Safe and Interactive Integration of Prediction and Planning for Autonomous Vehicles with Residual Chain Loss

The Integration of Prediction and Planning (IPP) methodology enhances interrelationships between autonomous vehicle modules while improving system interpretability. However, safety, interaction, and performance evaluation present significant challenges that may lead to safety risks. This paper introduces IPP* (Safe and Interactive Integration of Prediction and Planning with Residual Chain Loss), a framework that addresses these challenges through three key components. First, a safety judgment module employs rule-based methods to assess prediction trustworthiness and ensure stability before planning. Second, a Nash Equilibrium-based approach facilitates iterative refinement between prediction and planning modules, creating a game-theoretic framework that enables mutual learning and system optimization. Third, a Residual Chain Loss Function addresses covariate bias in imitation learning by analyzing disparities between predicted and planned trajectories. The results indicate that IPP* achieves an Average Displacement Error (ADE) of 0.74 m and a Final Displacement Error (FDE) of 0.89 m, and delivers an average L2 distance (L2) of 0.44 m with an average Collision Rate (CR) of 0.27%. Ablation results further verify the complementary contributions of the safety judgment module, the game-based interaction module, and the residual chain loss function, while sensitivity analyses over key thresholds and loss weights indicate stable behavior and clarify safety–performance trade-offs.