Paper | Weight | ProjectPage

GoalFlow: Goal-Driven Flow Matching for Multimodal Trajectories Generation in End-to-End Autonomous Driving
Zebin Xing^1,2*, Xingyu Zhang^2*, Yang Hu^1,2, Bo Jiang^4,2, Tong He⁵, Qian Zhang², Xiaoxiao Long³, Wei Yin^2✝
1 University of Chinese Academy of Sciences, ² Horizon Robotics, ³ Nanjing University, ⁴ Huazhong University of Science & Technology, ³ Shanghai AI Laboratory

Computer Vision and Pattern Recognition (CVPR), 2025

This is the official repo of 'GoalFlow: Goal-Driven Flow Matching for Multimodal Trajectories Generation in End-to-End Autonomous Driving (CVPR 2025)'. GoalFlow achieved PDMS of 90.3, significantly surpassing other baselines. Compared with other diffusion-policy-based methods, our approach requires only a single denoising step to obtain excellent performance.

News

• 20 Mar, 2025: We released our paper on arXiv. Code is coming soon.
• 27 Feb, 2025: GoalFlow was accepted at CVPR !

To Do

• Code for goal point construction module
• Goal Point scorer and cluster vocabulary cache
• Weights of model
• Code for validation
• Tutorial for installation
• Initial repo & main paper

Introduction

In autonomous driving, multiple optimal trajectories exist, like overtaking or following. (1) Traditional methods efficiently generate safe trajectories but struggle with multimodal ones. (2) Generative methods like diffusion models capture multimodal distributions but require heavy hardware and prior information. We propose GoalFlow, a goal-point-based method that guides trajectory planning. With a map-free evaluation and an efficient diffusion variant, Flow Matching, we reduce inference steps, achieving superior performance with just one denoising step.

Visualization

Comparison with Other Methods

❌ indicates that the trajectory results in a collision or goes beyond the drivable area, while ✅ represents a safe trajectory. The orange points are optimal goal points evaluated by the Goal Constructor, while the blue and yellow points correspond to samples from the vocabulary.

Driving Vedios

Driving Vedios generated by GoalFlow.

Goal Point Distribution

From top to down, they are respectively the distributions of DAC, distance, and the final score. The points with warmer color have higher score.

Results

Planning results on the proposed NAVSIM Test benchmark. Please refer to the paper for more details.

Method	SNC ↑	SDAC ↑	STTC ↑	SCF ↑	SEP ↑	SPDM ↑
Constant Velocity	68.0	57.8	50.0	100	19.4	20.6
Ego Status MLP	93.0	77.3	83.6	100	62.8	65.6
LTF	97.4	92.8	92.4	100	79.0	83.8
TransFuser	97.7	92.8	92.8	100	79.2	84.0
UniAD	97.8	91.9	92.9	100	78.8	83.4
PARA-Drive	97.9	92.4	93.0	99.8	79.3	84.0
GoalFlow (Ours)	98.4	98.3	94.6	100	85.0	90.3
Human‡	100	100	100	99.9	87.5	94.8

Getting started

• Download Datasets of NAVSIM
• Preparation of GoalFlow Environment
• Evaluation
• Training

Contact

If you have any questions or suggestions, please feel free to open an issue or contact us (xzebin@bupt.edu.cn).

Acknowledgement

1. We have gained valuable insights from Hydra-MDP, which provided many inspiring ideas referenced in our work.

2. We referred to an excellent GitHub project, tuplan garage, and incorporated aspects of its page design.

3. GoalFlow is also greatly inspired by the following outstanding contributions to the open-source community:

NAVSIM | TransFuser | Diffusion-ES | VAD-v2

Citation

If you find GoalFlow useful, please consider giving us a star 🌟 and citing our paper with the following BibTeX entry.

@article{xing2025goalflow,
	  title={GoalFlow: Goal-Driven Flow Matching for Multimodal Trajectories Generation in End-to-End Autonomous Driving},
	  author={Xing, Zebin and Zhang, Xingyu and Hu, Yang and Jiang, Bo and He, Tong and Zhang, Qian and Long, Xiaoxiao and Yin, Wei},
	  journal={arXiv preprint arXiv:2503.05689},
	  year={2025}}

(back to top)