Stranger Danger! Identifying and Avoiding Unpredictable Pedestrians in RL-based Social Robot Navigation

Abstract-- Reinforcement learning (RL) methods for social robot navigation show great success navigating robots through large crowds of people, but the performance of these learningbased methods tends to degrade in particularly challenging or unfamiliar situations due to the models' dependency on representative training data. Once an RL policy recognizes that it is in an unfamiliar I. Thus, it must distinguish between pedestrians spaces is yet to be achieved. Robots that interact with people who exhibit predictable behavior seen during training and are expected to navigate in a way that is predictable and those whose behavior is unpredictable, and learn to navigate unobtrusive, maintaining both the safety and comfort of efficiently around predictable pedestrians while maintaining surrounding people [1]. This would allow such is seeing a growing number of RL-based approaches that policies to generalize well to arbitrary pedestrian behavior. These RL-based into an existing RL policy called SARL [6] by (1) modifying approaches have achieved great success in enabling effective the training process to systematically inject significant deviations navigation around large crowds of people, outperforming into a model of pedestrian behavior, (2) augmenting traditional approaches [3]-[6]. However, the performance of the observation space of the value network algorithm to RL policies is contingent on having representative training recognize and quantify deviations of pedestrian behavior data, so these policies are sensitive to differences in from the assumed model, and (3) adding a term to the reward pedestrian behavior seen during deployment versus training function to encourage caution toward progressively more (a problem generally referred to as domain shift).