The model does this by taking into account common tennis strategies and the tendencies of each player. Through its training, it knows, for instance, that Djokovic likes to hit the ball toward his opponent’s weak side. Similarly, the model also considers how players tend to position themselves while waiting for their opponent to return the ball. It will consistently place Federer closer to the baseline than Rafael Nadal, reflecting how those two play the game in real life. According to the team that created the AI, it’s this behavioral aspect that separates their project from past attempts to create a system that can simulate tennis play.

Those capabilities allow the system to create potentially endless what-if scenarios. It can generate footage of Federer playing against himself or Serena Williams. It can even extrapolate how a match may have played out differently had a single shot landed in a different location. What’s more, the system allows you to control a player’s shot placement and recovery position, so there’s the potential a studio could adapt it for gaming.

Of course, it’s not perfect. While the researchers did their best to hide potentially distracting visual artifacts like changing lighting and player clothing, there are moments where the clips look more like they’re ripped straight from a 90s FMV game. As you can see from the clip above, fans and officials don’t move at all. Another factor that breaks the illusion is that the ball and players don’t cast shadows. It also looks like they’re skating across the surface of the court. Still, it’s an impressive system that could have a lot of fun applications.