TorchCraft 2

(aka StarCraft Gym)

The fastest, easiest way to use StarCraft as a reinforcement learning environment.

And all you have to do to get started is pip install it.

Motivation

Researchers are eager to use StarCraft-based environments for RL experiments. But the process of getting StarCraft up and running can be time-consuming, and most environments aren't easy to use out of the box.

TorchCraft 2 offers simple (one-touch!) installation, an easy-to-use Gym interface, and multiple challenges/baselines out of the box.

In addition, we've worked with Blizzard to allow to TorchCraft 2 to include first legal distribution of StarCraft: Brood War binaries. This will enable the public to use TorchCraft 2 without having to acquire their own copies of StarCraft, or having to install it.

See the project proposal for details.

Roadmap

With those ready, we're aiming to launch TorchCraft 2 internally for experiments and beta testing. From there, the plan is:

• Integration with TorchCraft (deprecating the original Lua/Python APIs)
• Pre-compiled binary distribution on PyPy
• Public release, PR, and tutorial content

The full roadmap lives on GitHub

Usage

Building TorchCraft 2 for development

This is how I recommend using TorchCraft 2 at the moment:

git clone --recursive https://github.com/fairinternal/TorchCraft2/ tc2
cd tc2

conda create --name tc2 python=3 pip
source activate tc2
conda install pip cmake pybind11 numpy
conda install -y -c conda-forge sdl2 zstd
conda install -y -c anaconda zeromq
pip install -e .

Downloading required Starcraft files

Once you have installed tc2, you have to download StarCraft data files (MPQ files). We provide you with a tool to do that just for you - we can't provide them right away because we need you to read and accept Blizzard's EULA first.

tc2-setup

Demo

This demo (run_demo.py) creates a StarCraft gym environment and controls an agent using a simple "attack the middle" policy.

import gym
import tc2
from tc2.agents import attack_middle

env = gym.make('tc2-demo-v0')
while True:
  env.reset()
  actions = []
  done = False
  while not done:
    observation, reward, done, info = env.step(actions)
    actions = attack_middle(observation)

Running the TorchCraft 2 demo:

• python3 run_demo.py