Now Google’s AI also plays football

Now DeepMind plays soccer. The artificial intelligence of the British group controlled by Alphabet, that is, by Google, has been put to the test with a different commitment, abandoning for a moment the glorious successes in games of chess, Go and video games.

The group has indeed published a paper and a post on his blog in which he explains in detail a new method of AI “training”, the so-called “npmp”, which stands for “neural probabilistic motor primitives”. Through this method an artificial intelligence can learn to intervene and operate on physical bodies. In short, moving agents in real life. But in this case it is still, for the moment, simulations on animated humanoids.

According to the post published by DeepMind, which tries to make the news clearer, “an npmp is a control module general-purpose engine that translates short-horizon motor intentions into low-level control signals, and is trained offline or via reinforcement learning (RL) by mimicking motion capture (MoCap) data, tracker-recorded on humans or animals that they carry out movements of interest “. In short, through this mechanism, the system has been trained to make simulated humanoids, therefore virtual, play two-on-two football. The videos released by DeepMind are indeed impressive: in just three days of learning to reinforce actions and movements they look like those of a football video game. At the beginning artificial intelligence is practically unable to move avatars on the field but by awarding a reward every time a goal was scored, the model was able to start getting the robots up and running, dribbling and shooting in 50 hours of work. After 50 days, however, the attitudes of the “virtual players” have become much more refined, almost tactical, based on anticipating the consequences of their actions.

It is precisely one simulation: To control robots around the world, researchers should set up machines adapted to reality. And the factors that could affect the result are many, from the force of gravity to slippery surfaces to other interferences. But the model shows that, theoretically, we can teach humanoid robots to play football with a certain class. Not so much to see them compete with each other in the future as to help the artificial intelligence itself and its developers to the usual goal of all time: to improve and optimize the ability to predict results. After all, from the parts of Boston Dynamics they know something, in terms of robots grappling with real life and the physical world.

“The result – explains the research team led by Siqi Liu – is a team of coordinated humanoid football players who exhibit complex behavior at different scales, quantified by a variety of analytics and statistics, including those used in real-world sports analysis. Our work is a complete demonstration of the integrated decision-making process learned on multiple scales in a multi-agent context ”. The road is obviously very long and complicated: if the model is theoretically able to move a robot, that is an agent, everything is still at the simulation stage. And then the rules have been greatly simplified: the ball is always kept in the field by invisible walls and there are no fouls or free kicks. Not bad: even in the real world there are those who still haven’t understood the offside rule.