Short answer
Manufacturers hold the authoritative product record: specifications, master data, CAD and original photography. Retailers hold breadth across many brands, customer-facing text and in-use or multi-angle imagery. Most perception projects need both, and rights differ: a retailer may display a brand's images without being able to license them for training, so ask who created each asset.
The choice in practice
A team assembling object data for a warehouse, kitchen or retail task usually asks, "Who has the product data?" The answer is layered. A product starts life in a manufacturer's systems: a part number, dimensions, bill of materials, CAD, studio photographs. It travels to distributors and retailers, which re-shoot some of it, rewrite the text, add reviews and map it to their own category tree. Each layer adds value and each changes who can say yes to a license.
This page compares the two supplier types on the dimensions that matter for robot perception. For what product catalog data does for perception in general, start with product catalog data for robot perception.
Comparison
| Dimension | Manufacturer | Retailer or marketplace |
|---|---|---|
| Breadth | One brand or line, deep | Many brands, shallow per brand |
| Specification accuracy | Authoritative: dimensions, materials, tolerances | Copied from manufacturers; may be rounded or wrong |
| Images | Studio photography; high resolution; consistent angles | Mixed: manufacturer packshots, own photography, user photos |
| 3D and CAD | Often holds engineering models | Rarely holds models; may have commissioned renders |
| Text | Technical and brand voice | Marketing copy, titles tuned for search, reviews |
| Taxonomy | Own scheme plus industry codes | Retailer tree; mapping to standards varies |
| Refresh | Follows product revisions | Follows assortment changes; lags |
| Rights to license | Holds copyright in most of its assets | May hold only display rights for brand images |
| Counterparty effort | Fewer, larger conversations per brand | One conversation, but rights may be split |
Retailer-derived research corpora illustrate the breadth: Amazon Berkeley Objects pairs 147,702 product listings and 398,212 unique catalog images with metadata of up to 18 attributes per product, and artist-made 3D models for 7,953 products (Collins et al., 2021). That is one marketplace's scale. It says nothing about what any other holder can license.
Where manufacturers are stronger
Authoritative measurements. GS1's Package and Product Measurement standard covers nominal measurement attributes for retail and non-retail products from consumer unit to pallet, and the GDSN lets trading partners share product master data through GS1-certified data pools and the GS1 Global Registry (GS1 PPM). Manufacturers that publish into that network hold a structured, versioned record. Note that these are nominal values: the declared size of a design, not the measured size of a unit in a robot's hand. See product dimensions and weights.
Models and drawings. Manufacturers can hold CAD and sometimes simulation-ready assets; see manufacturer 3D models and CAD. Retailers rarely do.
Rights clarity. The party that took the photograph or built the model usually owns it. For a narrow brand scope, one agreement may cover images, models and text.
Where retailers are stronger
Breadth across brands. One retailer may carry many brands; assembling the same range from manufacturers means many agreements.
Consumer-facing imagery. Retailers often hold lifestyle or in-use photos, size comparisons and views a shopper wants, which can be closer to real scenes than a studio packshot. See packshot vs in-situ images.
Language that matches how people ask. Titles and reviews carry the words customers use, which helps language grounding; see catalog text for language grounding.
Operational reality. Retailers and fulfillment operators also hold what actually moved: SKU velocity, returns, damaged goods photos. Those are different assets from catalog data and are covered elsewhere in the cluster.
What neither can supply
Neither supplier type holds the object as it appears in your facility, how it behaves when handled, or what a robot did with it. Catalog images show an ideal unit on a plain background; both retailers and manufacturers lack force, friction, grip and failure data, which come from robots or instrumented capture. Treat catalog data as the recognition layer and plan the rest separately, as laid out in object recognition data vs manipulation training data.
Rights: who can say yes
The same image can have several rights holders: the photographer, the brand, the retailer that commissioned the shoot, and a marketplace seller that uploaded it. A retailer's website terms usually describe what visitors may do with content, and training is not automatically included. The U.S. Copyright Office's report on generative AI training discusses where copying can occur in development, fair use, and licensing options (U.S. Copyright Office, 2025). This is not legal advice. Practical diligence is to require a statement of origin per asset class and a clear grant from the party who can give it. Details are in product image rights and licensing terms for robotics training data.
Illustrative approach: pick by the product set
# Illustrative decision aid - synthetic categories
product_set: 300 SKUs of household cleaning products from 40 brands
approach:
step_1: retailer for text, multi-angle and in-use images, and category tree
step_2: ask retailer which images are retailer-owned vs brand-supplied
step_3: for brand-supplied assets, approach top brands by SKU share for rights
step_4: manufacturers for measured dimensions, CAD, label artwork where needed
product_set_2: 60 machined parts from one maker
approach:
step_1: manufacturer first - CAD, drawings, studio images, master data
step_2: distributor for breadth of related parts, if required
Combining both sources
In most projects the stronger pattern is to use retailer data for breadth and manufacturer data for depth on the items that matter. Join the two on a global identifier such as a trade item number, or on the manufacturer part number where one exists, and keep both records with their provenance rather than merging them silently. Where the records disagree on a dimension or a color name, keep the manufacturer value as the reference and flag the retailer value.
Plan for the cases where no shared identifier exists, such as private-label goods that a retailer sells under its own brand. For those items the retailer or its supplier is the only source, and the rights position needs its own check.
Limitations and extra preparation
Either source needs a master-data audit before use (product master data quality audit), identifier matching across sources, deduplication, and a rights ledger. Assets from several suppliers will disagree; keep provenance per record. Records change when products are redesigned, so lock a dated snapshot.
Evaluating a sample
Request the same twenty products from each source and compare the records side by side. Count mismatched dimensions, differing color names, images that appear in both but at different resolution, and records present at one source only. The comparison costs little and shows quickly which source is more complete for your product mix, which is more accurate, and where the identifiers fail to join.
Questions to ask suppliers
- Which of these assets did you create, and which were supplied by manufacturers or sellers?
- Do your agreements with brands allow sublicensing images for AI training?
- Are dimensions nominal, measured, or supplied by the brand, and are they versioned?
- Do you hold any 3D models or CAD, and under what terms?
- Which product identifiers are present (GTIN, manufacturer part number, internal SKU)?
- How often do records change and can you provide a dated snapshot?
- Are user-contributed photos or reviews included, and who owns them?
- Can you exclude pricing, supplier terms and customer data?
Discuss your requirements
If you can name the product set and the assets you need, describe the object data requirement to SourceX. SourceX looks for US businesses that hold the data, checks the data and the supplier's licensing permissions, agrees pricing and allowed uses in a license, and coordinates delivery. A request does not guarantee a match, and nothing is contracted until a supplier agrees. Start from the humanoid training data guide or the object data hub.
Explore sourcing and licensing options with SourceX
Sources
- GS1 Package and Product Measurement Standard (formerly GDSN Package Measurement Rules) — nominal measurement attributes and master-data sharing through GDSN.
- Copyright and Artificial Intelligence, Part 3: Generative AI Training (pre-publication version) — copyright, fair use and licensing in AI training.
- ABO: Dataset and Benchmarks for Real-World 3D Object Understanding — scale of a retailer-derived catalog corpus.