Recording Assembly in Cleanrooms

The buyer's problem

Cleanrooms and static-controlled areas can hold very precise manual assembly, and footage from them can be hard to get. Outsiders may not be allowed in. Equipment brought in may need to be approved, cleaned or purchased specifically for the room. Operators may be covered head to toe. Existing footage, if any, was probably recorded for training or quality purposes and with a viewpoint no one planned for machine learning.

For a data-program lead, the answer is often a commissioned recording with a specification the site can review. This page gives a template for that specification. It does not describe what any particular cleanroom classification or electrostatic control program requires, because that belongs to the site and its documents; it tells you what to ask and what to write down. For the generic decision between archives and new recording, see existing workplace footage vs commissioned demonstration recording.

How gowning and the room change the data

• Faces are covered. Hoods, masks and eyewear reduce identifiable features but do not remove all identifying cues; voice, build, gait and badges may remain. Covered faces also remove facial expression and gaze, which some perception tasks use.
• Hands are gloved. Gloves change the appearance of hands and may hide skin-based tracking cues. Hand-pose models trained on bare hands can degrade. State the glove type in metadata.
• Everyone looks similar. Uniform garments make it harder to distinguish operators and to tell an operator from a technician or inspector; label roles explicitly.
• Lighting is controlled and often unusual. Some rooms use special lighting for particular processes, which can affect color, exposure and contrast in video.
• Motion is constrained. People may move deliberately and in set patterns, which makes the footage cleaner and also less like other plants.
• Surfaces are reflective. Stainless steel, glass and polished fixtures create glare.

None of these are reasons to avoid cleanroom data; they are reasons to record the conditions so that models are not mistaken about what they have learned.

Illustrative collection specification

# Illustrative only - synthetic values; every item is a request to the site, not a known rule
spec: cleanroom_assembly_capture_v0
site: site_071 (synthetic)
task_scope: [component_placement, fastening, inspection]   # chosen by buyer
site_approvals_needed:
  - camera and mount types approved for the room (site decides)
  - cleaning or preparation procedure for equipment before entry
  - areas where capture is permitted, and areas excluded
  - supervision by site staff during recording
equipment:
  fixed_cameras: [overhead, oblique], resolution and frame rate set by the buyer
  close_view: bench-mounted or fixed on fixture, avoid contact with the work surface
  head_mounted: only if the site permits; mass and fit with the hood are checked
  lighting: use room lighting; record its type and any process-specific light
  calibration: intrinsics and extrinsics recorded before and after, per the site's rules
operators:
  gowning: described in metadata (hood, mask, glove type), not edited out
  roles: operator | inspector | technician, labeled
  consent: written, specific to video use and downstream training
labels:
  steps: start/end per step, step_id from the station instruction
  events: part_placed, fastened, rejected, restart
  outcomes: accept | rework, source named
sync: shared clock or visible sync event per recording
data_handling:
  review: site reviews footage before it leaves the room's network or site
  removed: customer part numbers, visible screens, documents
excluded: robot actions, joint states, contact forces
rights: stated in the license; customer and site restrictions listed

Parameters that depend on you

The values above are placeholders. Decide these before approaching a site:

• Task scope. Which steps, parts and error cases you need, and whether non-energized or dummy parts can stand in for error cases.
• Viewpoint. Whether you need the operator's view, a bench view or both. The operator's view usually needs a head-mounted device and may be the hardest to approve.
• Hand data. Whether you need 3D hand tracking, which depends on glove appearance and marker policy.
• Volume. For arbitrary planning, 2 operators, 3 board types and 8 repetitions would give 48 recordings. These inputs are arbitrary and not minimums; the right amount depends on your architecture, task and embodiment.
• Train/eval split. Hold out operators, shifts or fixtures if your claim is about generalization; see requirements specification template.

Acceptance tests

1. Approval record. The site confirms in writing the equipment, areas and supervision used.
2. Viewpoint check. Each declared view shows the work area, and a sample of frames lets an annotator mark each step.
3. Hand visibility. A sampled set of frames shows hands and the part during contact events at a stated rate.
4. Glove and gown metadata. Present for every recording.
5. Sync test. A visible event appears in all streams, and offset is measured.
6. Label audit. A second annotator reviews a sample of step boundaries and outcome labels and the agreement is reported.
7. Leak check. The supplier confirms that nothing visible, such as screens or labels, breaches the restrictions agreed with the customer.

Limits: what this specification cannot support

A recording specification yields footage, labels and metadata. It does not yield robot actions, joint states or contact forces, and a humanoid or other robot needs robot or teleoperation data, or handheld-device capture, in its own action space to learn control; footage, SOPs, records and recognition or pose labels alone do not train it. Cleanroom recording also cannot guarantee that a robot would be permitted in the room or that its materials suit it; those are deployment questions for the site.

Further limits:

• Site approval is not assured. A site may decline cameras, restrict angles or require changes.
• Generalization. One room's procedures and garments will not represent others.
• Hidden steps. The most critical operations may be done inside equipment where no camera can see.
• Gowning effects. Models trained on bare-handed footage may not transfer to gloved hands.

Permissions and provenance

Gowned operators are still people who can be identified, and in a small team, even more so. Consent should state what is recorded, how long it is kept, who may see it and that the recording is for training or evaluating robot perception and not for assessing individual workers' speed or technique. Redaction or removing identifiers does not settle permissions: customer designs, process details and the site's own procedures remain restricted whether or not faces are masked. Written procedures may be controlled documents under the site's quality system; ISO 9001:2026 lists documented information among its generic requirements (ISO, 2026), which says nothing specific about cleanrooms. Ask what documents may be photographed or must be excluded. See worker consent and identifiable people in workplace recordings and manufacturing data confidentiality. This is not legal advice.

Questions to ask suppliers

1. What equipment may enter the room, and what preparation does the site require?
2. Can a head-mounted view be recorded, or only fixed cameras?
3. Who supervises the session, and which areas are excluded?
4. How are gowning, glove type and roles recorded in the metadata?
5. Which customer or process details are visible, and who reviews footage before release?
6. What written consent will operators give, and what use does it cover?
7. Can error cases be staged with dummy or non-energized parts?
8. Which acceptance tests will the site run, and who signs them off?

Discuss your requirements

If you need a recording plan for controlled-environment assembly, describe the task scope, viewpoints and permitted uses to SourceX. SourceX reviews the requirement and follows up about potential sources among US businesses; for a new recording, it would discuss whether a commissioned collection is feasible, since each site's rules and approvals vary. SourceX checks the data and the supplier's licensing permissions, and coordinates the license and delivery. A request does not guarantee a match. See the humanoid training data guide, the manufacturing hub and manual electronics assembly for the task side, and the machine safety constraints for recording page for equipment hazards.

Describe the workplace demonstration data your robotics project needs

Sources

• ISO 9001:2026 Quality management systems - Requirements — generic requirements including documented information.