Track Any Motions under Any Disturbances
Zhikai Zhang, Jun Guo, Chao Chen, Jilong Wang, Chenghuai Lin, Yunrui Lian, Han Xue, Zhenrong Wang, Maoqi Liu, Jiangran Lyu, Huaping Liu, He Wang, Li Yi
AI summary
Problem
Existing humanoid motion trackers lack the simultaneous capability to handle diverse, highly dynamic motions and adapt to real-world dynamics disturbances like terrain changes, external forces, and payload variations, limiting practical deployment.
Approach
The framework decouples motion tracking and dynamics adaptation into two stages: a base policy trained on diverse motions using canonicalized action spaces and specialist-to-generalist distillation, followed by a history-informed adapter that extracts dynamics features to adjust actions online without degrading tracking performance.
Key results
- Unified policy tracks diverse, highly dynamic, and contact-rich motions
- Online adaptation to terrains, external forces, and physical property changes
- Successful zero-shot sim-to-real transfer deployed on Unitree G1 hardware
- State-of-the-art robustness and motion expressiveness under simultaneous disturbances
Why it matters
Provides a foundational, deployable solution for humanoids to perform expressive, dynamic tasks reliably in unstructured real-world environments.
Abstract
A foundational humanoid motion tracker is ex- pected to be able to track diverse, highly dynamic, and contact-rich motions. More importantly, it needs to operate stably in real-world scenarios against various dynamics dis- turbances, including terrains, external forces, and physical property changes for general practical use. To achieve this goal, we propose Any2Track (Track Any motions under Any disturbances), a two-stage RL framework to track various mo- tions under multiple disturbances in the real world. Any2Track reformulates dynamics adaptability as an additional capability on top of basic action execution and consists of two key components: AnyTracker and AnyAdapter. AnyTracker is a general motion tracker with a series of careful designs to *Equal Contributions, :Corresponding Author 1Tsinghua University, 2Peking University, 3Galbot, 4Shanghai Qi Zhi Institute Paper website: https://zzk273.github.io/Any2Track/ track various motions within a single policy. AnyAdapter is a history-informed adaptation module that endows the tracker with online dynamics adaptability to overcome the sim2real gap and multiple real-world disturbances. We deploy Any2Track on Unitree G1 hardware and achieve a successful sim2real transfer in a zero-shot manner. Any2Track performs exceptionally well in tracking various motions under multiple real-world disturbances, as shown in Figure 1. For real-world demos, please refer to https://zzk273.github.io/Any2Track/.