Autonomous Robot Navigation and Exploration Using Deep Reinforcement Learning with Gazebo and ROS

In this paper, robot navigation and exploration methodologies are presented using Deep Reinforcement Learning (DRL). The methodology has two parts. Firstly, the waypoint selection towards the global goal is done from generated points of interest (POI) in the environment. Secondly, a local navigation and obstacle avoidance policy is learned using DRL with a single low-cost 2D LIDAR. Deep Deterministic Policy Gradient (DDPG) is the DRL algorithm used for local navigation policy training and the simulation environment is built in Gazebo with Robot Operating System (ROS). After training the robot, the learned policy and waypoint selection are integrated together to develop a complete autonomous navigation and exploration system. This overall system not only mitigates the necessity of building the map of the environment but also gives similar or better performance in comparison to planning-based methods.