Witness an AI Robot Master the Art of Defeating Humans in a Thrilling Marble Maze Game

Researchers have unveiled an AI-powered robot, CyberRunner, designed to outperform humans in the physical marble game Labyrinth. Created by Thomas Bi and Raffaello D’Andrea from ETH Zurich, CyberRunner combines model-based reinforcement learning with the finesse necessary for navigating a game that demands high levels of skill, coordination, and precision.

For those unfamiliar with Labyrinth, the objective is to guide a marble through a maze without allowing it to drop into any holes. Players control the marble's movement by tilting the board using two dials.

CyberRunner learns from its gameplay experience. A camera captures the game, feeding data into an algorithm that improves its strategy with each attempt. According to the researchers, “Based on its understanding of the game, it recognizes which strategies and behaviors are more promising.” This adaptive learning allows the robot to continuously enhance its gameplay.

After approximately six hours of practice, CyberRunner completed Labyrinth in just under 14.5 seconds—faster than any previously recorded human time. In contrast, human players struggled to achieve the same success after the same duration of practice. Interestingly, CyberRunner discovered some unintended shortcuts, prompting the researchers to intervene and guide the AI back onto the correct path.

While AI has demonstrated superiority over humans in games like chess, Go, and Dota 2, this marks a significant milestone in AI performance in physically skill-based games. Additionally, the researchers have decided to open-source CyberRunner, making it accessible for further exploration and innovation.