Train. Break.
Retrain. Ship.
The standard Python API for reinforcement learning. 60+ environments. Three lines to your first agent.
What will your agent learn?
Atari
Arcade Learning Environment
Classic Atari 2600 ROMs emulated at pixel level. From Pong to Montezuma's Revenge — the original benchmark for deep RL.
MuJoCo
Physics-Based Robotics
Multi-joint locomotion and manipulation in continuous state/action spaces. The gold standard for continuous control research.
Classic Control
Real-world Physics Problems
CartPole, Pendulum, MountainCar — small enough to understand completely, hard enough to build intuition about RL fundamentals.
Running in under 60 seconds
pip install gymnasiumimport gymnasium as gym# Initialize any environment in one lineenv = gym.make("CartPole-v1", render_mode="human")# Standard loop — works for every environmentobs, info = env.reset(seed=42)for _ in range(1000):action = env.action_space.sample() # your policy hereobs, reward, terminated, truncated, info = env.step(action)if terminated or truncated:obs, info = env.reset()env.close()
Gymnasium wins on every metric
| Metric | Gymnasium | Gym (legacy) | Competitor B |
|---|---|---|---|
| CartPole-v1 wall-clock (1M steps) | 18s | 47s | 61s |
| MuJoCo Ant-v4 steps/sec (single CPU) | 14200 | 8900 | 6300 |
| Atari Pong env reset latency | 0.8ms | 2.1ms | 3.4ms |
| Vectorized CartPole (1024 envs) steps/sec | 2.1M | 890k | 540k |
| API compatibility (Gym drop-in) | 100% | ~60% | ~40% |
| Python 3.13 support | ✓ | ✗ | ✗ |
Measured on Ubuntu 22.04, Intel Core i9-13900K, Python 3.12. Results may vary.
Steps / Second
faster than Sync vector env using NumPy CartPole. Parallel environments maximize GPU utilization.
Production-grade robotics environments
High-fidelity environments used in published robotics research. MuJoCo-powered, Gymnasium API-compatible.
24-DoF anthropomorphic manipulation
7-DoF arm: Reach, Push, Pick & Place
Object reorientation with 24 actuators
Terrain-adaptive locomotion
Cooperative locomotion, 4 agents
Domain randomization built-in
Pre-loaded CartPole environment. Watch your first agent learn in under 3 minutes. No install required.
Open in ColabGymnasium Pro Environments
Production robotics environments for serious research. Two fields, no friction.