FlashPlanner: Accelerating Diffusion-based Planner for Autonomous Driving via Globally Consistent Velocity Field and Redundancy Reduction

Qifeng Li, Yubing Gao, Xiaosong Jia, Zhiliu Liu, zuhao ge, Zuxuan Wu, Yu-Gang Jiang, Junchi Yan

Keywords: autonomous driving, flow matching, generative model

Abstract: Standard diffusion and flow matching approaches have recently been explored as imitation-based planners for autonomous driving due to their ability to produce multi-modal trajectories with high fidelity. However, these methods still suffer from limitations, e.g., low efficiency and reliance on post-processing. These issues are alleviated through practices from conventional imitation-based methods, but the principles of well-designed diffusion-based planners are still underexplored. In this paper, we propose FlashPlanner, a flow-matching-based planner for online planning of autonomous driving. FlashPlanner introduces a novel globally consistent velocity field as the training objective, which frames flow matching to model instantaneous dynamics in a consistent velocity field. This training objective manages to unleash the potential of diffusion-based planners and enables stable one-step generation of high-quality trajectories in closed-loop planning. Moreover, we systematically analyze the existing design choices of diffusion-based methods and prune inherent redundancy, which further accelerates the diffusion-based planning. It achieves state-of-the-art performance on the closed-loop nuPlan benchmark and delivers 12× faster inference (166FPS) compared to the existing best baseline (13FPS). We will open-source our project.

Primary Area: applications to robotics, autonomy, planning

Submission Number: 3145