Dynamic Leader Selection of a Multirobot System in Multiple Maze-like Environments Using Reinforcement Learning

This paper examines a leader selection problem of a multirobot leader-follower system in maze-like environments. Behavior-based control and a repulsive force method are applied to the leader and follower robots navigating the environments; a Q-learning algorithm combined with a fuzzy-based state approximation is developed to automatically assign a leader when robots are stranded; a cross-entropy exploration storing algorithm is proposed to exploit the information learned. Numerical analyses illustrate the effectiveness of the proposed leader selection approach based on reinforcement learning: the multirobot system can dynamically select a leader and navigate maze-like environments of interest to reach the destination. Index Terms—Reinforcement learning, Q-learning, mobile robot, multirobot system, maze, behavior based model, repulsive force method, fuzzy inference system.