[🎉CVPR 2026] MeanFuser

Fast One-Step Multi-Modal Trajectory Generation and Adaptive Reconstruction via MeanFlow for End-to-End Autonomous Driving

Junli Wang, Xueyi Liu, Yinan Zheng, Zebing Xing, Pengfei Li, Guang Li,

Kun Ma, Guang Chen, Hangjun Ye, Zhongpu Xia, Long Chen, Qichao Zhang📧

📧 indicates corresponding authors.
SKL-MAIS, CASIA   |   Xiaomi EV   |   AIR, Tsinghua University

---

📢 News

• [2026/3/20] We released code and checkpoints.
• [2026/2/25] We released our paper on arXiv.
• [2026/2/21] 🎉 Accepted to CVPR 2026.

📌 Table of Contents

• 📋 TODO List
• 🏛️ Model Zoo
• 🎯 Getting Started
• 📦 Data Preparation
  • Download Dataset
  • Set Up Configuration
  • Cache the Dataset
• ⚙️ Training and Evaluation
  • Evaluation
  • Training
  • Visualization
• ❤️ Acknowledgements

📋 TODO List

• HUGSIM code release (Apr. 2026).
• NAVSIMv2 navtest code release (Apr. 2026).
• Checkpoints release (Mar. 2026).
• Code release (Mar. 2026).
• Paper release (Feb. 2026).

🏛️ Model Zoo

Method	Backbone	Benchmark	PDMS	Weight Download
MeanFuser	ResNet-34	NAVSIM	89.0	Google Drive
MeanFuser + BeyondDrive	ResNet-34	NAVSIM	90.3	Google Drive
MeanFuser	ResNet-34	HUGSIM	-	Google Drive

🎯 Getting Started

1. Clone MeanFuser Repo

git clone https://github.com/wjl2244/MeanFuser.git
cd MeanFuser

2. Create Environment

conda create -n meanfuser python=3.9 -y
conda activate meanfuser
pip install -e .

📦 Data Preparation

NOTE: Please review and agree to the LICENSE file file before downloading the data.

1. Download Dataset

a. Download via NAVSIM offical installation.

Follow the instructions in the NAVSIM installation guide to download the dataset.

b. Download via Hugging Face

Alternatively, you can download the dataset using Hugging Face with the following commands:

export HF_ENDPOINT="https://huggingface.co"
# export HF_ENDPOINT="http://hf-mirror.com"  # Uncomment this line if you are in China

# Install the huggingface_hub tool
pip install -U "huggingface_hub"

# Download the OpenScene dataset
hf download --repo-type dataset OpenDriveLab/OpenScene --local-dir ./navsim_dataset/ --include "openscene-v1.1/*"

# Download the map data
cd download && ./download_maps.sh

2. Set Up Configuration

Move the download data to create the following structure.

navsim_workspace/
├── MeanFuser/
├── dataset/
│    ├── maps/
│    ├── navsim_logs/
│    │   ├── test/
│    │   ├── trainval/
│    ├── sensor_blobs/
│    │   ├── test/
│    │   ├── trainval/
└── cache/
     ├── navtest_v1_metric_cache/
     └── traintest_v1_cache/

3. Cache the Dataset

We provide a script to cache the dataset and metrics.

cd MeanFuser

# Cache the dataset. (navtrain and navtest)
python scripts/evaluation/run_dataset_cache.sh

# Cache the metric.
python scripts/evaluation/run_metric_cache.sh

⚙️ Training and Evaluation

1. Evaluation

Please download the pre-trained checkpoints from here and place them in the navsim_workspace/MeanFuser/exp/ directory.

cd MeanFuser

bash scripts/evaluation/run_meanfuser_evaluation.sh

2. Training

Please download the ResNet-34 pretrained weights from here. After downloading, update the corresponding path in the configuration file:navsim_workspace/MeanFuser/navsim/agents/meanfuser/meanfuser_config.py

cd MeanFuser

bash scripts/training/run_meanfuser_training.sh

3. Visualization

We provide a script to visualize the model's planned trajectory.

export NAVSIM_WORKSPACE="xxx/navsim_workspace"
python MeanFuser/tools/visualization_navtest_scenes.py

❤️ Acknowledgements

We acknowledge all the open-source contributors for the following projects to make this work possible:

• MeanFlow | NAVSIM | HUGSIM | GTRS