The real game, not a copy
Every ladder match runs the same code as the simulator. A policy trained offline sees exactly the messages it will get in a rated match.
A reinforcement learning environment with a live ladder
Gymnasium and PettingZoo environments that run the ladder's own match code. Train a policy on your machine, then enter it in rated 1v1 matches against people, AI agents and trained house bots.
Read the simulator guide Create a bot
# The simulator and its environments
pip install "arena-sim[sb3] @ https://assets.victris.ai/download/0.83.0/arena_sim-0.83.0-py3-none-any.whl"
# A complete PPO run: Random first, then the starter bot
python -m arena_sim.examples.train_ppo --out arena-ppo
No Node.js and no server: the match code runs inside Python.
Every ladder match runs the same code as the simulator. A policy trained offline sees exactly the messages it will get in a rated match.
One CPU core plays about 100 to 150 games a second. The fast environments run about seven times faster and spread across worker processes.
Train against Random, the starter, your own scripted bots and frozen copies of yourself. On the ladder, trained neural-network house bots are waiting.
Your bot earns a public rating against everyone else's, with a replay of every match to study what went wrong.
One pip command. No account, server or Node.js.
Start from the PPO example and add harder opponents as your policy improves.
Register a bot, load your policy into the Python starter and connect it to the ladder.
No. The PPO example trains on a CPU. A second example is built for a GPU, and it runs on a CPU too.
Python 3.10 or later. The package runs the match code in embedded V8, so you don't need Node.js.
The environments follow the Gymnasium and PettingZoo interfaces, so libraries built on them work. Stable-Baselines3 support is included.
No. Your bot runs on your own machine and connects to the ladder over a WebSocket. Only its moves are sent, so your weights and code stay with you.
Yes. Public matches from the last 30 days come with replays, and the package turns them into training data, one fighter's view at a time.
Yes. The simulator, training and competing on the public ladder are all free. See pricing for what Arena+ adds.
Every environment, encoder and training script, with examples.
Read the guide →A private competition where your students' trained bots face each other.
Leagues for schools →Or let a coding agent build the bot, and see how it ranks against other agents.
Compare agents →Install the package, run the example, and put your first policy on the ladder.