Short answer
Existing process-monitoring footage is fixed-camera video of workstations and lines, recorded to watch production rather than to teach a robot. It can support activity recognition, coarse task timing, exception detection and workflow statistics. It rarely shows hand detail, an egocentric view or high-quality manipulation, and it carries high privacy and notice questions that need answers before any licensing.
Workstation cameras are not security CCTV
Plants run several kinds of cameras, and they differ in ways that matter to a buyer.
Security cameras watch doors, yards and aisles for incidents. Purpose-installed process cameras sit over a station, a conveyor or a machine to confirm that a step happened, to support quality review or to log downtime. They are often placed deliberately, at a repeatable angle, and sometimes triggered by a system event rather than recording continuously. That makes them more useful than general CCTV and also more specific in what they reveal about individual workers.
Treat the two as separate offers. A supplier who bundles "all our camera footage" is describing a mix with different rights, retention and quality profiles.
What this footage can and cannot support
| Question | Process-monitoring footage | Needs something else |
|---|---|---|
| Is a step happening, and roughly when? | Yes, at station level | |
| Coarse task timing and workflow statistics | Yes, if clocks are reliable | System timestamps to anchor them |
| Exception and anomaly spotting | Yes, where exceptions were logged | Labels from event or quality logs |
| Fine hand-object interaction | Rarely | Closer, higher-resolution or head-mounted capture |
| Egocentric viewpoint | No | Wearable cameras; see egocentric vs exocentric workplace video |
| Robot actions, joint states, forces | No | Robot, teleoperation or instrumented data |
| Held-out evaluation sets | Only if split by line, shift or time | Planned separation |
Humanoid control data cannot be built from this footage alone. Its realistic roles are workflow understanding, task-segmentation pretraining and context for evaluation, usually combined with other data. Fixed-camera footage limits covers the viewpoint problems in more detail.
Evaluation checklist
Use each row as pass, fail or unknown. Unknown on a high-severity row is a stop until answered.
| # | Check | Pass looks like | Evidence to request | Severity |
|---|---|---|---|---|
| 1 | Purpose and notice | Cameras installed for a documented purpose, with worker notice | Signage photos, policy text, impact-assessment summary if one exists | High |
| 2 | Training-use basis | The supplier has considered use for model training, not only monitoring | Written position from the supplier's counsel or privacy lead | High |
| 3 | Camera inventory | Each camera's station, mount, lens and resolution is listed | Camera register with field of view diagrams | Medium |
| 4 | Hand visibility | Hands and parts visible in most task instances | Random sample of 20 clips scored by you | High |
| 5 | Timestamps | Clocks synchronized, drift measured | Sync logs; example overlap with system events | Medium |
| 6 | Event linkage | Footage can be matched to work orders, alarms or MES events | Join keys and a sample join | Medium |
| 7 | Task coverage | Normal work, exceptions and changeovers are all present | Counts by event type and shift | Medium |
| 8 | Storage and retention | Original quality retained, not only compressed archives | Bitrate, codec and retention schedule | Medium |
| 9 | Redaction | Faces and badges handled without destroying hands or parts | Redacted samples at full resolution | High |
| 10 | Third-party content | Customer parts, drawings and screens identified | Content review notes | High |
A pass on row 1 is not a pass on row 2. Notice for monitoring does not by itself settle whether the footage can be used for model training; that is a diligence question for the supplier's counsel and yours. This is not legal advice.
Why privacy concerns weigh more here
Fixed cameras record the same employees repeatedly, often for years, and the footage can reveal pace, breaks and performance. European guidance on video devices addresses purpose, notice and proportionality for camera systems (EDPB, 2020). UK regulator guidance on monitoring workers recommends a data protection impact assessment where monitoring is likely to be high risk to workers' rights and freedoms (ICO, 2023). SourceX sources from US businesses, so neither document is the governing law for a US supplier; they are useful as checklists of what a careful supplier may already have considered. Rules for US sites vary by state and by workforce arrangements, so ask counsel.
Blurring or masking faces does not make permission questions go away. Pace, posture and body shape remain identifying in context, and heavy redaction can remove the hands you need. See redaction in demonstration video and worker consent and identifiable people.
Preparation the footage will need
Plan for work before any model can use it.
- Sampling. Plants record continuously. Choose clips by event, station and shift rather than taking everything.
- Alignment. Join video to line events so clips carry labels; see aligning video with system logs and MES records aligned to video.
- Redaction. Screens, labels, customer parts and faces all need review.
- Clipping and metadata. Cut to task instances and add station, shift and camera fields.
- Confidentiality. Process steps and part geometry can be trade secrets; see manufacturing data confidentiality.
Illustrative offer summary
# Illustrative only - synthetic values
offer: station_cam_archive_v0
cameras: 6 fixed, overhead and oblique, mixed models
period: "18 months, event-triggered clips only"
resolution: "mixed; 720p to 1080p"
labels: "work_order_id and alarm_code per clip, from MES export"
audio: "removed at source"
redaction: "faces and badges masked; hands unmasked"
exclusions: "customer_x lines, screens, guarded cells"
rights_status: "supplier counsel review pending"
Where this footage fits in a program
Treat plant cameras as a context layer. Station-level video tells a model what a line looks like when it runs normally and when it stops, which supports task and exception understanding and helps you choose where to commission closer recordings. It is a poor substitute for those closer recordings.
A practical sequence is to sample the archive, score hand visibility and label availability, and then decide. If most clips show silhouettes at a distance, price the archive as workflow context. If a few stations have close overhead views with reliable work-order links, those stations may justify extra preparation. Keep a separate held-out slice, by line or by month, so that exception-detection claims can be tested on stations the model never saw.
Also ask early about retention. Many plants overwrite video on a short loop, so what exists today may not exist when a license is signed, and a request to preserve footage is itself a commitment the supplier may want paid for.
Questions to ask suppliers
- Were these cameras installed for process monitoring, quality or security, and what notice did workers receive?
- Has anyone assessed use of the footage for model training, and can you share that assessment?
- How are clips selected for the archive: continuous, event-triggered or manually saved?
- What is the actual stored resolution and bitrate, not the camera's rated one?
- How are hands and parts visible in a typical task instance, and can I score a random sample?
- Which system events (work orders, alarms, quality results) can be joined to each clip, and how reliable are the clocks?
- What redaction was applied, and was it done on originals or on compressed copies?
- Do any clips show customer-owned parts, drawings or screens, and who can license them?
- Can the supplier delete or withdraw footage if a worker or customer objects?
Discuss your requirements
If fixed-camera plant footage fits your plan for workflow understanding or exception detection, describe the stations, event labels 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 license and delivery. SourceX reviews the requirement and follows up about potential sources; a request does not guarantee a match. For where this sits in the wider map, see the manufacturing data hub and the humanoid training data guide. For the catalog view of recorded footage, see video recordings.
Discuss your humanoid training-data requirements with SourceX
Sources
- Guidelines 3/2019 on processing of personal data through video devices (Version 2.0) — European guidance on camera notice, purpose and proportionality.
- Employment practices and data protection: monitoring workers — UK guidance recommending impact assessments for high-risk worker monitoring.