BQA-Plan: Enhanced Imitation Learning-Based Planning via Behavior Quality Awareness

Imitation Learning (IL)-based planning has garnered widespread attention due to its potential in modeling human-like driving behaviors. However, constrained by the open-loop training paradigm, it is susceptible to causal confusion and distributional shift during closed-loop execution, leading to performance degradation. To address these challenges, we propose BQA-Plan, a method designed to enhance the closed-loop performance of IL-based planning through behavior quality awareness. Specifically, we develop a learnable behavior quality awareness module that comprehensively evaluates trajectory dynamics, interaction features, and safety features to assess behavior quality in real time. Furthermore, we introduce a dynamic feedback mechanism based on behavior quality deviation, which adaptively adjusts the training loss according to the quality deviation between the generated and expert trajectories, guiding the model to prioritize high-quality trajectory patterns while suppressing potential abnormal behaviors. We conduct systematic evaluations on the Test14-Random and Test14-Hard closed-loop simulation benchmarks from nuPlan. Experimental results demonstrate that BQA-Plan significantly outperforms representative learning-based methods in closed-loop simulation tests, achieving superior performance across multiple key metrics and exhibiting promising potential for real-world applications.