Hao Gao1, Shaoyu Chen2,†, Yifan Zhu2, Yuehao Song1, Wenyu Liu1, Qian Zhang2, Xinggang Wang1,📧
1 Huazhong University of Science and Technology, 2 Horizon Robotics
† Project lead 📧 Corresponding author
Hao Gao1, Shaoyu Chen1,2,†, Bo Jiang1, Bencheng Liao1, Yiang Shi1, Xiaoyang Guo2, Yuechuan Pu2, Haoran Yin2, Xiangyu Li2, Xinbang Zhang2, Ying Zhang2, Wenyu Liu1, Qian Zhang2, Xinggang Wang1,📧
1 Huazhong University of Science and Technology, 2 Horizon Robotics
† Project lead 📧 Corresponding author
-
[2026.10.31] RAD-2 has been accepted by TPAMI! 🎉🎉🎉
-
[2026.04.17] We have released the RAD-2 paper on arXiv. More details are available on our project homepage.
-
[2025.11.04] Good news! 🎉 The ReconDreamer-RL team has now open-sourced their reconstructed 3DGS environments based on nuScenes. You can find the release here: 👉 ReconDreamer-RL Environments
-
[2025.09.28] We have released core code for RL training.
-
[2025.09.18] RAD has been accepted by NeurIPS 2025! 🎉🎉🎉
-
[2025.02.18] We released our paper on Arxiv. Code are coming soon. Please stay tuned! ☕️
We have created a central discussion issue for RAD training details. You can view and participate in the discussion here: RAD Training Details Issue. We also hope that the experiences and tips shared there will be helpful for your own RL-related training, not limited to RAD.
- Project Structure
.
├── data/ # Action anchors for planning/control
├── compute_advantage.py # Script for computing RL advantages and evaluation metrics
├── generate_action_anchor.py # Script for generating action anchors for planning/control
├── planning_head.py # Planning head module
├── utils.py # Utility functions for training and evaluation
└── README.md- Run Key Scripts
# You can quickly test the core functionality by running the provided scripts.
# Generate action anchors
python generate_action_anchor.py
# Run the planning head module
python planning_head.py
# Compute advantage metrics
python compute_advantage.py- Using Your Own Data
To integrate this project into your pipeline and use your own data, follow these steps:
Replace the Planning Head
Useplanning_head.pyto replace the head of your end-to-end algorithm.Prepare the Closed-Loop Environment
Set up your closed-loop environment and collect closed-loop data.Compute Advantages and Train the Model
Usecompute_advantage.pyto calculate advantage values from the collected data, and then use them for model training.
If you find RAD useful in your research or applications, please consider giving us a star 🌟 and citing it by the following BibTeX entry.
@article{gao2026rad,
title={Rad: Training an end-to-end driving policy via large-scale 3dgs-based reinforcement learning},
author={Gao, Hao and Chen, Shaoyu and Jiang, Bo and Liao, Bencheng and Shi, Yiang and Guo, Xiaoyang and Pu, Yuechuan and Li, Xiangyu and Liu, Wenyu and Zhang, Qian and others},
journal={Advances in Neural Information Processing Systems},
volume={38},
pages={32551--32576},
year={2026}
}
@article{gao2026rad,
title={Rad-2: Scaling reinforcement learning in a generator-discriminator framework},
author={Gao, Hao and Chen, Shaoyu and Zhu, Yifan and Song, Yuehao and Liu, Wenyu and Zhang, Qian and Wang, Xinggang},
journal={IEEE Transactions on Pattern Analysis and Machine Intelligence},
year={2026},
publisher={IEEE}
}