An OpenAI Gym environment for the classic (imho) Terry Cavanagh game SuperHexagon
Note, that I have only tested on linux and for a variety of reasons, this probably won't work on other platforms.
The interface is the OpenAI gym environment interface. The rgb_array rendering
method is the only one supported right now. The action space is Dicrete(3) and
the observation space is Box(low=0, high=255, shape=(768*scale_factor, 480*scale_factor, 3), dtype=np.uint8) (scale_factor described later). You can
look in the tests directory for some example code. The following are the
environment keyword arguments accepted:
| argument | description |
|---|---|
superhex_dir |
the path to the root of your installation of Super Hexagon. I'm pretty sure it's not legal to distribute the game, so you have to provide it yourself. Can also be set through the SUPERHEXAGON_DIR environment variable. |
videodriver |
determines the SDL_VIDEODRIVER environment variable passed to the Super Hexagon process. Defaults to offscreen. Set to the empty string to make the game visible, assuming your SDL build supports that. |
start_process |
Defaults to True. Whether to actually start a new Super Hexagon game process or try to connect to an existing shimmed process. Useful for debugging the shim. |
superhex_log |
When false pipe the Super Hexagon subprocess output to /dev/null. Defaults to False |
scale_factor |
The Super Hexagon game will render at 768 by 480 by default. This determines the size of the observation made by the environment (useful to scale down when used as input to neural networks). Defaults to .16666 making the observed resolution 128 by 80 |
You need to have SDL (and development headers) and EGL libraries (see) installed in order to build the native code. Additionally, using the gym requires bspatch to be installed.
Run pip install -r requirements_test.txt
Running make install should attempt to build the native shim as well as the
python library. The python library is installed by running python setup.py install
Test that everything worked by running pytest tests/. This will run a few
pretty basic tests on the environment.
- Document other environment keyword args
- Document installation (don't forget to mention native dependencies like sdl and bspatch and ffmpeg for training)
- Document training with baselines zoo and tensorboard
- Document changes made in zoo fork
- Document how it works