OpenAI Gymnasium (formerly OpenAI Gym): the 80/20

TL;DR

Gymnasium is the standard Python API and set of ready-made environments for training and benchmarking reinforcement learning (RL) agents.gymnasium.farama+2
It supersedes OpenAI Gym and is now actively maintained by the Farama Foundation; use Gymnasium going forward.github+2
Core loop: reset the environment, step with an action, receive observation/reward, and repeat until the episode ends.cto2b+2

What it is: A maintained fork of OpenAI Gym that provides a simple, consistent interface to RL environments (CartPole, Atari, MuJoCo, etc.) plus a standard API for agents↔envs interaction.gymlibrary+2
Why it matters: It gives a common benchmark and interface so different RL algorithms can be developed, compared, and reproduced reliably.gymnasium.farama+1
Status: OpenAI handed over maintenance; future development is in Gymnasium, which is a drop-in successor for most use cases.gymnasium.farama+2
How you use it (conceptually):
- env = gymnasium.make("EnvName") to create an environment.gymnasium.farama
- obs = env.reset() to start a new episode.gymnasium.farama
- obs, reward, terminated, truncated, info = env.step(action) each time step.gymnasium.farama+1
- terminated vs truncated: terminated means a true terminal state; truncated means the episode was cut short (e.g., time limit).reddit+1
Key primitives: observation_space and action_space define what the agent sees and can do; they can be discrete or continuous.gymnasium.farama+1

Environment = task simulator exposing a standard API: reset(), step(action), render().gymnasium.farama+1
Agent = chooses actions based on observations to maximize accumulated reward over episodes.gymnasium.farama
Workflow = pick an environment, implement or use an RL algorithm, run the reset/step loop, and evaluate with consistent metrics across environments.github+1

Gymnasium docs for basic usage and API details.gymnasium.farama+2
Farama’s Gymnasium repo for the environment families and current maintenance status.github
If following older tutorials: most still work, but note Gym→Gymnasium and the split of “done” into “terminated” and “truncated”.github+2