How FC creatives made the most human-like goalkeepers

A last-second dive. A fingertip save that changes the match. A keeper reading the angle, adjusting in a split second, and reacting with instinctive precision.

In EA SPORTS FC 26, goalkeepers move with more human-like behavior, responding dynamically, adapting faster, and delivering moments that feel authentic and true-to-life.

The leap forward is powered by an innovative and creative solution developed by team members at EA SPORTS, Frostbite (which provides the model training integration and runs the model during gameplay), and the Search for Extraordinary Experiences Division (SEED), a pioneering group within Electronic Arts, combining creativity with applied research.

Let’s dive in.

“The results are the most life-like goalies in FC history, built entirely on an innovative and creative solution crafted in-house across teams of experts at EA.”

A designer-first approach to machine learning

By leveraging deep reinforcement learning, the team enabled faster training and more human-like behavior compared to traditional methods, improving realism while reducing training time.

“We used machine learning to train them in hundreds of thousands of in-game situations to find the best positioning, including those little micro steps world-class keepers use to close the angle,” says Mike Jones, Sr. Software Engineer, Electronic Arts. 

"The method utilizes a reinforcement training agent that learns to play the game by itself through a new training framework that was developed with the purpose of being data efficient,” says Alessandro Sestini, Research Scientist at SEED. “And finally, an evaluation and fine-tuning framework to allow designers feedback on the goalie behavior.”

Compared to the traditionally coded goalkeeper, the new system delivers clear improvements.

The reinforcement learning-driven goalkeeper achieves a 10% improvement in ball saving rate, trains 50% faster than standard reinforcement learning methods, and can be trained overnight. A robust validation system featuring over 300 “unit-test” scenarios ensures continuous evaluation and tuning.

The reinforcement learning approach produces more adaptive positioning, improved angle coverage, and behavior that feels more natural to the real sport.

The results are the most life-like goalies in FC history, built entirely through an innovative and creative solution crafted in-house across teams of experts at EA.

“We used machine learning to train them in hundreds of thousands of in-game situations to find the best positioning, including those little micro steps world-class keepers use to close the angle.”

More to come

The introduction of reinforcement learning-driven goalkeeper positioning in EA SPORTS FC 26 represents more than a single system upgrade; it reflects an evolution in how gameplay can be built, trained, and refined in production.

By combining machine learning with a designer-first framework, teams are able to iterate faster, validate continuously, and deliver more human-like behavior directly into player-facing experiences.

For players, that means smarter reactions, better positioning, and moments that feel closer to the real sport. For developers, it signals a scalable approach to applying machine learning in ways that are practical, measurable, and built for live production.

And for EA SPORTS FC, it marks another step forward in delivering more authentic and engaging football experiences.

Check out other amazing EA stories on ea.com/news.