teleoperated demonstration data

Teleoperated demonstration data is the training material a robot learns from when a human remotely drives the robot's own body through a task, and every motion is recorded. A person wears a glove, holds a controller, or moves a small twin of the arm, and the real robot copies their movements in real time; meanwhile the camera images, joint angles, and gripper states are all saved. The result is a recording that says, in effect, in this exact situation, here is what a skilled human chose to do with this robot, step by step.

This kind of data is prized because it is grounded in a way other sources are not. The demonstration is performed on the very kind of robot that will later be trained, so the recorded motions already respect that robot's real reach, speed, and clumsiness — there is no gap between a human hand and a metal gripper to paper over. Watching a video of a person cooking does not tell a two-fingered robot how to move its own joints, but teleoperated demonstrations speak the robot's native language of joint commands directly, which is exactly what learning methods like behavior cloning need in order to imitate.

The price is human time and effort. Every demonstration needs a person patiently steering the robot, so collecting the thousands of examples that modern learning craves is slow and expensive, and the data is only as good as the operators' skill and consistency. To stretch each precious demonstration further, teams gather varied examples across many objects and starting positions, share large teleoperation datasets, and increasingly let robots practise on their own once a base of human demonstrations has given them a head start.