I
First frame
The wearer
A head-mounted camera. It shows the hands and where the worker is looking. It can't show their body, the rest of the room, or anything behind them.
Thirdframe films real human work from four to eight calibrated, genlocked third-person cameras, all synced to the worker's own head cam. You get the body, the scene and the task, together in every frame.
A head-mounted camera. It shows the hands and where the worker is looking. It can't show their body, the rest of the room, or anything behind them.
Teleop and the robot's own sensors. The actions are exact, but it's slow and expensive to collect, and it only fits one robot body.
Fixed cameras all around the work. You get full-body pose, where every object sits, views around occlusions, and geometry that every camera can check. This is what we film.
Most egocentric datasets stop at the first frame. We film the third, locked to the first, so a model can learn the task and the world it happens in from the same moment.
Ego + exo · same clock