TWIST

Common

Stanford's 2025 whole-body teleoperation system for humanoids: the robot mirrors the operator's entire body in real time.

TWIST was released in May 2025 by Jiajun Wu and C. Karen Liu's teams at Stanford, working with Xue Bin Peng at Simon Fraser University, and published at CoRL 2025 with Yanjie Ze as first author. Earlier humanoid teleoperation usually controlled the upper and lower body separately, which couldn't produce coordinated whole-body moves like squatting to lift a box or kicking a ball. TWIST captures a person's full-body motion with OptiTrack optical motion capture, retargets it in real time (converting human motion into robot joint angles) onto a 29-degree-of-freedom Unitree G1, and tracks it with a single unified neural-network controller. That controller is trained in Isaac Gym simulation on about 42 hours of human motion-capture data: first a teacher policy that can see 2 seconds of future reference motion is trained, then it's distilled — using reinforcement learning plus behavior cloning — into a student policy that sees only the current frame, cutting latency. A later version, TWIST2, switches to a PICO VR headset and no longer needs motion capture, making it easier to collect data at scale.

ExampleAn operator on the motion-capture stage bends down to pick up a box from the floor, and the G1 mirrors the same whole-body motion in sync; when the operator kicks a ball or dances a waltz step, the robot follows along in real time.

Also called: TWIST: Teleoperated Whole-Body Imitation System
Related: Whole-Body Teleoperation · Motion Retargeting · Motion Tracking · Teacher-Student Distillation · TWIST2 · Unitree G1
Sources: TWIST: Teleoperated Whole-Body Imitation System (arXiv 2505.02833)
TWIST 项目页 (Chinese)
TWIST2: Scalable, Portable, and Holistic Humanoid Data Collection System (arXiv 2511.02832)
As of: 2025-11