Integrating ego-conditioned prediction and gap-driven motion planning for safe autonomous driving

Abstract

Efficient and safe trajectory planning in complex traffic remains challenging due to weak prediction-planning coupling in existing methods. This paper proposes a coupled prediction-planning framework that deeply integrates intention-aware multi-agent prediction with gap-driven trajectory optimization. We first design an ego-conditioned interaction prediction module that samples candidate ego plans and uses a transformer-based network to predict surrounding vehicle responses, generating consistent scenario plans. A gap generation algorithm is then introduced, explicitly incorporating lane-change dynamics and interactive influences. We further develop a gap transition cost function to jointly evaluate candidate gap reachability and efficiency. Finally, a coarse trajectory generated by combining the Intelligent Driver Model (IDM) and Dynamic Programming (DP) is refined via a constrained Quadratic Programming (QP) module to satisfy continuity, smoothness, and safety constraints. Experiments on the nuPlan dataset demonstrate that our approach significantly outperforms existing methods in planning stability, safety, and driving efficiency across multiple metrics.