A video game description language (VGDL) built on top pf pygame.
Latest commit 27a9086 Apr 11, 2014 @schaul Merge pull request #5 from dylski/master
Reinforcement learning environment interface added.



PyVGDL is a high-level video game description language (VGDL) built on top of pygame.

The aim is to decompose game descriptions into two parts: 1) a very high-level description, close to human language, to specify the dynamics, which builds on 2) an ontology of preprogrammed concepts for dynamics, interactions, control. Programmers extend the possibilities of (1) by writing modules in (2), and game designers can very quickly compose new games from those components without programming.


The original idea was discussed in the 2012 Dagstuhl seminar, with a full description presented at the IEEE CIG conference 2013 (this is also the reference paper to cite if you use PyVGDL for academic work).

Installation and Dependencies

  • Get the pygame package
  • (Alternative Method) Using Homebrew and virtualenv on Mac OSX

        brew install sdl sdl_image sdl_mixer sdl_ttf portmidi
        pip install mercurial
        pip install hg+http://bitbucket.org/pygame/pygame
  • For all reinforcement learning usage, also get the PyBrain machine learning library

  • For the upload to youtube functionality, you will need the gdata library

  • Download repository

      git clone git://github.com/schaul/py-vgdl.git
  • Try examples

      python -m examples.gridphysics.aliens
      python -m examples.gridphysics.frogs
      python -m examples.gridphysics.zelda


  • Language
    • A simple programming language of 2D video game design
    • A parser for the language
    • A parser for textual level descriptions
    • An ontology with numerous high-level building blocks for games
      • grid-based physics engine
      • continuous physics engine, including gravity, friction, etc.
      • stochastic events
  • Classic examples (simplified versions)
    • Space invaders
    • Frogger
    • Lunar lander
    • Zelda
    • Super Mario
    • Portal
    • Sokoban
    • PTSP
    • Pong
    • Tank wars
    • Dig-dug
    • Pacman
    • ...
  • Human play
    • Interactive play, either from bird-eye viewpoint, or from first-person viewpoint
    • Create animated GIFs from replayed action sequences
  • Bot play
    • Interface for artificial players (bots)
    • Conversion of game dynamics into the transition matrices of a Markov Decision Process (MDP)
    • Automatically generated local/subjective observation features
    • Reinforcement learning
      • Easy interface to RL algorithms from PyBrain
      • Classic grid world benchmarks