Home

 › 

Inventions

 › 

Software

 › 

DeepMind’s Training Its New AI to Play Soccer: How is That Possible?

books about artificial intelligence

DeepMind’s Training Its New AI to Play Soccer: How is That Possible?

Key Points:
  • One of the main goals of DeepMind is to provide more data to professional athletes and teams across many sports.
  • DeepMind is using their technology to teach robots how to push a ball toward a target. They also want to expand this technology past soccer and into other sports.
  • The biggest hold-up for DeepMind is that AI cannot learn from watching. This causes many of the simulated movements to be very alien-like, and they can’t be replicated outside of a simulation.

You may have seen AI play chess before or even star on Jeopardy. But Google’s Artificial Intelligence division DeepMind has been working on teaching their AI to play soccer.

Last month, DeepMind released a research paper debuting their new neural probabilistic motor primitives. Those last 4 words might seem like nonsense, but put more simply, it’s a method where AI learns how to operate physical bodies.

We are still not at the point where we can have full-on robot soccer leagues, as cool as that would be, but this is a giant leap in AI learning and technology.

Let’s dive into what we know so far!

DeepMind’s Goal

According to DeepMind’s abstract released with this research paper, one of their main goals was to provide more data to professional athletes and teams across many sports.

In their own words:

In this paper, we provide an overarching perspective highlighting how the combination of these fields, in particular, forms a unique microcosm for AI research while offering mutual benefits for professional teams, spectators, and broadcasters in the years to come.”

So, as of now, Mohamed Salah and Christiano Ranaldo do not have to fear competing with a robot for their top spots.

How is this Accomplished?

DeepMind has been using a method that can be described as trial and error for their AI soccer players. The researchers wanted to prepare these players for anything, so while learning how to play soccer, the AI had to deal with challenging factors like gravity, slippery surfaces, and other players interfering.

These AI players started by barely being able to walk, but after being shown motion capture data from real soccer players, the AI began to imitate their movements. After 1.5 simulated years, or 24 hours in real-time, the players started to dribble, shoot, and run. Although they could play the game at this point, teamwork was certainly lacking. But, after 20 to 30 simulated years, or 2 to 3 weeks in real-time, they had coordinated 2-on-2 games going on.

Now that DeepMind has this working with graphical representations, they are using the same technology to teach robots how to push a ball toward a target. They also want to expand this technology past soccer and into other sports.

AI robot in human posture
AI learns from the information they are fed, which takes time and dedication.

What Does This Mean for AI?

This is a huge step forward for AI research, but it is still firmly in the simulation phase. We feel that we will not see robot sports players for a reasonable amount of time. The brute-force method DeepMind has been using for this simulation is certainly effective, but it can only get them so far.

Right now, the biggest hold-up for DeepMind is that AI cannot learn from watching. This causes many of the movements we see in the simulations to be very alien-like and cannot be replicated outside of a simulation. If this goes the same way, Boston Dynamics has done with “Spot,” the search-and-rescue, dancing robot dog. However, with machine learning algorithms and pre-programmed movements, these AI soccer players can certainly manifest in physical form.

We will definitely be keeping an eye on this research. Even though it is still in its early phases, the results they were able to produce are quite exciting. We cannot wait to see what they are able to do next!

Up Next…

Frequently Asked Questions

Can AI predict sports results?

Artificial intelligence can predict in-game actions or results with up to 80% accuracy. With how many variables there are in sports, there is a good chance these AI will never reach 100% accuracy.

How is artificial intelligence used in soccer?

AI is used in soccer, from assisting referee decisions to improving physical performance and predicting future injuries to producing and implementing tactics.

Can robots replace athletes?

At this point and time, the answer is no. But, with how many leaps and bounds AI technology has been making lately, we would not be surprised to see some AI athletes at some point.

What data is collected in soccer?

There is a lot of detailed data usually collected in a soccer game. For example, a frame-by-frame basis of players from both teams, their locations, ball locations, action details, like tackles and passes, their direction, and in some cases, the level of intensity or speed.

What do soccer players run each game?

On average, soccer players run 7 miles per game, with a high of 10 miles. This can be as low as 2 miles for a goalkeeper.

To top