PEP: Policy-Embedded Trajectory Planning for Autonomous Driving

Autonomous driving demands proficient trajectory planning to ensure safety and comfort. This letter introduces Policy-Embedded Planner (PEP), a novel framework that enhances closed-loop performance of imitation learning (IL) based planners by embedding a neural policy for sequential ego pose generation, leveraging predicted trajectories of traffic agents. PEP addresses the challenges of distribution shift and causal confusion by decomposing multi-step planning into single-step policy rollouts, applying a coordinate transformation technique to simplify training. PEP allows for the parallel generation of multi-modal candidate trajectories and incorporates both neural and rule-based scoring functions for trajectory selection. To mitigate the negative effects of prediction error on closed-loop performance, we propose an information-mixing mechanism that alternates the utilization of traffic agents' predicted and ground-truth information during training. Experimental validations on nuPlan benchmark highlight PEP's superiority over IL- and rule-based state-of-the-art methods.