Gymnasium provides a standardized API for single-agent reinforcement learning environments, serving as a successor to OpenAI's Gym. It offers a comprehensive suite of reference environments and utilities, enabling consistent development and comparison of RL algorithms for researchers and practitioners.

How It Works

Gymnasium models environments as Python classes, abstracting the interaction loop into simple reset() and step() methods. This design promotes interoperability between diverse RL algorithms and environments, facilitating reproducible research and easier algorithm benchmarking. The library enforces strict environment versioning (e.g., "-v0") to ensure reproducibility.

Quick Start & Requirements

• Install base library: pip install gymnasium
• Install environment dependencies: pip install "gymnasium[atari]" or pip install "gymnasium[all]"
• Supported Python versions: 3.8-3.12 on Linux and macOS. Windows support is accepted via PRs but not officially tested.
• Documentation: https://gymnasium.farama.org/
• Discord: https://discord.gg/bnJ6kubTg6

Highlighted Details

• Includes diverse environment families: Classic Control, Box2D, Toy Text, MuJoCo, and Atari.
• Strict environment versioning for reproducibility.
• Integrates with related libraries like CleanRL and PettingZoo.
• Offers a citation for academic use.

Maintenance & Community

Maintained by the Farama Foundation, the project actively coordinates development on Discord. Financial contributions are welcomed to support ongoing development.

Licensing & Compatibility

The library is released under the Apache 2.0 license, permitting commercial use and integration with closed-source projects.

Limitations & Caveats

While Gymnasium supports Linux and macOS, official support for Windows is limited to PR acceptance. Some environment dependencies may be challenging to install on certain systems.