Occupancy Reward-Driven Exploration with Deep Reinforcement Learning for Mobile Robot System

Abstract

This paper investigates the solution to a mobile-robot exploration problem following autonomous driving principles. The exploration task is formulated in this study as a process of building a map while a robot moves in an indoor environment beginning from full uncertainties. The sequence of robot decisions of how to move defines the strategy of the exploration that this paper aims to investigate, applying one of the Deep Reinforcement Learning methods, known as the Deep Deterministic Policy Gradient (DDPG) algorithm. A custom environment is created representing the mapping process with a map visualization, a robot model, and a reward function. The actor-critic network receives and sends input and output data, respectively, to the custom environment. The input is the data from the laser sensor, which is equipped on the robot. The output is the continuous actions of the robot in terms of linear and angular velocities. The training results of this study show the strengths and weaknesses of the DDPG algorithm for the robotic mapping problem. The implementation was developed in MATLAB platform using its corresponding toolboxes. A comparison with another exploration algorithm is also provided.