We use Gym Retro to conduct research on RL algorithms and study generalization. Prior research in RL has mostly focused on optimizing agents to solve single tasks. With Gym Retro, we can study the ability to generalize between games with similar concepts but different appearances.

This release includes games from the Sega Genesis and Sega Master System, and Nintendo’s NES, SNES, and Game Boy consoles. It also includes preliminary support for the Sega Game Gear, Nintendo Game Boy Color, Nintendo Game Boy Advance, and NEC TurboGrafx. Some of the released game integrations, including those games in the data/experimental folder of Gym Retro, are in a beta state — please try them out and let us know if you encounter any bugs. Due to the large scale of the changes involved the code will only be available on a branch⁠(opens in a new window) for the time being. To avoid breaking contestants’ code we won’t be merging the branch until after the contest concludes.

The ongoing Retro Contest⁠(opens in a new window) (ending in a couple weeks!) and our recent technical report⁠(opens in a new window) focus on the easier problem of generalizing between different levels of the same game (Sonic The Hedgehog™). The full Gym Retro dataset takes this idea further and makes it possible to study the harder problem of generalization between different games. The scale of the dataset and difficulty of individual games makes it a formidable challenge, and we are looking forward to sharing our research progress over the next year. We also hope that some of the solutions developed by participants in the Retro Contest can be scaled up and applied to the full Gym Retro dataset.