Abstract
We propose GoalFlow, an end-to-end autonomous driving method for generatinghigh-quality multimodal trajectories. In autonomous driving scenarios, there israrely a single suitable trajectory. Recent methods have increasingly focusedon modeling multimodal trajectory distributions. However, they suffer fromtrajectory selection complexity and reduced trajectory quality due to hightrajectory divergence and inconsistencies between guidance and sceneinformation. To address these issues, we introduce GoalFlow, a novel methodthat effectively constrains the generative process to produce high-quality,multimodal trajectories. To resolve the trajectory divergence problem inherentin diffusion-based methods, GoalFlow constrains the generated trajectories byintroducing a goal point. GoalFlow establishes a novel scoring mechanism thatselects the most appropriate goal point from the candidate points based onscene information. Furthermore, GoalFlow employs an efficient generativemethod, Flow Matching, to generate multimodal trajectories, and incorporates arefined scoring mechanism to select the optimal trajectory from the candidates.Our experimental results, validated on the Navsim\cite{Dauner2024_navsim},demonstrate that GoalFlow achieves state-of-the-art performance, deliveringrobust multimodal trajectories for autonomous driving. GoalFlow achieved PDMSof 90.3, significantly surpassing other methods. Compared with otherdiffusion-policy-based methods, our approach requires only a single denoisingstep to obtain excellent performance. The code is available athttps://github.com/YvanYin/GoalFlow.