A fork of gym-retro ('lets you turn classic video games into Gymnasium environments for reinforcement learning') with additional games, emulators and supported plateforms. Since gym-retro is in maintenance now and doesn't accept new games, plateforms or bug fixes, you can instead submit PRs with new games or features here in stable-retro.

Currently added games on top of gym-retro:

• Super Mario Bros 2 Japan (Lost Levels) - NES
• Hang On - SMS
• Punch Out - NES
• WWF Wrestlemania the Arcade Game - Genesis
• NHL 94 - Genesis
• NHL 94 (1 on 1 rom hack) - Genesis
• Super Hang On - Genesis
• Tetris - GameBoy
• Virtua Fighter 2 - Genesis
• Virtua Fighter - 32x

PvP games that support two models fighting each other:

• Samurai Showdown - Genesis
• WWF Wrestlemania the Arcade Game - Genesis
• Mortal Kombat II - Genesis
• NHL 94 - Genesis

As well as additional states on already integrated games.

• Fixed UI flickering issue in OpenAI integration tool
• fix compile with c++ >=17

pip3 install git+https://github.com/Farama-Foundation/stable-retro.git

• NOTE: The Game Boy (gambatte) emulator is not supported on Apple Silicon

Build from source

1. pip install cmake wheel
2. brew install pkg-config [email protected] libzip qt5 capnp
3. echo 'export PATH="/opt/homebrew/opt/qt@5/bin:$PATH"' >> ~/.zshrc
4. export SDKROOT=$(xcrun --sdk macosx --show-sdk-path)
5. pip install -e .

Build Integration UI

1. build package from source
2. cmake . -DCMAKE_PREFIX_PATH=/usr/local/opt/qt -DBUILD_UI=ON -UPYLIB_DIRECTORY
3. make -j$(sysctl hw.ncpu | cut -d: -f2)
4. open "Gym Retro Integration.app"

Video on how to setup on Ubuntu and Windows: https://youtu.be/LRgGSQGNZeE

Docker image for M1 Macs: https://github.com/arvganesh/stable-retro-docker

'Nature CNN' model trained using PPO on Airstriker-Genesis env (rom already included in the repo)

Tested on Ubuntu 20.04 and Windows 11 WSL2 (Ubuntu 20.04 VM)

sudo apt-get update
sudo apt-get install python3 python3-pip git zlib1g-dev libopenmpi-dev ffmpeg

You need to install a stable baselines 3 version that supports gymnasium

pip3 install git+https://github.com/Farama-Foundation/stable-retro.git
pip3 install "stable_baselines3[extra]>=2.0.0a9"

Start training:

cd retro/examples
python3 ppo.py --game='Airstriker-Genesis'

@misc{stable-retro,
  author = {Mathieu and Poliquin},
  title = {Stable Retro, a fork of OpenAI's gym-retro},
  year = {2021},
  publisher = {GitHub},
  journal = {GitHub repository},
  howpublished = {\url{https://github.com/MatPoliquin/stable-retro}},
}

Game Integration tool: https://www.youtube.com/playlist?list=PLmwlWbdWpZVvWqzOxu0jVBy-CaRpYha0t

Join here: https://discord.gg/dXuBSg3B4D

Stable Retro lets you turn classic video games into Gymnasium environments for reinforcement learning and comes with integrations for ~1000 games. It uses various emulators that support the Libretro API, making it fairly easy to add new emulators.

Supported platforms:

• Windows 7, 8, 10
• macOS 10.13 (High Sierra), 10.14 (Mojave)
• Linux (manylinux1)

CPU with SSSE3 or better

Supported Pythons:

• 3.6
• 3.7
• 3.8

Each game integration has files listing memory locations for in-game variables, reward functions based on those variables, episode end conditions, savestates at the beginning of levels and a file containing hashes of ROMs that work with these files.

Please note that ROMs are not included and you must obtain them yourself. Most ROM hashes are sourced from their respective No-Intro SHA-1 sums.

Documentation is available at https://retro.readthedocs.io/en/latest/

You should probably start with the Getting Started Guide.

See CONTRIBUTING.md

There is an effort to get this project to the Farama Foundation Project Standards. These development efforts are being coordinated in the stable-retro channel of the Farama Foundation's Discord. Click here for the invite

See CHANGES.md

• Atari
  • Atari2600 (via Stella)
• NEC
  • TurboGrafx-16/PC Engine (via Mednafen/Beetle PCE Fast)
• Nintendo
  • Game Boy/Game Boy Color (via gambatte)
  • Game Boy Advance (via mGBA)
  • Nintendo Entertainment System (via FCEUmm)
  • Super Nintendo Entertainment System (via Snes9x)
• Sega
  • GameGear (via Genesis Plus GX)
  • Genesis/Mega Drive (via Genesis Plus GX)
  • Master System (via Genesis Plus GX)

See LICENSES.md for information on the licenses of the individual cores.

The following non-commercial ROMs are included with Stable Retro for testing purposes:

• the 128 sine-dot by Anthrox
• Sega Tween by Ben Ryves
• Happy 10! by Blind IO
• 512-Colour Test Demo by Chris Covell
• Dekadrive by Dekadence
• Automaton by Derek Ledbetter
• Fire by dox
• FamiCON intro by dr88
• Airstriker by Electrokinesis
• Lost Marbles by Vantage

Please cite using the following BibTeX entry:

@article{nichol2018retro,
  title={Gotta Learn Fast: A New Benchmark for Generalization in RL},
  author={Nichol, Alex and Pfau, Vicki and Hesse, Christopher and Klimov, Oleg and Schulman, John},
  journal={arXiv preprint arXiv:1804.03720},
  year={2018}
}