RAP: Role-Aware Joint Prediction and Planning in Autonomous Driving

Xiaolong Tang, Meina Kan, Shiguang Shan, Xilin Chen

Abstract: Planning the trajectory for an autonomous vehicle (AV) requires considering both the AV’s objectives and the behaviors of surrounding agents, which interact in a dynamic process. Joint prediction and planning (JPP) methods have been proposed to model this bidirectional interaction and generate compatible joint futures. However, existing methods often overlook the fundamental differences between the two tasks and treat them almost identically, thus limiting the effect of planning. In this work, we revisit JPP from a role-aware perspective and identify three inherent asymmetries between prediction and planning: information, objective, and feedback. Based on these insights, we propose the Role-Aware Joint Prediction and Planning (RAP) framework, which explicitly differentiates the treatment of the ego vehicle and surrounding agents through three lightweight asymmetric designs. Specifically, RAP introduces a route–identity token pairing mechanism for route-aware yet agent-agnostic conditioning, ego-only vectorized auxiliary losses for safety-constrained planning, and dropout and perturbation applied to the ego’s history and kinematic state to improve closed-loop robustness. Comprehensive experiments on the nuPlan benchmark demonstrate that RAP achieves state-of-the-art performance on both open-loop and closed-loop metrics, validating the effectiveness of asymmetric modeling in unified prediction and planning.

External IDs:doi:10.1109/lra.2025.3641815