Event
Random Motion Tokenization for Humanoid Whole-Body Control
1 reports · 1 sources
Michael Drolet, Oleg Arenz, and Jan Peters published a paper proposing the use of a stochastic motion tokenizer to generate concise discrete motion representations that are fed directly into a whole-body tracking policy. The authors claim that, at equal capacity, this tokenizer surpasses FSQ and VQ-VAE in reconstructing Unitree G1 motions. In real-robot demonstrations, this discrete interface drives the whole-body tracking policy to execute generated dance, combat, and walking motions on a physical Unitree G1, each lasting 16.8 seconds, and completes a 35-second combat-to-dance sequence.
Report timeline
- Whole-Body Humanoid Control via Stochastic Motion Tokenization
This study proposes using a stochastic motion tokenizer to generate concise discrete motion representations that can be fed directly into a whole-body tracking policy to execute dance, fighting, and walking motions on the Unitree G1. At matched capacity, the tokenizer reconstructs Unitree G1 motions better than FSQ and VQ-VAE. In real-robot demonstrations, each generated motion lasted 16.8 seconds, and a 35-second fighting-to-dance sequence was completed.
arXiv cs.RO ·