Egocentric (First-Person) Video for AI and Robotics Training

Egocentric video is footage filmed from a person's own viewpoint, usually by a head-mounted camera or smart glasses, and AI developers acquire it to teach robots and models how human hands handle real tasks. The Ego4D consortium states 3,670 hours of publicly available video, Claru states USD 10 to 120 per capture hour, and Mindrift states USD 6 to 15 per hour to contributors (both checked November 9, 2026).

This page sets out, for first-person video in AI training data licensing, who collects it, the two routes in, the dated terms and stated pay, the public datasets it competes with and what must be cleared. Not licensing is a valid outcome.

The DataPacts Editorial Team explains how we source, date and check every figure.

What Egocentric Video Is and Why Robotics Developers Want It

Egocentric video shows what the wearer sees and does, hands included, which is close to what a robot's own camera sees while it works.

The term covers point-of-view footage from head-mounted cameras, smart glasses and phone head mounts. The person filming is the wearer; anyone else in frame is a bystander. The Ego4D consortium published Ego4D on February 17, 2022, with 3,670 hours from 923 participants in 74 locations and 9 countries. Build AI states that its Egocentric-10K set comes from head-mounted cameras worn by factory workers.

Developers pair first-person video with other signals. Midcentury states that it sells egocentric vision together with UMI gripper data, and Figure collects egocentric human video of everyday activities in Brookfield residential environments itself (announced September 18, 2025). Robot teleoperation data, where an operator drives a robot, is a different data type.

The parent page lists every one of the data types AI companies buy, and first-person video is one of them.

Who Collects and Licenses First-Person Video

First-person video reaches AI developers through five kinds of companies: vendors that record it to order, licensees that also buy existing archives, contributor apps and marketplaces, intermediaries that license on the owner's behalf, and labs that film it themselves.

The table lists each company's role for first-person video as of November 10, 2026 (programs checked November 9, 2026).

Figure and Build AI collect footage for themselves, and neither states a purchase of datasets. Newsflare and Sureform license or manage licensing on the owner's behalf; they do not buy footage. Midcentury and Mecka publish no seller or contributor terms.

Two Routes In: Licensing Existing First-Person Footage or Recording to Order

A company either licenses first-person footage it already holds, such as smart-glasses, body-cam or training recordings, or lets a collector record its staff at work; individuals take part only through contributor programs.

Claru states that it takes archive footage, including smart-glasses recordings, as well as body-cam footage, ships the cameras itself and starts capture about 2 weeks after a signed agreement. Asimov states that it covers the salaries of staff while they wear its recording kit. Sieve states that it pays businesses to record physical work, and Sureform lists egocentric worker video under a separate contributor agreement.

Toloka states that its Physical AI Data Partner Program works with companies and capture teams, with an intro call, a sample and a partner agreement before recording. Camera choice, capture protocol and quality filtering are builder topics and stay outside this page.

Individuals record through contributor programs: Mindrift states there is no assessment and that setup time is unpaid. Contributor work is compared on the page of AI training jobs and platforms compared.

First-Person Video Terms by Program: Pay Unit, License Grant and Recall

First-person programs differ most on three terms: what unit they pay for, how wide a license the recorder grants, and whether footage can be pulled back once delivered.

The table compares nine programs as of November 10, 2026; every value is the program's own statement, checked November 9, 2026.

Claru states non-exclusive training rights, while Kled states a transferable, fully sub-licensable, worldwide, perpetual and irrevocable license that is exclusive for training third-party AI models. Asimov states a worldwide, transferable, sublicensable, royalty-free, paid-up grant.

None of the nine programs publishes a deletion right that reaches models already trained. Terms for other data types are collected in the program terms compared, dated across the site.

How First-Person Footage Is Paid: Archive Fees, Capture Hours, Clips and Salary Cover

No public rate card exists for egocentric video; the figures on record are a few programs' own ranges, each tied to one unit.

Five pay units appear, each as stated by the program (checked November 9, 2026):

• Archive license: Claru states a typical first deal of USD 5,000 to 50,000, paid net-15 on a signed license.
• Capture hour: Claru states USD 10 to 120 per hour, paid weekly (checked November 9, 2026).
• Contributor hour: Mindrift states USD 6 to 15 per hour and a market range of USD 3 to 9+ per hour of accepted footage (checked November 9, 2026).
• Per task: Luel lists USD 25 for a 5-minute indoor walkthrough, with payouts processed daily (checked November 9, 2026).
• Salary cover and revenue share: Asimov states it covers salaries while staff wear its kit and shares dataset revenue in larger partnerships, without a percentage.

The ranges are not averaged or converted, because the units differ. Human Archive does not publish a per-hour rate, Sieve states no amount, and Toloka sets compensation in a statement of work by data type, volume and quality.

Price points for every other data type are collected on the page of AI training data prices per work, hour, image.

Open Egocentric Datasets That Compete With Paid Footage

Large egocentric datasets are already publicly available, so paid footage has to offer what they lack: new settings, tasks, rights clearance or exclusivity.

The table compares four egocentric collections as of November 10, 2026, each size and license as its publisher states it.

Build AI states that Egocentric-10K is released under Apache 2.0. Its dataset card shows 16.4 TB in one table and 18 TB as total file size, so both figures appear above, and the number of workers is not stated.

