Can we teach a robot to be as good as Messi or Ronaldo?

Messibot or Ronaldobot approaches the area dangerously, shoots on goal and … gooool! This could be the predominant scene in the dreams of a team of engineers from the California Institute of Technology (CalTech). His research focuses on designing algorithms that know both football and Leo Messi and other superlative quality players. If they were used in a robot, silicon pichichis could be manufactured.

Algorithms that learn how to play a good game

The objective pursued by these experts in artificial intelligence, supported by Disney itself, still poses great technological and scientific challenges.

However, they have discovered a method for algorithms to learn the ingredients of a good match.

The team, led by CalTech Hoang Le researcher , uses machine learning techniques so that artificial systems assimilate football details and make predictions about the best game strategy.

The novelty of their work, which they detail in a recent study, is that they develop algorithms that they learn on their own. Robots have plenty of human lessons.

Learning systems alone

The traditional strategy in machine learning is to control algorithms by defining a series of rules that must be followed and goals that must be achieved. Its developers support the baton throughout the process, guiding the training and evolution of their particular army of technological soldiers.

However, in the face of this controlling approach, the new fashion in artificial intelligence is called ‘unsupervised learning’ unsupervised learning, in Spanish). A scheme that is based on allowing algorithms to run at ease without marking the path. This is the modus operandi that Le and his team have decided to adopt in their studies.


The revolutionary strategy can not only serve to generate new products, software tools and drive technological innovation. It could even become the foundation to make robots as self-sufficient and smart as those that appear in science fiction movies come true.

From basketball to football

Actually, researchers have taken the idea of ​​another sport: basketball.

In 2013, the Toronto Raptors coach, an NBA team, decided to analyze the images captured by cameras in games to study the behavior of the players. It had a team of experts that developed a program to detect the movements of athletes and identify their skills.

But the software went further: it was able to predict what each player was going to do based on previous observations.

In the videos generated by the tool, these predictions were represented as shadows that accompanied the figures of the athletes. That is why they were baptized as ‘ghost players’ and their idea as ‘ghostling’ (from ‘ghost’, ghost in English).

The analyzes served the Toronto Raptors coach to detect failures and gaps in both the defense of his team and that of the opponents.

However, the program had some limitations. Although he could overcome a human being when analyzing the plays, he could not invent new interpretations of the movements of athletes.

The CalTech experts and the Disney research division found the idea of ​​Canadians very interesting, but they realized these weaknesses. That is why they decided to improve the tool and expand their capabilities using ‘machine learning’ tools.

Unleash algorithms

Although scientists have long been interested in unsupervised learning in artificial intelligence systems, their results are too vague to be useful in solving problems. It would be like releasing a hound without telling him the trail he should follow.

field_robots_futbolistas_caltech: disney

However, the ability of these tools to identify patterns on their own from data can provide interesting findings in some cases. Using them is the only way to get those results that their creators would never look for, but which can be extremely relevant.

CalTech researchers began by collecting a sufficient amount of data on football.

Later, they designed a machine learning system that could dive into such an ocean of information and find some hidden treasure. The bad thing is that they could not teach you previously what were the different types of players or how attacks or defenses develop.

Soccer self-taught

But the artificial intelligence software did not need previous indications. He was able to categorize soccer players by analyzing their real movements. That is, despite the fact that the program does not know what a right end is, it can be realized that an eleventh part of the detected points is very close and that they follow different patterns from the others, which can also be grouped according to their distribution.

In addition, the tool is not confused by changes in positions that occur when players move across the field or concentrate on the lines of defense or attack.

Thus, their algorithms would quickly learn the strategy of a new team from a series of disordered data.

The coaches could use the artificial intelligence system to analyze the matches and disengage the tactics of their opponents.

They could also show their players how they should act to improve or what they should have done to win their last game.

And the creators of the system believe that, with the right data, the tool has possibilities in other sports such as basketball and hockey.

Robot 1 – human 0

The potential of machine learning based on unsupervised learning goes beyond football and basketball. Another example of their ability to surprise their own creators is AlphaGo . This artificial intelligence system created by Google has managed to defeat the millennial playing the game go to various human opponents using moves that no one had taught him.

Analyzing data on football matches, the tool conceived in CalTech was based on factors such as team distribution and coordination between players from different positions.

Although they are not yet perfected enough, their algorithms, such as those of AlphaGo, could generate new movements and strategies for the teams.

This technology could also be used so that a robot traveling to Mars can explore its craters autonomously or improve the ability of drones to avoid obstacles.

But, returning to the field, it may not be convenient for flesh and blood players to improve their algorithms. The speed and leg play of Messi or Cristiano Ronaldo would not be enough to defeat a team led by an intelligent robotic trainer. And less to one composed of silicon athletes.