Enhancing Closed-Loop Performance in Learning-Based Vehicle Motion Planning by Integrating Rule-Based Insights

Abstract

This letter introduces an innovative vehicle motion planning method that leverages the integration of rule-based insights to significantly improve closed-loop performance within a learning-based framework. We first employ rule-based methods to heuristically search and generate a diverse set of trajectory proposals. Then, we filter these trajectories using prediction and kinematic constraints, and subsequently select the highest-scoring trajectory based on metrics such as similarity to ground truth trajectories. We utilize a model to learn the mapping from observations to selected trajectories. Finally, we optimize the model's output trajectories using the lane centerlines. Validated through closed-loop simulations in Highway-Env and nuPlan, our method demonstrates a higher success rate and reduced computation time relative to the rule-based method we used, achieving competitive performance with the state-of-the-art method in nuPlan.