Mobile Robot Navigation Using Learning-Based Method Based on Predictive State Representation in a Dynamic Environment

Abstract

Mobile robot navigation in a dynamic environment with pedestrians is essential for service robots operating in a living environment. Accordingly, the robot needs to understand and predict the behavior of pedestrians. However, predicting pedestrian behavior in advance is difficult because human behavior may be affected by factors that cannot be directly observed or modeled in advance, such as intentions and environmental influences. In addition, pedestrian behavior may be affected by the behavior of the robot.In this study, we apply a deep reinforcement learning method based on a novel predictive state representation (PSR) model to mobile robot navigation for realizing a navigation method considering the changes in pedestrian behavior caused by robot actions and other pedestrians. In addition, we propose two methods for integrating the states of the PSRs corresponding to each pedestrian and evaluate these methods in situations where the number of pedestrians differs between learning and testing.