Map-less End-to-end Navigation of Mobile Robots via Deep Reinforcement Learning

Abstract

A deep reinforcement learning model with a unique long-term memory capability is proposed in this paper, which addresses the map-less navigation of mobile robots in dynamic environments. Based on the recurrent neural network, the proposed model takes continuous historical states as input and better handles dynamic obstacles. Furthermore, a novel reward function is designed to ensure smooth navigation trajectories and satisfactory navigation results in dynamic environments. The proposed approach is evaluated on the Gazebo simulation platform, and higher navigation success rates in dynamic environments are achieved.