Code for "BeT-AIL: Behavior Transformer-Assisted Adversarial Imitation Learning from Human Gameplay in Gran Turismo Sport." Website: link, Paper link forthecoming.
This repository contains pseudocode detailing the training process and BeT-AIL algorithm as a supplement to our original manuscript. The agent interface in Gran Turismo Sport is not enabled in commercial versions of the game. Please refer to this article for more information.
In this repository, we provide an implementation in the "MountainCarContinuous-v0" environment. The implementation is identical to the results presented in the manuscript in the Gran Turismo Sport environment. The hyperparameters are set to those used in the GTS experiments in the paper, and have not been optimized for the Mountain Car environment.
We provide a requirements.txt file with the required packages which can be installed with
pip install -r requirements.txt
To train a BeT-AIL policy, run:
python main.py --algorithm=BeT-AIL
To train an AIL policy, run:
python main.py --algorithm=AIL
To train an BeT policy, run:
python main.py --algorithm=BeT
Our AIL implementation is based on the pypi imitation library available here: imitation. Our Behavior Transformer implementation is based on the Decision Transformer implementation available here: online-dt. Please cite the respective authors if you employ their code in your research.