Teleoperation puts a human in the robot’s control loop: the operator moves a leader device (a twin arm, a VR controller, a hand-held gripper) and the robot follows, while sensors record the observations and actions as training data. Low-cost rigs like ALOHA made this cheap enough that collecting hundreds of demonstrations per task became routine, which is what powered the imitation-learning renaissance in manipulation.
Its economics are also its limit: every hour of data costs at least an hour of human labor, so teleoperation scales linearly where web data scaled for free. Closing that gap, whether with deployed robot fleets, simulation, or egocentric human video, is one of the central open questions in robot learning.
