What Is First-Person Video in AI Training?

A policy trained on video from a viewpoint it will never occupy at inference time learns an observation distribution that does not match deployment. First-person video closes that gap by recording from the actor’s own point of view, so the resulting footage matches what a robot with a head or chest camera actually sees — hands entering the frame from below, the target object approaching from the wearer’s own angle, occlusion as a hand closes around an object. Third-person or fixed-camera footage does not have this property: it shows the task from an external angle a robot’s own camera will never occupy.

Firsthand captures first-person video from a head-mounted global-shutter camera plus two wrist cameras, hardware-triggered onto one PTP clock domain so the first-person stream is temporally aligned with depth, 3D hand pose, and action-segment labels rather than recorded in isolation. Alongside the instrumented head-mounted rig, Firsthand also offers an Insta360-style head-mount capture tier for lighter-weight collection. Consent is tracked through a signed consent chain, faces and identifying details outside the task are anonymized, and custom first-person collections run against a written skill spec, are QA-checked against it, and are delivered under a buyer-owned commercial license in RLDS, LeRobot format, or WebDataset. Firsthand does not capture third-person or fixed external-camera footage.

Yes. "First-person," "POV," and "egocentric" describe the same capture viewpoint — footage recorded from a camera worn by the person doing the task, so the camera moves with their head and shares their field of view. "Egocentric" is the term used in robotics and computer-vision research; "first-person" and "POV" are the everyday phrases people search for and use in casual conversation. On this site, the terms are used interchangeably, and both route to the same underlying capture method: a head-mounted (and, for manipulation, wrist-mounted) camera rig rather than a fixed or external one.

A robot with a head or chest camera observes the world from roughly the same vantage point a human head occupies. First-person video recorded from a human’s own head position produces a similar observation distribution — hands and tools entering frame from the wearer’s own reach envelope, occlusion patterns during grasp, and gaze-driven framing of the task — which is closer to what the robot will see at inference than footage filmed from across the room or from a fixed overhead rig. That is the core argument for using first-person video as pretraining or fine-tuning data rather than relying solely on third-person demonstrations.

As Firsthand captures it, first-person video comes from a head-mounted global-shutter camera, paired with two wrist cameras for close-up hand-object detail that the head view alone would miss or occlude. All streams are hardware-triggered onto one PTP clock domain, so the first-person video is temporally aligned with synchronized depth, 3D hand pose, and action-segment labels rather than delivered as video alone. Consent is tracked through a signed consent chain, and faces or identifying details outside the task are anonymized before delivery.

Third-person video is recorded from an external observer’s viewpoint or a fixed camera in the room, not from the performer’s own position. That viewpoint does not match what a robot’s own head or chest camera will see at inference, and it is a different data source from what this site provides — Firsthand captures first-person (egocentric) video and does not capture third-person or fixed external-camera footage. See the egocentric vs third-person comparison for how the two viewpoints differ on hand visibility, occlusion, body coverage, and labels.

Custom first-person video collection starts with a written skill spec describing the tasks, settings, and contributor requirements, is typically validated with a smaller Spec Pilot before scaling, and is QA-checked against the spec before delivery. Cost depends on scope — see the pricing page for cost drivers rather than fixed figures — and delivery uses the same formats, licensing, and consent handling described above.

What does first-person video mean in AI training?

It means footage recorded from the viewpoint of the person performing a task — usually from a head-mounted camera — rather than from an external observer’s or a fixed camera’s position. It is the everyday name for what robotics and computer-vision research calls egocentric video.

Is first-person video the same as egocentric video?

Yes. Both terms describe footage captured from a camera worn by the person doing the task, so the camera shares their viewpoint. "Egocentric" is the technical term used in research; "first-person" and "POV" are the common search terms for the same thing.

Is POV video the same as first-person video?

Yes, "POV" (point-of-view) and "first-person" are used interchangeably to describe the same capture viewpoint — a camera worn by the performer, typically on the head.

Why is first-person video better than third-person video for robot learning?

It isn’t universally better, but for training a policy that will run on a robot with a head or chest camera, first-person video produces an observation distribution closer to what the robot will actually see at inference — hands entering frame from the wearer’s own angle, matching occlusion during grasp — which third-person footage does not provide.

What labels come with first-person video?

As captured on this site, first-person video is delivered alongside synchronized depth, 3D hand pose, and action-segment labels — not as video alone — because all streams are hardware-triggered onto one clock domain.

Does Firsthand capture third-person video?

No. Firsthand captures first-person (egocentric) video from head-mounted and wrist-mounted cameras. It does not capture third-person or fixed external-camera footage.

How is first-person video delivered?

Custom first-person video collections are delivered under a buyer-owned commercial license in RLDS, LeRobot format, or WebDataset, after QA against the written skill spec used to scope the collection.

How do you order a custom first-person video dataset?

Start with a written skill spec describing the tasks, settings, and contributor requirements, typically validated with a smaller Spec Pilot before scaling to the full collection. See the pricing page for cost drivers.