A robust motion planning approach for autonomous driving in urban areas

Abstract

This thesis presents an improved sampling-based motion planning algorithm, Robust RRT, that is designed specifically for large robotic vehicles and uncertain, dynamic environments. Five main extensions have been made to the original RRT algorithm to improve performance in this type of applications. The closed-loop system is used for state propagation, enabling easy handling of complex, nonlinear, and unstable dynamics. The environment structure is exploited during the sampling process, increasing the probability that a given sample will be reachable. Efficient heuristics are employed in the expansion of the tree and a risk penalty is incorporated to capture uncertainty in the environment and keep the vehicle a safe distance away from hazards. The safety of the vehicle is guaranteed with the assumption of no unexpected changes in the environment, which is achieved by requiring that every trajectory sent for execution ends in a state with the vehicle stopped. Finally, risk evaluation follows a lazy evaluation strategy, allowing the algorithm to spend most of the computation time in the expansion step. The effectiveness of the Robust RRT algorithm for planning in an urban environment is demonstrated through numerous simulated scenarios and real data corresponding to its implementation in MIT's robotic vehicle that competed in the DARPA Urban Challenge.