Physics-based control tasks. CartPole, MountainCar, Pendulum — the canonical RL benchmarks.
Gymnasium vs. the alternatives
Compare environment support, API quality, and reproducibility features across the RL ecosystem. Gymnasium is the actively maintained fork of OpenAI Gym — the standard that won't leave you stranded.
SB3 = Stable Baselines 3 (algorithm library, not env library). Last updated Feb 2026.
The complete API surface
Six methods. That's the entire interface. Every RL algorithm in the literature maps to this contract — your code works across all 60+ environments without modification.
gym.make(id, **kwargs)env.reset(seed=None, options=None)env.step(action)Execute one timestep. Returns observation, reward, terminated flag, truncated flag, and info dict. Termination and truncation are now clearly separated.
tuple[Obs, float, bool, bool, dict]Example
obs, rew, term, trunc, info = env.step(action)gym.vector.make(id, num_envs, **kwargs)gym.register(id, entry_point, **kwargs)gym.wrappers.WrapperClass(env)pip install gymnasium# base install[atari][mujoco][box2d][jax]Benchmark results: CartPole-v1
DQN agent trained for 500K steps. Gymnasium's clean termination/truncation separation and reproducible seeding results in faster convergence and higher final performance.
seed=42 · 5 runs averaged
DQN · lr=1e-3 · ε-greedy
Reproducible with gymnasium 1.2.3
Contributing institutions
Your next experiment
starts in 3 lines.
Gymnasium 1.2.3 is production-stable, MIT-licensed, and ready for your ablation studies, sim-to-real pipelines, and 2am debugging sessions. Join researchers from Mila, Brown, and Southampton.
— or request early access to advanced features —