LCSim

LCSim: A Large-Scale Controllable Traffic Simulator

[ Webpage | Code | Paper ]

Getting Started

step 1: clone the repository

git clone https://github.com/tsinghua-fib-lab/LCSim.git

step 2: create a virtual environment with conda or virtualenv

# with conda
conda create -n lcsim python=3.10
conda activate lcsim
# with virtualenv
virtualenv -p python3.10 lcsim
source lcsim/bin/activate

step 3: install the dependencies

pip install -r requirements.txt

Training the diffusion model

Codes for training the diffusion model are in the motion_diff directory.

step 1: dataset preparation

• Download the Waymo Open Motion Dataset, we use version 1.2. We use the 9s scenario for training the diffusion model.
• Run the following command to preprocess the dataset

python3 motion_diff/dataset/process/process.py --data-dir /path/to/waymo_open_motion_dataset_v1.2 --output-dir /path/to/output_dir --dataset training/validation

This will generate the preprocessed dataset in h5 format (training.h5/validation.h5) in the output directory.

step 2: training the diffusion model

• Modify the configuration file motion_diff/configs/config.yml to specify the dataset path and other hyperparameters.

• Run the following command to train the diffusion model

python3 experiments/diff/train_md.py --config motion_diff/configs/config.yml --save /path/to/log_dir

The trained model will be saved in the log directory and you can check the training process in tensorboard by running tensorboard --logdir /path/to/log_dir. We trained our model for 200 epochs on the whole training set of WOMD, which takes about one week on 4 NVIDIA 4090 GPUs, the hyperparameters are the same as the ones in the config file.

Running the simulation

step 1: scenario data preparation

We provide scenario construction tools from multiple sources, including the Waymo Open Motion Dataset (WOMD), the Argoverse dataset, and the MOSS scenarios. You can convert them into the unified format for simulation by scripts in the lcsim/scripts/scenario_converts directory. Example scenarios are provided in the examples/data directory.

step 2: running the simulation

The whole simulation process is in the exapmles/simulation.ipynb notebook. You can run the notebook to see the simulation results.

Reinforcement learning

For multi-style reinforcement learning, you need to train the diffusion model first, and then specify the path to the trained model in the configuration file experiments/rl/configs/waymo_ppo.yml. Then you can run the following command to train the reinforcement learning model.

python3 experiments/rl/train_waymo.py 

By modifying the environment configuration in the configuration file, you can enable diffusive simulation and guidance for constructing driving scenarios with different styles.

Citation

If you find this work useful, please consider citing our paper:

@misc{zhang2024lcsimlargescalecontrollabletraffic,
    title={LCSim: A Large-Scale Controllable Traffic Simulator}, 
    author={Yuheng Zhang and Tianjian Ouyang and Fudan Yu and Cong Ma and Lei Qiao and Wei Wu and Jian Yuan and Yong Li},
    year={2024},
    eprint={2406.19781},
    archivePrefix={arXiv},
    primaryClass={cs.RO},
    url={https://arxiv.org/abs/2406.19781}, 
}

License

This project is licensed under the MIT License - see the LICENSE file for details.