Kostas Bekris, Safe and Fast Online Replanning of Trajectories for Individual Vehicles and Vehicular Networks

Prototype systems have been recently developed that allow us to imagine the deployment of autonomous vehicles in real life. The primary requirement for such vehicles is the capability to autonomously plan and coordinate their motion. In many realistic tasks, however, there is only partial information available for the environment or the environment itself is dynamic. Such problems require interleaving sensing, planning and execution. This talk focuses on the planning module of this problem. An important challenge in this planning task is that realistic vehicles have to respect complex constraints in their motion. For example, an automobile cannot stop or rotate instantaneously. The first part of this talk describes an algorithm that is designed to be called frequently and has finite time to replan a trajectory for a vehicle given new sensory input. The planner respects the vehicle's constraints and the produced trajectory can be followed in practice. The planner also provides safety guarantees given these constraints under limited computation time. Safety in this context means that the vehicle avoids collisions. The second part of the talk focuses on the problem of coordinating the motion of multiple vehicles that form communication networks. A fully distributed, message-passing algorithm is presented that achieves similar safety guarantees for teams of vehicles and retains network connectivity. Experimental results on a distributed simulator show that both techniques achieve favorable computational performance compared to alternatives.