This artificial intelligence generates surreal tennis matches

Little by little, major sporting events begin to wake up after a few months of hibernation. But that will not make us forget the great competitions sacrificed on the altar of the coronavirus, starting with the 2020 edition of the Wimbledon tournament, the mecca of lawn tennis. But let the fans rest assured: thanks to researchers at Stanford University, it is still possible to have your dose of tennis. This team had fun simulating the tournament with the help of a machine learning. To do this, they have built a large database of annotated video clips of the best players in the world. It was then used to create a statistical model, in order to train a neural network to understand the way they play… and to simulate it! Judge for yourself with this demonstration video featuring Novak Djokovic, Serena Williams, Rafael Nadal and Roger Federer.

The cyclical nature of tennis has made it possible to produce a fairly comprehensive algorithm without much effort. For example, we know that a player will hit the ball, get back, wait for the opponent’s strike, then start over. The system was therefore able to focus on points such as the player’s position, his speed and posture, his way of hitting… And we must admit that these trends are quite realistic! We can see that Nadal is systematically placed further from the line than Federer. The latter also gratifies us with his famous one-handed backhand. Djokovic’s numerical double sends chestnut to chestnut on his opponent’s weak side.

Make your wildest sports dreams come true

With a system like this, the possibilities are endless. For example, what tennis fan has never dreamed of a Williams final against Federer? The authors indulged in the experience and these few extracts really make you want to see an entire game. This system would also make it possible to replay iconic matches which could have ended very differently. For example, what would have happened if the sulphurous John McEnroe had not lost his means against Ivan Lendl, in the final of the Roland Garros 1984? Failing to be able to answer them, this system could allow us to simulate some possible scenarios.

The logical continuation of the data in the sport ?

We can also imagine other applications. For example, players could use such a system to analyze a confrontation a posteriori. It could be a formidable tool for debriefing a game or analyzing a bad habit. A coach would only need a few minutes to know if this move was the right one, or if this positioning would have worked better. There is also the obvious advantage of being able to simulate the possible response (s) to different strokes: this would allow the player to arrive on the court with a significant strategic advantage. One can even imagine that this system could be applied to other sports! However, there is one caveat: if the example of tennis works well, it is above all thanks to its cyclical nature. Applying it to more “free” sports will require overcoming new obstacles.

A potential gold mine for video games

Predictive systems of this kind could even be imposed in the world of video games to animate characters. Currently, most sports or fighting games are based on a mixture of two principles: a part of classic animation, and a part of procedural animation. The first is often carried out on the basis of motion capture, where the movements of an actor are digitized with specialized equipment. The second, on the other hand, is generated in real time by the software from rules specified in advance.

By mixing these two aspects, we can therefore have a realistic movement (mocap) but reactive, organic and adaptive (procedural). A system like this could do wonders in sports games based on this hybrid approach, like NBA 2K or FIFA. Admittedly, there are still several small bugs; special mention to the lighting and to the feet that float above the lawn! But these flaws will certainly be corrected fairly quickly, and we are already looking forward to seeing how this system will be used in practice.