Learning to Navigate in Human Crowds

Social robot navigation requires that the robot follows social norms while navigating towards its goal. Current algorithms model pedestrians as independent agents, making the problem computationally intractable and degrading overall performance in dense human crowds. In this work, we explore various ways to enhance the robot’s performance in environments with a higher number of pedestrians and aim to achieve well-behaved scaling. Specifically, we compare different approaches common in the literature such as state reduction, reward shaping, and curriculum learning. We find that the use of curriculum learning closely approximates optimal (human-like) behavior. The report serves as supplemental information, while the presentation includes the results of the experiments including animated visualizations.