Motion planning and decision-making for autonomous systems : from quadrotors to autonomous vehicles
Wenchao Ding · 2020
Autonomous systems have been widely deployed in various field applications, such as inspection, exploration, and aerial videography using micro aerial vehicles (MAVs), and highway pilot (HWP), rush-hour pilot (RHP), robo-taxi in field of intelligent vehicles. For accomplishing such missions in challenging environments, navigation functionality is essential. In this thesis, we study practical and efficient planning methods for autonomous systems, which can achieve full autonomy in real-world complex environments. Our study covers both low-level motion planning problems and high-level decision-making (i.e., behavior planning) problems, targeting for safe and smooth autonomous navigation with minimum user intervention. We investigate the navigation problem for two different platforms, namely, quadrotors for MAVs, and autonomous vehicles (AVs), which have significant practical impacts while illustrating different focuses and functionalities. We start with a novel kinodynamic motion planning method for quadrotors, which achieves safe and efficient replanning in unknown cluttered 3-D environments. Leveraging the knowledge of motion planning for quadrotors, we further study the planning problem for autonomous vehicles in highly dynamic environments. Different from the motion planning for quadrotors, planning for AVs requires reasoning about other dynamic traffic participants and social compliance. To this end, a novel behavior prediction method is proposed to understand the behaviors of other traffic participants in the real world. The behavior prediction is then tightly incorporated into our proposed decision-making framework which generates robust decisions efficiently. The output decision is subsequently utilized by our novel motion planning module to generate a smooth and safe trajectory for closed-loop execution. We assemble all the individual modules into a complete and robust planning system which is validated in real-world dense city traffic.