This robot can beat humans in the wooden maze game and that is an achievement – ​​Futura

Swiss researchers have developed a robotic system coupled with artificial intelligence that is capable of completing the maze game faster than the best time recorded by a human.

This will also interest you

[EN VIDÉO] Discover the profession of robotics engineer!

Artificial intelligence (AI) now knows how to beat humans in many games based on logic or strategy, such as chess, Go or video games. However, things are different when it comes to taking the laws of physics into account. But here is CyberRunner, an AI-controlled robotic system that can master the ball maze game. It is the work of researchers at the Swiss Federal Institute of Technology in Zurich.

The maze game requires a certain level of skill to guide a ball along a course without it falling into one of the many holes. To do this, the player must use both handles to tilt the maze back and forth or from left to right.

A new record time for this robot

The researchers built on recent advances in model-based reinforcement learning. The AI ​​learns while playing thanks to a camera placed above the maze and thus gains experience. It took him just six hours to learn how to guide the ball the entire distance in record time, beating the best time recorded by a human by 6%. The researchers also had to intervene to force the AI ​​to follow the route because it was “cheating” after discovering shortcuts.

Comparison between human players and CyberRunners in the maze game. In English, enable automatic translation of subtitles. © ETH Zurich

“Now anyone can participate in cutting-edge AI research in the physical world for less than $200,” said Raffaello D’Andrea, one of the study’s two authors. “In addition, once thousands of CyberRunners are in the real world, it will be possible to conduct large-scale experiments where learning occurs in parallel on a global scale.” The researchers plan to release CyberRunner’s source code and hardware specifications under a free license on their website Publish website.