google-deepmind/lab

Features

• Reinforcement Learning Environments - Exposes 3D simulated levels through a uniform reinforcement learning API so agents and training code control every level the same way.
• Reinforcement Learning Simulators - Provides a customisable 3D simulation platform where researchers train and evaluate learning agents on navigation, puzzle and cognitive tasks.
• Navigation and Puzzle Tasks - Supplies a suite of challenging 3D navigation and puzzle-solving tasks as a testbed for deep reinforcement learning research.
• Reinforcement Learning Training - Supplies challenging 3D navigation and puzzle tasks as a testbed for training and evaluating deep reinforcement learning agents.
• AngelScript and Lua Game Scripting - Extends Lua level scripting with tensor and image libraries, event creation, entity retrieval and post-load texture updates.
• Environment Scripting Interfaces - Hooks level scripts into game events through callback functions the environment calls periodically, with defaults when a callback is missing.
• Lua Scripting Frameworks - Defines level behavior in Lua scripts whose callbacks the engine invokes at game events, with defaults when a callback is missing.
• Level Generation Scripts - Lets researchers define their own levels as Lua scripts, instantiated through the same scripting and Python interfaces as shipped levels.
• Agent Action Spaces - Lets level scripts declare their own discrete action sets through callbacks so agents can act beyond built-in controls.
• Standardized Training Interfaces - Connects agents to simulated environments through a general reinforcement learning adapter so standard training code works without custom glue.
• Cognitive Ability Test Levels - Ships ready-made levels that each test a specific cognitive ability of an agent, startable from scripts or Python.
• Environment Observation Retrieval - Lists every supported observation type, including level-specific ones, and returns requested observations as arrays for the agent at each step.
• Environment Stepping with Actions - Steps the environment forward a chosen number of frames, applying the supplied action array on every frame per the action specification.
• Episode Reset and Seeding - Restarts the environment for a new episode, optionally selecting an episode and seeding the random generator for reproducible runs.
• Frame-Stepped Environment Loops - Advances the simulation a chosen number of frames per call, applying actions each frame and returning observation arrays to the agent.
• Procedural Generators - Compiles playable levels at runtime from plain text layouts and reuses previously compiled maps through a level cache.
• RL Interface Standards - Exposes the environment through a general reinforcement learning interface and packages it as an installable module for reuse.
• Python Scripting Environments - Wraps the native simulation engine in a Python module so researchers can create, reset and step environments from Python scripts.
• Cognitive Paradigms - Provides a platform for implementing classical cognitive psychology paradigms, with sample tasks such as visual search and sequential comparison.
• First-Person Shooter Engines - Builds on a modified first-person game engine that supplies rendering, physics, entity management, console commands and map loading.
• Game Environment Setups - Loads a chosen game level and builds an environment object with configurable resolution, frame rate, level directory, seeding and level options.
• Python and C++ Game Engines - Offers a native C++ engine built with Bazel and pinned dependencies, exposed to Python and installable as a module.
• Quake III Engine Forks - Builds on a modified Quake III Arena engine adapted to render first-person 3D worlds for agents instead of human players.
• Text-Based Map Generation - Builds playable levels on the fly from plain text layouts, with themes, skyboxes, appearance variations and bot navigation support.
• Rendering Backends - Chooses between software and hardware rendering, or a native window, via a build-time graphics setting and a per-environment renderer option.
• Python Environments - Exposes simulated 3D environments through an installable Python module so researchers can create, reset and step them directly from scripts.
• Python-C++ Bindings - Wraps the compiled C++ simulation core in a Python module so environments can be created, stepped and reset from scripts.

Star history