Who is the best tennis player in all history? Many may bet on Federer, others on Pete Sampras and others on McEnroe, but what if they could face each other when they were in the best sporting moment of their life? That is what this AI allows us to do.
Playing at being a tennis god
Researchers from Stanford University have created an artificial intelligence that is capable of generating Wimbledon matches with the help of image resources recorded in previous years. Thanks to this, the system (called Vid2Player) is capable of emulating matches, allowing players to face that never did in the tournament, and even being able to change the course of history by taking the ball inside the capo at points that meant the Full Time.
Artificial intelligence has been trained so that, with the help of all the graphic material obtained from the broadcasts, it is able to create a statistical model in charge of predicting the movement and hits of certain players in random situations.
Thus, for example, they have emulated a match between Roger Federer and Serena Williams in the green of London, something that has never happened and that in the video you can see that it happens quite realistically (leaving aside the graphic quality of the sprite from Serena).
Everything is based on statistics
In the end we are just a simple number. The system is based on the way each tennis player plays, statistically predicting what type of hit and direction he will apply to the ball. And not only that, since it will also place the player on the field according to their way of playing, closer or further from the net according to the profile of each one, and will even return the ball according to their way of playing.
What would happen to Federer playing against himself? That is exactly what this technology offers us. What if we could define where to direct the ball? This last idea obviously makes us think of a video game that allows us to have control of the players in the purest Virtua Tennis style. It is something that has not yet been seriously considered, but that would undoubtedly seem an extremely interesting application and that would succeed.
Much remains to be improved
But as we have commented previously, visually it is not perfect, and the system returns visual errors that are too distracting and break that bar of the real and the simulated. Anyway, this is simply due to the way the original images were captured, so with the help of controlled conditions the results could be infinitely superior, and it would not look like an old cinematic game of which we were able to test some in the era of the Laser-Disc.
