Short answer
Part master data from PLM and ERP systems gives a perception model what images cannot: a stable part identity, nominal dimensions, material, revision and classification. Paired with photos or 3D scans it supports recognition and grasp planning. It does not show real-world appearance, wear or how a part behaves when handled, and it often carries design IP that limits what manufacturers can license.
The buyer's problem
A perception team can collect thousands of images of bracket-like objects and still not know which image shows which part number, revision or variant. Without a reliable identity link, labels are guessed from appearance, and near-identical parts with different tolerances get merged. The manufacturer's systems already hold that link: every part has a record, usually in an ERP item master and in a PLM system that tracks its design.
This page covers what those records contain, how they join to images, and what to check before asking a manufacturer to share them. For object recognition versus manipulation training generally, see object recognition data vs manipulation training data.
What PLM and ERP records hold
The two systems overlap but answer different questions.
- ERP item master. Part number, description, unit of measure, supplier, make-or-buy flag, lot and serial tracking, bill of materials position, sometimes weight and packaging data.
- PLM records. Revision history, CAD file references, drawings, material and finish specifications, tolerances, change orders, approved alternates.
For trade items moving between companies, GS1 publishes standards worth knowing as vocabulary. The GS1 Package and Product Measurement standard covers nominal measurement attributes of retail and non-retail products from consumer unit to pallet, and GDSN lets trading parties share product master data through GS1-certified data pools (GS1). GS1's Global Product Classification groups products in four tiers (Segment, Family, Class and Brick), each brick carrying attributes and attribute values (GS1 GPC). These apply to packaged goods and trade units; internally made industrial parts often use a company numbering scheme, so do not assume a GS1 mapping exists.
How master data supports perception
| Use | What the record contributes | What images or scans must still add |
|---|---|---|
| Identity labels | Part number and revision for each image | Many views, lighting and backgrounds |
| Class hierarchy | Family and category labels for coarse recognition | Examples of each class in context |
| Size and shape priors | Nominal dimensions, CAD model | Real-world deviation, finish, wear |
| Grasp planning inputs | Mass, material, feature locations | Friction, compliance, how it sits in a tray |
| Synthetic data | Exact geometry for rendering | Real sensor noise and clutter |
| Simulation assets | Geometry converted to neutral formats | Physical properties unless recorded |
Neutral geometry exports can be converted into simulator descriptions. Robot description formats such as URDF describe rigid links and joints (ROS URDF); a part model usually needs separate collision geometry and physical parameters before it behaves sensibly in simulation.
What master data cannot support
- Real appearance. CAD renders are clean. Machined, cast, painted or oily parts look different on a bench.
- Physics. Mass distribution, friction and stiffness are rarely recorded. Treat any physics columns as a check: who measured them, and how?
- Handling. A record does not show how a part is picked, presented or fumbled. See parts presentation and identification for what handling data adds.
- Condition. Burrs, scratches and variants outside the drawing are not in the record.
- Robot actions. There are no joint states or grasps. Master data supports perception and planning, not control training.
Sourcing routes
Existing exports from a manufacturer. A manufacturer can export part tables, classification fields and neutral-format geometry. This is the usual route, and it is high-rights: design IP, customer-owned designs and supplier contracts all apply.
Paired collection. Photograph or scan parts on the manufacturer's floor and attach the master record at capture time, so identity labels come from the system rather than from annotators. This is new collection, with its own site access and consent questions, and works best when scoped with pilot collection scoping style limits.
Public or industry catalogs. For standard components such as fasteners and connectors, manufacturer catalogs can substitute for internal records; see product catalog data for robot perception.
Preparation before licensing
- Export to neutral formats rather than native CAD or PLM dumps.
- Strip proprietary features, internal cost fields, supplier names and customer designs.
- Replace internal part numbers with stable pseudonymous IDs if identity must be hidden, and keep a controlled mapping.
- Add physical properties where they matter, with measurement method.
- Record revision and effective dates so images map to the right design state.
Illustrative field list
# Illustrative only - synthetic values
part_record:
part_id: part_00417 # pseudonymous, stable
revision: C
class_gpc_style: "family_x / brick_y" # only if a mapping exists
description: "formed bracket, two-hole"
material: "steel, zinc plated"
nominal_dims_mm: [60, 25, 12]
mass_g: 38 # measured, method: bench scale
geometry: ["part_00417_revC.step", "part_00417_revC_collision.obj"]
variants: ["revB", "revD"]
image_links: ["img_batch_021/part_00417_*.jpg"]
physics_properties: "not recorded"
excluded_fields: ["unit_cost", "supplier_name", "customer_design_flag"]
Evaluating an export
Request a small sample before committing: perhaps fifty part records spanning easy and awkward cases, with geometry and any image links. Check three things.
First, identity quality. Can two near-identical parts be told apart from the record alone, and do duplicates, superseded numbers or placeholder descriptions appear? Bad identity is worse than none because it silently mislabels images.
Second, fit to the task. For bins of loose small parts, the records matter most alongside industrial bin-picking data; for fixtures and stations, geometry should be consistent with workcell layouts and workstation geometry.
Third, exposure. Design data is among the most sensitive assets a manufacturer holds. Review what a competitor could infer from the fields before asking for them, and read manufacturing data confidentiality for how suppliers tend to think about it. A reduced export, such as nominal dimensions and class labels without CAD, may be enough for recognition work and far easier to license.
Questions to ask suppliers
- Which system is the source of truth for each field, PLM or ERP, and how often do they disagree?
- Which parts are your own designs and which belong to customers or suppliers?
- Can geometry be provided in a neutral format, and what proprietary features would be removed?
- Are mass, material and finish measured or copied from drawings?
- How are revisions handled, and can an image be tied to a specific revision?
- What classification scheme do you use, and is there a mapping to a public one?
- Do any records contain personal data, such as designer or inspector names, and how are they removed?
- Which fields do contracts with customers or suppliers restrict from sharing? This is not legal advice; ask counsel.
Discuss your requirements
If you need part records to anchor image labels or simulation assets, describe the parts, fields and permitted uses to SourceX. SourceX looks for US businesses that hold the data you describe, checks the data and the supplier's licensing permissions, and coordinates license and delivery. SourceX reviews the requirement and follows up about potential sources; a request does not guarantee a match. For wider context, see the manufacturing data hub, the humanoid training data guide, and the catalog entries on CAD and PCB engineering files.
Discuss your humanoid training-data requirements with SourceX
Sources
- GS1 Package and Product Measurement Standard (formerly GDSN Package Measurement Rules) — nominal measurement attributes and product master data sharing.
- GS1 Global Product Classification (GPC) — four-tier product classification.
- URDF (Unified Robot Description Format) — rigid-link robot description format used in ROS tools.