Robust motion planning for autonomous vehicles based on environment and uncertainty-aware reachability prediction

Planning and navigation in real-time traffic is challenging, since the driving environment (e.g., road network and infrastructure) is complex and the accurate prediction of surrounding vehicles is hard. To address this, this paper proposes an environment and uncertainty-aware robust motion-planning strategy. The method achieves environment awareness by considering road-geometry constraints in the reachability prediction of surrounding vehicles, and uncertainty awareness by online learning the intended control set of the surrounding vehicles. By integrating this dual awareness, the method effectively predicts the forward reachability of surrounding vehicles, which is applied in the design of collision-avoidance constraints in the optimal motion-planning strategy. The motion planner then computes the reference trajectory for the autonomous ego vehicle using a receding-horizon approach to fit variations in the dynamic traffic. The effectiveness of the strategy is demonstrated through simulations in roundabout scenarios by comparing with alternative methods, further validated in a traffic scenario from a dataset recorded in the real world. Additionally, the feasibility of real-time implementation is verified through hardware experiments using car-like mobile robots.