Aprendizado por reforço profundo distribucional paralelo para navegação sem mapa de robôs móveis terrestres

Abstract

This paper presents a study of deep reinforcement learning techniques that uses parallel distributional actor-critic networks to navigate terrestrial mobile robots. The proposed approaches are developed taking into account only a couple of laser range findings, the relative position and angle of the mobile robot to the target as inputs to make a robot reach the desired goal in an environment. Was used a sim-to-real development structure, where the agents trained in a robot simulator are deployed in real scenarios to enhance the evaluation. The obtained results show that parallel distributional deep reinforcement learning algorithms, with continuous actions, are effective for the decision-make of a terrestrial robotic vehicle and outperform the classical behavior-based algorithm approach in terms of speed and navigation capability.