A hierarchical trajectory planning framework based on diffusion models for autonomous vehicles

Diffusion models have shown remarkable potential in modeling complex and multi-modal trajectory distributions for achieving human-like driving for autonomous vehicles. Contemporary learning-based methods such as imitation learning often suffer from the covariate shift issue in complex scenarios. Due to the limitations of explicitly integrating physical and dynamic constraints into network space, this paradigm also lacks robustness in long-tail traffic situations. In this paper, we propose a novel hierarchical progressive planning framework (HPP) for trajectory generation, which introduces a two-stage strategy for coarse-to-fine planning, reducing the modeling of redundant information. In the first stage, a diffusion model is employed to generate sparse key points that capture high-level motion intentions while avoiding redundant motion details and slow denoising convergence. To incorporate vehicle physical constraints, in the second stage, we design the MPC-based refinement module to expand generated key points into continuous and drivable trajectories. By combining the expressive generative capability of diffusion models with the physical interpretability of model-based optimization, the proposed framework achieves safe, robust, and adaptable trajectory planning across diverse driving scenarios. We show that the proposed framework achieves competitive performance on nuPlan, a large-scale planning benchmark for autonomous driving. Our approach outperforms several representative rule-based, learning-based, and hybrid planners. Additionally, validation in the MetaDrive simulator illustrates the framework’s capability in realistic driving scenarios. The source code is available at https://github.com/HITXCI/xy_hpp .