Any2Track

A humanoid motion tracker from Tsinghua, Peking University, and Galbot that keeps tracking a motion even when pushed, pulled, or loaded down.

Any2Track was proposed jointly by Tsinghua University, Peking University, Galbot, and the Shanghai Qi Zhi Institute (corresponding author Li Yi), released in September 2025. Motion tracking means having a humanoid reproduce a reference motion in real time, a core capability underlying whole-body control and teleoperation; most existing trackers had only been validated on flat ground with no external forces. Any2Track uses two-stage reinforcement learning. First, a general tracker called AnyTracker is trained on LAFAN1 and AMASS motion-capture data, learning highly dynamic, multi-contact motions of many kinds with a single policy. It's then frozen, and an adaptation module called AnyAdapter is added, extracting dynamics features from a recent window of state-action history (trained with an auxiliary task of predicting future states, like a world model) to compensate online for changes in terrain, external force, and load — without degrading the original tracking ability. The policy is trained in simulation and deployed zero-shot to a real Unitree G1.

ExampleIn real-robot tests, a loaded-down G1 tracks a reference motion over uneven ground while someone pulls it with a rope and pushes it with a foot, and the robot stays stable and keeps completing the motion.

Also called

Track Any Motions under Any Disturbances, AnyTracker, AnyAdapter

Related

Motion Tracking · Rapid Motor Adaptation · Sim-to-Real Transfer · Unitree G1 · AMASS (Archive of Motion Capture as Surface Shapes) · GMT

Sources

Track Any Motions under Any Disturbances (arXiv 2509.13833)
Any2Track 项目页 (Chinese)

As of

2025-09