This repo contains source code for the natural signal Atari environments, introduced in the paper Natural Environment Benchmarks for Reinforcement Learning.

Instructions

1. Install dependencies with pip install gym[atari] pygame scikit-video opencv-python
2. Prepare a directory of images or videos
3. Play with new versions of Atari games with the following commands:

# Inject gaussian noise to the observations
./natural_env.py --env BreakoutNoFrameskip-v4 --imgsource noise

# Inject some video signals to the observations
./natural_env.py --env SpaceInvadersNoFrameskip-v4 --imgsource videos --resource-files "~/my/videos/*.mp4"

License

This project is under the CC-BY-NC 4.0 license. See LICENSE for details.