An idea conceived by its two promoters
Football played by autonomous minds.
A football competition in which 22 virtual players observe, decide and cooperate on their own. No one steers them during the match, and no one writes the result in advance. Every match is computed, not scripted.
Inventing a video is not the same as playing a match.
Today anyone can generate a convincing football clip. That is not what this project proposes. The difference decides whether a result can be trusted.
Generated or scripted video
- Someone decides what happens
- Images are produced to show it
- The "result" is whatever was written
Cannot be verified. Cannot be a sporting competition.
Computed match
- 22 players decide, each from what it can observe
- A simulator applies physics and the laws of the game
- The result emerges and is sealed in a record
- Only then are images produced from the record
Repeatable, auditable, and impossible to change after the fact.
A precomputed match is acceptable; a scripted match is not. This is the project's first approved architectural decision (ADR-001): simulate first, seal the record second, produce the visuals offline third.
02 How autonomous football works
Twenty-two decision-makers. One shared pitch.
Each virtual player runs the same loop roughly ten times per second. It sees the match from its own position, decides what it intends to do, and lets the simulator determine what actually happens.
The player receives only what it could observe: the positions and movement of everyone, the ball (including its height), the score, the clock and its own team's instructions.
- Autonomous
- Each player's action depends only on its own observation, memory, attributes, its team's instructions and its own seeded randomness.
- Cooperative
- Teammates follow a shared plan from the team manager. They may not read each other's decisions made in the same instant.
- Fair
- No player can read future states, the opponent's internal data or the simulator's randomness. This is tested, not assumed.
03 The five-layer production pipeline
From a decision to a broadcast, in one direction only.
Five separate systems, connected by data contracts the project owns. Select a layer to see what it is responsible for.
Binding rule The visualisation never alters the result. Images are produced from the sealed record; nothing downstream can change what happened.
04 Inside the AI Brain
A football mind in three layers.
The Brain is not one giant model. It is a hierarchy: a team manager that sets the plan, a tactical decision for every player, and a skill layer that turns intentions into precise actions.
Team manager
Formation, height of the block, pressing, substitutions
every few seconds and at stoppages
Player tactical decision
Offer, mark, press, which pass, when to shoot
≈ 10 times per second, per player
Skill layer
Turns the intention into the exact action the engine accepts
≈ 10 times per second; the only layer tied to the engine
The two upper layers are designed to survive a change of engine or animation system Engineering hypothesis
No large language model per player
Each player does not need a chatbot in its head. The first Brain uses rules and utility scoring, an approach that has played complete autonomous matches for decades in academic robot football (RoboCup 2D) Verified from sources. The 22 players may share one decision program; what makes them individual is what each one observes, remembers and is physically able to do.
- 1Rules and utility, in a hierarchyBase of the first product
- 2Value models tuned on the simulator's own matchesBetter passes and positions
- 3Imitation of positioning from real tracking dataHuman style; data with clear licences is scarce
- 4Multi-agent reinforcement learningOptional improvement; risk of exploiting engine flaws
05 The Interactive Football Lab
Watch the decisions, not just the movement.
Choose a situation, press play, and select any player to see what it observes, what it is weighing and what it chooses. Switch on the tactical layers to read the match like an analyst.
Paused at a decision. Read the chain below, then press Play.
decision with a full chain other event
Tactical layers
How to read the layers
- Passing options: lines from the ball carrier to teammates. Bright = open lane, dashed amber = tight, red = blocked by an opponent close to the line. "Chosen" marks the pass the scenario plays next.
- Free space: each zone is tinted by the team whose player is nearest: a simplified picture of who controls space.
- Pressing: the nearest defenders converging on the ball.
- Cover lines: the defending back line and the "cover shadow" behind each pressing player, meaning the lanes his body blocks.
- Positioning: dashed ring = the zone the team plan assigns; thin line = how far the player has moved from it.
- Movement routes: where each player goes in the next three seconds.
- The cone around the selected player is a 160° field of view, an optional limit that the candidate engine documents.
Brain → Simulator → Record
1Observation received by the Brain
—AI Brain
2Intention selected
—AI Brain
3Action sent to the simulator
—
Contract 1 · illustrative format
4Result calculated by the simulator
—Simulation engine
5Event registered
—
Contract 2 · illustrative format
The Brain only intends. Whether a pass arrives, a tackle succeeds or a shot is saved is computed by the simulator, then written to the record before any image exists.
