Embodied AI requires novel data generation and training methods—real-world teleoperation, motion capture, and simulation are critical infrastructure bottlenecks for scaling robot capabilities

“There is no source of data for how you should coordinate all of the joints to work together. There's no source of that data, so we have to sort of generate that. We have to find a way to create that... skills are something where maybe you would train it using motion capture data or animation or tele-operation, a person remote controlling a robot or something, some way of showing the robot how to do a thing like grasp a door handle or pick up a mug or some skill that you want to show it how to do.”