On-Site Consent and Compliance for Egocentric Capture

Egocentric stereo capture puts cameras on a person’s head. That is a different compliance problem from tripod video on a factory line. Faces of bystanders, badges, screens, whiteboards, customer premises, and worker identity can all enter the frame without anyone intending to “film people.” Buyers who only specify resolution, fps, and schema discover too late that legal and site-access constraints—not stereo residuals—block delivery.

This essay is a buyer-facing guide to on-site consent and compliance for egocentric human-demonstration programs. It is not legal advice and does not invent jurisdiction-specific statutes. It translates common procurement failure modes into SOW language you can attach next to technical acceptance gates. Technical numbers below come only from the public sample envelope: 2 episodes, 4,914 frames, 163.47 s (EP001 PC wiring 121.57 s; EP002 restocking 41.90 s), 1920×1200/eye, 30 fps, hardware-synced stereo, timestamps recorded (not computed), LeRobot v3, sample license CC BY-NC 4.0. Scene planning may reference a library of 89 clips / 8 scenes; only EP001 and EP002 are our-device captures in that narrative.

Why egocentric demos need a compliance block in the SOW

Third-person robot cells can sometimes be staged in empty bays. Egocentric demos cannot: the demonstrator moves through real work—wiring boards, restocking shelves, crossing aisles—while dual-eye colour video and IMU/pose streams run continuously. Compliance risk concentrates in five places:

1. Demonstrator consent (the wearer)
2. Bystander and co-worker exposure (people who did not wear the rig)
3. Site and customer confidentiality (premises, clients, screens, documents)
4. Data license and redistribution (what buyers and partners may do with the files)
5. Provenance honesty (what was captured on the contracted device stack versus scoping library material)

Technical excellence does not erase those risks. A perfect timebase—median/p99/max ≈ 33.28 ms, 0 out of spec, 100% device-side timestamps, IMU ~300.48 Hz with 0 gaps >10 ms—still ships whatever the lenses saw. Write consent and site rules before capture days, in the same SOW appendix as stereo and schema gates.

Layer 1 — Demonstrator (wearer) consent

Treat the wearer as a research/work participant with documented consent, not as anonymous “operator footage.”

Buyer-ready requirements

• Written consent covering: purpose (robot learning / imitation / evaluation), sensors (dual-eye RGB, IMU, head pose), retention, and who may access the data.
• Clear statement whether the wearer’s face, voice (if any audio exists in a given program), hands, and work performance may appear in training, eval, marketing stills, or only in restricted research use.
• Withdrawal and incident process: who to contact on site if the wearer stops a session.
• Role separation: consent forms are not the same document as the purchase order.

Tie consent IDs to episode manifests so EP-style deliveries remain auditable. The public sample’s episode labels (EP001 wiring, EP002 restocking) illustrate why task context belongs next to identity controls: different sites and tasks often need different consent annexes.

Layer 2 — Bystanders, co-workers, and incidental PII

Egocentric FOV will catch other people. Pretending otherwise is how programs get shut down mid-week.

Practical controls to put in the SOW

• Site rules for occupied areas: capture windows, restricted zones, escort requirements.
• Signage and advance notice where the site requires it.
• Exclusion lists: no capture in restrooms, medical areas, or other banned zones defined by the host.
• Handling of incidental faces and badges: blur policy, reshoot policy, or controlled-access retention—pick one and fund it.
• Stop-the-line authority: any co-worker or supervisor can pause capture without career penalty.

Do not rely on “we will fix it in post” as the only control. Dual-eye 1920×1200 stereo at 30 fps produces a lot of frames quickly (4,914 frames in 163.47 s on the public sample scale). Blur pipelines must be scheduled as part of accepted-hour cost, not as unpaid cleanup.

Layer 3 — Site, customer, and trade-secret hygiene

Wiring and restocking demos often happen where schematics, inventory systems, or customer brands are visible.

SOW controls

• Pre-capture walkthrough with the site owner: which benches, screens, and documents are in-frame OK.
• Screen policy: login desktops, customer PII, and unreleased product UI either powered down, mocked, or framed out.
• Badge and label policy: when employee IDs or shipment labels must be covered.
• Export review: site owner right to review a sample clip before bulk acceptance—time-boxed so it cannot become infinite delay.
• Physical security: who holds the capture kit overnight; chain of custody for raw vs accepted deliveries.

Compliance here is operational. It pairs with technical packaging: accepted deliveries should be LeRobot v3 loads via LeRobotDataset(), aligned with pi-data-sharing checks, so reviewers inspect the same artifact trainers will use—not a one-off MP4 dump that bypasses the schema.

Layer 4 — License, sharing, and commercial boundaries

The public sample is released under CC BY-NC 4.0. That is an inspectable reference for non-commercial sample use—not an automatic license for every production hour you buy.

