PlanCiLQR: Hierarchical trajectory planning for autonomous driving using constrained iterative linear quadratic regulator

Trajectory planning equips autonomous vehicles with advanced cognitive capabilities, enabling holistic environmental understanding in dynamic scenarios. However, traditional approaches are often validated only in limited scenarios, lacking rigorous large-scale benchmarking under standardized metrics. While learning-based approaches facilitate fair comparisons, they are typically compared against other learning models and have not yet consistently surpassed well-established traditional planners in complex real-world settings. To address these limitations, this paper introduces PlanCiLQR, a hierarchical optimization-based framework that decouples the planning process into a behavior layer and a motion layer: the former handles environmental inequality constraints, while the latter refines trajectories using a Constrained Iterative Linear Quadratic Regulator (CiLQR) within a time-indexed convex spatial corridor. This formulation enables efficient resolution of convex optimization problems, ensures dynamic feasibility, and compensates for system delays. By integrating a robust evaluation framework, PlanCiLQR supports objective performance assessment across diverse large-scale scenarios. Extensive experiments on the real-world nuPlan dataset demonstrate that PlanCiLQR achieves state-of-the-art performance among traditional planners in closed-loop evaluation, while maintaining real-time computational efficiency and high reliability.