Format. Open the files in your own pipeline; a neutral format with units stated avoids silent scale errors.
Limitations: what scans cannot support
Geometry is not behavior. A scan does not say whether a bag is soft, a bottle is full, or a lid is stiff; mass, friction and compliance appear only if the supplier measured them, as in physical property annotations. A scan of an isolated object on a turntable also lacks the clutter, partial views and lighting a robot faces, so scans typically feed rendering or simulation, with real in-situ images added to close the gap; see synthetic renders vs real images.
Scans are one sample per physical object. Two units of the same SKU can differ, and a model trained on one scan may over-learn its flaws. They also describe the product at scan date, so redesigns need new scans. Finally, scans support recognition and grasp planning but not manipulation policies; for that distinction, read object recognition data vs manipulation training data.
Rights and provenance
Scans of a product copy its shape, and shape can embody a design that someone owns. Ask whether the scan owner is the designer or manufacturer, whether customer-owned designs were scanned under a services contract, and whether packaging artwork, logos or trademarks appear in textures. Strip or restrict proprietary features where the license requires it. Image rights are covered in product image rights; this is not legal advice, and counsel should review the grant.
Choosing between scan sources
Match the source to the job. For pose estimation on real shelves, scans of used and damaged samples matter more than pristine ones. For simulation, watertight meshes with measured mass matter more than texture. For a long-tail catalog, a manufacturer's existing metrology scans may cover far more parts than any commissioned run, but they were built for measurement and may need color, scale and format conversion. Ask each candidate for three sample objects in your target format, load them in your pipeline, and compare against a caliper reading before agreeing a larger scope.
Questions to ask suppliers
- Did you scan physical samples, or are the meshes produced from CAD or artist work?
- Which scanner or sensor, and what measured accuracy against reference dimensions?
- How many units per SKU were scanned, and in what condition?
- How are transparent, reflective and dark regions handled, and are interpolated areas flagged?
- Are raw depth frames, intrinsics and per-frame poses retained, or only the merged mesh?
- Are mass and other physical properties measured, and with what instrument?
- Who owns the underlying product design, and can you grant AI-training rights over its geometry?
- Do textures contain logos, barcodes or customer information that must be removed?
- Can scans be linked to SKU, GTIN or part-number records?
- What neutral export formats, units and coordinate conventions do you deliver?
Discuss your requirements
If you need measured 3D geometry for specific products, describe the object data your robotics project needs. SourceX investigates whether US businesses hold suitable scan material, checks the data and the supplier's licensing permissions, and coordinates the license and delivery. If no existing asset fits, SourceX can scope a commissioned collection of new recordings. A request does not guarantee a match.
Start from the humanoid training data guide or the product and object data hub. Related catalog pages cover CAD and engineering drawings and CAD and PCB engineering files.
Describe the object data your robotics project needs
Sources
- Google Scanned Objects: A High-Quality Dataset of 3D Scanned Household Items — scanned household items, structured-light rig, simulator-ready preprocessing.
- OmniObject3D: Large-Vocabulary 3D Object Dataset for Realistic Perception, Reconstruction and Generation — real-scanned objects with meshes, point clouds and multi-view data.
- Benchmarking in Manipulation Research: The YCB Object and Model Set and Benchmarking Protocols — scans released with physical properties and models.