Teleoperation

Teleoperation is full remote control of a robot by a human operator, typically with low-latency video feedback and continuous joystick or motion-capture input. Distinct from remote assistance (high-level decision-making layered over autonomous behavior) and from autonomous control (no human in the loop).

The distinction matters because teleoperation is the standard data-collection method for imitation-learning datasets in humanoid manipulation. Figure's Helix training corpus, Physical Intelligence's π series, and Apptronik's Apollo training all rely heavily on teleoperated demonstrations. Every hour of teleoperation is direct human labor; if the resulting policy generalizes poorly, the cost per task may not improve over the labor it replaces. The strongest justification for VLA models is that they amortize teleoperation cost across many downstream tasks.