UniTracker

A three-stage whole-body motion tracking framework that lets a single policy make a humanoid robot track a wide range of human motions.

UniTracker was released in July 2025 by Shanghai Jiao Tong University, the Shanghai Artificial Intelligence Laboratory, and other institutions. Motion tracking means having a robot imitate a reference human motion in real time; the difficulty is that one policy has to cover thousands of different motions, and a real robot cannot access the complete state information available in simulation. UniTracker has three stages: first, a teacher policy is trained in simulation using privileged information (full state information unavailable on the real robot); then it is distilled into a real-robot-ready student policy, which learns a global latent variable for each motion via a conditional variational autoencoder (CVAE) to reduce drift in global quantities like orientation when only partial observations are available; finally, a fast-adaptation module fine-tunes individually or in batches on motions that are hard to track. The training data is 8,179 human motion clips filtered from AMASS, validated both in simulation and on a real Unitree G1.

ExampleGiven a dance motion clip taken from AMASS, the same UniTracker policy can make a real Unitree G1 perform it; for a particularly hard motion sequence the robot cannot track directly, the third-stage fast-adaptation module fine-tunes on it individually or in a batch.

Also called

UniTracker: Learning Universal Whole-Body Motion Tracker for Humanoid Robots

Related

Motion Tracking · Conditional Variational Autoencoder · Teacher-Student Distillation · Privileged Information · AMASS (Archive of Motion Capture as Surface Shapes) · Unitree G1

Sources

UniTracker (arXiv 2507.07356)
UniTracker 论文 HTML 版（含机构与实验设置） (Chinese)

As of

2025-09