DeepMind researchers have developed AI agents that can play virtual football.

According to the engineers, the AI ​​agents learned the game almost from scratch. First, virtual robots learned to walk, then run and hit the ball.

At each new stage, videos with real football players were shown to artificial intelligence systems. This allowed them to learn the basics of the game, imitating the moves of professional athletes while participating in real high-level events.

As soon as the robot learned to play alone, it was paired with another AI agent. As the skills of virtual football players improved, so did the number of players.

As a result, the researchers formed small groups that competed with each other.

“The result was a team of coordinated humanoid football players exhibiting complex behaviors at different scales, quantified by a range of analyzes and statistics, including those used in real-life sports analytics,” the study says.

According to the engineers, each footballer robot makes a decision on his own during the game. However, they did accept some simplifications in the rules. For example, the system does not take into account fouls and an invisible boundary is drawn around the field to prevent the ball from going outside the boundary.

The researchers also noted that it takes a long time to train agents. This could make it difficult to scale the technology to physical robots, they added.

Recall that in July, DeepMind’s AlphaFold algorithm predicted nearly all proteins known to science.

In August 2021, the lab presented a universal architecture for building artificial intelligence.

In June, scientists at DeepMind said that reinforcement learning is enough to achieve artificial general intelligence.