lab
A customisable 3D platform for agent-based AI research
A customizable 3D game-world from Google DeepMind for training AI agents using reinforcement learning, where agents navigate mazes and solve puzzles to earn rewards inside a Quake III-based simulation.
DeepMind Lab is a 3D simulation environment built for training and testing AI research agents. It was created by Google DeepMind and is based on the ioquake3 game engine, which is itself derived from id Software's Quake III Arena. The result is a first-person 3D world that an AI agent can navigate, observe, and interact with, rather than a human player.
The environment provides a collection of 3D levels involving navigation tasks and puzzle-solving. These levels are designed to challenge AI agents that learn from experience by receiving rewards when they do the right thing, a technique called deep reinforcement learning. Researchers write agents in Python that interact with the environment through an API: the agent reads observations from the world (like what the camera sees), sends actions, and receives reward signals back.
Levels can be customized using Lua scripts, so researchers can design new tasks without rebuilding the whole system. The project ships with example levels and a random agent as a starting point. It also supports human play through a keyboard and mouse interface, which makes it easier to understand what an agent is experiencing.
Building the project requires Bazel, a build tool from Google, and runs on Linux. Several system libraries also need to be installed. The source code is provided directly rather than as a pre-built binary, so users compile it themselves.
The project is not an official Google product. It has an accompanying research paper on arXiv for those who want to cite it in academic work.
Where it fits
- Train a deep reinforcement learning agent to navigate 3D mazes and collect rewards in a controlled environment.
- Design custom research tasks by writing new levels in Lua without modifying the underlying game engine.
- Test a new RL algorithm against the included navigation and puzzle-solving benchmark levels.
- Play through a level yourself with keyboard and mouse to understand what your AI agent is experiencing.