This repository is the implementation of the paper "Deep Reinforcement Learning for Human-Like Driving Policies in Collision Avoidance Tasks of Self-Driving Cars" and "Example-guided learning of stochastic human driving policies using deep reinforcement learning", an extended paper of the former. Note that this project also provides the Generative Adversarial Imitation Learning (GAIL) (Ho, J., & Ermon, S., 2016) algorithm implementation as a benchmark. The results will be published soon under the revised version. It also contains a unique simulator environment and data (which will be discussed later) for training so you can replicate the results. In short, this project is designed to find human-like driving policies using imitation learning approach we have developed. To this end, we demonstrate our approach to relatively simple tasks - static/dynamic obstacle avoidance tasks. In the article, you will find further discussions and limitations.
Tested on python 3.7.9 version. At the moment, you can only run this project on a Windows OS.
- Clone the repository and install the requirements:
pip install -r requirements.txt
-
Extract data.zip. You will find there two files:
- Under expert->gp (stands for Gaussian Process) you will find the GP_expert.csv to run the "PLAIN" and "MDN" algorithms mentioned in the paper.
- Under expert->demonstrations you will find the demonstrations.csv to run the "GAIL" algorithm that used as a benchmark comparison to our method.
-
Download simulator environment run files from the following link: simulator_run_files.zip
-
Extract simulator_run_files.zip to any preferred folder. You will find there two types of road (described in the paper):
- training road (track 0)
- generalization road (track 1)
While the training road contains one distribution of obstacles, the generalization road contains three - Random, Gauss, and Batch.
To replicate the paper results you sholud train the agents on track 0 and test it on track 1. To train the agent open the track 0 file and run on your terminal:
python main.py --algorithm MDN --train
Please be patient, this may take up to a week.
To test your agent open any track you want and run on your terminal:
python main.py --algorithm MDN
If you find this work useful in your research, please cite:
@article{emuna2022example,
title={Example-guided learning of stochastic human driving policies using deep reinforcement learning},
author={Emuna, Ran and Duffney, Rotem and Borowsky, Avinoam and Biess, Armin},
journal={Neural Computing and Applications},
pages={1--14},
year={2022},
publisher={Springer}
}