Buyer checklist

• Sample vs production: state explicitly that CC BY-NC 4.0 applies to the public Hugging Face sample, while production lots use the commercial terms in the SOW.
• Downstream sharing: may the buyer’s partners train on the data? May eval forks leave the buyer’s VPC?
• Marketing stills: default deny unless a separate release exists.
• Retention and deletion: how long raw capture vs accepted LeRobot deliveries are kept; who certifies deletion after a site requests it.
• Capacity context: 10,000+ accepted hours/month only helps if each hour is licensable under the contract you actually signed.

License ambiguity is a compliance failure mode equal to missing consent forms.

Layer 5 — Provenance and over-claim boundaries

Compliance includes truthful description of what was captured.

• The scene library (89 clips / 8 scenes) is for scoping coverage, not proof that every clip shares the our-device envelope.
• Only EP001 and EP002 are our-device public reference captures in that narrative.
• Do not attribute third-party or mixed library footage to the contracted stereo stack (hardware-synced dual-eye, device-side timestamps, shipped calibration K / KB4 / R1/R2 / P1/P2 / Q / IMU↔cam / version, residuals at the public reference scale).
• Do not invent customers, brands, or site names in datasets or marketing.

Mislabeling provenance creates legal and scientific debt: models and auditors both assume the metadata is true.

How compliance gates sit next to technical acceptance

A complete egocentric program accepts an hour only when both columns pass.

If either column fails, the hour is not an accepted hour—even if the stereo looks beautiful.

Paste-ready compliance appendix (non-legal template)

EGOCENTRIC CAPTURE — CONSENT & COMPLIANCE APPENDIX
Program / PO: ____________  Site: ____________  Dates: ____________

A. Wearer
[ ] Written consent executed before first powered capture
[ ] Purpose, sensors (stereo RGB, IMU, head pose), retention, accessors described
[ ] Withdrawal / stop process posted on site
[ ] Consent ID linked on episode manifest

B. Site & bystanders
[ ] Site authorization letter / ticket on file
[ ] Restricted zones listed; capture schedule agreed
[ ] Signage / notice completed if required by site
[ ] Incidental face/badge policy chosen: blur / reshoot / controlled access
[ ] Stop-the-line authority documented

C. Confidentiality
[ ] Screen and document rules agreed in walkthrough
[ ] Export review window (if any) time-boxed: ____ business days
[ ] Chain of custody for raw vs accepted deliveries

D. License & sharing
[ ] Public sample license (CC BY-NC 4.0) not confused with production terms
[ ] Production license / partner-sharing rules attached
[ ] Marketing-still policy: default deny unless release attached
[ ] Retention & deletion SLAs written

E. Provenance
[ ] Library scoping (89 clips / 8 scenes) not labeled as our-device
[ ] Only EP001/EP002 cited as our-device public reference captures
[ ] No fabricated customer or site claims in dataset cards

F. Technical coupling (accepted hour requires pass)
[ ] Hardware-synced stereo envelope + recorded timestamps + IMU continuity
[ ] Calibration package shipped (K, KB4, R1/R2, P1/P2, Q, IMU↔cam, version)
[ ] LeRobot v3 load via LeRobotDataset(); pi-data-sharing checks
[ ] Sign-off: Capture lead ____  Site rep ____  Buyer ____  Date ____

Use counsel in your jurisdictions to harden this template. The point for procurement is structural: compliance is a first-class acceptance column.

Operational tips that keep capture days moving

1. Consent before kit power-on. Never “start rolling and sign later.”
2. One compliance owner on site with authority to pause—parallel to the technical capture lead.
3. Manifest discipline. Episode IDs, consent IDs, site IDs, calibration version, and license flag in one row.
4. Review on the delivery schema. Prefer spot-checks on LeRobot v3 artifacts so legal/site review sees what trainers see.
5. Separate scoping from supply. Plan task mix with 89 clips / 8 scenes; count supply in accepted hours (capacity reference 10,000+/month) that cleared both technical and compliance gates.
6. Keep the core technical claim honest. Device-side timestamps and hardware sync are evidence of capture integrity; they are not a substitute for consent.

Soft next step

If you are drafting an egocentric stereo demo SOW and need compliance language that sits beside timing, calibration, and schema gates, start from the inspectable public sample—EP001 / EP002, 4,914 frames, 163.47 s, hardware-synced dual-eye capture, recorded timestamps, continuous IMU, versioned calibration, LeRobot v3—then attach a consent/site appendix before the first capture day.

• Sample: https://huggingface.co/datasets/tatezhou/ego-stereo-cn-v1
• Tools: https://github.com/TateZhouSiu/ego-stereo-cn-v1-tools
• Product: https://10khoursdata.com
• Contact: tate@10khoursdata.com

Ask for a walkthrough of how accepted hours couple technical QA with consent and site controls. That is how egocentric programs scale without discovering compliance blockers after the data is already in the training cluster.

10K HoursData — egocentric stereo human-demonstration data with timestamps recorded, not computed; technical acceptance and on-site compliance as paired gates.