Anticipate Before Act: Prediction Based Constrained Reinforcement Learning Framework for Skiing Robot Control

Abstract

Enabling a robot to ski with agility presents an exciting yet complex challenge, primarily due to the intricate dynamics arising from ski-snow interactions. Existing robotic simulators are unable to accurately model the non-rigid, highly dynamic contact between skis and deformable snow surfaces. Hence, reinforcement learning frameworks cannot obtain such locomotion polices. In this work, we introduce the Prediction based Constrained Reinforcement Learning (PCRL) framework for skiing robots. PCRL integrates a learned prediction module with a constrained policy optimization, enabling the robot to anticipate its future states and make informed decisions under delayed and uncertain contact conditions. We also develop a skiing simulator with realistic ski-snow interaction modeling to support effective policy training. Experimental results on a hexapod skiing robot confirm the accuracy of the simulator and demonstrate that PCRL achieves great performance in both simulated and real-world skiing tasks. Compared to model-based approaches, our method significantly improves agility across a wide range of skiing speeds, underscoring its promise for advancing dynamic locomotion in complex, deformable environments.