Gym-retro's afterlife: 1,000+ games under active maintenance

What it does Stable-Retro wraps fourteen classic console and arcade emulators into Gymnasium environments for reinforcement learning. It bundles integrations for over 1,000 games—including memory locations, reward functions, and level savestates—so you can train agents on titles ranging from Super Mario World to Street Fighter II. ROMs are not included, but a hash-matching importer and an integration tool let you wire in your own legally obtained copies.

The interesting bit This is community life-support for OpenAI’s gym-retro, which is now in maintenance mode. The Farama Foundation took the original, expanded the emulator roster to systems like the Sega Saturn and Nintendo DS, and opened the door for new game submissions—turning a static research artifact into an actively maintained benchmark suite.

Key highlights

• 14 emulated systems, from Atari 2600 through Sega Dreamcast and Arcade machines
• Over 1,000 games with pre-defined reward functions and savestates
• Cross-platform support for Linux, macOS (including Apple Silicon), and Windows via WSL2
• Python 3.10 through 3.14 compatibility
• Built-in tooling for integrating new ROMs and emulator cores

Caveats

• N64 and Dreamcast cores are gated behind build flags and platform-specific rendering requirements; Dreamcast hardware rendering is currently Linux-only
• Apple Silicon builds skip the Gameboy (Gambatte) core by default
• You must source your own ROMs and BIOS files; nothing is bundled except a single test ROM

Verdict A solid choice for RL researchers who want a maintained, expanding library of classic game benchmarks. Less appealing if you need turnkey ROM distribution or guaranteed builds for every niche core on every platform.