Humanoids 2026 paper adopts stochastic motion tokenization for whole-body control on the Unitree G1

SUMMARY

Michael Drolet, Oleg Arenz and Jan Peters have posted a paper on arXiv, submitted on 27 October 2026, that proposes controlling humanoids through compact motion representations instead of velocity commands. The paper has been accepted to the 2026 IEEE-RAS International Conference on Humanoid Robots (Humanoids 2026).

The authors start from two limitations. Velocity-commanded reinforcement learning policies work well for quadrupeds, but they are a limited interface for a humanoid that has to perform complex whole-body motions. Tracking policies for such motions also tend to know little about the desired future motion, which makes it hard for them to deviate in a controlled way to succeed at a long-term task, for example when preparing a transition between behaviours.

Their approach has a high-level controller generate concise motion representations from a behaviour label or a text command. These representations are passed straight to a tracking policy, where they serve as a detailed motion reference. The authors say the method handles real-world motions such as dancing, fighting and walking, depending on the conditioning information.

The authors adopted a stochastic motion tokenizer for these motion representations. They report that, at matched capacity, it reconstructs motions from the Unitree G1 at a level that exceeds finite scalar quantization (FSQ) and the vector-quantized autoencoder (VQ-VAE). The interface that results is discrete, and on a physical Unitree G1 it conditions whole-body tracking policies that execute generated dance, fight and walk motions lasting 16.8 s each, plus a 35 s fight-to-dance sequence.

Drafted with AI assistance from the source and reviewed by WBH. Follow the source link for the full text.