Socially Aware Robot Navigation Using Deep Reinforcement Learning

In this study, we propose a socially aware navigation framework for mobile service robots in dynamic human environments using a deep reinforcement learning algorithm. The primary idea of the proposed algorithm is to incorporate obstacles information (position and motion), human states (human position, human motion), social interactions (human group, human-object interaction), and social rules, e.g. minimum distances from the robot to regular obstacles, individuals, and human groups into the deep reinforcement learning model of a mobile robot. We then distribute the mobile robot into a dynamic social environment and let the mobile robot automatically learn to adapt to an embedded environment by its experiences gained through trial-and-error social interactions with the surrounding humans and objects. When the learning phase is completed, the mobile robot is able to navigate autonomously in the social environment while guaranteeing human safety and comfort with its socially acceptable behaviours.