Project Page

Bionic Human-Motion Style Transfer for Physically Executable Whole-Body Control of Humanoid Robots

Exemplar-driven expressive motion generation for humanoid robots, combining physics-aware latent diffusion with deployable whole-body tracking on Unitree G1.

Tianchen Huang* Mingkuan Zhao* Yang Gao Feiyang Yuan Junchi Gu Xiaohu Zhang Dongdong Zhao Shi Yan Yu Wang Wei Gao Shiwu Zhang

* Equal contribution. Yu Wang and Wei Gao are corresponding authors.

University of Science and Technology of China; Xi'an Jiaotong University; Lanzhou University

Paper arXiv Video Code (Coming Soon)

Abstract

Short human motion exemplars can become reusable bionic style sources for expressive humanoid behavior.

Expressive whole-body motion is important for humanoid robots operating in human environments, where robots need to remain stable while presenting readable and adjustable body behaviors. This work proposes a bionic generation-to-control framework that transfers style from a short human motion exemplar to a target content motion.

The generated stylized references are regularized for contact consistency and temporal smoothness, converted into G1-compatible robot references, and executed by a preview-based whole-body tracking policy trained with a cluster-and-distill strategy.

Physics-aware multi-condition latent diffusion architecture — A physics-aware multi-condition latent diffusion model transfers short human style exemplars while preserving target motion content and trajectory.

Generation To Control

The pipeline connects expressive style generation with robot-oriented conversion, tracking, and deployment.

Cluster-and-distill training for stylized whole-body tracking — Cluster-and-distill tracking trains style-diverse experts and distills them into one deployable controller.

Style Source

Short human exemplars

Human-derived cues such as gait rhythm, posture, arm swing, and body sway are transferred to new motion contents.

Generator

Multi-condition diffusion

Content, style, and trajectory streams are fused as conditional context for whole-body reference generation.

Executability

Physics-aware regularization

Contact-consistency and temporal-smoothness terms reduce stance-foot drift and high-frequency joint artifacts.

Deployment

Preview-based tracking

Generated motions are converted to 29-DoF Unitree G1 references and tracked with a distilled whole-body policy.

Results

Evaluation combines perceptual motion metrics with robot-oriented feasibility and closed-loop execution.

96.0% hardware success rate

120/125 successful Unitree G1 trials

96.02% simulated success rate

0.004722 foot sliding factor

Method	FMD	CRA (%)	SRA (%)	FSF	Apos	Jpos	Sim. SR (%)
MotionDiffuse	-	-	-	0.006233	0.497825	0.339368	27.93
Motion Puzzle	113.31	26.31	46.33	0.005853	0.568928	0.822193	67.14
MCM-LDM	34.78	33.62	58.66	0.006273	0.372012	0.316419	59.81
Ours	39.81	35.18	55.55	0.004722	0.307976	0.283434	96.02

Exemplar-driven style transfer across content motions and style exemplars — Exemplar-driven style transfer across different contents and human style sources.

Style Intensity And Deployment

Classifier-free guidance adjusts style intensity while the controller keeps the motion inside an executable range.

Style intensity control in simulation and real Unitree G1 deployment

Increasing the guidance scale strengthens recognizable human-like cues such as slower gait rhythm and upper-body sway. The reported hardware evaluation covers 25 Unitree G1 trials for each guidance scale in {0, 0.25, 0.5, 0.75, 1.0}.

Across 125 total trials, the robot completed 120 trials without falling, safety shutdown, or unrecoverable tracking divergence. Most failures occur at larger guidance scales, where stylization can increase sway and stance-timing difficulty.

Citation

@misc{huang2026bionic,
  title  = {Bionic Human-Motion Style Transfer for Physically Executable Whole-Body Control of Humanoid Robots},
  author = {Huang, Tianchen and Zhao, Mingkuan and Gao, Yang and Yuan, Feiyang and Gu, Junchi and Zhang, Xiaohu and Zhao, Dongdong and Yan, Shi and Wang, Yu and Gao, Wei and Zhang, Shiwu},
  year   = {2026},
  eprint = {2606.03536},
  archivePrefix = {arXiv},
  primaryClass = {cs.RO},
  url    = {https://arxiv.org/abs/2606.03536}
}