Counter-Strike: Global Offensive environment for OpenAI Gym on Linux
To use the gym environment, steam for Linux with Counter-Strike: Global Offensive installed needs to be available.
As the native (Linux, using OpenGL) version of Counter-Strike: Global Offensive
does not get hardware acceleration in virtual X servers like Xvfb or Xephyr,
it is necessary to run the game in compatibility mode, to get reasonable
performance (frames per second) in the gym environment: Using the steam client,
in the Properties of Counter-Strike: Global Offensive navigate to
Compatibility and check Force the use of a specific Steam Play
compatibility tool and select Proton 5.13-6 (others might work but are not
tested) from the drop-down menu below.
As of 1 October 2021 Proton 6.3-6 is longer available in the steam client. Using 6.3-7 the game keeps crashing just after startup, thus 5.13-6 seems to be the best option for now.
With the recent addition of experimental Vulkan support, it might actually be possible to run the native Linux version of Counter-Strike with sufficient performance. Please try this by adding
-vulkanto your command line.
It should be possible to launch Counter-Strike: Global Offensive (App ID 730) from the terminal (this might take some time, especially the first start after updating or setting the compatibility):
steam -applaunch 730 -insecure -untrusted -novid -nojoy
Counter-Strike: Global Offensive Game State Integration is necessary to
communicate information about the current game state to the python interface.
This needs to be set up in the game configurations: Copy the game state
integration configuration file from
the cfg directory of the repository into the cfg directory of the
Counter-Strike: Global Offensive installation.
To find out more about the Counter-Strike: Global Offensive Game State Integration and its configuration look at the Valve Developer Community.
The gym environment executes the game on a virtual X server display, either
inside a window on the pre-existing X display (Xephyr) or invisible in the
background (Xvfb). To install the required packages on Ubuntu:
sudo apt install xvfb xserver-xephyr
Note: This package is still in early stages of development, installing might miss dependencies or does not work at all.
pip install --upgrade gym-csgo
Running a Deathmatch (game mode) environment with default configuration and random actions per step until it is done (the match is done after 10 minutes):
# gym_csgo registers the envs (to gym.make(...))
import gym_csgo
# Gym environments
import gym
# Open new environment context (automatically closes env at end of scope)
with gym.make('csgo_dm-v0') as env:
# Reset the environment
env.reset()
# Env is not done yet
done = False
# Until the environment is done
while not done:
# Get random action from environment
action = env.action_space.sample()
# Execute the random action and collect observation
obs, rew, done, info = env.step(action)Programs showing demo actors in the environment are provided in the
gym_csgo.demo subpackage. There are random and noop actors which sample
random actions from the environment's action space or do no action at all: These
might be useful for testing the environment in general (esp. functionality,
startup, graphics, etc.) or experiment with configuration options which can
passed to the game. Start a random actor playing a deathmatch and show the
frames per second to evaluate the performance:
python -m gym_csgo.demo.random csgo_dm-v0 de_dust2 +cl_showfps 1
A special case is the manual actor which allows to actually play the game
through a pygame display interacting with the gym environment which itself wraps
the game on the virtual display. This is merely a technical demonstration but
might as well be suited as a starting point for collecting human demonstration
data. Play a game of casual game mode on the map train (the pygame actor will be
available once the match starts after the warmup period):
python -m gym_csgo.demo.manual csgo_casual-v0 de_train