A Deep Reinforcement Learning Approach for the Pursuit Evasion Game in the Presence of Obstacles

Abstract

In a pursuit-evasion game, the pursuer tries to capture the evader, while the evader actively avoids being captured. Traditional approaches usually ignore or simplify kinematic constraints by using a grip world discrete model and they assume that the game is played in free space without obstacles. In this paper, a curriculum deep reinforcement learning approach is proposed for the pursuit-evasion game, which considers the kinematics of mobile robot in practical applications and the influence of static obstacles in the environment. To improve the system performance, we use the mechanism of self-play to train the pursuer and the evader at the same time. In addition, the method of curriculum learning is used, making the agent learn simpler tasks before learning more complicated ones. Comparative simulation results show that the proposed approach presents superior performance for both pursuer and evader when playing against intelligent opponents.