Short answer
Technician body camera footage is an existing first-person asset: it shows repair work from the worker's view, sometimes with narration or remote guidance. Evaluate it on framing, step completeness, labels and, above all, rights. Footage recorded inside customer premises, with customer voices and data on screen, carries consent and purpose questions that must be answered before any AI-training use.
Why this asset looks better than it is
Field-service organizations increasingly equip technicians with body-worn or helmet cameras. The footage is first-person, shows real equipment and real faults, and may include the technician talking through a diagnosis. For a buyer looking for repair demonstrations, that looks like an existing, large and relevant archive.
The reality is mixed. Cameras are often switched on for liability protection, quality audit or remote support, not for training data. Mounting differs, recordings start and stop at the technician's discretion, and the footage was made inside other people's homes and businesses. Each of those facts changes both the training value and the permission position.
This page is a checklist for judging an offer. It sits beside the repair video demonstrations page and the remote assistance session recordings page, which cover related first-person and guided footage. See the maintenance hub and the humanoid data sourcing guide for the wider picture.
What this footage can support, and what it cannot
It can support: visual pretraining on real repair scenes, step and tool recognition, progress estimation, vision-language pairs where narration exists, and evaluation of procedure understanding.
It cannot supply: robot joint states or actions, contact forces, consistent framing, or complete task coverage. Footage from existing cameras tends to cover the visible, memorable part of a visit: the opening, the fault, the new part going in. The middle steps may be off-screen or stopped. Human video lacks the action space a humanoid needs, and turning it into control signals requires retargeting, inferred actions or co-training with robot data. The distinction between existing and purpose-built capture is set out in existing workplace footage versus commissioned recording.
By contrast, research datasets are designed around consent. Ego4D reports 3,670 hours of daily-life video from 931 camera wearers across 74 locations and 9 countries (latest version), recorded with consenting participants and de-identification procedures (Grauman et al., 2021). Technician body camera archives usually differ on exactly those points: the wearer is an employee, the people around are customers, and the purpose was operational.
Evaluation checklist
Score each item pass, fail or unknown. Unknown on any rights item should be treated as fail until the supplier answers.
A. Rights and consent
| # | Check | Pass looks like | Evidence to request |
|---|---|---|---|
| A1 | Original purpose documented | Written purpose (safety, quality, support) and a statement of whether AI training was ever in scope | Policy, signage text, notices |
| A2 | Worker notice and agreement | Technicians informed about recording and any secondary use | Employee notices, agreements |
| A3 | Customer notice | Customers told of recording at the time of the visit | Visit scripts, booking terms |
| A4 | Third parties in frame | Process for bystanders, children, neighbors | Redaction rules, sample frames |
| A5 | Authority to license | Supplier can show it controls rights, including as against platform vendors | Contracts with camera and cloud vendors |
| A6 | Impact assessment | For monitoring of workers likely to be high risk, an assessment was done | DPIA or equivalent |
Regulators have written about these areas. The European Data Protection Board's guidelines on processing personal data through video devices (Version 2.0 adopted 29 January 2020) address video processing generally (EDPB, 2020), and the UK Information Commissioner's Office guidance on monitoring workers (published October 2023) recommends a data protection impact assessment for monitoring likely to result in a high risk to workers' rights and freedoms (ICO, 2023). Those are European and UK sources; US suppliers are subject to different federal and state rules, and requirements vary by location. This is not legal advice, and neither source is a US legal standard. Use them as prompts for diligence questions, and take advice for your jurisdiction. The page on worker consent and identifiable people goes further.
Blurring faces or muting speech does not resolve a permission problem. It reduces some exposure and may reduce training value.
B. Customer premises and content
| # | Check | Pass looks like | Evidence |
|---|---|---|---|
| B1 | Interiors | Policy on private homes, offices, medical or secure sites | Site exclusion list |
| B2 | Screens and documents | Rules for monitors, mail, paperwork in frame | Sample frames, redaction notes |
| B3 | Audio | Customer speech removed or excluded, or consented | Audio handling statement |
| B4 | Identifiers | Addresses, plates, nameplates, serials handled | Redaction logs |
| B5 | Hazardous or restricted spaces | Excluded or recorded under agreement | See recording in hazardous spaces |
C. Technical and task quality
| # | Check | Pass looks like | Evidence |
|---|---|---|---|
| C1 | Framing consistency | Known camera model, mount and field of view | Device list, sample from each |
| C2 | Hands and objects visible | Share of frames with working area in view meets your threshold | Reviewer sample count |
| C3 | Step completeness | Whole task or a stated subset is recorded | Segment map, start/stop logic |
| C4 | Timing | Constant or logged frame rate and clock | Metadata export |
| C5 | Narration | Present or absent, labeled | Audio track notes |
| C6 | Task diversity | Task and asset types listed | Visit-type counts |
| C7 | Outcome labels | Fixed, not fixed, escalated, linked to work order | Work-order join |
Red flags
- The supplier describes the footage by hours but cannot say how many visits, technicians or customers it covers.
- Cameras were mandatory and the notice mentions only safety.
- Footage lives on a vendor's cloud and the supplier cannot show the vendor's terms.
- Customer faces, voices or screens are visible and the redaction plan is unspecified.
- Recordings begin only after a fault is found or stop before the fix.
- Only successful visits were retained.
- The supplier offers all-rights buyout of material it did not create.
Illustrative acceptance record
# Illustrative only - synthetic values
offer_id: bodycam_offer_007
visits: reported_by_supplier
technicians: pseudonymous IDs
checks:
A1_purpose_documented: unknown # treat as fail
A3_customer_notice: pass
B3_audio: fail # customer speech present, no consent record
C1_framing: pass (two camera models, chest-mounted)
C3_step_completeness: partial # 40% of sampled visits missing the install step
decision: hold pending A1, B3
next_step: request policy, customer terms, redaction sample
Limitations and preparation
Even after rights are cleared, expect preparation: segment visits into tasks, remove idle time, label steps, link to work orders, and mask private content. Masking can occlude exactly the hands and tools you need. A realistic outcome is that a minority of the archive meets the bar, and that some tasks may be better served by commissioned recording with consent designed in.
Questions to ask suppliers
- Why were cameras introduced, and what did workers and customers sign or hear?
- How many visits, technicians and distinct sites does the archive cover?
- Which camera models and mounts were used over time?
- Does footage include audio, and is customer speech retained?
- What proportion of visits is complete from arrival to completion?
- Are visits linked to work orders and outcomes?
- What redaction has been applied, and how was it checked?
- Which vendors host the video, and what do their terms say about reuse?
Discuss your requirements
If you are weighing an offer of technician footage or want to describe what you need, tell SourceX the tasks, viewpoints and permitted uses. SourceX looks for US businesses that hold suitable material, checks the data and the supplier's licensing permissions, and follows up about potential sources. A request does not guarantee a match. The catalog entry for first-person video of manual work and the page on corporate training video archives describe neighboring routes.
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 video processing, Version 2.0 adopted 29 Jan 2020.
- Employment practices and data protection: monitoring workers — UK guidance on worker monitoring and impact assessments.
- Ego4D: Around the World in 3,000 Hours of Egocentric Video — consent-based egocentric video collection and de-identification.