Egocentric and first-person video data for AI training
Egocentric video is footage recorded from the point of view of the person doing the work, through a camera on the head, the chest or a pair of glasses. AI labs and robotics teams look for it because it shows how hands-on tasks are really done. It is also footage of people, so what can be licensed depends on who agreed, who else appears, and where it was filmed.
Why it is different from other training data
A support ticket or a chat thread records what was said. Egocentric video records what was done: the order of the steps, where the hands go, which tool is picked up and how a slip is corrected. That is the part a written procedure leaves out, and it is what robots, embodied AI and computer vision models need to learn manual and screen-based work. Vision-language-action (VLA) models, imitation learning and hand-object interaction research all start from footage like this, and it is the core of what people call physical AI. Screen recordings of software work belong to the same family.
What a usable set contains
| What | Why it matters |
|---|---|
| Hours of footage | More hours of real tasks give a model more to learn from. A short sample comes first. |
| Different people | Many wearers show many ways of doing the same job. One person filmed for 100 hours is a narrower set. |
| Task and place variety | Different tasks, tools and settings matter more than the same task repeated. |
| Camera and quality | Camera type and position, resolution and frame rate, and whether the picture is steady and the hands stay in view. |
| Audio and narration | Sound is optional. A wearer who says what they are doing, or a step list kept with the clip, makes the footage easier to use. |
| Labels and metadata | Task name, tool, date, outcome and any reviewer note. Without them a clip is a picture with no lesson attached. |
| Consent and ownership | A written agreement from each wearer, a plan for anyone else who appears, and a clear owner of the files. |
Nothing here is a fixed minimum. What a buyer needs depends on the model it is training, so we ask each side for the specifics.
What can be considered
- Footage the wearer agreed in writing to share for this use
- Footage recorded on purpose for training, with a signed agreement
- Screen recordings of work, with customer data masked
- Files with a clear owner, and a record of where and when they were made
What stays out
- Footage recorded in secret or without an agreement
- Video taken from the web or from other people's accounts
- Clips where bystanders, patients or customers cannot be removed
- Restricted sites and home interiors that cannot be cleaned
Go deeper
- Egocentric video for robotics and physical AI. What models learn from it and the tasks labs ask for.
- Egocentric video annotation. Narration, hand trajectories, keypoints and object grounding.
- How it is captured. Head mounts, glasses, headsets and phones, and what makes footage usable.
- Data collection. What to check before licensing a collected set, and regions and languages.
Consent and privacy come first
Faces, voices, addresses and screens can all identify a person. Before a sample leaves, personal information is removed with automated tools and checked by people, and the seller approves the sample. The rules differ by place, including recording-consent and biometric laws, so counsel reads the ones that apply where each clip was made. We do not give legal advice. See how de-identification works and the privacy rules by region.
More on first-person video
- Company-held video. Where it sits and how it is cleaned, for a company with footage of its own staff.
Questions about first-person video
What is egocentric video?
Video recorded from the point of view of the person doing a task, usually with a camera on the head, the chest or a pair of glasses. It shows what the wearer sees and what their hands do. First-person video is the same thing in plainer words.
Why do AI labs want it?
Robots and AI agents need to learn how hands-on work is done: where the hands go, which tool comes next, how a mistake is fixed. Written instructions leave most of that out. Video from the worker's point of view keeps it.
Is it the same as company records?
No. Tickets, chat and CRM history are text written down at the time. First-person video is footage of a person, so consent, faces, voices and places decide whether it can be licensed at all.
Can any video be licensed?
No. Footage taken in secret, scraped from the web, or recorded with no agreement from the people in it is left out. A clip needs a clear owner and a written yes from the wearer, and anyone else who can be seen or heard has to be handled.
What does a lab ask to see first?
A short summary: how many hours, how many different people filmed, which tasks and places, what camera and what resolution, whether there is audio or narration, and how consent was collected. Then a cleaned sample that the seller has approved.
Do you hold the footage?
No. We work from a summary and an approved sample. Footage goes to a buyer only after the seller signs, and only through us.
Didn't find your question? Ask us directly.