Evaluate Optimal Redirected Walking Planning Using Reinforcement Learning

Abstract

Redirected Walking (RDW) is commonly used to overcome the limitation of real walking locomotion while exploring virtual worlds. Although a few machine learning-based RDW algorithm is proposed, most of the system did not go through live user evaluation. In this work, we evaluated a novel RDW controller proposed by Chang et al., in which the formatted steering rule is replaced with reinforcement learning(RL), by simulation and live user experiment. We found the RL-based RDW controller reduced boundary collisions significantly in both simulation and user study comparing to the heuristic algorithm, Steer-to-Center(S2C); also, there are no noticeable differences in immersiveness. These results indicate that the novel controller is superior to the heuristic method. Furthermore, as we conducted experiments in a relatively simple space and still outperformed the heuristic method, we are optimistic that the RL-based controller can maintain the high-performance in complicated scenarios in the future.