Jowa Football AI v3

The AI

Jowa Football AI v3 is a full 11v11 football match where all 22 players are controlled by AI. Two separate brains play each other: red always plays team 0 and blue always plays team 1. Each brain has its own network, its own training and its own pool of past opponents.

Each brain is a multi-agent PPO (MAPPO) model. One shared entity transformer drives all eleven players of its team: every player sees the ball, its ten teammates and the eleven opponents as separate tokens, and attention lets each decision take the whole pitch into account. A coach head picks the team instruction (block height and style), and a centralised critic scores the play during training.

The policy decides what to do — who presses, who covers, who runs in behind, whom to pass to and when to shoot — while a tactical layer executes the movement and keeps the team in shape. For every pass option the brain is given an estimate of its chance of success.

A real match: the red brain against the blue brain on the full 105 × 68 m pitch

Tactical overlay: defensive lines, anchors and every pass option coloured by its success estimate

Both pictures are real matches of the red brain against the blue brain with the default weights. The second one shows the tactical overlay (key O): the team lines, the formation anchors and every pass option, coloured by its estimated chance of success.

What's New in v3

From 6v6 to a full 11v11 match on a real 105 × 68 m pitch, 4-4-2 (4-3-3 optional)

AI rebuilt from DQN to MAPPO with an entity transformer and a centralised critic

Two independent brains, red and blue, each with its own model, optimizer and opponent pool

Learned coach and a tactical layer: build-up, attack, counter-attack, pressing, high / mid / low block

Laws of the game: offside, fouls, free kicks, corners, throw-ins, goal kicks and penalties

Ball physics with height: lofted passes, headers, bounces, shots over the bar; stamina for every player

Self-play against past versions, a seven-stage curriculum and crash-safe checkpoints with automatic resume

Ready-to-run Windows and Linux apps that ship with the default weights

Model & Game Info

Mode	11v11, red brain vs blue brain
Pitch	105 × 68 m · physics 20 Hz · decisions 5 Hz
Algorithm	MAPPO — PPO with GAE, value normalisation, self-play
Network	Entity transformer: 3 layers, d = 192, 6 heads
Parameters	About 3.1M per brain (actor 1.62M + critic 1.50M)
Observation	24 tokens × 56 features per team
Actions	19 per player (move, press, mark, cover, tackle, shoot, …) + pass target
Default weights	Red & blue brains at 52,783,408 training steps
Device	Apps: CPU · Source code: CUDA (GPU) or CPU
Build	2.1.0

The first 3M training steps imitate the built-in rule-based AI to get started; after that the brains learn only from playing. They are still learning: against the rule-based AI they lose more matches than they win.

Download

The Windows and Linux apps come as a ZIP with the default weights inside — the red and blue brains at 52.8M training steps — so the AI plays right away. The source code is the full project without weights.

ZIP · 208 MB · weights included Windows app Windows x86-64 · runs on the CPU Download ZIP · 294 MB · weights included Linux app ELF x86-64 · Ubuntu 22.04+, Debian 12+ Download ZIP · 154 KB Source code Python · no weights Download

The apps include PyTorch for the CPU, so no Python or GPU is needed; to train on an NVIDIA GPU, use the source code. SHA-256 checksums

How to Run

Windows or Linux app

1

Extract the ZIP

Keep the extracted folder together: the weights sit next to the app.

2

Watch the two brains play

Start the app from a terminal in the extracted folder with --play. On Linux, mark the ELF file as executable first.

# Windows
jowa-football-ai.exe --play
# Linux
chmod +x ./jowa-football-ai
./jowa-football-ai --play

3

Or let them train

Without options (or with a double-click on Windows) the app starts in normal mode: a visible match at real-time speed while both brains keep training from the default weights. New checkpoints, logs/ and screenshots/ are saved in the app's folder. --help lists every option.

From the source code

1

Install the dependencies

Python 3 with PyTorch (the CUDA build for an NVIDIA GPU), pygame and NumPy.

pip install torch pygame numpy

2

Add the weights (optional)

To use the default brains, copy the checkpoints folder from the Windows or Linux ZIP (latest_red.pt, latest_blue.pt and their .sha256 files) next to jowa-football-ai.py. Without it, training starts new brains.

3

Run it

# watch the red brain play the blue brain
python jowa-football-ai.py --play
# red brain against the rule-based AI
python jowa-football-ai.py --play --blue scripted
# visible training at max speed, resumes automatically
python jowa-football-ai.py --train

Controls

Match window

Take control of a player

TAB / ↑ ↓ select | Enter take control | ESC give it back to the AI

WASD / arrows move | Space shoot | F pass | T tackle