HumanoidNDM: Closing the Training-Deployment Gap of Motion Tracking via Humanoid Neural Dynamics Model

SUMMARY

A paper posted to arXiv on 25 October 2026 by Chao Chen, Zhenrong Wang, Jilong Wang, Yusen Feng, Jiangran Lyu, Li Yi and He Wang describes HumanoidNDM, an online fine-tuning framework for humanoid motion tracking. The authors start from the observation that robots in real deployment meet dynamics gaps, caused either by extra physical disturbances or by inaccurate simulation modelling, and that these gaps open a significant difference between training and deployment. They argue that using real-world data to fine-tune policies is the last mile for robust deployment.

The framework has three stages. For data collection, the authors introduce a low-frequency reference perturbation method with an adaptive intensity mechanism, which they say balances deployment feasibility against exploration efficiency. For model training, they train a history-based, proprioceptive neural dynamics model in Euclidean space. It incorporates differentiable forward kinematics, which the authors say removes the reliance on costly motion capture systems.

In the final stage, the policy is fine-tuned directly through backpropagated model gradients, which the authors say gives high data efficiency. The paper reports physical validation on the Unitree G1 robot and says HumanoidNDM enables rapid real-world adaptation to various perturbations. The authors describe it as the first online, motion-capture-free fine-tuning framework for humanoid tracking policies.

The arXiv page notes that the paper has been accepted to CoRL 2026.

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