PREPRINT

Make Tracking Easy:

Neural Motion Retargeting for Humanoid Whole-body Control

Qingrui Zhao1, Kaiyue Yang1, Xiyu Wang1, 2, Shiqi Zhao1, Yi Lu1, Xinfang Zhang2, Wei Yin3, Qiu Shen1, Xiao-Xiao Long1*, Xun Cao1*
*Corresponding authors.
1Nanjing University, 2Huawei Technologies, 3Horizon Robotics
Paper ArXiv Code HuggingFace Ascend

Boosting RL learning with motion reconstructed from video

Less Error, Better Real-World Performance

Method

Optimization-based methods suffer from local optima and usually sensitive to initial value.

Our Core Idea: reformulate retargeting from static optimization over frame-wise states into dynamic mapping between motion distributions.

Clustered - Expert Physics Refinement (CEPR).

Clustered - Expert Physics Refinement (CEPR)

Network Structure.

Network Structure

PRE-TRAINING:

Kinematic alignment with large-scale data

POST-TRAINING:

Physical grounding with CEPR data

More Results

Floating Feet

Joint Jump

Self Intersection

Reference

  1. Liao Q, Truong T E, Huang X, et al. Beyondmimic: From motion tracking to versatile humanoid control via guided diffusion[J]. arXiv preprint arXiv:2508.08241, 2025.
  2. Araujo J P, Ze Y, Xu P, et al. Retargeting matters: General motion retargeting for humanoid motion tracking[J]. arXiv preprint arXiv:2510.02252, 2025.

Citation

@article{zhao2026maketrackingeasy,
      title={Make Tracking Easy: Neural Motion Retargeting for Humanoid Whole-body Control},
      author={Qingrui Zhao and Kaiyue Yang and Xiyu Wang and Shiqi Zhao and Yi Lu and Xinfang Zhang and Wei Yin and Qiu Shen and Xiao-Xiao Long and Xun Cao},
      journal={arXiv preprint arXiv:2603.22201},
      year={2026},
      url={https://arxiv.org/abs/2603.22201}
}