Online POMDP Planning for Vehicle Navigation in Densely Populated Area

Abstract

Technologies for autonomous vehicles have advanced dramatically in the last decade. However, it remains a challenge for these vehicles to navigate safely, reliably, and smoothly among pedestrians and other human-driven vehicles in densely populated urban centers. One major difficulty is the uncertainties arising from unknown pedestrian intentions, unexpected changes in the environment, sensor noise, and imperfect vehicle control. By leveraging a state-of-the-art online POMDP algorithm and constructing the model suitably to take advantage of its strengths, we demonstrate an successful application of the POMDP framework to autonomous vehicle navigation among pedestrians. In simulation, we analyze the performance of our POMDP approach in comparison with alternative models and algorithms in uncertain, dynamic environments. We further show that our autonomous golf-cart under the POMDP controller is able to navigate safely and smoothly in a dense crowd on the U-Town Plaza at our university.