All articles
Guide··6 min read

Egocentric Video for Robot Training: How We Source First-Person Footage (2026)

For robotics and physical AI teams that need egocentric, head-mounted POV video of everyday tasks: what buyers specify, what our contributor pool can film today, and what it can't.

Some links in this article are referral links — we may earn a small commission if you sign up, at no cost to you. This keeps the site free.

Robots learn hand skills from watching people do ordinary work from the person's own point of view: washing dishes, folding laundry, wiping a counter. That footage is called egocentric video, and robotics and physical AI teams are short of it. This post is for those teams. It explains what buyers ask for, what our contributor pool can film today with a head-mounted phone, and, just as important, what it cannot do yet.

Why egocentric video, and what the public datasets look like

A camera on the head sees what the hands see, so the footage shows reach, grasp and sequencing from the viewpoint a humanoid or manipulator would need. Two public datasets show the range. Ego4D holds 3,670 hours of daily-life video from 931 camera wearers across 74 locations in 9 countries, mostly unscripted. Apple's EgoDex holds 829 hours of tabletop manipulation across 194 tasks, recorded on Vision Pro with hand and finger poses captured at recording time, not labelled afterwards. They sit at opposite ends: Ego4D is broad and mostly unscripted, EgoDex is narrow and instrumented. Our contributors film the first kind, with a phone, and we would not claim to match the second.

What buyers are asking contributors to film

We list jobs from AI training platforms, so we can see how the market describes this work. On November 22, 2026 our feed held these egocentric collection listings:

Three things stand out. Capture is a consumer phone on a head strap. Payment is per accepted hour, so quality control decides what the contributor earns. And buyers are already splitting supply by country, which is the gap a worldwide pool can fill.

What we can provide today

We run a physical AI list on aitrainingjobsfinder.com. Members opt in with a city and the devices they can record with, and they get a one-page filming guide and a personal clip code. As of November 22, 2026, 98 people have joined since the list opened on October 24.

WhatWhere it stands today
Pool size98 members who opted in
Geography35 countries on member profiles. Brazil is largest at 32, then Ethiopia (7), Egypt (6), Kenya (6), Nigeria (4) and Portugal (4). Country is blank on 8 profiles.
Devices (members can pick several)Recent Android 48, iPhone 12 or newer 30, older Android 19, iPhone 11 or older 11, other 16, smart glasses 3, GoPro 2
Capture standardPhone on the head or forehead pointing down at the hands at about 30 to 45 degrees, landscape, ultra-wide lens, 1080p at 30 fps or higher, stabilization off, good light
Framing rulesBoth hands in frame almost all the time, slow head movement, mistakes kept in, one task per clip
ScreeningEach member films a 60-second kitchen task as a test clip, and we watch every one
IdentityEach clip starts with the member's code on paper held to the camera for 3 seconds, so we know the footage is theirs
LanguagesGuide and emails in English, and in Portuguese for Brazil, Portugal, Angola and Mozambique

Privacy rules built into the brief

Contributors film alone, never other people and never children, and only in their own home. They are told never to film documents, cards, letters or screens with personal information, or their own face, including in mirrors and shiny surfaces. These rules are in the filming guide every member receives.

What we cannot do yet

How a pilot could work

This is the shape we would propose, not a finished service. You send the brief: the tasks, the environment, the hours you want, the video spec and any annotation you need. We recruit from the pool by country, city and device, invite only members whose test clip passed, and agree the pay per accepted hour with you so contributors see it before filming. Footage is checked against your acceptance rules, and you pay for accepted hours only. Before any filming starts we would agree in writing the consent wording, licence terms and how long footage is kept. Start with a small batch, measure your acceptance rate, then scale.

Common questions

Do you have existing footage to license?

No. Test clips are for screening contributors. We recruit for new capture against your brief.

Can contributors use a GoPro or smart glasses?

A few members have them (2 and 3 of the 98), but most use phones, so plan your spec around a phone and treat other devices as a bonus.

How much does footage cost?

We have not set a price. The listings above show what platforms state to contributors, up to $15 per accepted hour on Handshake AI's and $20 to $30 an hour on Rex's paired recordings. Your budget also needs to cover quality control and the cost of rejected footage.

How do I get in touch?

Email info@aitrainingjobsfinder.com or use our contact page, and tell us the tasks, countries, volume and spec you need. If you want to film rather than buy, join the list from the banner or your account settings after you sign up. For other kinds of AI training work, see what AI training jobs are.

Browse live listings →