Short answer
Fixed-camera and CCTV footage can support activity recognition, coarse task timing, exception detection and workflow statistics. On its own it is a poor fit for manipulation learning: hands are small or occluded, the viewpoint is far from a robot's, and no actions or forces are recorded. Surveillance recordings also carry heavy privacy and notice questions. Evaluate any offer against the checklist below before treating it as robot training data.
Why this footage keeps being offered
Businesses hold large volumes of camera footage from warehouses, kitchens, retail floors and facilities. It looks like free training data for workplace robots. It was recorded for security, safety or monitoring, not for learning manipulation, so its viewpoint, resolution and consent basis were set for other purposes. That is the core mismatch behind the distinction between existing footage and purpose-built collection; see existing workplace footage vs commissioned recording.
What fixed-camera footage can support
| Use | Why fixed cameras can serve it | Caveat |
|---|---|---|
| Activity recognition at coarse level | Whole-body motion and zone entry are visible | Label scheme needed; see action taxonomies |
| Task timing and throughput | Start and end of work in a zone | Needs time alignment with system events |
| Exception and anomaly detection | Long continuous coverage | Rare events are rare; labels required |
| Workflow statistics and layouts | Where people and goods move | Not manipulation detail |
| Scene and layout understanding | Stable background, repeated views | Single viewpoint per camera |
Ego-Exo4D shows what synchronized multi-view capture looks like when it is designed for learning: skilled activities recorded simultaneously from egocentric and several exocentric cameras, with 740 participants in 13 cities, 123 scene contexts and 1,286 hours (arXiv v4), plus audio, eye gaze, point clouds, camera poses and IMU (Grauman et al., 2023). Business CCTV often has a single exocentric view per area, none of those extra channels and no designed coverage. See egocentric vs exocentric workplace video.
What it cannot support
- Hand detail. Finger placement, grasp type and tool contact need close, unobstructed views.
- Egocentric viewpoint. A ceiling camera does not show what a humanoid's head camera would see.
- Manipulation-quality demonstrations. No actions, no calibration, no force; humanoid control cannot be trained from it alone.
- Object state detail. Resolution and compression often blur labels, parts and small tools.
- Coverage you did not choose. Cameras show where they were put, not where your task data would be best.
Using it to learn control would require inferring actions, which is a research problem rather than a procurement shortcut. For where video does help, see what human demonstration video can and cannot do.
Privacy and rights diligence
CCTV is personal data about identifiable people in many jurisdictions, and the permission basis is usually narrow. This is not legal advice; these are questions to put to the supplier and to counsel.
- The European Data Protection Board's guidelines on personal data processing through video devices (Version 2.0, adopted 29 January 2020) address video surveillance and its purposes (EDPB, 2020). Ask what purpose notices covered the cameras and whether training use was within them.
- The UK Information Commissioner's Office guidance on monitoring workers (October 2023) recommends a data protection impact assessment for monitoring likely to result in a high risk to workers' rights and freedoms (ICO, 2023). Ask whether the business performed one, and whether the recordings fall within it.
- Illinois' Biometric Information Privacy Act defines biometric identifiers to include a scan of hand or face geometry (Illinois General Assembly). Ask whether any processing, including analytics run by the supplier, creates such identifiers. See biometric privacy and workplace video.
Customers, visitors, minors and other non-employees may appear. Blurring faces or redacting does not by itself settle these questions, and heavy redaction can remove the cues a model needs; see redaction in demonstration video and worker consent and identifiable people.
Evaluation checklist
| # | Check | Evidence to request | Red flag |
|---|---|---|---|
| 1 | Original purpose of recording | Camera policy, signage text, retention policy | "Security" only, with no notice covering AI use |
| 2 | Authority to license | Contract with site owner, camera vendor and staff terms | Footage hosted by a third-party monitoring vendor |
| 3 | Camera geometry | Camera list with mount height, lens, resolution, frame rate | Unknown or varying cameras |
| 4 | Hands visible | Random sample of clips, sized by you; count usable frames | Hands below a size or visibility you set |
| 5 | Compression | Codec, bitrate, whether motion-triggered recording | Clips only when motion triggers; gaps |
| 6 | Time base | Timestamp source and sync to system logs | Wall-clock only, drift unknown |
| 7 | Labels | Existing labels, schema, quality checks | None, or alarm labels only |
| 8 | People and sensitive areas | Redaction plan, areas excluded | Restrooms, break rooms, customer areas included |
| 9 | Sampling plan | Proposed selection by site, shift, date | Only incident clips offered |
| 10 | Coverage | Counts by site, camera and time of day | One site, one shift |
A pass on the checklist means the footage is evaluable for coarse activity and workflow uses. It does not turn it into manipulation data.
Narrow cases where fixed cameras are useful
Some uses do fit. Overhead or side views of a bounded workstation can support task-level timing, counting and sequence checks, particularly where the camera was installed for process monitoring rather than security and the subject area is limited to equipment. Footage from a dedicated test cell, recorded with a signed notice and consent from the staff in it, is different in kind from general site surveillance. Even then, expect coarse labels. If a supplier proposes fixed-camera footage as a manipulation dataset, ask them to show hand-sized detail in a random sample, and compare it with the viewpoint guidance in matching camera viewpoint to humanoid sensors.
Preparation if you proceed
Plan sampling to select clips, heavy redaction decisions agreed before delivery, alignment with system events such as scanner or machine logs (see aligning video with system logs), and a label plan. Budget for review of every clip sample by people with the right access. For robot manipulation, commission recordings designed for it, or consider egocentric collection; see head-mounted camera capture specification.
Questions to ask suppliers
- Which cameras recorded this footage, where were they mounted, and what are their resolution and frame rate?
- Why was it recorded, and what notices did workers and visitors receive?
- Who owns the recordings, and does any monitoring vendor hold rights?
- Did anyone assess data protection impact for the cameras, and can you share the outcome?
- Do recordings contain customers, minors or sensitive areas, and how are those excluded?
- Is footage continuous or motion-triggered, and what gaps exist?
- What labels or event logs exist, and how reliable are they?
- Can you share a random, not hand-picked, sample for review?
- What redaction is planned, and how does it affect hand and tool visibility?
- How many sites, shifts and seasons does the footage cover?
Discuss your requirements
If you are weighing fixed-camera footage against purpose-built recordings, describe the tasks, viewpoints 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. A request does not guarantee a match. Start at the humanoid training data guide or the workplace task demonstration hub, and see the catalog page on 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) — EDPB guidelines on personal data processing through video devices.
- Employment practices and data protection: monitoring workers — ICO guidance on monitoring workers, including DPIA recommendation for high-risk monitoring.
- Ego-Exo4D: Understanding Skilled Human Activity from First- and Third-Person Perspectives — Ego-Exo4D records skilled activity with synchronized egocentric and exocentric cameras.
- Biometric Information Privacy Act (740 ILCS 14) — Illinois BIPA scope for hand or face geometry scans.