Gym (OpenAI Gym)

0.26.2 · deprecated · verified Sat Apr 11

Gym (formerly OpenAI Gym) is a Python library that provided a universal API for developing and comparing reinforcement learning (RL) algorithms across a diverse collection of environments. While it was historically the standard for RL environments, the `gym` library is no longer actively maintained. All future development and support have transitioned to its successor, `gymnasium`, a drop-in replacement. The last major release of `gym` was version 0.26.2, released in October 2022, which introduced significant breaking API changes.

Warnings

breaking The `gym` library is no longer maintained; all future development and support have moved to `gymnasium`. Users are strongly encouraged to migrate to `gymnasium` for continued updates, bug fixes, and compatibility with modern Python and NumPy versions.
Fix: Migrate your code to use `gymnasium`. The API is largely a drop-in replacement with `import gymnasium as gym`, but review `gymnasium` migration guides for version-specific changes, especially if upgrading from older `gym` versions.
breaking The `env.step()` method now returns a 5-tuple: `(observation, reward, terminated, truncated, info)`. The old `done` flag is split into `terminated` (agent's action led to termination) and `truncated` (e.g., time limit reached).
Fix: Update your `step()` calls to unpack 5 values. Use `terminated or truncated` where you previously used `done`.
breaking The `env.reset()` method now returns a 2-tuple: `(observation, info)`. The `return_info` parameter has been removed.
Fix: Update your `reset()` calls to unpack 2 values: `observation, info = env.reset(...)`. Access additional information from the `info` dictionary.
breaking The `env.seed()` method has been removed. Environment seeding is now handled by passing a `seed` argument to `env.reset()`.
Fix: Replace `env.seed(my_seed)` with `env.reset(seed=my_seed)` when initializing or restarting an episode.
breaking The `render_mode` should be specified during `gym.make()` (e.g., `gym.make('Env-v1', render_mode='human')`) and is no longer passed to the `env.render()` method.
Fix: Provide `render_mode` when creating the environment with `gym.make()`. The `env.render()` method should then be called without arguments if rendering is enabled.
gotcha Many environments require additional dependencies beyond the base `pip install gym`. Attempting to `gym.make()` such an environment without its extras will result in `ModuleNotFoundError`.
Fix: Install the necessary environment extras, e.g., `pip install 'gym[atari]'` for Atari environments, or `pip install 'gym[mujoco]'` for MuJoCo environments. Use `pip install 'gym[all]'` for all extras, though this can be substantial.

Install

pip install gym Base installation
pip install 'gym[atari]' # Example for Atari environments With environment extras
pip install 'gym[all]' # Install all supported environments All environment extras

Imports

gym
```
import gym
```
make
```
env = gym.make('CartPole-v1')
```
Many environments have been versioned up (e.g., v1, v2) over time, and older versions may be removed or behave differently.

Quickstart

This example demonstrates how to create a CartPole-v1 environment, reset it with a seed, take random actions, and handle the new 5-tuple return value from `step()` and 2-tuple from `reset()` in Gym 0.26.x+. The environment is rendered to a human-viewable window.

import gym

env = gym.make("CartPole-v1", render_mode="human")

# Reset returns (observation, info) in 0.26.x+
observation, info = env.reset(seed=42)

for _ in range(1000):
    action = env.action_space.sample()  # Agent selects an action
    # Step returns (observation, reward, terminated, truncated, info) in 0.26.x+
    observation, reward, terminated, truncated, info = env.step(action)

    if terminated or truncated:
        print(f"Episode finished after {_+1} timesteps.")
        observation, info = env.reset(seed=42) # Reset for a new episode

env.close()

view raw JSON →