Efficient and Versatile Quadrupedal Skating: Optimal Co-design via Reinforcement Learning and Bayesian Optimization

Abstract

In this paper, we propose an effective mechanical and algorithmic solution to enabling skating motion with passive wheels on state-of-the-art quadrupedal robots. The skating locomotion enables a hybrid combination of wheeled and legged mobility without the necessity of motorization at the feet, which simultaneously promote efficiency, speed, and mechanical simplicity. To realize these potential advantage of skating, we employ a bilevel optimization approach with an upper level optimization via Bayesian Optimization (BO) to search for the best mechanical design and a lower level Reinforcement Learning (RL) to find an optimal motor policy. The end results not only provide optimal mechanical and control designs but also show versatile locomotion behaviors such as hockey stop (a rapid braking maneuver by turning sideways to maximize friction) and self-aligning behavior (automatically adjusts its orientation to move more efficiently in the commanded direction), providing the first comprehensive study on quadrupedal robotic skating.

Hanwen Wang 王瀚文

PhD Student & UW-Madison

Hanwen joined the lab in the Fall of 2024 as a PhD student. His research interest focuses on the intersection of learning-based and model-based methods for legged and humanoid robots.

Zhenlong Fang 方振龙

Research Assistant

Zhenlong joined the lab in the Summer of 2025 as a visiting student.

Xiaobin Xiong 熊晓滨

Associate Professor @ SII

Prof. Xiong is a full-stack roboticist who develop rigorous theories for practical applications.