The core structure of this repo is as follows:

├── fusion
│   ├── agent       # the training configs of each algorithm
│   ├── configs     # the evaluation escipts
│   ├── envs        # the training and testing environments
├── utils           # the globally shared utility functions
├── train           # the training scripts
├── tools           # tools for debugging and visualization
├── scripts

The implemented offline safe RL and imitation learning algorithms include:

git clone https://github.com/HenryLHH/fusion
# install fusion package in the virtual env
cd fusion
conda create -n fusion python=3.8
conda activate fusion
pip install -e .[all]

# train the FUSION agents
bash scripts/run_fusion.sh
# train the other agents
bash scripts/run_all.sh
# evaluate the trained model
bash scripts/run_eval_task.sh
# visualization
bash scripts/run_vis.sh

Our dataset to train the offline RL and imitation learning baselines is available on this Google Drive Link.

We acknowledge the following related repositories which contributes to some of our baselines in this offline RL and imitation learning libraries for autonomous driving in metadrive:

• ICIL: https://github.com/ioanabica/Invariant-Causal-Imitation-Learning
• OSRL: https://github.com/liuzuxin/OSRL

For more information about implementation, you are welcome to check our RAL paper:

@article{lin2024safety,
  title={Safety-aware causal representation for trustworthy offline reinforcement learning in autonomous driving},
  author={Lin, Haohong and Ding, Wenhao and Liu, Zuxin and Niu, Yaru and Zhu, Jiacheng and Niu, Yuming and Zhao, Ding},
  journal={IEEE Robotics and Automation Letters},
  year={2024},
  publisher={IEEE}
}