Grandmaster Zerg bots you can fight, train, or dissect

What it does DI-star is an open-source reinforcement-learning framework purpose-built for StarCraft II. It packages pre-trained supervised and reinforcement-learning agents—ranked by the authors at diamond through grandmaster skill on the Zerg mirror matchup—alongside the full training stack. You can fight the AI locally, stage agent-versus-agent bouts, or run distributed supervised and reinforcement learning across multiple GPUs.

The interesting bit The architecture treats actors and environments as shared infrastructure and isolates each agent’s logic in a self-contained directory, so new bots can be dropped in and immediately played against each other. The project also fields named personality models—Abathur favors mutalisks, Brakk rushes ling-bane—suggesting the authors are enjoying the meta, and the agents have even been tested against the streamer Harstem.

Key highlights

• Pre-trained SL and RL agents are included; the RL models are claimed to reach master or grandmaster level on Zerg versus Zerg.
• Supports distributed training across multiple machines and GPUs, with a documented single-PC baseline for limited resources.
• You can play against the agent, watch two agents fight, or pit one against the built-in elite bot.
• Named RL models (Abathur, Brakk, Dehaka, Zagara) each favor distinct unit compositions and strategies.
• The evaluation build is locked to StarCraft II patch 4.10.0 because later balance patches significantly changed performance.

Caveats

• Only Zerg versus Zerg is available in the current pre-trained release; other races and matchups are not yet supported.
• The README notes that in rare cases two StarCraft II instances can lose connection during play, leaving the agent frozen until you restart.
• It targets Python 3.6–3.8 and PyTorch 1.7.1, which are now several years behind the current ecosystem.

Verdict Worth a download if you want a ready-made StarCraft II research opponent or a distributed RL training scaffold to hack on. Skip it if you need Terran or Protoss agents, or if babysitting legacy PyTorch versions sounds like a chore.