Deep-learned pedestrian avoidance policy for robot navigation

Abstract

Being able to avoid obstacles and pedestrians in particular, is essential for robots to function in dynamic environments. In contrast with model based methods utilizing primarily computer vision, this project proposed a learning-based approach. Two deep neural networks were trained with images labeled with movement decisions, for pedestrian avoidance and path following tasks, where computer vision labeling and camera order labeling techniques were applied respectively. Together with ultrasonic sensors for static obstacle avoidance, the three components cooperatively contributed to our robot navigation policy. Comparing to existing experiments and research with sophisticated sensors, for instance LIDAR, the project utilized a monocular RGB camera and exploited its capability. Focusing on pedestrian avoidance, the project explores limitations and advantages of deep neural network method. A robot integrating above components was built, and performed satisfactorily in relevant test runs.