Gymnasium: A Standard Interface for Reinforcement Learning Environments

Gymnasium is the universal interface for reinforcement learning environments. Originally a fork of OpenAI Gym (which was unmaintained after 2021), Gymnasium is now the standard API used by virtually every RL library: Stable Baselines3, RLlib, CleanRL, Tianshou, and hundreds of research codebases. The `env.step(action)` → `obs, reward, terminated, truncated, info` API is so ubiquitous that any new environment or algorithm is expected to support it. Gymnasium ships with 100+ built-in environments: classic control (CartPole, Pendulum), box2d physics (LunarLander, CarRacing), Atari (via ALE), MuJoCo robotics (HalfCheetah, Ant, Humanoid), and third-party environments via the Gymnasium API. The 1.0 release adds vectorized environments (`AsyncVectorEnv`, `SyncVectorEnv`) that run 1000+ parallel episodes on multi-core CPUs, plus first-class wrappers for observation normalization, frame stacking, and reward clipping. The ecosystem is vast: `gymnasium-robotics` adds dexterous manipulation tasks (Shadow Hand, Fetch), `gymnasium-mujoco` provides advanced physics, and `shimmy` translates DeepMind, PettingZoo, and other APIs to Gymnasium. This standardization means that any RL algorithm written for Gymnasium works on any compatible environment with zero code changes—a huge productivity boost for research. For robotics specifically, Gymnasium is the integration layer: IsaacGym, MuJoCo, PyBullet, and SAPIEN all provide Gymnasium-compatible wrappers, so you can prototype a policy in a simple CartPole environment and deploy it to a humanoid robot with only environment ID changes. The `RecordVideo` and `RecordEpisodeStatistics` wrappers make experiment tracking and visualization trivial.