Short answer
High-mix, low-volume assembly makes many product variants in small batches, so each variant has few repetitions and fixed automation is hard to justify. Data for it should be scoped by variation, not by hours: variants, parts, fixtures and sequence changes, with recordings split so evaluation covers variants and sites the model never saw in training.
Why variation is the whole question
A high-volume line repeats one job thousands of times. Dedicated machines and purpose-built grippers pay back, and a humanoid has to compete with them. A job shop or contract assembler builds many variants, each in small lots, and the setup, changeover and learning cost is spread over few units. That is why this setting is attractive for a general-purpose robot, and also why it is hard to supply data for: the robot must cope with product variants it has seen little or never.
The buyer's central question changes accordingly. It is not "how many hours of assembly?" but "how many distinct variants, parts, fixtures and sequences, and how do they relate?" Hours of one product repeated tell you little about what happens on the next one.
What variation looks like in practice
| Source of variation | Example | What it changes for a robot | Data that exposes it |
|---|---|---|---|
| Variant | Same enclosure with different connector layouts | Which parts, orientation, steps | Recordings per variant, with variant ID |
| Part | Alternative fastener type or supplier | Appearance, grasp, tool bit | Part images and parts lists |
| Sequence | Reordered steps for a configuration | Plan, preconditions | Sequence labels, assembly sequence data |
| Fixture and workstation | Different jigs per job | Reach, occlusion, camera view | Fixture photos, workcell layouts |
| Operator | Different technique across people | Motion, speed | Multiple operators per variant |
| Changeover | Retooling between jobs | Setup tasks, tool selection | Changeover procedures |
What published work offers
Assembly101 is a useful reference for how a dataset can represent variation. It contains 4,321 videos of people assembling and disassembling 101 toy vehicles, recorded with 8 static and 4 egocentric cameras, with more than 100,000 coarse and 1 million fine-grained action segments and 18 million 3D hand poses (Sener et al., 2022). The recordings are unscripted and include mistakes, and the authors introduce a mistake-detection task. Its lessons for a buyer are about structure: many distinct items rather than one, several viewpoints, fine step segmentation, hand tracking, and natural errors.
It is a toy-vehicle dataset recorded for research, so it says nothing about industrial parts, tolerances or force, and it does not tell you how much variety suffices for any model. Use it as a template for what to ask for.
Scoping by variation, not volume
A practical request lists variation axes and sets a target for each:
- Count distinct variants and how many recordings exist per variant. A set with many variants and few recordings each is closer to the deployment than the reverse.
- Describe variant relationships: shared subassemblies, shared fasteners, shared fixtures. They determine what a model can reuse.
- Record variant metadata: bill of materials, revision, and work instruction version per recording.
- Capture the long tail: rare variants, first-article builds and rework.
- Plan splits: hold out variants, not just recordings. To test transfer to new variants, a variant seen in training must not appear in evaluation, even from another operator.
- Cover changeover: setup and fixture change are real work in this setting and often not recorded.
Whether a given quantity of variants is enough depends on architecture, task and embodiment. Pilot with a small held-out variant set before committing to a large purchase.
Illustrative scoping sheet
# Illustrative only - values are synthetic
scope_id: hmlv_scope_v0
site_type: job shop (contract assembler)
variants:
families: 3
variants_per_family: "buyer-set; enough to hold out a distinct subset"
per_variant_recordings: "few, by design"
metadata_per_recording: [variant_id, bom_revision, work_instruction_version, fixture_id, operator_id]
labels: [step_segments, part_ids, tool_ids, errors_and_rework]
modalities: [fixed_video, head_video, work_instruction_pdf, parts_list]
splits:
train: variants A-F
eval_unseen_variant: variants G-H
eval_unseen_site: all variants at site_009
Limitations and additional preparation
Human assembly recordings do not contain robot actions, forces or contact signals, and they show a human body and hand. The embodiment gap remains, and published work reports human data as helping, not replacing, robot data. Variant-rich data compounds this: with few repetitions per variant, a model learns from variety, but supervised labels for each variant are expensive.
Existing job-shop footage is hard to use. It was recorded for training or quality review, not for variant coverage, and mapping each clip to a variant and revision may be manual. Operators change setups mid-job, and one-off builds are often not documented with a work instruction. Commissioned capture lets you define the variants, but needs site access and consent, and still reflects the host's product mix. For the supplier side see contract manufacturers as suppliers, and for the wider scoping see scoping manufacturing tasks.
Production data also raises confidentiality. Variants may be customer-owned designs. Ask who owns the design shown, and whether the customer has consented. Blurring labels does not resolve the question. This is not legal advice.
Quality and acceptance criteria
For a variant-rich set, acceptance should look at distribution, not totals.
- Variant inventory. A table lists every variant with recording count, operator count and site, so you can see long-tail variants with a single recording.
- Metadata completeness. Variant ID, BOM revision and work instruction version are present for each recording, with the missing share reported.
- Sequence fidelity. Step labels follow the work instruction for that variant, and deviations are flagged.
- Visibility. On a random sample, the hands and the insertion point are visible during part placement and fastening.
- Error content. Rework, wrong parts and dropped components are included and labeled, as a mistake-detection use requires.
- Split readiness. Variant families are identified so that unseen-variant and unseen-site evaluation sets can be created without leakage.
If a package shows many hours but few variants, treat it as depth on a narrow slice and price and scope it accordingly.
Plant documents can describe each variant's steps; see standard work and job breakdown sheets, and for recorded work, assembly demonstration video.
Questions to ask suppliers
- How many distinct variants and part families does the set contain, and how many recordings per variant?
- Is each recording linked to a variant ID, BOM revision and work instruction version?
- Are rare variants, first-article builds, rework and errors included?
- How many operators and workstations appear for the same variant?
- Are changeover and setup tasks recorded?
- Who owns the product designs visible in the footage, and has the customer consented?
- Can you identify variant families, so that unseen variants can be held out cleanly?
- Which camera positions were used, and are hands visible during insertion and fastening?
- Could new recordings be made for a named list of variants at your site?
Discuss your requirements
If you are scoping data for variant-rich assembly, describe the variants, fixtures and tasks to SourceX. SourceX reviews the requirement and follows up about potential sources among US businesses; a request does not guarantee a match. For orientation, read the manufacturing data hub and the humanoid training data guide, plus the catalog entry for manufacturing quality records.
Discuss your humanoid training-data requirements with SourceX
Sources
- Assembly101: A Large-Scale Multi-View Video Dataset for Understanding Procedural Activities — multi-item, multi-view assembly video with hand poses and mistakes.