Teleoperation as a data interface
Teleoperation can provide demonstrations in which a human directly produces the behavior a robot should learn. These trajectories can contain observations, actions and task outcomes.
Why it matters
Physical tasks often have complex contact dynamics and long-tail edge cases that are difficult to specify with hand-written programs. Human demonstrations provide a practical route to collecting grounded experience.
Trade-off
Teleoperation can be expensive and may encode operator-specific behavior. Dataset diversity, coverage and quality therefore remain important.