Package Homepage

Modular Object-Oriented Games


Project Website


This Modular Object-Oriented Games (MOOG) library is a general-purpose python-based platform for interactive games. It aims to satisfy the following criteria:

  • Highly customizable. Environment physics, reward structure, agent interface, and more are customizable.
  • Easy to rapidly prototype tasks. Tasks can be composed in a single short file.
  • Usable for both reinforment learning and psychology, with DeepMind dm_env and OpenAI Gym interfaces for RL agents and an MWorks interface for psychology and neurophysiology.
  • Light-weight and efficient. Most tasks run quickly, almost always faster than 100 frames per second on CPU and often much faster than that.
  • Facilitates procedural generation for randomizing task conditions every trial.

See moog_demos for a variety of example tasks.

Intended Users

MOOG may be useful for the following kinds of researchers:

  • Machine learning researchers studying reinforcement learning in 2.5-dimensional (2-dimensional with occlusion) physical environments who want to quickly implement tasks without having to wrangle with more complicated game engines that aren't designed for RL.
  • Psychology researchers who want more flexibility than existing psychology platforms afford.
  • Neurophysiology researchers who want to study interactive games yet still need to precisely control stimulus timing.
  • Machine learning researchers studying unsupervised learning, particularly in the video domain. MOOG can be used to procedurally generate video datasets with controlled statistics.


The core philosophy of MOOG is "one task, one file." Namely, each task can be implemented with a single configuration file. This configuration file should be a short "recipe" for the task, containing as little substantive code as possible, and should define a set of components to pass to the MOOG environment. See the MOOG README for more details.

We also include an example MWorks interface for running psychophysics experiments, as well as a python demo script for testing task prototypes.

Features Compared to Existing Platforms

Compared to professional game engines (Unity, Unreal, etc.) and existing visual reinforcement learning platforms (DM-Lab, Mujoco, VizDoom, etc.):

  • Python. MOOG tasks are written purely in python, so users who are most comfortable with python will find MOOG easy to use.
  • Procedural Generation. MOOG facilitates procedural generation, with a library of compositional distributions to randomize conditions across trials.
  • Online Simulation. MOOG supports online model-based RL, with a ground truth simulator for tree search.
  • Psychophysics. MOOG can be run with MWorks, a psychophysics platform.
  • Speed. MOOG is fast on CPU. While the speed depends on the task and rendering resolution, MOOG typically runs at ~200fps with 512x512 resolution on a CPU, which is faster than one would get with DM-Lab or Mujoco and at least as fast as Unity and Unreal.

Compared to existing python game platforms (PyBullet, Pymunk, etc.):

  • Customization. Custom forces and game rules can be easily implemented in MOOG.
  • Psychophysics, Procedural Generation, and Online Simulation, as described above.
  • RL Interface. A task implemented in MOOG can be used out-of-the-box to train RL agents, since MOOG is python-based and has DeepMind dm_env and OpenAI Gym interfaces.

Compared to existing psychophysics platforms (PsychoPy, PsychToolbox, MWorks):

  • Flexibility. MOOG offers a large scope of interactive tasks. Existing psychophysics platforms are not easily customized for game-like tasks, action interfaces, and arbitrary object shapes.
  • Physics. Existing psychophysics platforms do not have built-in physics, such as forces, collisions, etc.
  • RL Interface, as described above.

MOOG can interface with MWorks, allowing users to leverage the MOOG task framework while also allowing for precise timing control and interfaces with eye-trackers, joysticks, and electrophysiology software.


  • Not 3D. MOOG environments are 2.5-dimensional, meaning that they render in 2-dimensions with z-ordering for occlusion. MOOG does not support 3D sprites.
  • Very simple graphics. MOOG sprites are monochromatic polygons. There are no textures, shadows, or other visual effects. Composite sprites can be implemented by creating multiple overlapping sprites, but still the graphics complexity is very limited. This has the benefit of a small and easily parameterizable set of factors of variation of the sprites, but does make MOOG environments visually unrealistic.
  • Imperfect physics. MOOG's physics engine is simple. It uses Newton's method to effect action-at-a-distance forces. MOOG does include a collision module that implements rotational mechanics, but it is not as robust as more professional physics engines and can have instabilities (particularly if multiple objects collide simultaneously). See moog_demos/example_configs/ for an extreme example of unstable physics.

Getting Started

See the project website for API documentation about every file and function in MOOG.


If you would like to install this library as a package, you can install using pip:

pip install moog-games

This will install moog and moog_demos packages. Be sure to use python 3.7 or later.

Running The Demo

Tasks can be played by running the run_demo script, in which the --config flag indicates the task config to demo. For example, to demo the pong task, you would run:

python3 -m moog_demos.run_demo --config='moog_demos.example_configs.pong'

When this command is run, the demo will produce an interactive display. At the top of the display is the rendered environment state, in the middle of the display is a histogram of recent rewards, and at the bottom of the display is a top-down view of a cartoon joystick. You can click and drag the joystick around to control the agent avatar. The demo can be terminated by pressing escape.

The pong task looks like this:

You can change the config flag to point to any of the example configs. They will all run except for cleanup, which is multi-agent so cannot be played by a single-agent demo, (though see multi_agent_example for more about that).

Implementing Tasks

Before implementing your own tasks, please read the MOOG README.

To begin implementing your own task, we recommend first looking at all the example configs in moog_demos and copying one with some similarities to your task into a working directory. Then modify it incrementally to your specification.

To demo your config, copy into your working directory and run it with

$ python3 --config='' --level=$your_config_level

Contact and Support

Please see for information about support. Please email Nick Watters at with questions and feedback.


If you use this library in your work, please cite it as follows:

author = {Nicholas Watters and Joshua Tenenabum and Mehrdad Jazayeri},
title = {Modular Object-Oriented Games: A Task Framework for Reinforcement Learning, Psychology, and Neuroscience},
url = {},
eprint = {2102.12616},
journal = {arXiv preprint arXiv:2102.12616},
year = {2021}

Some parts of this codebase are derived from Spriteworld. See the Spriteworld license in LICENSE-spriteworld.

Expand source code
.. include::



Automatically Generating Documentation …


Modular Object-Oriented Games …


MOOG Demos …


Multi-Agent Example …


MWorks with Modular Object-Oriented Games …

tests …