Scenario positions, decisions, values and texts are authored for explanation only and are not measurements of any prototype. Geometric overlays are computed from those authored positions as visual aids. Players and teams are fictional. Keyboard: K play/pause · . step · R restart · arrow keys switch scenario tabs.
06 The simulation engine
The referee, the physics and the clock live here.
The simulation engine is the match itself. It receives 22 intentions and computes what really happens. It runs without graphics, repeats exactly with the same seed, and exports everything. Select a component.
Leading candidate engine
MESSI, footballworld line
Pending experiment Read, never run by us. A candidate, not an approved dependency.
Why it fits
- Players have body and gaze orientation as separate states: a player can run one way and look another Verified
- Actions are intentions (move, control, pass, shoot, clear, challenge) with continuous parameters, matching the proposed agent interface Verified
- 3D ball with drag, spin, bounce and goal-frame contact; physics at 80 Hz, decisions at 10 Hz Verified
- Exports tracking and typed events, a base for the match record
What it still lacks
- Fouls, cards and the winner of each contest are decided by a draw
- In the fouls file read, advantage is not applied and handball is not emitted
- No jump action (relevant for headers and saves)
- Its authors state it is not a laws-conformance product and that realism is not established
The repository contains three unrelated programs. Only the footballworld line (commit above) is the proposed candidate; the soccerworld line (b04b15c5…) is a different engine, and the main branch (722bc62a…) has no visible licence and is excluded. Alternatives studied: RoboCup 2D, Google Research Football (archived in August 2026), open-football, AgentPitch, or an own engine. Compare all engines in the Blueprint →
07 The authoritative match record
The match is decided before the video exists.
Every instant of the match is written to a record: where everyone was, what each player decided, what the ball did, what the referee ruled. The record is chained with cryptographic hashes, so any later change is detectable.
tick 412 · 00:41.2
Illustrative example, not output from any run.
Locked by the record
Positions · ball trajectory · contacts · decisions · referee rulings · score · clock
The 3D production must reproduce these exactly.
Free for the production
Body posture between contacts · gestures · celebrations · camera choices · lighting
Creative choices that cannot change the match.
For an official match Planning assumption: parameters to decide
- Registerthe match identifier
- Freezethe manifest: engine version, teams, rules
- Publishits hash before the match
- Derivethe seed from a future public randomness beacon
- Run onceno retries, no choosing
- Sealthe record and its hash chain
Experimental simulations can be repeated freely and never become official. The protocol only prevents "re-run until we like it" if the simulation is reproducible with the same seed, an open condition to verify Pending experiment.
08 The first real milestone
A five-minute, 11-versus-11 autonomous match in 2D.
The first thing worth showing is not a beautiful image. It is proof that our own Brain, our contracts, our record and our integrity rules produce real football, without anyone touching it.
It must demonstrate
- 11 v 11 for about five minutes, two halves with change of ends
- Every player, goalkeepers included, moved by our Brain through Contract 1
- Decentralised decisions that pass the four autonomy tests
- No human intervention after kick-off
- A result that was not predetermined, across a batch of 20 or more seeds
- Repeatable: same manifest and seed, same record hash
- Legal play with zero illegal states and zero deadlocks
- Organised football within bands taken from real matches, plus a blind review by three football-literate people
- Better attributes measurably favour the better team
- A 2D viewer that replays the match from the record alone
It does not yet demonstrate
- Anything about human bodies, animation or 3D
- A full 90-minute match
- A complete referee: fouls and cards would still be decided by the engine's draw, and declared as such
0–0
is the most likely score. At real-match pace, football averages roughly 2.7 goals per 90 minutes, so about 0.15 expected in five minutes Approximate. Honest football over five minutes is quiet; the plan is to show a batch of matches.
How it will be evaluated
- Autonomy
- Re-evaluation, forbidden-field audit, perturbation test, per-player state isolation.
- Repeatability
- Identical canonical hash for the same manifest and seed.
- Football quality
- Structural metrics against real-football bands and a blind human review, with thresholds declared before the run.
09 The development roadmap
Twelve steps, each one earned by evidence.
Progress is driven by verified results, not by the calendar. Each stage has a deliverable, its dependencies and an acceptance test declared before it starts. Select a stage.
