Mohammad Reza Samsami, Mohammadhossein Bahari, Saber Salehkaleybar, Alexandre Alahi. arXiv 2021.
[Project Website]       [Paper]

This repo provides implementations of Causal Imitative Model (CIM). The idea is to explicitly discover the causal model of environment and utilize it to improve the autonomous driving system. All code is written in Python 3, using PyTorch, NumPy, and CARLA.

The project is built on OATomobile, a framework for autonomous driving research which wraps CARLA in OpenAI gym environments. The main part of our contribution is gathered in oatomobile\baselines\torch\cim.

To install requirements, refer to OATomobile github repo.

To train the perception model, you would run with

python -m oatomobile.baselines.torch.cim.perception.train --dataset_dir=dataset_dir --output_dir=output_dir --in_channels=1 --num_epochs=num_epochs --beta=6

After training the perception model and obtaining representations of scenarios' observations, you could train the speed predictor with

python -m oatomobile.baselines.torch.cim.predictor.train --dataset_dir=dataset_dir --output_dir=output_dir --num_epochs=num_epochs

To perform the model on a task:

python -m test --task=task --model_dir=model_dir --predictor_dir=predictor_dir --output_dir=output_dir --alpha=alpha --gamma=gamma

If you find this code useful, please cite:

@misc{samsami2021causal,
   title={Causal Imitative Model for Autonomous Driving}, 
   author={Mohammad Reza Samsami and Mohammadhossein Bahari and Saber Salehkaleybar and Alexandre Alahi},
   year={2021},
   eprint={2112.03908},
   archivePrefix={arXiv},
   primaryClass={cs.RO}
}

This code was developed using OATomobile and disentanglement_lib.