MARRC: A Multi-Agent Reinforcement Redirection Controller for Redirected Walking

Redirected walking technique can provide users with immersive virtual experiences, but frequent resets can reduce user immersion. In multi-user environments, existing redirection controllers can guide all users to move in the optimal direction to reduce the number of resets, but they usually ignore the collaborative relationship between users, resulting in mutual interference of user movements and thus affecting the redirection effect. To tackle this problem, this paper proposes a Multi-Agent Reinforcement Redirection Controller (MARRC) , which minimizes the number of resets through collaboration among users by taking into account physical obstacles and dynamic movement of multiple users. MARRC is trained using multi-agent reinforcement learning, which enables the model to make optimal decisions in combination with spatial geometries by observing minimum distance information in the surrounding 24 directions. At the same time, a ”nearest user distance inverse ratio” penalty term is introduced in the reward function to effectively avoid collisions between users. We compare MARRC with other redirection controllers through simulation and user experiments and find that MARRC significantly reduces the number of resets by about 15 % on average and increases the distance between resets by about 32 % on average in a multi-user environment.