facebookarchive/mazebase

MazeBase is a tool that lets researchers build simple 2D maze-like games and train AI models to play them. The core benefit is providing a flexible sandbox where you can quickly design games with specific challenges, tune their difficulty, and then teach neural networks to solve those tasks through trial and error. The games take place on a 2D grid. An agent (the player character) navigates this grid, which can contain various items like walls, water that slows movement, colored switches that open matching doors, and pushable blocks. The games are presented to the AI not as pixels on a screen, but as a list of text sentences describing what is where. For example, the model might see: "Block at [-1,4]. Switch at [+3,0] with blue color." The agent then takes actions, the game updates, and the model tries to figure out the best moves to reach its goals. This project would be used by AI researchers or engineers studying reinforcement learning. For example, if you want to test how well an AI model understands language instructions, you could use the "Conditional Goals" game where the agent must read a sentence telling it which goal to visit based on a switch's color. Ten games are already included, with tasks ranging from pushing blocks to navigating around walls through doors. The project also includes code to train neural networks to play these games, along with a visual display so you can watch the trained models play in your browser. Notably, it was built with Torch, a machine learning framework that is now largely retired, so the code is written in Lua rather than the Python most modern AI tools use.