README.md

May 9, 2019 ยท View on GitHub

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.