Mobile Robot Obstacle Avoidance Based on Deep Reinforcement Learning

Obstacle avoidance is one of the core problems in the field of mobile robot autonomous navigation. This paper aims to solve the obstacle avoidance problem using Deep Reinforcement Learning. In previous work, various mathematical models have been developed to plan collision-free paths for such robots. In contrast, our method enables the robot to learn by itself from its experiences, and then fit a mathematical model by updating the parameters of a neural network. The derived mathematical model is capable of choosing an action directly according to the input sensor data for the mobile robot. In this paper, we develop an obstacle avoidance framework based on deep reinforcement learning. A 3D simulator is designed as well to provide the training and testing environments. In addition, we developed and compared obstacle avoidance methods based on different Deep Reinforcement Learning strategies, such as Deep Q-Network (DQN), Double Deep Q-Network (DDQN) and DDQN with Prioritized Experience Replay (DDQN-PER) using our simulator.