README.md

March 26, 2025 · View on GitHub

🌌 AlphaDrive: Unleashing the Power of VLMs in Autonomous

Driving via Reinforcement Learning and Reasoning

Bo Jiang1, Shaoyu Chen1,2, Qian Zhang2, Wenyu Liu1, Xinggang Wang1,📧

1 Huazhong University of Science and Technology, 2 Horizon Robotics, 📧 corresponding author

arxiv paper License

https://github.com/user-attachments/assets/71695178-90ca-4f5c-acc2-ab2e13c63e43

✨ Highlights

  • To the best of our knowledge, AlphaDrive is the first to integrate GRPO-based RL with planning reasoning to autonomous driving, significantly boosting both performance and training efficiency.

  • We are excited to discover that, following RL training, AlphaDrive exhibits some emergent multimodal planning capabilities, which is promising for improving driving safety and efficiency.

📋 News

[2025-3-26]: We have released the training and evaluation scripts of AlphaDrive.

[2025-3-11]: AlphaDrive arXiv paper released. Code are coming soon. Please stay tuned! ☕️

🎮 Getting Started

Installtion

git clone git@github.com:hustvl/AlphaDrive.git
conda create -n alphadrive python=3.11 -y
conda activate alphadrive
sh setup.sh

Data Preparation

We provide the prompt templates used in AlphaDrive for training and generating planning reasoning data, and an example QA is provided in example.json.

Training

For Supervised Fine-tuning Phase:

sh train_tools/run_train_sft.sh

For Reinforcement Learning Phase:

sh train_tools/run_train_grpo.sh

Evaluation

You can evaluate the meta-action planning accuracy using the script below.

sh eval_tools/qwen2vl_plan_cmd_eval.sh

📊 Qualitative Results

❤️ Acknowledgements

This repo is built on open-r1 and R1-V. We sincerely thank the contributors for their great work!

📚 Citation

If you find AlphaDrive useful in your research or applications, please consider giving us a star 🌟 and citing it by the following BibTeX entry.

@article{jiang2025alphadrive,
      title={AlphaDrive: Unleashing the Power of VLMs in Autonomous Driving via Reinforcement Learning and Reasoning}, 
      author={Bo Jiang and Shaoyu Chen and Qian Zhang and Wenyu Liu and Xinggang Wang},
      year={2025},
      eprint={2503.07608},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2503.07608}, 
}

Check out our other awesome projects:

VAD & VADv2: Vectorized Scene Representation for Efficient Autonomous Driving.

Senna: Bridging Large Vision-Language Models and End-to-End Autonomous Driving.

DiffusionDrive: Truncated Diffusion Model for End-to-End Autonomous Driving.

RAD: Training an End-to-End Driving Policy via Large-Scale 3DGS-based Reinforcement Learning.

MapTR: An End-to-End Framework for Online Vectorized HD Map Construction.