Minigrid Documentation

Minigrid provides a suite of environments generated using the Wave Function Collapse (WFC) algorithm. These environments are designed to create diverse, procedurally generated layouts such as mazes, dungeons, rooms, and obstacle-filled spaces.

Available WFC environment types include:

• Mazes: WFCMazeEnv, WFCMazeSimpleEnv, WFCMazeKnotEnv, WFCMazeWallEnv, WFCMazeSpiralsEnv, WFCMazePathsEnv, WFCLikeEnv.
• Dungeons: WFCDungeonEnv, WFCDungeonMazeScaledEnv, WFCDungeonRoomsEnv, WFCDungeonLessRoomsEnv, WFCDungeonSpiralsEnv.
• Rooms: WFCRoomsFabricEnv, WFCRoomsOfficeEnv, WFCRoomsMagicOfficeEnv.
• Obstacles/Skewed Layouts: WFCObstaclesBlackdotsEnv, WFCObstaclesAngularEnv, WFCObstaclesHogs3Env, WFCObstaclesHogs2Env, WFCSkew2Env, WFCSkewCaveEnv, WFCSkewLakeEnv.
• General: WFCEnv.

Minigrid provides a variety of grid-world environments implemented in the minigrid/envs directory. Each environment is registered with OpenAI Gym (or Gymnasium) and can be instantiated using standard environment IDs.

Key features of these environments include:

• Gym Registration: Environments are accessible via standard gym.make() calls using their specific IDs.
• Programmatic Tuning: Environments are tunable in terms of size and complexity, allowing developers to adjust difficulty levels for tasks like curriculum learning or fine-tuning agent performance.

BabyAI environments are imported from the BabyAI project and are designed for research on grounded language learning.

Key characteristics include:

• Foundation: They are derived from the standard Minigrid grid-world environments.
• Language Instructions: They include functionality to generate synthetic, natural-looking instructions (e.g., “put the red ball next to the box on your left”).
• Tasks: Agents must navigate the world and move objects to specified locations to fulfill these linguistic commands.

Minigrid environments are discrete grid-world environments designed for Reinforcement Learning research. They follow the Gymnasium standard API.

Key characteristics include:

• Agent: A triangle-like agent with a discrete action space.
• Map: A 2D map containing various obstacles such as Walls, Lava, or Dynamic obstacles.
• Missions: Tasks are defined by a mission string returned in the agent's observation. Missions can include goal-oriented or hierarchical tasks like picking up boxes, opening doors with keys, or navigating mazes.
• Customization: Environments are registered with Gymnasium configurations and are programmatically tunable in terms of size and complexity, making them suitable for curriculum learning.

The Wrapper class is the base class for all environment transformations in MiniGrid. It allows you to modularly modify the step and reset methods of an environment without altering the original environment code.

When creating a custom subclass of Wrapper, if you override the __init__ method, you must call super().__init__(env) to ensure the environment is correctly wrapped.

Minigrid provides 2D grid-world environments with goal-oriented tasks using a Gymnasium interface. You can initialize and interact with environments using gym.make(). The agent uses a discrete action space and interacts with objects like doors, keys, or boxes within maze maps.

To use Minigrid, ensure you have minigrid and gymnasium installed, then follow the standard Gymnasium loop pattern.

import gymnasium as gym
import minigrid

env = gym.make("MiniGrid-Empty-5x5-v0", render_mode="human")
observation, info = env.reset(seed=42)
for _ in range(1000):
   action = policy(observation)  # User-defined policy function
   observation, reward, terminated, truncated, info = env.step(action)

   if terminated or truncated:
      observation, info = env.reset()
env.close()

To build the Minigrid documentation locally, you must first install the project in editable mode along with the documentation requirements. After that, you need to run the scripts that generate environment-specific documentation before building the HTML files.

1. Install dependencies

pip install -r docs/requirements.txt
pip install -e .

2. Generate environment documentation

Run these scripts to populate the documentation with environment details:

python docs/_scripts/gen_env_docs.py
python docs/_scripts/gen_envs_display.py

3. Build the HTML

To perform a one-time build of the documentation using dirhtml format:

cd docs
make dirhtml

You can install the minigrid package directly using pip for standard usage.

python3 -m pip install minigrid

If you are actively making changes to the documentation and want to see them reflected immediately, use sphinx-autobuild. This will watch for file changes and rebuild the documentation automatically.

Run the following command from the docs directory:

cd docs
sphinx-autobuild -b dirhtml . _build