EG
Human2Humanoid framework diagram showing the physics-aware motion retargeting pipeline from human to Unitree G1 robot
ResearchJune 11, 2026University of Science and Technology of China

Human2Humanoid: Physics-Aware Framework Enables Unsupervised Motion Retargeting to Unitree G1

Researchers present Human2Humanoid, an unsupervised motion retargeting framework that transfers human motions to humanoid robots with high fidelity. Using a CycleGAN-based architecture with skeleton-aware graph convolutional networks and physics-aware feasibility constraints, the framework successfully retargets human motion to Unitree G1 without paired data, outperforming existing methods in controllability and physical feasibility.

#humanoid-robot#motion-retargeting#physics-aware#unitree-g1#unsupervised#arXiv#cyclegan
Reading in English

Researchers have introduced Human2Humanoid, an unsupervised motion retargeting framework designed to bridge the morphological gap between humans and humanoid robots.

The Challenge of Cross-Morphology Retargeting

Retargeting human motion to humanoid robots is critical for teleoperation, imitation learning, and human-robot interaction. However, substantial morphological discrepancies—including skeletal topology differences, limb proportions, and degrees of freedom—make this challenging without paired motion data.

Technical Approach

The framework employs three key innovations:

  1. CycleGAN-based architecture with skeleton-aware graph convolutional networks to capture topology-dependent motion features
  2. Morphology-invariant end-effector consistency loss that aligns normalized end-effector trajectories to preserve motion semantics
  3. Physics-aware feasibility constraints that encourage reproduction of contact patterns in source motion

Experimental Results

The framework successfully retargets human motion to Unitree G1 without paired data and outperforms existing methods in:

  • Downstream controllability
  • Physical feasibility
  • Contact pattern preservation

Significance

This work enables rapid deployment of human-like motions to humanoid platforms without expensive data collection, significantly accelerating the development of expressive and natural humanoid robots.

Source: arXiv:2606.03476
Language: English- Showing content in English