End-to-End Mobile Robot Navigation using a Residual Deep Reinforcement Learning in Dynamic Human Environments

Abstract

Safe navigation through human crowds is key to enabling practical mobility ubiquitously. The Deep Reinforcement Learning (DRL) and the End-to-End (E2E) approaches to goal-oriented robot navigation have the potential to render policies able to tackle localization, path planning, obstacle avoidance, and adaptation to change in unison. In this paper, we report an architecture based on convolutional units and residual blocks being able to enhance adaptability to unseen and dynamic human environments. In particular, our scheme outperformed the state-of-the-art baselines SOADRL and NAVREP by about 13% and 18% on average success rate, respectively, throughout 27 unseen and dynamic navigation instances. Furthermore, our approach avoids the explicit encoding of positions and trajectories of moving humans compared to the standard models. Our results show the potential to render adaptive and generalizable policies for unknown and dynamic human environments.