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 simplified 2D maps where they must complete goal-oriented tasks without the computational overhead of complex physics simulators.

How it works

The library uses the Gymnasium standard API to provide environments where a triangle-shaped agent navigates a 2D map. The agent interacts with obstacles like walls, lava, and dynamic obstacles. Tasks are defined by a "mission" string that guides the agent (e.g., picking up boxes or opening doors with keys). It includes two main groups of environments:

• Minigrid: Standard goal-oriented and hierarchical missions.
• BabyAI: Environments focused on grounded language learning, providing synthetic natural-language instructions (e.g., "put the red ball next to the box on your left").

Who it’s for

It is primarily for RL researchers and developers who need a controlled, efficient environment to benchmark agents on navigation, task completion, and grounded language understanding.

Highlights

• Gymnasium API: Fully compatible with the industry-standard API for RL environments.
• Customizable: Environments are programmatically tunable in size and complexity, making them ideal for curriculum learning.
• Grounded Language Support: Through BabyAI, agents can be trained to follow complex natural language commands.
• Lightweight: Designed for high-speed execution and fast iteration.