EG
Bionic motion style transfer framework showing the physics-aware multi-condition latent diffusion architecture
ResearchJune 11, 2026University of Science and Technology of China, Xi'an Jiaotong University, Lanzhou University

USTC Researchers Achieve 96% Success in Bionic Human Motion Style Transfer to Unitree G1

Researchers from the University of Science and Technology of China present a bionic generation-to-control framework for exemplar-driven style transfer on humanoid robots. The approach generates stylized whole-body references using a physics-aware multi-condition latent diffusion model, achieving a 96.0% success rate over 125 real-robot trials on Unitree G1. The framework preserves intended motion content while transferring human-derived movement styles for natural and expressive humanoid locomotion.

#humanoid-robot#motion-generation#style-transfer#diffusion#unitree-g1#expressive-motion#arXiv
Reading in English

Researchers from the University of Science and Technology of China have developed a bionic generation-to-control framework for exemplar-driven style transfer on humanoid robots, enabling expressive and natural whole-body motion generation.

The Challenge of Expressive Motion

Expressive whole-body motion is crucial for humanoid robots operating in social, service, and human-coexistence environments. Most expressive motions are still obtained from fixed demonstrations or manually designed scripts, making it difficult to reuse demonstrated styles across different motion contents.

Technical Innovation

The framework introduces:

  1. Physics-aware multi-condition latent diffusion model that fuses style, content, and trajectory conditions
  2. Classifier-free guidance to adjust style intensity without retraining
  3. Contact-consistency and temporal-smoothness regularization to improve hardware executability
  4. Cluster-and-distill training strategy for the whole-body tracking policy

Key Results

  • 96.0% success rate over 125 real-robot trials on Unitree G1
  • Reduces contact and jitter artifacts compared to animation-oriented baselines
  • Transfers short human style exemplars to diverse robot motion contents
  • Preserves intended motion content while expressing human-derived styles

Applications

This technology enables humanoid robots to adopt natural human movement styles for various tasks, from service interactions to industrial applications, significantly enhancing their ability to integrate seamlessly into human environments.

Source: arXiv:2606.03536
Language: English- Showing content in English