Source-linked AI summary
DeepMind Lab
Charles Beattie, Joel Z. Leibo, Denis Teplyashin, Tom Ward, Marcus Wainwright, Heinrich Küttler, Andrew Lefrancq, Simon Green, Víctor Valdés, Amir Sadik, Julian Schrittwieser, Keith Anderson, Sarah York, Max Cant, Adam Cain, Adrian Bolton, Stephen Gaffney, Helen King, Demis Hassabis, Shane Legg, Stig Petersen
TL;DR
DeepMind Lab addresses the challenge of studying autonomous agents in rich, realistic, and diverse environments without relying on physical robotics hardware. It provides a first-person 3D game platform with a flexible reinforcement learning API, varied observations and actions, and customizable levels. The platform supports research across navigation, exploration, planning, strategy, and fine control in visually complex worlds.
Problem
AI research needs realistic, perceptually rich environments for studying general intelligence, while physical robots are expensive and slow and prior virtual-world platforms required considerable engineering.
Method
DeepMind Lab combines a Quake III Arena-based first-person 3D platform with a reinforcement learning API, rich observations and actions, and customizable levels.
Results
DeepMind Lab provides diverse tasks spanning navigation, procedural exploration, fruit collection, laser tag, planning, strategy, and fine control in rich visual environments.
Takeaways & Limitations
The platform enables creative task development and the construction of varied environments and intelligence tests for research.
Abstract
from arXiv · showhide
DeepMind Lab is a first-person 3D game platform designed for research and development of general artificial intelligence and machine learning systems. DeepMind Lab can be used to study how autonomous artificial agents may learn complex tasks in large, partially observed, and visually diverse worlds. DeepMind Lab has a simple and flexible API enabling creative task-designs and novel AI-designs to be explored and quickly iterated upon. It is powered by a fast and widely recognised game engine, and tailored for effective use by the research community.
Introduction
DeepMind Lab addresses the engineering and speed constraints of building realistic virtual worlds for AI research. It combines rich 3D perception and physics with autonomous learning challenges in a flexible research platform.
- Motivation: Realistic virtual worlds can combine near-realistic perception and physics with the speed and flexibility of software.This offers an alternative to studying embodied intelligence directly with expensive, slow physical robots.
- Platform: DeepMind Lab is a first-person 3D platform built on Quake III Arena, with rich visuals, 3D actions, and diverse tasks.Examples include navigation, fruit collection, dangerous traversal, procedural environments, laser tag, and neuroscience-inspired tasks.
- Research use: DeepMind Lab has supported research on asynchronous reinforcement learning, unsupervised auxiliary tasks, and navigation.
- Platform comparison: Compared with other game-based platforms, DeepMind Lab provides richer visuals, more naturalistic physics, and fine-grained pointing in a fully 3D world.It is also more removed from its first-person-shooter origins than VizDoom.
- Research challenge: The platform combines raw-pixel 3D vision, egocentric viewpoints, fine motor control, navigation, planning, strategy, time, and autonomous exploration.Putting these frontier challenges together in one platform is described as a significant challenge for the field.
DeepMind Lab Research Platform
DeepMind Lab provides a game-engine-based research platform with flexible observations, rich actions, customizable levels, and varied tasks. Its levels cover collection, navigation, procedural exploration, and laser-tag control challenges.
- Platform foundation: DeepMind Lab is built from Quake III Arena and related open-source tools, with custom assets tailored for machine learning.
- API: The reinforcement learning API provides complex observations and accepts a rich set of actions.Interaction is lock-stepped, so the engine advances after the agent supplies its next action.
- Observations: Agents receive reward and pixel observations, with optional depth and velocity information for analysis.RGBD observations add per-pixel depth, while translational and angular velocities are exposed as separate three-dimensional vectors.
- Actions: Agents can simultaneously control movement, looking, and laser-tag actions in three-dimensional environments.
- Example levels: The platform includes fruit-gathering, static and random-goal navigation, procedurally generated mazes, and laser-tag levels.These tasks test collection, goal finding, exploration, aiming, planning, strategy, and robustness to visual complexity.
Technical Details
DeepMind Lab combines a minimally modified Quake III Arena engine with C and Python interfaces, flexible level-generation tools, and Lua customization for creating environments and mechanics.
- APIs and execution: The platform provides a simple C API with Python bindings while preserving compatibility with future engine changes.The original engine is modified only where necessary.
- APIs and execution: DeepMind Lab supports windowed or headless execution, with OpenGL rendering through either a GPU or software renderer.Headless mode supports higher performance and non-windowed environments such as remote terminals.
- APIs and execution: Instances are initialized with a level name, screen resolution, and frame rate before interaction through a simple reinforcement-learning-style API.The environment can be run through the Python interface illustrated in Figure 4.
- Level generation: Levels use Quake III Arena packages containing geometry, navigation information, and textures, with tools for generating maps from .map files.The platform also supports freely available level editors such as GtkRadiant.
- Level generation: Text Levels provide human-readable files for specifying walls, spawn points, and other game mechanics.The text specification can be rendered as a generated level, as shown in Figures 5 and 6.
- Level generation: The Lua-based level API customizes bots, item pickups, observations, restarts, reward schemes, messages, and other level aspects.This flexibility requires minimal changes to the original game engine.
Results and Performance
The paper reports platform frame-rate benchmarks on representative navigation and laser-tag levels, measured with agents connected through the Python API; earlier machine-learning results are cited separately.
- Benchmark setup: Frame-rate performance was measured on two typical included levels using an agent performing random actions through the Python API.The agent's overhead was insignificant, so measurements were dominated by engine simulation and rendering.
- Benchmark setup: The benchmarks ran on a Linux desktop with a 6-core Intel Xeon 3.50GHz CPU and an NVIDIA Quadro K600 GPU.The reported tables cover nav_maze_static_01 and lt_space_bounce_hard.
- Prior results: Earlier machine-learning results from DeepMind Lab versions are reported in Mnih et al. (2016), Jaderberg et al. (2016), and Mirowski et al. (2016).
Conclusion
DeepMind Lab supports research in rich 3D environments with game-like physics and enables flexible development of tasks, environments, and intelligence tests.
- DeepMind Lab enables research in a 3D world with rich science fiction visuals and game-like physics.
- The platform supports creative task development across a wide range of environments, tasks, and intelligence tests.The authors anticipate varied uses by the research community.