DeepMind Lab2D

A learning environment written in C++ and Lua for the creation of grid worlds.

About

DeepMind Lab2D is a system for the creation of 2D environments for machine learning. The main goals of the system are ease of use and performance: The environments are "grid worlds", which are defined with a combination of simple text-based maps for the layout of the world, and Lua code for its behaviour. Machine learning agents interact with these environments through one of two APIs, the Python dm_env API or a custom C API (which is also used by DeepMind Lab). Multiple agents are supported.

If you use DeepMind Lab2D in your research and would like to cite it, we suggest you cite the accompanying whitepaper.

Getting started

We provide an example "random" agent in python/random_agent, which performs random actions. This can be used as a base for creating your own agents, and as a simple tool to preview an environment.

bazel run -c opt dmlab2d/random_agent -- --level_name=clean_up

External dependencies, prerequisites and porting notes

DeepMind Lab2D currently ships as source code only. It depends on a few external software libraries, which we ship in several different ways:

• The dm_env, eigen, luajit, lua5.1, lua5.2, luajit, png and zlib libraries are referenced as external Bazel sources, and Bazel BUILD files are provided. The dependent code itself should be fairly portable, but the BUILD rules we ship are specific to Linux on x86. To build on a different platform you will most likely have to edit those BUILD files.

• A "generic reinforcement learning API" is included in //third_party/rl_api.

• Several additional libraries are required but are not shipped in any form; they must be present on your system:

  • Python 3.6 or above with NumPy, PyGame, and packaging.

The build rules are using a few compiler settings that are specific to GCC/Clang. If some flags are not recognized by your compiler (typically those would be specific warning suppressions), you may have to edit those flags.

Disclaimer

This is not an official Google product.