The National Football League (NFL) has given researchers at Carnegie Mellon University access to official game footage and data on player movement on the field. The goal is to develop artificial intelligence (AI)-based technologies that will help coaches, players and fans annotate, analyze and better understand the game. Tech Xplore reports this.
The project is led by Chenyan Xiong, an associate professor at the Language Technologies Institute of the Kenneth C. Griffin School of Computer Science, and Ron Yurko, director of the Carnegie Mellon Sports Analytics Center (CMSAC). Xiong describes himself as a football fan, and Yurko is an assistant teaching professor in the Department of Statistics and Data Science. Together they are building an AI model that learns to understand American football.
"Access to official NFL film is an incredible opportunity," Yurko says. "We are the only institution conducting research with this data." Yurko had previously worked with the league: his work was incorporated into the NFL IQ AI chatbot, presented at the 2026 NFL Draft. In his work at Carnegie Mellon, Xiong creates AI technologies that analyze large volumes of heterogeneous data to solve problems in healthcare. His interest in football prompted him to build a tool for the NFL as well.
The researchers applied to the NFL Forward Initiative program, which allows entrepreneurs and scientists to propose technological solutions for competition, training, player development and popularizing the sport. The application was accepted. The NFL handed over several terabytes of data to the university: game recordings and the Next Gen Stats system - player tracking data collected using radio-frequency identification (RFID) chips embedded in shoulder pads, equipment and the ball itself. These chips transmit information about the location of objects on the field in real time.
"It is an honor for us to work with such a prestigious institution as CMU, which is looking for ways to make the game more effective," said Mike Lopez, the NFL's senior director of football data and analytics. "Innovation is the foundation of the game's development, and we look forward to the results."
The volume of data required the use of so-called foundation models - deep neural networks (multilayer algorithms trained on large arrays of information) that are capable of solving a wide range of tasks. The more data for training, the more the model can absorb and the more complex the tasks it can handle. However, before combining video and tracking, the scientists created two pre-trained systems - essentially building blocks for a future large model. One learns from game recordings, the other from data on player movement. Both have the same task: to understand what is happening on the field.
The next stage is more difficult. "For example, after understanding what happened in a play, can the model predict events before the next play? Will there be a blitz? Will the safety shift to a different coverage zone?" Xiong explains. A blitz is a massed defensive attack on the quarterback, and a safety is a defensive player responsible for the deep zone. For the model to learn to work with video and tracking at the same time, more data and time may be needed. But in the long term this will yield more accurate forecasts and efficiency: for example, the model will be able to simulate the development of a play under certain decisions or propose tactical combinations based on evaluated outcomes.
"This will help coaches and scouts focus on what matters most and will make the league more effective," Yurko believes. For now this is a research project, not a ready-made product for teams or fans. No timeline for a commercial version has been announced.