pysc2

Why we included this project

For reinforcement-learning researchers who want a benchmark that genuinely stresses an agent, PySC2 is the standard way to drive StarCraft II from Python. DeepMind created it alongside Blizzard, wrapping the game's machine-learning API into a familiar RL interface: the agent reads observations, sends actions, and steps through a match like any other environment. A full game is a long planning problem, mixing resource management, scouting, and imperfect information, so it exposes weaknesses that simple grid worlds usually don't. The package ships pre-built maps, mini-game scenarios tied to the companion research paper, a scripted random agent for baselines, and a replay tool for watching what agents actually did. It is still the reference implementation people reach for when reproducing or extending published results on this game.