Title:	MARS: Multi-Agent Representation learning for Sports data
DNr:	Berzelius-2025-179
Project Type:	LiU Berzelius
Principal Investigator:	Atsuto Maki <atsuto@kth.se>
Affiliation:	Kungliga Tekniska högskolan
Duration:	2025-06-01 – 2025-08-01
Classification:	20208
Keywords:

Abstract

This project seeks to advance automatic interpretation of football tracking data using machine learning and computer vision techniques. Our goal is to develop foundation models for football language that can understand game dynamics, focusing on modeling the complex spatial and temporal correlations of 22 players and the ball on the pitch. We plan to employ self-supervised learning on large graph structures that map the spatial and temporal connections between players' joints and the ball. Given the scale and complexity of this data, substantial computational resources will be required to support effective training. At the current stage, we aim to develop a self-supervised representation learning technique for multi-player skeleton tracking data. We plan to explore self-supervised learning methods such as player joint position prediction and motion prediction, aiming to maximize the utility of existing data for football understanding. We will expand our research to cover a broader range of tasks that evaluate the model’s capability in understanding football dynamics. These tasks include player motion prediction, shot detection, throw-in recognition, and further improvements in expected goal estimation. This project is related to TRACAB at EA Sports and KTH, and partially supported by the Wallenberg AI, Autonomous Systems and Software Program (WASP) funded by the Marianne and Marcus Wallenberg Foundation.