Egocentric data

Most robot training footage used to come from cameras bolted to a wall or to the robot itself. Egocentric data flips that around. The camera rides on your head, so it sees what your eyes see and, crucially, what your hands are doing.

An example: you wear the rig while making tea. The recording shows the kettle coming into view, your hand closing round the handle, the angle of the pour. A robot arm can be trained to copy that sequence without a robot ever being in the room.

The catch is that heads move fast, hands block the view, and the camera misses what the fingers feel. That is why companies add wrist sensors or touch-sensing grips, and why a lot of raw footage gets thrown away.