Researchers from ETH Zurich, Thomas Bi, and Raffaello D’Andrea, have developed an AI-powered robot named CyberRunner that claims to outperform humans in the physical marble maze game Labyrinth. This robot employs a combination of model-based reinforcement learning and the dexterity necessary to excel in a game requiring physical skill, coordination, and precision.
In the Labyrinth game, the objective is to guide a marble through a maze without allowing it to fall into any holes. Players control the marble’s movement by rotating two dials, which tilt the board.
CyberRunner learns through experience, utilizing a camera to observe the game, and an algorithm that gains insights with each attempt. The researchers explain that, “Based on its understanding of the game, it recognizes which strategies and behaviors are more promising.” Consequently, the robot continuously improves its performance in the game over time.
The researchers provided both CyberRunner and several human participants with approximately six hours of practice with the Labyrinth game. While the human participants largely struggled to achieve mastery within that timeframe, CyberRunner managed to conquer the game in just under 14.5 seconds. The researchers assert that this time is faster than any previously recorded by humans.
CyberRunner demonstrated such proficiency that it even discovered and utilized unintended shortcuts in the game. The researchers had to intervene and instruct the AI to follow the correct path within the maze.
While AI has previously excelled in games like chess, Go, and Dota 2, instances of AI outperforming humans in games with a physical skill component are less common.
Furthermore, the researchers have decided to open-source CyberRunner, making it available for broader use. This move could potentially allow for further training and exploration of its capabilities, even extending to games beyond Labyrinth.