No stage is authorised by this document. Indicative durations exist only for the early stages, are planning assumptions, and are documented in the Blueprint. Beyond the first 2D match, no duration is given: it depends on measurements that do not exist yet.
10 From simulation to human motion
The hardest problem is the foot on the ball.
A simulation can calculate that a player touches the ball at a precise point and instant. Making a human body arrive there, with a natural stride, without sliding feet, at exactly that moment, is a different technological problem. It is the project's second decisive risk.
Three ways to connect bodies and simulation. None is selected yet.
Stage 4a, an animation feasibility study on real match records, will decide with evidence. If the choice moves away from abstract bodies, the engine tests and the Brain's evaluation ladder are repeated on the new engine. Architecture detail in the Blueprint →
11 The visual production pipeline
Software engineering meets the craft of animation.
Turning a record into credible 3D football requires twelve components. Some are software engineering; several depend on specialised animation expertise that a general developer (or a video editor) does not have. Select a component.
The critical, scarce profile is a lead technical animator with Unreal Engine, Motion Matching, inverse kinematics, rigging and sports animation experience. No single profile solves the visual problem on its own: an editor does not replace the technical animator, and the technical animator does not replace motion data.
12 From a match to a broadcast
A director's cut of a match that already happened.
Because the match is sealed first, the production can be planned like a film: every camera knows in advance where the action will be. Select a camera to see its role.
- Match Record
- Camera Planning
- 3D Animation
- Stadium
- Render
- Commentary
- Scoreboard
- Highlights
- Final Broadcast
Generated commentary would be produced from recorded events, and an automatic check would compare the final video against the record: zero discrepancies is the acceptance test for a publishable pilot.
13 Engineering gates
Tests that decide whether to continue, change course or stop.
Thresholds are declared and dated before each stage runs, and are not changed after seeing results. Two technical stop points and three gates protect the project from building on assumptions.
1Stop point after the Brain
Is it football?
Do rule-based agents produce organised, credible play?
- Evidence
- Scenario ladder, real-football bands, blind review by three people.
- If not
- Stop, or fund specific research on intelligence.
AGate after the 2D match
Open the body question
Did the five-minute match pass all ten criteria?
- Opens
- Only the animation feasibility study and the consolidation of the full-length simulation.
- If not
- Iterate within the declared maximum effort, or stop.
BGate after feasibility and consolidation
Allow one small visual test
Is there an architecture with evidence in its favour and reachable motion data? Does a full simulated match hold up?
- Opens
- Only a 20 to 30 second visual test.
- If not
- A stylised product, or stop.
2Stop point at the first 3D test
Does the body obey the record?
A randomly chosen sequence with reception, turn, passes, shot and save.
- Evidence
- Foot-to-ball distance, timing error, foot sliding, interpenetrations, without frame-by-frame manual fixes.
- If not
- Change architecture.
CGate after the first 3D test
Consider larger production
With measured production effort per second of video and per frame, can a short match be planned?
- Opens
- The short 3D match and later stages.
- If not
- Stay at 2D or stylised.
No gate renews automatically. Mandatory order: validate the simulator → build the Brain → prove the 2D match → assess animation feasibility → one small visual test → only then, larger production.
14 The autonomous league
From one match to a living competition.
The promoters' vision: fictional professional clubs, persistent autonomous players with their own attributes and careers, full seasons and championships, with a governing function that protects sporting integrity.
Promoters' initial proposal Future vision, not built
8 founding clubs, designed to grow to 16 or 20. 22 registered players per club. 14 league rounds. 56 regular-season matches. A top-four playoff, 59 matches in all, and no relegation in the first season.
Season table template: no match has been played
| # | Club | P | W | D | L | GD | Pts |
|---|
Original identities, never real teams or players.
Persistent individuals; attributes fixed within a season.
Every number derived from sealed match records.
Scheduled premieres of precomputed, sealed matches.
Championships, playoffs and, later, promotion or expansion.
A repeatable league requires producing matches without manual correction, a capability that cannot be defined until the earlier visual stages are measured.
15 Explore the engineering
Every claim here has a source, a hypothesis or a test behind it.
The Engineering Blueprint documents each decision in depth: architecture, contracts, engines, the Brain, autonomy tests, the match record, risks, acceptance criteria and pending experiments, with an executive explanation and an engineering deep dive for each.
Explore the Full Engineering Blueprint