Glue code that stops RL environments from speaking in tongues

What it does

Shimmy is a collection of API adapters—essentially polite, well-typed glue—that wraps external reinforcement learning environments in Farama’s standard interfaces. It converts OpenAI Gym, DeepMind Control Suite, DM Control Soccer, and OpenSpiel environments into either Gymnasium or PettingZoo APIs. The goal is straightforward: let researchers use existing simulators without learning new calling conventions for each one.

The interesting bit

The project treats API fragmentation as a systems integration problem rather than a user-education problem. Instead of forking legacy environments, Shimmy registers them under familiar gym.make-style IDs and handles the translation layer internally. It is unglamorous infrastructure, which is exactly why it matters.

Key highlights

• Converts OpenAI Gym (v21 and v26), DeepMind Control Suite, DM Control Soccer, and OpenSpiel environments
• Exposes converted environments through standard Gymnasium and PettingZoo registration interfaces
• Python version support is strictly bounded by upstream dependencies (e.g., gym-v21 tops out at Python 3.10 due to setuptools constraints)
• Maintained by the Farama Foundation as a companion project to Gymnasium and PettingZoo

Caveats

• Several extras are pinned to older Python versions by transitive dependencies the project cannot control
• The README is upfront about these ceilings, but they effectively fragment your Python environment strategy if you need bleeding-edge interpreters

Verdict

Worth a look if you are migrating research code between RL frameworks or need to benchmark against legacy environments without maintaining multiple API patterns. Skip it if you only use native Gymnasium environments and have no cross-ecosystem headaches.