Farama-Foundation/Shimmy

What it solves

Shimmy addresses the incompatibility between different reinforcement learning (RL) environment APIs. It allows developers to use popular external environments—such as those from OpenAI Gym, DeepMind Control, and OpenSpiel—within the standardized Gymnasium and PettingZoo frameworks without needing to rewrite the environment logic.

How it works

Shimmy acts as an API conversion tool. It provides bindings (wrappers) that translate the interface of external RL libraries into the Gymnasium (for single-agent) and PettingZoo (for multi-agent) formats. This enables a seamless transition between different environment libraries while maintaining a consistent API for the RL agent.

Who it’s for

It is designed for reinforcement learning researchers and developers who need to access a wide variety of benchmark environments across different libraries but want to maintain a single, consistent API for their training and evaluation pipelines.

Highlights

• Broad Compatibility: Supports conversion for OpenAI Gym (v21 and v26), DeepMind Control (single and multi-agent), and OpenSpiel.
• Gymnasium Integration: Converts external environments into Gymnasium-compatible formats.
• PettingZoo Integration: Converts multi-agent environments into PettingZoo-compatible formats.
• Flexible Installation: Offers specific installation extras (e.g., shimmy[dm-control]) to manage dependencies for different environment suites.