Gymnasium is an open source Python library that provides a standard API for developing and comparing reinforcement learning algorithms. It serves as the maintained successor to OpenAI's original Gym library, with the Farama Foundation taking over stewardship after OpenAI handed off maintenance responsibilities. The library establishes a unified interface for communication between learning algorithms and environments, enabling researchers and practitioners to build, test, and benchmark RL solutions consistently.
The repository includes multiple families of built-in environments spanning different complexity levels and problem domains. Classic Control environments address real-world physics problems, while Box2D environments provide toy games using physics simulation and PyGame rendering. Toy Text environments offer simple discrete state and action spaces suitable for algorithm debugging. MuJoCo environments deliver more complex multi-joint control tasks using advanced physics simulation. The Atari family provides emulated 2600 ROM games with high complexity ranges, and third-party environments extend the ecosystem further for specialized use cases. Installation is flexible, allowing users to install the base library or specific environment families as needed, with support for Python 3.10 through 3.14 on Linux and macOS.
The API models environments as simple Python classes with straightforward interaction patterns. Gymnasium maintains strict versioning for all environments using suffixes like "-v0" to ensure reproducibility, incrementing version numbers when changes might impact learning results. This versioning discipline was inherited from the original Gym library and remains central to the project's design philosophy.
The repository's contributor network overlaps with major projects including Microsoft's VSCode and TypeScript implementations as well as the Rust language project, suggesting broad influence across the development ecosystem.
The Farama Foundation maintains a complementary ecosystem of related libraries built on the Gymnasium API, including PettingZoo for multi-agent reinforcement learning and CleanRL for reference implementations. The project actively welcomes community contributions and provides detailed contribution guidelines. Users can support development through GitHub Sponsors, and the project has published an associated research paper for citation purposes. The library's comprehensive documentation is available at gymnasium.farama.org, with community coordination occurring through a public Discord server.
