Humanoid robot motion tracking adapts faster to real-world changes
HumanoidNDM: Closing the Training-Deployment Gap of Motion Tracking via Humanoid Neural Dynamics Model
RoboticsArtificial IntelligenceMachine Learning
Summary
Humanoid robots often struggle when their trained skills don’t match real-world conditions due to physical disturbances or imperfect simulations. The authors created HumanoidNDM, a system that uses real-world data to quickly adjust robot movements during use without needing expensive motion capture tools. This system balances trying new movements with what the robot can feasibly do and trains a smart model that predicts future movements based on past steps. It then fine-tunes the robot’s control policy efficiently using this model’s information. Tested on a real robot, the method showed rapid adaptation to different real-world challenges.
What this means in practice
- •For robotics engineers: Fine-tune humanoid robot motion policies online using real-world data without motion capture systems.
- •For robotics hardware developers: Improve robot physical design and control synergy by rapidly adapting motion tracking to disturbances encountered in physical deployments.
Authors
Chao Chen, Zhenrong Wang, Jilong Wang, Yusen Feng, Jiangran Lyu, Li Yi, He Wang
Abstract
Humanoid robots have recently achieved remarkable success in motion tracking. However, during real-world deployment, robots inevitably encounter dynamics gaps arising from either extra physical disturbances or inaccurate simulation modeling, resulting in a significant training-deployment gap. Therefore, effectively leveraging real-world data to fine-tune policies represents the crucial last mile for robust deployment. To bridge this gap, we propose HumanoidNDM, an online fine-tuning framework for humanoid motion tracking driven by a real-world-trained neural dynamics model. First, in the data collection stage, we introduce a low-frequency reference perturbation method with an adaptive intensity mechanism, balancing deployment feasibility and exploration efficiency. Second, for the model training stage, we train a robust history-based, proprioceptive neural dynamics model in Euclidean space by incorporating differentiable forward kinematics, eliminating the reliance on costly motion capture systems. Finally, in the policy fine-tuning stage, we directly fine-tune the policy via backpropagated model gradients, achieving high data efficiency. Through physical validation on the Unitree G1 robot, we demonstrate that HumanoidNDM enables rapid real-world adaptation to various perturbations, representing the first online, MoCap-free fine-tuning framework for humanoid tracking policies.