Midcentury is a commercial catalog, listed for contrast. Footage from a seller competes with these sets on setting, task, clearance and exclusivity, not on hours alone.

Clearing First-Person Footage: Wearers, Bystanders and Places

First-person footage involves three groups: the staff who wear the camera, the people in frame and the premises filmed. This section is information, not legal advice.

Staff Who Wear the Camera

A head camera worn at work records the employee's own conduct, so EU employers face workplace-monitoring rules and, in Germany, works council co-determination before recording starts.

GDPR Art. 88(2) (Regulation (EU) 2016/679, applies from May 25, 2018) requires national or collective employee rules to include suitable measures on transparency and on monitoring systems at the workplace. BetrVG § 87(1) Nr. 6 gives the German works council co-determination over technical monitoring devices. Spain's LOPDGDD art. 89 (in force December 7, 2018) allows workplace video only with prior express information.

Notice and works council steps are set out under employee notice and consent when licensing.

Faces, Hands and Voices of People in Frame

Everyone the wearer looks at is recorded too, so collectors blur faces and identifying marks, remove voices and bar re-identification by licensees.

Claru states that it auto-scrubs faces, cards and plates, and Toloka states that it blurs faces, palms, fingerprints and tattoos and removes voices (checked November 9, 2026). Human Archive states that its written license agreements prohibit re-identification and that minors may not be the primary subject.

Methods are covered under de-identification before licensing.

Illinois BIPA, 740 ILCS 14/15(b), requires written notice and release before biometric identifiers are collected, and 15(c) bars selling or profiting from them. Va. Code 59.1-575 excludes photographs and audio or video recordings from biometric data (in force January 1, 2023), while Texas CUBI as amended by HB 149 (in force January 1, 2026) exempts biometric identifiers used to train AI unless the system uniquely identifies a specific individual. GDPR Art. 9(1) treats biometric data that uniquely identifies a person as special-category data (applies from May 25, 2018).

Homes, Customer Sites and Footage You Do Not Own

Footage filmed in a client's premises or a private home carries the rights and privacy of whoever controls that place, not only of the camera owner.

Claru states that owners of non-public footage can apply, and Figure films in Brookfield residential environments. Under CDPA 1988 s.11(2) in the UK, the employer is first owner of a film an employee makes in the course of employment, subject to agreement.

Each clip is checked in a rights audit of the footage before licensing.

Recording First-Person Video for AI: Benefits, Risks, Downsides and Alternatives, Including Not Licensing

Recording or licensing first-person footage can pay for work that happens anyway, but licenses are often wide, delivered copies usually stay with the licensee, and the people filmed carry most of the privacy exposure.

Benefits:

• Work that happens anyway can be paid: Asimov states it covers salaries while staff wear its kit.
• Ownership stays with the partner under Claru's and Sieve's stated terms.
• Claru's non-exclusive terms leave other licensing open.

Risks:

• Delivered copies are not recalled, as Claru, Asimov and Human Archive state.
• Kled's license is exclusive, perpetual and irrevocable for training third-party AI.
• Wearers and bystanders carry privacy exposure under workplace-monitoring rules and biometric statutes.

Downsides:

• Mindrift states that footage can be rejected and setup time is unpaid.
• Public datasets such as Ego4D and Egocentric-10K already cover generic settings.

Alternatives:

• License an existing archive only, as Claru accepts.
• Grant non-exclusive terms.
• Start with a sample under a statement of work, as Toloka's process allows.
• Not license at all, as set out in the alternatives to selling data to AI labs, including not selling at all.

The grant width and the rights position decide which option fits.

Neighboring Data Types and the Seller Steps for First-Person Footage

This page covers footage filmed from a person's viewpoint; existing third-person footage, robot teleoperation, dashcams and motion capture each have their own page. Footage of people also falls under the AI training data law by jurisdiction.

Neighboring pages:

• Selling video for AI training: existing footage and archives
• Robotics data for AI training: teleoperation and manipulation data
• Driving and dashcam footage: footage from vehicles
• Motion capture and 3D data
• Computer-use and agent trajectory data: screen recordings of computer work

Seller steps for first-person footage:

1. Read how to sell or license company data for the whole process.
2. Clear the footage for wearers, bystanders and places, as above.
3. Agree the license, clause by clause, using AI data licensing agreements.
4. Check payment timing under when AI data buyers pay.

Frequently Asked Questions

What is egocentric video?

Egocentric video is first-person footage from a head-mounted camera or glasses, showing the wearer's hands and actions. Ego4D holds 3,670 hours of it as a publicly available dataset.

Is egocentric video recording legit?

Real programs publish terms, and the terms can be read before recording. Mindrift states there is no assessment, footage is accepted or rejected and setup is unpaid, and Asimov states payment is not guaranteed (checked November 9, 2026). Checking a platform is covered under is AI training work legit.

How much does egocentric video collection pay?

Stated pay differs by unit and no rate card exists. Mindrift states USD 6 to 15 per hour and Luel lists USD 25 per walkthrough (checked November 9, 2026). Rates elsewhere are listed under pay for AI training work by platform.

Can first-person footage be withdrawn after delivery?

Usually not. Claru, Asimov and Human Archive state that delivered or licensed footage is not recalled, and Kled's license is irrevocable (checked November 9, 2026). Model consequences are covered under retraining and model retention rights.

Advisers and agents can use the introducer form.