Autonomous Driving Planning Based on Interaction-Aware Enhancement and Motion-Collaborative Query

Achieving human-like driving behavior in complex and dynamic environments has been a critical objective for autonomous driving. Although learning-based planning methods have made significant progress, existing models often overlook explicit modeling of the correlation between lateral and longitudinal motions, and their scene encoding rely primarily on historical observations of surrounding agents, thereby underutilizing future interaction information and limiting the learning of complex multimodal human driving behaviors. To address these limitations, a trajectory decoder with motion-collaborative queries was proposed to jointly model lateral-longitudinal motion dependencies so as to improve planning coordination and diversity. Moreover, an interaction-aware enhancement mechanism augmenting scene encoding with predicted trajectories of surrounding agents as additional future context was designed to enhance the capability to perceive potential dynamic interactions. Experiments were conducted on the nuPlan dataset. The results demonstrate that the proposed method effectively captures multimodal human driving behavior, enables flexible trajectory planning, and improves closed-loop planning performance.