SANG: Socially Aware Navigation Between Groups

Abstract Although robots are able to navigate both static and dynamic environments, they lack the desired level of social awareness to move among humans. Previous work has explored solutions combining social force models and reinforcement learning to address this problem, but socially aware navigation has not been truly achieved. Furthermore, these works are hardly comparable due to the wide variety of metrics used for evaluation. Our work, Socially Aware Navigation Between Groups (SANG), aims to improve socially aware navigation by introducing a real-to-sim simulation for the training and testing of a reinforcement learning algorithm that has a more realistic group-based social reward calculation compared to previous works. To provide a thorough and unified evaluation practice, we propose a wide range of objective and subjective metrics including the use of a Navigation Turing Test. Based on our evaluation, the SANG algorithm achieves human-level smoothness scores, and better Social Information Processing scores as compared with the state-of-the-art. On average it was judged to be more human-like 43% of the time in a Navigation Turing Test, significantly improving the st