Autonomous RL: Autonomous Vehicle Obstacle Avoidance in a Dynamic Environment using MLP-SARSA Reinforcement Learning

摘要

This paper presents a Multi-Layer Perceptron-State Action Reward State Action (MLP-SARSA) based reinforcement learning methodology for dynamic obstacle detection and avoidance for autonomous vehicle navigation. MLP-SARSA is an on-policy reinforcement learning approach, which gains information and rewards from the environment and helps the autonomous vehicle to avoid dynamic moving obstacles. MLP with SARSA provides a significant advantage over dynamic environment compared to other traditional reinforcement algorithms. In this study, a MLP-SARSA model is trained in a complex urban simulation environment with dynamic obstacles using the pygame library. Experimental results show that the trained MLP-SARSA can navigate the autonomous vehicle in a dynamic environment with more confidences than traditional Q-learning and SARSA reinforcement algorithms.