Event

HumanoidNDM: An Online Fine-Tuning Framework for Humanoid Motion Tracking

1 reports · 1 sources

Chao Chen and other authors propose HumanoidNDM, an online fine-tuning framework for humanoid motion tracking that uses a neural dynamics model trained on real hardware to narrow the dynamics gap between training and deployment, and they completed physical validation on a Unitree G1 robot; the authors state that it can quickly adapt to a variety of perturbations. Methodologically, the team collected data using low-frequency reference perturbations with an adaptive intensity mechanism, and trained a dynamics model based on proprioceptive history in Euclidean space combined with differentiable forward kinematics, thereby avoiding reliance on expensive motion capture systems; they then directly fine-tuned the policy via model gradient backpropagation. The work has been accepted by CoRL 2026. The above adaptation results come from the authors' validation on a single robot model; they are the authors' claims and have not yet been independently reproduced.