Autonomous Robot Navigation in Dynamic Environment Using Deep Reinforcement Learning

摘要

Compared to traditional control methods, deep reinforcement learning (DRL) has the ability to learn how to solve complex tasks in a dynamic environment simply by collecting experience. In this paper, we study the application of DRL method in robotic autonomous control with detection capability in simulated dynamic environment. More specifically, we have adopted Deep Q Network (DQN), double DQN and dueling DQN algorithms in DRL. As with fixed reward settings, these original DRL algorithms do not perform well while navigating a robot in dynamic environment. To address the problems, we designed a novel reward shaping method and conducted a series of experiment with all three improved DRL algorithms. The results show that the new reward shaping method can significantly improve the DRL performance when they are applied in robot navigation settings.