Design Planning Framework Based on Bidirectional Refinement Interaction for Autonomous Driving

Classical autonomous driving pipelines typically decompose the task into three sequential stages: perception, prediction, and planning, with each module optimized independently. In contrast, end-to-end approaches, which jointly train all modules within a unified pipeline, have gained increasing attention in both academia and industry due to their promising planning performance. However, autonomous driving in complex and uncertain scenarios requires not only accurate perception but also a strong capacity to model interaction patterns among agents. A key intuition is that driving behavior is inherently dual: while the ego vehicle plans based on the behavior of surrounding agents, those agents simultaneously respond to the ego vehicle's actions. Despite this, most existing end-to-end methods primarily rely on modeling statistical correlations in the scene via a one-way attention paradigm, often overlooking the inherently reciprocal nature of agent interactions. To address this limitation, we propose the Bidirectional Refinement framework, which explicitly models mutual interactions during planning. Specifically, we design a stacked graph cross-attention module to effectively capture bidirectional interactions between the ego vehicle and surrounding vehicles. The scene is first transformed into instance-level representations, after which a bidirectional refinement module is applied to model reciprocal attention among agents. In addition, we incorporate a sequence queue to encode temporal context, followed by a dedicated planning decoder to generate driving actions. Extensive experiments demonstrate that our framework achieves superior performance in planning, mapping, and tracking tasks, as shown in Fig. 1, highlighting the advantages of jointly modeling perception and planning through reciprocal interaction mechanisms.