GMT: General Motion Tracking for Humanoid Whole-Body Control
Zixuan Chen, Mazeyu Ji, Xuxin Cheng, Xuanbin Peng, Xue Bin Peng, Xiaolong Wang
2025-06-19
Summary
This paper talks about GMT, a new system that helps humanoid robots track and imitate a wide variety of body motions by learning a general control policy.
What's the problem?
The problem is that humanoid robots have a hard time following complex, whole-body movements because motions can vary a lot over time, different parts of the body need to work together, and training to do this well is difficult.
What's the solution?
The researchers created GMT with two main ideas: adaptive sampling to balance easy and hard motions during training, and a motion mixture-of-experts architecture that lets different parts of the system specialize in tracking different types of movements. They also use a method that looks at multiple future motion frames to help the robot understand and predict movements better.
Why it matters?
This matters because better motion tracking makes humanoid robots more capable and versatile in performing tasks that require natural and coordinated body movements, bringing us closer to useful robots in everyday life.
Abstract
GMT, a unified motion-tracking framework, addresses challenges in tracking diverse humanoid robot motions through adaptive sampling and a motion mixture-of-experts architecture, achieving state-of-the-art performance.