From teleoperated session to training data
When an operator performs a task through an XR headset purely to generate training data, every frame of the session, the robot's camera views, joint positions, and gripper states, can be recorded alongside the task being performed. Each recorded session becomes one demonstration, and a large collection of these demonstrations forms a dataset that a robot learning model can be trained on. This connects XR teleoperation directly to imitation learning, the family of techniques covered in the Imitation Learning and Learning from Demonstration course, where a model learns a policy by studying human demonstrations rather than being explicitly programmed.
What makes XR particularly well suited to this pipeline is throughput. Because the interface feels natural, a single operator can produce many demonstrations of a task in a session, and because the data being recorded is just the robot's own sensor and joint readings, it requires no separate translation step between how the demonstration was performed and how the robot will later act.