Learning Surfing-like Balance without Water Simulation

Our method learns surfing-like balance control without water simulation. The learned policy maintains stable balance on a moving board under wave conditions at runtime.

Video

Abstract

Controlling characters in fluid environments such as water remains challenging due to complex dynamics and high simulation cost. In particular, surfing requires maintaining balance on a continuously moving support, making stable control difficult without accurate physical modeling. In this work, we present a method for learning surfing-like balance control without relying on water simulation. Our approach combines a hierarchical control framework with a stage-based training scheme that progressively introduces dynamic board behavior. The low-level policy generates full-body motion from partial joint targets, while the high-level policy adapts these targets based on environmental conditions. Despite being trained entirely in a non-fluid setting, the learned policy maintains stable balance on a moving board under wave conditions at runtime. These results suggest that surfing-like balance can be achieved without explicitly modeling fluid dynamics through appropriate control structures and training strategies.