Whole-Body Humanoid Control via Stochastic Motion Tokenization

A stochastic motion tokenizer gives a high-level generator a compact interface to a whole-body tracking policy; reports better G1 motion reconstruction than FSQ/VQ-VAE and real dance/fight/walk sequences.用随机运动分词器把行为标签或文本生成的动作压缩成离散表示，直接作为全身跟踪策略的参考；报告在 Unitree G1 动作重建上优于 FSQ 与 VQ-VAE，并在真机上执行舞蹈、格斗、行走及 35 秒的格斗转舞蹈序列。

@misc{drolet2026whole,
  title = {{Whole-Body Humanoid Control via Stochastic Motion Tokenization}},
  author = {Michael Drolet and Oleg Arenz and Jan Peters},
  year = {2026},
  eprint = {2610.36452},
  archivePrefix = {arXiv},
  primaryClass = {cs.RO},
  url = {https://arxiv.org/abs/2610.36452},
}