Modeling Socially Normative Navigation Behaviors from Demonstrations with Inverse Reinforcement Learning

摘要

Navigation in an efficient and socially normative manner is essential for the robot to operate in human populated environments. Traditional methods treat the pedestrians as dynamic obstacles and design a manual cost function for collision avoidance, but neglect social norms in navigation and do not generalize well to new environments. In this paper, we propose a mixture model to capture the human navigation behaviors in terms of the features of the continuous trajectories and discrete navigation decisions, such as passing on the left or right. The lower level of the model aims to generate socially normative trajectories. To this end, we extend inverse reinforcement learning (IRL) framework to a motion planner called Timed Elastic Band to learn from demonstrations. The upper level comprises a discrete distribution over the homotopy classes of the trajectories. IRL algorithm is employed to find the parameters of distribution that match demonstrations best. Experiments demonstrate that our learning algorithm has the capacity to recover the human navigation behaviors that respect social norms, which makes our approach outperform state-of-the-art methods in social navigation scenarios.