Guide

What is egocentric video data? First-person video of real work, with the hands in frame

Egocentric video data is first-person video of people doing real tasks, filmed at eye level with both hands in frame; Dovra captures about 5,000 hours of it per 4-week cycle. Robotics and world-model teams use it because it shows a task the way a robot's own camera would see it: the hands, the tools, the objects and the order in which things happen.

Updated 5 November 2026 · 7 min read

What does "egocentric" actually mean?

Egocentric means filmed from the person's own point of view. The camera sits at eye level, usually on the head, and looks where the person looks. In practice that means the hands, the tool and the object being worked on stay in frame for the whole task. Exocentric video, by contrast, is filmed from outside and looks at the person. Both have uses, but only the egocentric view matches what a robot's head or wrist camera sees while it works.

Who uses egocentric video data?

Teams building models that have to act in the physical world. The most common uses are:

• Vision-language-action (VLA) models, which learn to map what they see and an instruction to the next action, and need demonstrations of people doing the task.
• Humanoid and mobile-manipulation robots, which have to handle tools and objects in places built for people.
• World models, which learn how scenes change when someone acts on them.
• Video generation, which needs realistic, unstaged footage of hands and objects in motion.

In all four cases the value is the same: a large number of real demonstrations, seen from the viewpoint the model will act from.

What makes egocentric video data useful for training?

Four things: hands, real tasks, real workplaces and diversity. Hands matter because manipulation is the hard part, so the footage has to show grips, hand-object contact and tool use clearly. Real tasks matter because purposeless or staged actions teach the wrong sequence. Real workplaces matter because cramped, cluttered and badly lit commercial spaces are where models fail after training on clean scenes. And diversity across countries, scene types and trades is what reduces distribution shift when the model is deployed somewhere new.

Real work or staged work: why does it matter?

Unscripted footage of people doing their actual job carries what a policy needs and staged footage leaves out. Real shifts include hesitations, slips and corrections, the true order of steps, and the clutter around the work surface. Scripted demonstrations look tidy but generalise poorly. That is why Dovra captures only the human layer, first-person capture of real work, and not simulation or teleoperation. Footage with purposeless actions, near-static shots, screen recordings or third-person views is rejected at review and never paid.

How is egocentric video captured?

At Dovra, with the contributor's own phone in a head mount. The phone sits on the forehead at eye level, tilted down and in landscape, with both hands fully in view and good light; chest-mounted footage is not accepted. Each clip covers one real, complete task and runs from 1 to 30 minutes. The exact device model and OS are logged per clip, so a buyer can filter or stratify by sensor. Most rejections come down to clarity: poor light or a badly placed phone.

What metadata should come with each clip?

Enough to filter, stratify and audit the dataset without watching it. At minimum:

What consent and provenance does egocentric video need?

More than a contributor checkbox, because the camera is inside someone's workplace. A defensible dataset rests on a signed agreement per contributor, written permission from each business where filming happened, a payment record per approved hour, a training licence and a provenance document. Exclusions belong in the agreement and are enforced at human review: no minors, documents, screens, internal pricing or identifiable customers. If you are comparing suppliers, use our 10-question vendor checklist.

How does Dovra supply egocentric video data?

• Standing capacity of about 5,000 hours per 4-week cycle, across 32 countries, deepest in Latin America.
• 15 work domains and 84 catalogued task types; see dataset coverage.
• 72 hours from brief to first clips in a running domain; a new domain in 2–3 weeks.
• Every session rated by a human before it counts; priced per approved hour.
• Annotation on request, priced per hour on top. Delivery by encrypted transfer to your bucket with a manifest and dataset card.

Dovra is backed by Simera. Read why Simera backs Dovra.