HomeOur Blog

Two Decades of Robot Soccer, +8400 Matches, and a Record Nobody Can Edit

Share to Socials

At six in the morning UTC on November 4, two autonomous soccer teams named Versus Projects and Wayzz played a match on FoxLeague, and Wayzz won five goals to nil. Eleven minutes after kickoff the broadcast ended, and later that day a file of roughly 830 bytes was written to Auto Drive that will still be there, unchanged, long after the tournament, the teams, and the website have gone.

The file names the teams, the score, the competition, the kickoff time, and the exact fingerprint of the 23 megabyte replay, so anyone who downloads that replay from FoxLeague can prove it is the one the match actually produced, without an account, without special software, and without asking Foxsy AI to vouch for anything. The league has now done this for every match it has ever aired.

It started as a graduation project

The work behind FoxLeague is older than most companies in this industry. In July 2002 a student named Sebastian Marian started a team called Oxsy as a graduation project in multi-agent systems in the Department of Computer Science in Sibiu, Romania, and he kept at it after graduating, entering the RoboCup Soccer Simulation League year after year. By the time Oxsy filed its 2023 team description paper, it was preparing for a nineteenth consecutive appearance. RoboCup is the research competition where autonomous systems play soccer against one another, organized around a stated long-term goal of fielding a robot team capable of beating the human World Cup champions, and nothing in it is scripted, so a team wins only because its code reads a situation it has never seen and responds better than the other team’s does.

Foxsy AI grew out of that lab. In the company’s own description:

Foxsy AI is building the evolution engine for robotics. Through FoxLeague, autonomous competitions become a live evolution loop: every match played by the community produces competition data and makes the next generation of intelligent machines smarter. Foxsy AI grew out of Oxsy, a RoboCup team since 2003, and is taking that engine from simulation toward real robots.

The practical form of that is a public league. Developers upload a RoboCup-compatible team binary through the Foxsy AI Developers Hub, run a verification match against a base team, review the replays and logs, and then enter tournaments that are broadcast live while the community watches.

Why a league cannot be its own witness

Every competition keeps two things, the result and the recording that shows how the result happened, and in almost every league both of them sit on servers the organizer owns. The arrangement is fine while nobody is suspicious, and it has no answer at all for the first person who is, because a score in a database can be corrected quietly and a replay file can be re-encoded, trimmed, or swapped without leaving a mark that a viewer could find. An organizer’s assurance that nothing was touched is not evidence, however honest the organizer happens to be.

In a league of human players that is a governance problem, handled well enough by referees and reputations. In a league of autonomous ones it turns into something closer to a scientific problem, because each match records how a set of independently written systems behaved under conditions nobody controlled, and those records feed the design of the systems that come next. Rather than ask anyone to take its word for it, Foxsy AI put the evidence somewhere it cannot reach.

Live in a day, complete in a week

Records ran on a test network on November 3 and went live on the production network the next day, with seventy-five matches written in the first two and a half hours. By November 10, more than 8,400 records covered the complete history of everything FoxLeague had aired, with none pending and none failed.

A record is written only after the live broadcast ends, which keeps provisional results out of the public history. Here is what one holds, using the November 4 match as the example:

‍

The record itself is public and readable right now at gateway.autonomys.xyz. Every match page on FoxLeague carries a “Permanent record on Autonomys” link beneath the replay.

Checking a replay takes three steps: download the replay log from the match page, compute its SHA-256 fingerprint (shasum -a 256 on macOS or Linux, Get-FileHash on Windows), and compare that fingerprint and the file size against the record. Matching values mean the file in your hands is the one that existed when the broadcast ended. The record behind that check cannot be rewritten, because a file written through Auto Drive is addressed by its own content, so an altered version would resolve to a different address while the original stayed exactly where it was put. Foxsy AI can add new records as matches air, but changing the ones already written is not something it or anyone else can do.

The part that matters in ten years

What exists today is a set of small files describing replays. The next phase stores the matches themselves. From November, the full replay file of every tournament match FoxLeague broadcasts and every 1v1 challenge, close to 300 matches a day, is written to Auto Drive once its broadcast has ended, linked to the match’s permanent record. That includes every Champions FoxLeague match, from the first round on December 16 to the final on December 24. After that Foxsy AI intends to assemble an open match dataset for AI and robotics research out of the history the league has already accumulated, while continuing to move the engine from simulation toward physical robots.

That dataset would be a continuous, time-stamped account of how thousands of independently written autonomous systems behaved in a competitive environment, produced by matches nobody could arrange in advance, with every entry fingerprinted before anyone had a reason to argue about it. Material like that is rare, and it grows rarer as the share of machine-generated content rises, since the cost of producing a convincing fabrication keeps falling while the cost of proving something genuine does not fall with it.

A corpus assembled after the fact inherits that problem, because the only thing standing behind it is the word of whoever assembled it, while one whose every entry was committed to a public record on the day it was created does not. That is the quiet value of what is being written now. The records are small, cheap and individually unremarkable, and their worth lies in when they were written and in the fact that nobody can adjust them later.

Why this integration matters to us

Sebastian Marian has been on the same problem since a graduation project in 2002, which is longer than most companies in this industry have existed, and he had permanent records running on the live network for the league’s entire history before anyone asked him for it. Patience measured in decades and delivery measured in days do not often turn up in the same team. Sebastian has been a pleasure to work with throughout, and we are looking forward to the next phase as much as to this one.

Auto Drive’s role is narrow and permanent. It keeps each record exactly as written, for as long as anyone cares to check it, with no pinning, no renewal cycle, and no subscription whose lapse would quietly erase the archive. As Foxsy AI moves from describing replays to storing them, and from stored replays to an open research dataset, every one of those steps depends on the records staying exactly as written.

The robots are still playing in simulation, and by the time they are not, there will be a verifiable record of every match that got them there.

We look forward to sharing more about Foxsy AI and the team behind it as the integration deepens, and we are proud to provide the permanent storage holding the record of their work.

Auto Drive is free to start at ai3.storage.

‍