Aprendizado por reforço profundo para navegação de robôs móveis

Abstract

This work presents a study of deep reinforcement learning techniques that uses the Deep Deterministic Policy Gradient network and the Soft Actor-Critic network for application in navigation of mobile robots. In order for the robot to arrive to a target on a map, the networks have as input: 10 laser range findings, the previous linear and angular velocity, and relative position and angle of the mobile robot to the target. As output, the network has the linear and angular velocity. From the results analysis, it is possible to conclude that the deep reinforcement learning algorithms, with continuous actions, are effective for the decision-making of robotic vehicles and the Soft Actor-Critic networks present superior results, in less episodes, than the Deep Deterministic Policy Gradient. However, it is necessary to create a good reward function for the intelligent agent to accomplish its objectives. In order to show the performance of the Deep Reinforcement Learning Algorithms, they were applied in experiments with a simulated robot in three different environments and in a real robot in two environments.