ICP-AD: Interactive-Cognition Planning for Autonomous Driving

Trajectory planning is critical for ensuring the safety and efficiency of autonomous vehicles. While traditional planning methods excel in structured environments by avoiding static and dynamic obstacles, they often struggle in dense urban scenarios due to inadequate modeling of traffic participants' intentions and interactions. To address the challenges, we propose ICP-AD, an end-to-end trajectory planning algorithm driven by multi-agent behavioral interaction modeling and intention cognition. Our approach preserves agent identities through spatial cross-attention while capturing and learning multi-agent behaviors and intentions via temporal self-attention. By constructing instance-level dense scene representations coupled with dynamic occupancy prediction, our approach enables robust trajectory planning optimized for both safety and efficiency. Moreover, the proposed end-to-end trajectory planning framework achieves global joint optimization from perception to planning, effectively reducing information loss and suboptimal performance inherent in conventional modular architectures. We conducted experiments on the challenging real-world driving scenario of the nuScenes dataset. Our method demonstrates competitive performance across multiple metrics, validating its robust multi-object tracking and prediction capabilities in complex environments and efficiency in safety-oriented planning.