Motion Planning Using Feasible and Smooth Tree for Autonomous Driving

Weihuang Chen,Yuwei Chen,Shen'ao Wang,Fanjie Kong,Xuchong Zhang,Hongbin Sun

DOI: https://doi.org/10.1109/tvt.2023.3343694

IF: 6.8

2024-01-01

IEEE Transactions on Vehicular Technology

Abstract:Motion planning is a crucial and challenging step in autonomous driving navigation tasks. Since traditional motion planning methods largely rely on complex hand-crafted designs, recent research has shifted focus to simple and effective learning-based methods. However, most of the existing learning-based methods primarily employ end-to-end networks to directly generate trajectories or control commands, which may not satisfy vehicle kinematic constraints. These kinematically infeasible outputs cause autonomous driving vehicles deviating from the intended trajectories, significantly impacting both ride comfort and task completion. In this paper, we propose a new Feasible And Smooth Tree (FAST) to concentrate on the generation and utilization of trajectory tree. This trajectory tree structure possesses curvature continuity and kinematic feasibility, making it a cluster of feasible and smooth trajectories that can represent various intentions for traffic participants. Moreover, by employing the trajectory tree as the outputs for both motion planning and auxiliary trajectory prediction tasks, FAST effectively integrates the trajectory tree into a multi-task framework to improve its performance. We evaluate FAST in the closed-loop urban driving setting using the different scenarios provided by CARLA simulator. Experimental results demonstrate that FAST outperforms other learning-based methods while maintaining high efficiency.