PAPER_TITLE

FIRST_AUTHOR_LAST, FIRST_AUTHOR_FIRST; SECOND_AUTHOR_LAST, SECOND_AUTHOR_FIRST

Physics filtering favors the generalization of robot learning

Jindou Jia^1,3†, Shixuan Han^3†, Meng Wang^3†, Gen Li¹, Zihan Yang⁴, Sicheng Zhou², Kexin Guo^4*, Jianfei Yang^1,2*, Xiang Yu³, Wei Wang⁵, Lei Guo³

¹School of Mechanical and Aerospace Engineering, Nanyang Technological University, 50 Nanyang Avenue, 639798, Singapore. ²School of Electrical and Electronic Engineering, Nanyang Technological University, 50 Nanyang Avenue, 639798, Singapore. ³School of Automation Science and Electrical Engineering, Beihang University, Xueyuan Road, Beijing, 100191, China. ⁴School of Aeronautic Science and Engineering, Beihang University, Xueyuan Road, Beijing, 100191, China. ⁵China Aerospace Science and Technology Corporation, Fucheng Road, Beijing, 100048, China.
^†Indicates Equal Contribution

Paper Supplementary Code

Overview of the presented PhyFilter. a The learning output of RL or SL is enhanced by a physics-informed filter which helps narrow the large generalization gap. b The implementation details of the PhyFilter, which is lightweight and analytic. More theoretical details refer to Methods “Filtering learning residual”. c The PhyFilter is validated across four distinct empirical examples. i Policy learning for quadruped locomotion. The quadruped is trained in a simplified simulational environment; with the PhyFilter, it generalizes to unseen velocity, payload, and real-world terrains. ii Dynamical learning for drone maneuvering flight. A specific machine learning technique learns mass uncertainty, while the PhyFilter helps handle unseen wind disturbance. iii Dynamical learning for aerial manipulation. A designated machine learning technique learns coupling uncertainties between drone and manipulator, with the PhyFilter aiding in the handling of unseen wind disturbances and mass uncertainties. iv Kinematics learning for acceleration perception. An NN-based differentiator with the input of velocity sequence, is trained to obtain the real-time acceleration for a drone. The PhyFilter can favor its generalization to inputs with distribution shift.

Experimental results of the quadruped example.

The employed hardwares.

Abstract

Living organisms exhibit extraordinary adaptability to unseen environments through their intrinsic physical structures and lifelong feedback-driven learning. Endowing robots with comparable generalization is critical for reliable operation in the real world. While recent approaches attempt to improve generalization by scaling training data, such strategies remain impractical for robotics, where collecting real-world demonstrations at the scale of large language models is prohibitively costly and slow. Contrary to this reliance on massive datasets, we show that robots can generalize effectively even with limited training data by leveraging a feedback mechanism, namely PhyFilter, that corrects learning outputs with physics-filtered learning residuals. PhyFilter operates as a lightweight, model-agnostic module whose parameters can be automatically optimized through an auto-learning algorithm, eliminating manual tuning and enabling seamless integration with diverse robot policies. We validate PhyFilter across four representative robotic systems, demonstrating that it enables quadruped robots to generalize to unseen terrains, payload variations, and speed ranges; drones to flight under unseen wind disturbances; aerial manipulators to achieve centimeter-level in-air capture despite wind and mass uncertainties; and acceleration differentiators to remain robust with distribution shift. These results reveal a scalable and data-efficient pathway toward generalizable robot learning, showing that physics-filtered feedback can serve as a powerful alternative to massive data scaling.

Quadruped Experiments

Quadruped Simulations

Drone Experiments

Aerial Manipulation Experiments

Contact

If you have any questions, please contact us. jdjiabuaa@126.com; mwangbuaa@126.com