Egocentric video of people doing real tasks, cut into single actions and described hand by hand. We collect it, annotate it and review it four times, for teams training VLA models, world models and robot policies.
Use one part or all of it. Each stage works on footage we capture or footage you already have.
"Washes the bowl" is not enough supervision. A vision-language-action model has to resolve which bowl, with which tool, and what the other hand is doing. So we describe each hand separately and ground every description in the objects in view.
Camera motion lets you weight or filter clips where apparent object motion is really head motion. Scene comes from a controlled vocabulary, so it works as a conditioning signal rather than a free-text string.
Clip boundaries decide what a robot learns an action is. These rules are applied to every clip, and three violations remove an annotator from the queue.
The full pipeline, special-case handling and record schema are on the home page.
The environment decides the action vocabulary, how cluttered the scene is and where annotation gets hard.
Each approved clip is one structured record: clip index, start and end timecode (capped at 10 seconds), which hand acts (left, right or both), a separate natural-language description for each hand, camera motion (small or large), scene from a controlled vocabulary, and review state with the full correction history.
Yes. We scope the environments, tasks and hours with you, along with resolution, frame rate and field of view, then capture with head-mounted cameras at real workplaces and homes. Collection is live in Pakistan today.
Yes. That is how our live VLA program works: we annotate the client's own footage, 30,000+ hours so far, through the same segmentation and description pipeline, rule set and four-pass review.
Every record passes four human reviews: an internal and an external review of the clip boundaries, then an internal and an external review of the descriptions. The external team is independent and blind to who did the work. Annotators who break the rule set three times are removed from the queue.
With a pilot. Send about two hours of video and your taxonomy, or use ours. You get labelled clips back, a QA report with an error taxonomy, and the questions we would ask before scaling.
Get labelled clips back in 72 hours, with a QA report and the questions we would ask before scaling. Your taxonomy or ours.