Farama-Foundation/Gymnasium

What it solves

Gymnasium provides a standardized way for reinforcement learning (RL) researchers and developers to build, test, and compare different learning algorithms. It eliminates the need to write custom communication code for every new environment, ensuring that algorithms can be easily ported across different simulation tasks.

How it works

It defines a standard Python API that acts as a bridge between the learning algorithm (the agent) and the environment. By modeling environments as Python classes with consistent methods like reset() and step(), it allows agents to interact with a world, receive observations and rewards, and perform actions in a uniform way regardless of whether the environment is a simple text game or a complex physics simulation.

Who it’s for

It is designed for developers and researchers working on reinforcement learning, ranging from beginners using reference implementations to experts building complex agents for physics or gaming environments.

Highlights

• Diverse Environment Library: Includes a wide range of built-in environments, including Classic Control, Box2D physics, Toy Text for debugging, MuJoCo multi-joint control, and Atari 2600 emulations.
• Strict Versioning: Uses a versioning system (e.g., "-v0") for environments to ensure that changes affecting learning results do not break reproducibility.
• Extensible Ecosystem: Supports a vast number of third-party environments that comply with the Gymnasium API.
• Broad Compatibility: Supports multiple Python versions (3.10 through 3.14) on Linux and macOS.