HiTail: Hierarchical Neural Planner for Adaptive and Flexible Long-Tail Trajectory Planning  

A planner for autonomous vehicles must be capable of operating in diverse and complex real-world environments. However, learning-based planners often struggle with limited generalization due to the long-tail distribution in datasets. Moreover, the black-box nature of neural networks limits their interpretability and complicates the integration of explicit rules. In this letter, we propose a hierarchical neural trajectory planner that takes the bird's-eye view (BEV) rasters as input. The planner operates in two phases: in the first, spatial proposals are sampled from a policy generated from interpretable learned reward maps, and in the second, learnable temporal velocity profiles are assigned to the spatial proposals using clothoid curves. We conduct training and closed-loop simulation on the nuPlan dataset, and the results demonstrate that our proposed planner outperforms other learning-based methods, exhibiting superior adaptability in long-tail scenarios. Additionally, we explore the scalability of our planner in flexibly integrating manually defined rule sets.