TWIST2

A portable teleoperation system that collects whole-body humanoid data with a VR headset, with no motion-capture studio needed.

TWIST2 was released in November 2025 by a joint team from Stanford, Amazon FAR, USC, UC Berkeley, and CMU (authors including Yanjie Ze, Pieter Abbeel, Guanya Shi, Jiajun Wu, and C. Karen Liu), an upgrade to the earlier full-body teleoperation system TWIST. Whole-body humanoid data is hard to collect, and its predecessor relied on optical motion capture tied to a fixed studio. TWIST2 instead captures the operator's full-body motion with a PICO 4 Ultra headset plus two PICO body trackers, and adds a roughly $250, 2-degree-of-freedom active neck to the Unitree G1 so the robot has its own first-person view; the whole setup can be carried into any environment. The paper reports that 15 minutes of use can collect about 100 demonstrations at close to 100% success. On top of this, a hierarchical policy is trained: a low-level whole-body motion tracker learned through simulated reinforcement learning, and a high-level visuomotor policy trained with imitation learning on the collected data. The system, hardware design, and dataset are all open-sourced.

ExampleAn operator wears the PICO headset and straps on the trackers to perform pick-and-place motions, and the Unitree G1 reproduces the whole-body motion in real time, recording over a hundred bimanual pick-and-place demonstrations in just 15 minutes.

Also called

TWIST 2, TWIST2: Scalable, Portable, and Holistic Humanoid Data Collection System

Related

Twist · Whole-Body Teleoperation · VR Teleoperation · Unitree G1 · Hierarchical Architecture · Motion Tracking

Sources

TWIST2 (arXiv 2511.02832)
TWIST2 project page

As of

2025-11