Footbar's generative artificial intelligence comes into play

Published on March 29, 2024

At the start of 2024, more than a year has passed since generative AI applications in the field of soccer were first introduced. For example, the match summary paragraph provided in our mobile app after each match has been written since March 2023 by Open AI's GPT 3.5 model, which synthesizes the performances measured by the tracker Footbar.

However, the model presented today by Footbar is radically different. It has all the characteristics of a classic language model, but unlike most of them, it doesn't deal with text, but directly with the content of a soccer match. It can, therefore, generate matches, just as the main templates generate text.

In this article, we explore how this technological feat was made possible, and the applications it opens up.

You'll also find a link at the end of the article to join the waiting list, so you can test this model in advance.

The technical solution

The two essential ingredients for the creation of a language model are: a good architecture based on transformers, and a dataset that is both qualitative and quantitative. At Footbar, the architecture is a classic transformers-based model, capable of estimating the most likely token to complete the sequence provided as input. The creation of the token, on the other hand, required more engineering, since in this case it's not a question of cutting out a text, but of sequencing the practice of soccer.

The dataset is the fruit of almost a decade's work by the Footbar team in collecting and annotating data.

A tokenizer that sequences human activity

The main challenge in adapting LLM architecture to soccer was to create a tokenizer capable of interpreting physical movements rather than words. Here are the pillars of Footbar's approach:

  • A tracker activity monitor strapped around the leg, recording movements over time.

  • Movement sequencing: Unlike conventional LLMs, which are based on a breakdown of text into words or groups of characters called tokens, our technology sequences each distinct stride or group of strides.

  • Encoding and embedding: These tokens are not only encoded according to their category (walking, running, technical gesture etc.), but also receive position information, as is the case in traditional LLMs, as well as other unique data points such as the speed, direction or time of flight of each stride, enabling a complex understanding of the player's movements.

  • Efficient vocabulary and model size: Unlike traditional language models, which handle a vocabulary of hundreds of thousands of words, Footbar's model benefits from a much smaller token set of just a few dozen actions. This efficiency translates into a model that requires fewer parameters and a more compact data set, without compromising the depth and quality of the analysis.
A rich and diversified data set

Footbar's technology is powered by a vast dataset comprising around 1 million matches played by a diverse range of players using Footbar's tracker activity over the past 10 years:

  • Diversity of players: The data set includes professional athletes, amateurs, children and adults, both male and female, covering a wide range of skill levels and playing styles.

  • Different game formats: from 5-a-side to 11-a-side, from training exercises to informal practice, the data set covers all forms of soccer.

  • Comprehensive data collection: Traditional data collection in professional soccer often comes up against limitations, either because it consists of positional tracking data, without detailed technical actions, or because it focuses only on the players directly involved in the actions. Footbar's dataset overcomes these obstacles, providing a comprehensive view of matches and player interactions.

Applications

If the first concrete use of this language model is to power Footbar's services, many other applications can be imagined, categorized according to the two axes that have been Footbar's DNA since its creation: helping all footballers, from the field to the professional training center, to understand their performance and progress, as well as contributing to making them want to play more often thanks to a playful interface, inspired by video games and which reinvents the way of practicing this sport.

Footbar takes a look at the transformations that will be taking place in soccer over the next few years.

Performance enhancement
  • The coach's tactical companion : A ChatGPT-style command prompt allowing each coach to enter a tactical setup, configure a few parameters and receive a simulated match result. This feature enables coaches to strategically analyze the impact of selecting different players or using varied tactics, effectively simulating scenarios before actual matches.

  • Game strategy simulator: Players and coaches can experiment with different set-piece scenarios, refining strategies for corners, free kicks and penalties. This tool enables detailed analysis of potential outcomes, increasing the probability of scoring.

  • Modeling recruitment scenarios: Recruiters can simulate an entire season with different players on their team, providing a comprehensive forecast of each player's potential impact. This enables data-driven decisions to be made to strengthen the team.

  • Injury prevention and physical preparation: By exploiting the training load measurements collected by our sensors, physical trainers can anticipate and optimize players' physical condition, in order to optimize performance and reduce the risk of injury throughout the season.

  • Improved betting insights: brokers and betting enthusiasts can use our LLM to get more accurate predictions of match outcomes, reshaping the betting landscape with advanced AI insights.
Gaming experience
  • Simulation model: The traditional engines used in soccer simulation games and fantasy soccer leagues can be replaced by Footbar's LLM, offering an unprecedented level of realism and engagement.

  • A World Cup with no carbon footprint : Players from all over the world can play on their local pitch, while our AI generates a coherent virtual match, enabling them to compete as if they were in the same stadium.

  • Rewriting historic soccer moments: Our technology allows users to virtually participate in historic soccer matches, offering a unique opportunity to change the outcome of these matches.

Become a player in the AI-powered soccer revolution

There are 3 ways to become a pioneer by testing this tool yourself.

First of all, if you are a player, coach, physical trainer or have any other connection with soccer, you can purchase our tracker, either individually or for a group, directly from our online store.

If you're a designer of a product or application aimed at sports enthusiasts, you'll no doubt be interested in our our API which will enable you to design applications based on players' statistics.

Finally, if you are only interested in the language model aspect and want to test its capabilities, you are invited to join the waiting list to be the first to know when it will be made available to the general public.