Short answer
Industrial distributors hold structured records for very large ranges of parts, fasteners, tools and consumables, often with dimensions, materials, standards and classification codes such as UNSPSC. That suits parts recognition and attribute grounding. The images are usually manufacturer-supplied packshots, so rights, in-situ appearance and handling behavior need separate attention.
Why distributors are a distinct supplier type
A general retailer sells items a shopper can name. An industrial distributor, often called an MRO (maintenance, repair and operations) supplier, sells a very wide range that a technician specifies by number: hex bolts in many lengths, hose fittings by thread type, bearings by bore, gloves by cut level. The catalog is the product. Its value for robot perception is that attributes are deep and standardized, and that the items resemble what a humanoid would meet in a plant, a workshop or a tool crib.
That depth is why distributors earn their own page. A retailer record may have a title, a picture and a few attributes. A distributor record may list a dozen engineering parameters, the governing standard, material, finish and a classification code. For recognizing and describing a part, that is closer to a label than most images ever get.
What the records contain
| Record element | Typical content (Illustrative) | Use for perception |
|---|---|---|
| Part number and manufacturer | Distributor SKU plus maker's part number | Joins records across suppliers |
| Classification | UNSPSC code, internal category tree | Hierarchical labels, taxonomy mapping |
| Dimensional attributes | Length, diameter, thread pitch, bore | Size grounding, measurement prediction |
| Material and finish | Stainless 316, zinc plated | Material classes |
| Standards | Named norms the part meets | Attribute text, filtering |
| Images | Packshot, sometimes angle or detail view | Recognition, retrieval |
| Drawings | Dimensional line drawings | Shape priors; check drawing rights |
| Packaging data | Pack quantity, box size | Not about the item itself |
The classification layer is worth a closer look. UNSPSC is an open, global standard for classifying products and services, with eight-digit codes arranged in four levels (Segment, Family, Class, Commodity) and an optional two-digit business-function level; management of the code passed from GS1 US back to the UN Development Programme after 31 December 2024 (UNSPSC). A distributor that tags products with UNSPSC gives you a shared label hierarchy, which makes it easier to compare suppliers and define held-out classes. Confirm which edition the supplier uses, because codes change over time. For more on hierarchies, see product taxonomies.
What this data can support, and what it cannot
Distributor catalogs can support recognition of part families, retrieval of the right part number from an image, attribute prediction (is this M8 or M10?), grounding of technical terms, and the construction of fine-grained evaluation sets in which similar parts must be told apart.
They cannot show the parts in use. Packshots are lit, centered and isolated; a real bin holds a jumble in bad light, with oil, scratches and paper labels. They contain no force, friction or grip data, and a dimension listed in a record is a nominal value, not a measurement of the specific unit in a robot's hand. They also say little about assemblies or tool use. To go from catalog recognition to handling, a team needs the data described in object data for hands vs grippers and in the maintenance cluster. Small parts raise their own resolution issues; see fasteners and small hardware.
Rights and provenance
In many distributor catalogs the images and drawings come from manufacturers, so the distributor may hold a license to display them and not to sublicense them for training. The reverse can be true for images the distributor shot itself, and for text and attribute normalization it created. Ask for a split by origin: distributor-created, manufacturer-supplied, third-party. Part numbers and trademark marks appear in the images and can raise their own questions; see trademarks in training images and product image rights. Pricing, customer lists and order history are commercially sensitive and should be excluded from any data deal for perception purposes unless explicitly intended.
Illustrative evaluation sample
A small paid or free sample tells you more than a brochure. Ask for a slice such as the following and score it.
# Illustrative sample request - synthetic values
sample:
families: [fasteners, fittings, bearings, hand_tools]
records_per_family: 200
fields: [part_number, mfr_part_number, unspsc_code, title, attributes, image_ids]
images: all views per record, original resolution, with origin tag
checks:
duplicate_images_across_part_numbers: report share
attribute_completeness: share of records with >= 6 populated attributes
classification_consistency: same part type -> same code?
image_resolution: median short side in pixels
rights_origin: distributor | manufacturer | third_party per image
discontinued_items: included? with dates?
Quality issues specific to MRO data
Variants multiply: one image may stand for thirty lengths of the same bolt, so the picture does not show the attribute that distinguishes the record. Metric and inch variants can be near-identical in photographs. Attribute fields are free text in many catalogs ("3/8-16 x 1 1/2" versus "3/8 in. x 1.5 in."). Records are updated in place, so a snapshot without version dates mixes old and new. An audit before purchase, as described in product master data quality audit, exposes these issues cheaply.
How to choose among suppliers
Choose by the parts you must recognize. A general MRO distributor covers breadth: fasteners, safety supplies, abrasives, janitorial. A specialist in bearings, hydraulics or electrical connectors holds depth in a narrow area. Automotive aftermarket suppliers use a different standard family, covered in automotive parts catalog data. Manufacturers hold the most accurate models and specifications for their own lines; the trade-offs are discussed in object data from retailers vs manufacturers.
Matching the catalog to the task
Be specific about which kind of recognition your robot needs. A tool-crib task needs to tell a 10 mm combination wrench from a 12 mm one; a kitting task may only need to recognize the bag the fasteners come in. Catalog depth pays off in the first case, where attributes such as size and drive type are the labels. In the second, packaging images and label text matter more, and barcode or label data is the more useful field; see labels, barcodes and text on packages.
Also consider the fraction of the catalog that is relevant. Most of a distributor's range will not appear in your facility, so a license that lets you select by category or code range is more useful than an all-or-nothing file. Check that the supplier can filter by UNSPSC segment or family, or by their own category, before you commit to a package. For the broader supplier context see product catalog data for robot perception.
Limitations and extra preparation
Expect to dedupe images, resolve variants to specific part numbers, map categories to your own scheme, and add in-situ imagery from your own facilities or from a facility that uses the parts. If your task needs pose, add models and calibration (6D pose annotation). Document what you received in a datasheet in the style of Gebru et al. so that later users know its origin and limits.
Questions to ask suppliers
- How many active part records exist, and how many have at least one image?
- Which images did you shoot, which came from manufacturers, and under what license do you hold each?
- Do you tag records with UNSPSC or another scheme, and which edition?
- Is one image reused across variants, and are variant-specific attributes recorded?
- Can you deliver dated snapshots of the catalog, including discontinued items?
- Are drawings and CAD files part of the record, and who owns them?
- What share of attribute fields are free text rather than controlled values?
- Can pricing, customer-specific part numbers and order data be excluded cleanly?
Discuss your requirements
If you can list the part families and attributes your robot must recognize, describe them to SourceX. SourceX looks for US businesses that hold suitable material, checks the data and the supplier's licensing permissions, and coordinates license and delivery; nothing is contracted until a supplier agrees, and a request does not guarantee a match. Start from the humanoid training data guide or the object data hub.
Explore sourcing and licensing options with SourceX
Sources
- United Nations Standard Products and Services Code (UNSPSC) — structure, levels and stewardship of the classification code.
- Datasheets for Datasets — documenting composition and collection process.