OPTIMAL OBSTACLE AVOIDANCE STRATEGY USING DEEP REINFORCEMENT LEARNING BASED ON STEREO CAMERA

Abstract

Mobile robots (MRs) are exerting a significant influence in industrial as well as residential settings. Concurrently with the escalating advancement of technology, the incorporation of artificial intelligence algorithms into the obstacle evasion issue of MRs is gaining increasing attention. The paper utilizes Deep Reinforcement Learning (DRL) to a MR that is furnished with a camera. Images captured by a stereo camera will be inputted into the YOLO-v8 model to identify obstacles situated in the path of the MR. Subsequently, the distances to these obstacles will be regarded as the state of the MR. The information was utilized to train a Deep Q-Network. Throughout this training process, the system acquires the capability to determine suitable actions for the MR to advance towards the destination while circumventing obstacles. Each action executed by the MR is accompanied by a reward, with the path yielding the most desirable outcome receiving the highest reward. The outcomes of the simulations conducted on the Robot Operating System 2 (ROS2) corroborate the effectiveness of this Deep Reinforcement Learning technique for the task of obstacle avoidance.