Short answer
Bills of materials, routings and work instructions record which parts go into a product, in what order, at which station and with what tools. From ERP, MES and PLM systems they give structured assembly plans for planning models. They contain no motion, force or perception data, and they are tightly bound to a manufacturer's product designs and process know-how.
The buyer's problem
Assembly planning needs three answers: what are the parts, what order do they go in, and what constrains the order. Videos show a person doing the task but leave the structure implicit. Instruction manuals give structure as pictures. Factories already hold the structure as data: a bill of materials lists components, a routing lists operations and work centers, and work instructions describe each operation.
A team that wants to teach a model assembly reasoning, or to align video with the operations it shows, may find this structured data more valuable than more footage. It is also the most commercially sensitive material in a plant. Getting it requires understanding what each record is.
The records and what each gives
| Record | System | What it holds | What it adds for a model |
|---|---|---|---|
| Bill of materials (BOM) | PLM / ERP | Parts, quantities, hierarchy, revisions | Part identity, sub-assemblies, kit lists |
| Routing / process plan | ERP / MES | Ordered operations, work centers, standard times | Operation order and station assignment |
| Work instruction | MES / document system | Step text, images, tools, torque values, checks | Step content and success criteria |
| Assembly drawings and CAD | PLM | Geometry, mating relations | Spatial relations between parts |
| Production records | MES | Actual start/stop, operator, lot, scrap | Real timings and deviations |
| Change orders | PLM | Revisions and effectivity | Which sequence applied to which unit |
Each layer answers a different question. For part master data and identifiers, see PLM and ERP part master data. For the human-readable instruction layer, see standard work and job breakdown sheets.
What the data can support - and what it cannot
Sequence data supports learning and evaluating assembly plans: predicting the next operation, checking whether a sequence is feasible, matching video segments to operations and building task graphs (see task graphs from procedures). It supplies vocabulary and constraints for language-conditioned models.
It does not show how parts are grasped, aligned or pressed together. It has no force data and no camera view of the work. Planned sequences also differ from actual ones: operators reorder steps, rework, or skip. Treat the routing as intent and the MES record as closer to reality. A model trained only on sequences cannot execute an assembly; it needs demonstrations or robot data for the physical steps. See assembly demonstration video and what SOPs can and cannot teach a humanoid for the same boundary applied to documents.
What published work shows
Public research datasets illustrate how sequence structure is used. IKEA-Manual pairs 102 IKEA objects with their assembly manuals and annotates decomposed parts, assembly plans, manual segmentation and 2D-3D correspondence (Wang et al., 2022). MEPNet converts a sequence of manual images into assembly steps, giving the parts to add and their 3D poses at each step, evaluated on LEGO manual datasets and a Minecraft house dataset (Wang et al., 2022).
Both start from consumer instruction material. Factory BOMs, routings and work instructions are structured differently, usually as database records and controlled documents, with variants, effectivity and revision control that toy and furniture manuals lack. Methods developed on consumer manuals may need adaptation, and their results do not predict performance on industrial data.
Supplier types and routes
- Product manufacturers hold the full set but are the most protective; expect NDAs and heavy scoping.
- Contract manufacturers hold routings and instructions for customer-owned products. They rarely own the rights.
- Product-lifecycle and MES software holders, on the customer's behalf, may be able to export records, subject to the customer's permission.
- Industrial training organizations and system integrators sometimes hold anonymized or generic assembly instructions.
Realistically, buyers should expect to start from a narrow request: a product family, a station, or a set of operations, with part names and dimensions abstracted. Confidentiality is covered in manufacturing data confidentiality.
Preparation before delivery
- Export in neutral formats (CSV/JSON, with a schema) rather than native system dumps.
- Map internal part numbers to pseudonymous identifiers while keeping hierarchy.
- Remove customer names, supplier names, prices and costs.
- Generalize or remove proprietary parameters, such as exact tolerances or process settings, if the owner requires.
- Keep effectivity: which revision applied to which period.
- Link records to video or images if recordings exist; see MES records aligned to video.
Illustrative package
# Illustrative only - synthetic values, not an available dataset
package: assembly_sequence_v0
scope: one product family, 3 variants, 2 stations (synthetic)
bom:
format: JSON tree
fields: item_id (pseudonymized), parent_id, qty, uom, material_class, revision
routing:
fields: operation_id, sequence_no, station_id, tool_class, standard_time_s, predecessor_ids
work_instruction:
fields: operation_id, step_no, text, image_ref, check_type, torque_class
production_log:
fields: unit_id_hash, operation_id, start_ts, end_ts, operator_id (pseudonymized), rework_flag
excluded: prices, supplier names, customer names, exact tolerances
rights: training and internal evaluation; no redistribution
Evaluating an offer
- Check that BOM hierarchy and routing sequence reconcile: every routed operation consumes listed parts.
- Check that production logs include deviations and rework, not only nominal times.
- Verify revision control so each record states which design it applies to.
- Test whether the work-instruction text describes the observable action, not only the outcome.
- Check the variant coverage: sequences differ by option set.
- Confirm any images can be released with the text.
Questions to ask suppliers
- Which systems hold the BOMs, routings and instructions, and who can export them?
- Do the products and processes belong to the supplier, or to customers?
- Which fields will be removed or generalized before delivery?
- Are production logs available, and do they include rework and operator deviations?
- How are revisions and variants represented?
- Can records be linked to video, images or CAD for the same operations?
- What is the confidentiality status of the process: trade secret, customer-restricted or public?
- Are work instructions available in multiple languages?
Discuss your requirements
If assembly planning is part of your roadmap, describe the product families, record types and permitted uses to SourceX. SourceX looks for US businesses that hold suitable records, checks the data and the supplier's licensing permissions, and follows up about potential sources; a request does not guarantee a match, and nothing is contracted until a supplier agrees. See the humanoid training data guide, the manufacturing and assembly data hub, and the catalog entries for manufacturing quality records and CAD and PCB engineering files. For exact geometry, see manufacturer 3D models and CAD.
Discuss your humanoid training-data requirements with SourceX
Sources
- IKEA-Manual: Seeing Shape Assembly Step by Step — manuals paired with assembly plans and part annotations.
- Translating a Visual LEGO Manual to a Machine-Executable Plan — manual-to-plan translation with parts and 3D poses.