A real-time search-based motion planning framework with risk assessment for urban environments
Kailin Tong · 2023
A reliable autonomous driving system should be capable of handling driving tasks in real-time while ensuring user safety. However, many search-based planners suffer from high computation time, which makes them unsuitable for embedded applications. Some simple methods, like Roll-out from OpenPlan-ner, achieve high computation efficiency, but lack robustness and might be unsafe in certain conditions. To solve this problem we proposed an efficient search-based motion planning framework incorporating risk evaluation. Firstly, based on necessary inputs from perception, map, and global planning stack, a 3D grid map is generated to capture environmental semantics for our extended hybrid A* search algorithm. We proposed a heuristic function consisting of risk evaluation, which not only reduces A* search time but also plans a safer trajectory. A two-step optimization process refines the results from the hybrid A* search into a smooth and continuous trajectory. Our proposed framework was tested using realistic simulation scenarios within the open-source Autoware autonomous driving stack, showing superior responses and more robustness than the baseline OpenPlanner. Additionally, the proposed hybrid A* search framework shows high computation efficiency. It has a worst-case run-time of 55 ms for a 12-second planning horizon and a planning horizon of 200 meters, allowing for real-time application.