Learning to Navigate for Mobile Robot with Continual Reinforcement Learning

Abstract

Autonomous navigation of mobile ground robots is a promising research topic due to its extensive applications. Existing works are difficult to deal with obstacle-cluttered problems and often do not generalize well because of the differences between training scenarios and practical environments. In this paper, we present an end-to-end motion planning model that has the ability to navigate the mobile robot to the given destination safely in an unknown environment without any prior knowledge. Specifically, with the application of continual reinforcement learning, our approach is able to learn subtasks in a sequential fashion by decomposing the arbitrarily complex navigating task. The experiments show that our approach is more effective and can be directly transferable to other unseen environments.