Paired egocentric and allocentric video of real work, across thousands of tasks and settings, to make robot learning generalise.
Scroll ↓A robot acts from its own point of view: egocentric. Most of the world's footage of people working is allocentric: a camera in the corner, a security feed, a phone on a tripod. Translating between the two is what lets a robot learn a task by watching it.
Video foundation models are now strong enough that a small amount of well-paired data goes a long way. What is missing is that data. The largest public ego–exo sets total a few hundred hours, and the most-used benchmark covers about ten tasks.
We think the holy grail is a robot that learns a job from the cameras already watching it. Getting there runs through one dataset nobody has built yet.
Coverage is what makes a model generalise. So we capture simple, repeatable tasks in as many settings as we can reach, with the external camera moving and placed at random — near, far, side, above — rather than bolted to a wall.
Leave your email. We'll write when the first release is out.