Short answer
Changeover data documents how a production line is switched from one product to the next: tooling swaps, adjustments, cleaning, verification and restart. Procedures, checklists and video exist at many plants, but changeovers are infrequent, varied and often involve hazardous-energy control. They help describe tasks and exceptions; they do not provide robot actions.
Why changeover is a hard data problem
A line that runs one product for a week changes over once. The work is a sequence of unlike jobs: stop the line, isolate energy, remove guards, swap tooling or dies, change fixtures, adjust guides and settings, clean, load new material, run first-off samples, release. A single changeover can involve a dozen people and a dozen tools, and no two are identical because the starting state and the target product both vary.
For robotics this combination of infrequent, high-variation and tool-heavy work is the opposite of a repetitive pick-and-place cycle. That does not make it irrelevant. It makes the written and recorded record of how experienced technicians do it valuable as task structure, while physical learning is still a robot-data problem. Related reading: high-mix low-volume assembly, which shares the variation issue.
What records exist
| Record | Typical content | Use |
|---|---|---|
| Changeover checklist | Ordered steps, responsible role, verification | Task structure and success criteria |
| Setup sheet | Settings per product: speeds, gaps, positions | Parameters a task depends on |
| Lockout/tagout procedure | Isolation points and steps | Safety constraints on the task |
| Standard work or job breakdown | Time per element, key points | Step timing; see standard work and job breakdown sheets |
| Video of a changeover | Technician actions over 20 minutes to hours | Visual task examples |
| Downtime records | Start, end, reason code | Timing labels; see andon and downtime logs |
| Improvement notes | Before/after changes, time saved | Variation and rationale; see kaizen and improvement records |
Lean manufacturing practice calls the effort of shortening changeover "single-minute exchange of dies," and the search terms for this page include SMED. Plants that have run SMED projects often have recorded and timed changeovers, split into work that can be done while the line runs and work that requires a stop. That decomposition is valuable as annotation; ask for it.
The hazardous energy dimension
Where a changeover involves servicing or adjusting equipment, lockout/tagout may apply; confirm against the standard and the plant's own procedures. OSHA's standard states that employers must develop, document and use procedures to control potentially hazardous energy when employees service equipment, and that these procedures cover scope, purpose, authorization, rules and techniques, along with specific steps for shutdown, isolation and securing (OSHA, 29 CFR 1910.147). The standard has a narrow documentation exception, so check the details with the plant's safety lead.
For a buyer, that creates two points. First, a plant's energy-control procedures are task specifications: they list isolation points and ordering that any automated system working there would need to honor. Second, recording a changeover means cameras near equipment that is being isolated and re-energized, so the site's safety approval matters. See machine safety constraints for recording. This is not legal advice.
What the data can support, and what it cannot
Changeover records and video can support:
- Task decomposition. Ordered steps, dependencies and verification points.
- Language and procedure understanding. Checklists paired with video for instruction following.
- Exception and variation analysis. What differs between two products or two technicians.
- Planning constraints. Isolation, verification and release steps a planner must not skip.
- Evaluation items. Procedure-compliance questions and ordering checks.
They cannot support:
- Dexterous control. Loosening a clamp or aligning a die by feel is not captured in video. Force, torque and tool data would come from instrumented capture.
- General competence. One plant's changeover is one equipment configuration. A model trained on it will not infer another plant's machines.
- Statistical learning from few examples. Because changeovers are infrequent, existing recordings may number in the tens, not thousands. Treat that as a limit on what you can train, and do not infer a threshold from it; requirements depend on architecture and use.
Another point: the written procedure is what should happen. Technicians often work from experience, so recorded practice and checklist can diverge. That divergence is useful labeling material. See also what SOPs can and cannot teach a humanoid.
Illustrative package: changeover record set
# Illustrative only - synthetic values, not an available dataset
package: changeover_v0
events: 12 changeovers across 3 lines, product A -> B variants
per_event:
checklist_version, setup_sheet_ref, loto_procedure_ref
video: 2 fixed cameras + optional head-mounted, continuous
downtime_record: start, end, reason_code
element_times: internal vs external work (SMED split)
annotations:
step_segments, tool_used, adjustment_made, verification_check
deviations_from_checklist: free text + code
roles: technician_A03, operator_A11 (pseudonymous)
excluded: proprietary tooling drawings, customer part numbers
rights: training and evaluation use; deletion on request
Quality and acceptance checks
- Confirm each recording covers the full changeover from stop to first good part, not extracts.
- Check that the checklist version matches the recording date.
- Verify that deviations are annotated with a reason where possible.
- Check that hazardous-energy steps are visible or documented, including who performed them.
- For evaluation, hold out whole changeover events or lines.
Limitations and additional preparation
Long recordings need segmentation and indexing before they are useful, which is annotation labor. Technicians move around the machine, so one fixed camera misses much of the work. Tool-heavy tasks are poorly captured at ceiling level, and wearable cameras raise consent and safety issues. Setup sheets often hold commercially sensitive process parameters, which a manufacturer may withhold or mask. Rework and repair share many properties; see rework and repair operations.
Choosing between recordings and written procedure
If your goal is a planner or language model that understands changeover structure, written checklists, setup sheets and a modest number of well-annotated recordings are a sensible start. If your goal is a robot that performs part of a changeover, such as moving a tooling cart or placing a fixture, you will need robot-specific data for those steps, and the procedure data will serve as the task plan and safety envelope. Expect to scope those steps narrowly: changeovers are not one skill but many.
Since each event is long, consider annotating a subset densely instead of many events thinly. A dense annotation of three events with step boundaries, tools and checks can serve better than light labels over thirty.
Questions to ask suppliers
- How many changeovers are recorded, on which lines and product pairs?
- Do recordings cover the full event, and with how many camera views?
- Are checklists, setup sheets and lockout/tagout procedures available with version dates?
- Has the plant split internal and external changeover work?
- Are deviations from the checklist annotated or logged?
- Do cameras capture isolation and verification steps?
- Which process settings or tooling details are excluded for confidentiality?
- Who appears in the footage, and what notice did they receive?
- Could a planned changeover be recorded with a defined camera setup and safety approval?
Discuss your requirements
If you need changeover procedures or recordings, describe the lines, task types and permitted uses to SourceX. SourceX looks for US businesses that hold suitable material, checks the data and the supplier's licensing permissions, and coordinates the license and delivery. Specialist or commissioned capture is scoped case by case, and a request does not guarantee a match; SourceX follows up about potential sources. For the wider map, start at the humanoid training data guide or the manufacturing and assembly hub. The site's entry on SOPs and playbooks describes related document types.
Discuss your humanoid training-data requirements with SourceX
Sources
- 29 CFR 1910.147 The control of hazardous energy (lockout/tagout) — employer duty to document and use energy-control procedures during servicing.