If ball maze games remind us of our childhood, they now have a link with the most recent technologies. This skill game consists of moving a ball from point A to point B by crossing a board generally made of wood, dotted with obstacles and holes, by tilting it on all axes using wheels. One of these trays was motorized and connected to a robot powered by artificial intelligence. Filmed, the experience is disconcerting to say the least and is reminiscent of the first time a world chess champion was beaten by an IBM computer in 1996.
An AI that allows itself to cheat
This robot, codenamed CyberRunner, has been trained beforehand. For this, the two wheels allowing the tray to be tilted have been motorized and connected to a learning algorithm which uses artificial intelligence. A camera has been added to allow the AI to follow the movements of the ball.
Three humans were invited to the experiment and given six hours to practice on set. In the interest of equal opportunities, the CyberRunner robot was given the same amount of time.
The video, shot by the team behind the experiment, shows the robot brilliantly following the layout of the maze in precisely 14.48 seconds. He thus beat the human record of 15.41 seconds, which was established in 2022 by maze virtuoso Lars Göran Danielsson.
More surprising than mastering the ball, the AI discovered shortcuts on its own: certain movements allowed it to cheat by passing the ball right next to a hole, without it falling. As part of the CyberRunner experiment, the developers were forced to prohibit this practice from the AI, in order to give its three flesh-and-blood rivals a chance.
The project should soon be published under an open source license, to allow anyone to improve the AI model used.