HackAtari

This repository relies on OCAtari, a fork of the OpenAI Gym Atari environment. The OCAtari repository can be found here.

Installation

To install the OCAtari environment, please follow the instructions in the OCAtari repository, or simply run the following command:

pip install ocatari

To install the hackatari environment, run

pip install hackatari

or clone this repository and run the following command:

git clone https://github.com/k4ntz/HackAtari
cd HackAtari
pip install -e .

Usage

To use the HackAtari environment, simply import it as you would any other OpenAI Gym environment: You can run the run.py file to start the original game or any of the modified versions. E.g.:

python run.py -g Freeway # Starts normal Freeway (random agent)
python run.py -g Freeway -hu # Starts Freeway with the cars being invisible (interactive/human playing mode)
python run.py -g Freeway -m color8 # Starts Freeway with the cars of color #8 being (i.e. invisible) (random agent)
python run.py -g Freeway -m stop3 # Starts Freeway with stopping mode #3 (i.e. static cars) (random agent)
python run.py -g Seaquest -m oxygen disable_enemies gravity # Starts Seaquest with infinite oxygen, no enemy, gravity (random agent)
python run.py -g Kangaroo -m random_init disable_monkeys # Starts Kangaroo with random initial floor and no monkeys (random agent)

See the documentation or this markdown file for more information on the available modifications.

Cite our work

@article{delfosse2024hackatari,
  title={HackAtari: Atari Learning Environments for Robust and Continual Reinforcement Learning},
  author={Delfosse, Quentin and Bl{\"u}ml, Jannis and Gregori, Bjarne and Kersting, Kristian},
  journal={arXiv preprint arXiv:2406.03997},
  year={2024}
}