WAM-Flow: Parallel Coarse-to-Fine Motion Planning via Discrete Flow Matching for Autonomous Driving
1Fudan University 2Yinwang Intelligent Technology Co., Ltd
## 📰 News
- **`2026/02/01`**: 🎉🎉🎉 Release the pretrained models on [Huggingface](https://huggingface.co/fudan-generative-ai/WAM-Flow).
- **`2025/12/06`**: 🎉🎉🎉 Paper submitted on [Arxiv](https://arxiv.org/pdf/2512.06112).
## 📅️ Roadmap
| Status | Milestone | ETA |
| :----: | :----------------------------------------------------------------------------------------------------: | :--------: |
| ✅ | **[Release the SFT and inference code](https://github.com/fudan-generative-vision/WAM-Flow)** | 2025.12.19 |
| ✅ | **[Pretrained models on Huggingface](https://huggingface.co/fudan-generative-ai/WAM-Flow)** | 2026.02.01 |
| 🚀 | **[Release the evaluation code](https://huggingface.co/fudan-generative-ai/WAM-Flow)** | TBD |
| 🚀 | **[Release the RL code](https://github.com/fudan-generative-vision/WAM-Flow)** | TBD |
| 🚀 | **[Release the pre-processed training data](#training)** | TBD |
## 📸 Showcase
## 🏆 Qualitative Results on NAVSIM
### NAVSIM-v1 benchmark results
### NAVSIM-v2 benchmark results
## 🔧️ Framework
Our method takes as input a front-view image, a natural-language navigation command with a system prompt, and the ego-vehicle states, and outputs an 8-waypoint future trajectory spanning 4 seconds through parallel denoising. The model is first trained via supervised fine-tuning to learn accurate trajectory prediction. We then apply simulatorguided GRPO to further optimize closed-loop behavior. The GRPO reward function integrates safety constraints (collision avoidance, drivable-area compliance) with performance objectives (ego-progress, time-to-collision, comfort).
## Quick Start
### Installation
Clone the repo:
```sh
git clone https://github.com/fudan-generative-vision/WAM-Flow.git
cd WAM-Flow
```
Install dependencies:
```sh
conda create --name wam-flow python=3.10
conda activate wam-flow
pip install -r requirements.txt
```
### Model Download
Download models using huggingface-cli:
```sh
pip install "huggingface_hub[cli]"
huggingface-cli download fudan-generative-ai/WAM-Flow --local-dir ./pretrained_model/wam-flow
huggingface-cli download LucasJinWang/FUDOKI --local-dir ./pretrained_model/fudoki
```
### Inference
```sh
sh script/infer.sh
```
### Training
```bash
sh script/sft_debug.sh
```
## 📝 Citation
If you find our work useful for your research, please consider citing the paper:
```
@article{xu2025wam,
title={WAM-Flow: Parallel Coarse-to-Fine Motion Planning via Discrete Flow Matching for Autonomous Driving},
author={Xu, Yifang and Cui, Jiahao and Cai, Feipeng and Zhu, Zhihao and Shang, Hanlin and Luan, Shan and Xu, Mingwang and Zhang, Neng and Li, Yaoyi and Cai, Jia and others},
journal={arXiv preprint arXiv:2512.06112},
year={2025}
}
```
## ⚠️ Social Risks and Mitigations
The integration of Vision-Language-Action models into autonomous driving introduces ethical challenges, particularly regarding the opacity of neural decision-making and its impact on road safety. To mitigate these risks, it is imperative to implement explainable AI frameworks and robust safe protocols that ensure predictable vehicle behavior in long-tailed scenarios. Furthermore, addressing concerns over data privacy and public surveillance requires transparent data governance and rigorous de-identification practices. By prioritizing safety-critical alignment and ethical compliance, this research promotes the responsible development and deployment of VLA-based autonomous systems.
## 🤗 Acknowledgements
We gratefully acknowledge the contributors to the [Recogdrive](https://github.com/xiaomi-research/recogdrive), [Janus](https://github.com/deepseek-ai/Janus), [FUDOKI](https://github.com/fudoki-hku/FUDOKI) and [flow_matching](https://github.com/facebookresearch/flow_matching) repositories, whose commitment to open source has provided us with their excellent codebases and pretrained models.
