Human-aware subgoal generation in crowded indoor environments

With mobile robots becoming more prevalent in our daily lives, it is crucial that these robots navigate in a safe and socially-aware manner. While recent works have shown promising results by using Deep Reinforcement Learning (DRL) techniques to learn socially-aware navigation policies, most approaches are limited to local, short-term navigation. For more complex settings, DRL approaches rely on subgoals often computed using traditional path planners. However, these planners are not necessarily suitable for social navigation since they rarely consider the pedestrians in the scene and often disregard the long term cost of a path or the pedestrian dynamics. In this paper, we present an alternative global planner that uses a learnt local cost predictor to generate subgoal guidance to help a DRL robot to make progress towards its goal while also taking into account the pedestrians in the environment. We evaluate the proposed approach in simulation. We consider several environments of varying complexity as well as different pedestrian behaviours. Our results show that a DRL robot using the proposed planner is less likely to collide with pedestrians and exhibits improved social awareness when compared to a baseline approach using traditional path planner methods.