DeepMind trains AI agents to play football

Researchers at DeepMind have developed AI agents capable of playing virtual football.

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

At each new stage, AI systems were shown video footage of real football players. This allowed them to learn the basics of the game, imitating professional athletes’ movements during real high-level events.

As soon as a robot learned to play solo, another AI agent was paired with it. As the virtual footballers’ skills improved, the number of players grew.

Ultimately, the researchers formed small groups that competed against one another.

“The result was a team of coordinated humanoid footballers displaying complex behaviour at various scales, quantitatively characterized by a range of analyses and statistics, including those used in real-world sports analytics,” the study says.

According to the engineers, each robot footballer makes decisions autonomously during the game. However, they acknowledged some simplifications in the rules. For example, the system does not account for fouls, and there is an invisible boundary around the field preventing the ball from going out of bounds.

The researchers also noted that training the agents took a long time. This could complicate scaling the technology to physical robots, they added.

In July, DeepMind’s AlphaFold algorithm predicted almost all proteins known to science.

In August 2021, the lab unveiled a universal architecture for creating artificial intelligence.

In June, researchers from DeepMind stated that for achieving artificial general intelligence reinforcement learning is sufficient.

Subscribe to ForkLog’s Telegram news: ForkLog AI — all the news from the world of AI!