Teleoperation and Robot Data Collection
Robot learning runs on demonstrations, so how they are collected matters a lot. A leader-follower setup, where the operator moves a small copy of the robot arm, feels natural and gives precise joint data. VR controllers are cheap and flexible. Camera-based hand tracking needs no hardware at all but is noisier. Each choice trades cost, quality, and speed.
What makes a good demonstration
Policies learn whatever is in the data, including the interface's quirks. Three properties matter: delay between the operator's intent and the recorded motion, jitter, the noise in the recorded path, and throughput, demonstrations per hour. A simple model of a recorded signal is the intended motion, delayed and noisy:
What to notice
The leader-follower arm records smooth motion with little delay, and with force feedback the operator can feel contact, which matters for tasks like insertion. Hand tracking is the noisiest and loses the hand when it turns. VR sits in between and is easy to scale to many operators. The numbers here are illustrative of each interface's typical character, not measurements of particular products. Datasets such as Open X-Embodiment mix data from many interfaces, which is one reason cross-embodiment learning matters.