Deep Reinforcement Learning for Mapless Robot Navigation Systems

• Iure Rosa L. Oliveira UFV
• Alexandre S. Brandão UFV

Resumo

DRL has emerged as a promising approach for mobile robot navigation in unknown environments without a prior map. However, the performance of DRL methods for this task varies greatly, depending on the choice of algorithm, state representation, and training procedure. In this paper we explore various cutting-edge DRL algorithms, such as policy-, value-, and actor-critic-based approaches. Our results demonstrate the effectiveness of the ranging sensor approach, which achieves robust navigation policies capable of generalizing to unseen virtual environments with a high success rate. We combine Behavior Cloning with Imitation Learning to expedite the training process, leveraging expert demonstrations and reinforcement learning. Our methodology enables faster training while enhancing the learning efficiency and performance of the robot, obtaining better results in terms of crash and success rate, and being able to reach a cumulative reward of approximately 12000.

Palavras-chave: Mapless Navigation, Imitation Learning, TD3