Farama-Foundation/Minigrid

What it solves

Minigrid provides a collection of lightweight, fast, and customizable discrete grid-world environments designed for Reinforcement Learning (RL) research. It allows researchers to test agents in controlled 2D maps with obstacles and goal-oriented tasks without the overhead of complex simulations.

How it works

The library follows the Gymnasium standard API. It features two main groups of environments:

• Minigrid: Agents (represented as triangles) navigate 2D maps containing walls, lava, or dynamic obstacles. They must complete tasks described by a "mission" string, such as picking up boxes or unlocking doors with keys.
• BabyAI: These environments focus on grounded language learning by generating synthetic, natural-language instructions (e.g., "put the red ball next to the box on your left") that the agent must follow to complete a task.

Who it’s for

It is designed for researchers and developers working on Reinforcement Learning, grounded language learning, and curriculum learning, as environments can be programmatically tuned for size and complexity.

Highlights

• Gymnasium API: Fully compatible with the standard RL environment API.
• Customizable: Environments can be tuned for difficulty and complexity to support curriculum learning.
• Grounded Language: Includes BabyAI environments for testing how agents respond to natural language commands.
• Lightweight: Designed for high-speed execution and low resource usage.