Watch an AI robot learn how to demolish humans at a marble maze game

In a groundbreaking display of technological advancement, researchers at ETH Zurich have unveiled CyberRunner, an AI-powered robot designed to conquer the physical marble game Labyrinth with unparalleled speed and precision. Developed by Thomas Bi and Raffaello D’Andrea, CyberRunner integrates model-based reinforcement with the dexterity necessary to excel in a game demanding skill, coordination, and accuracy.

For those unfamiliar with Labyrinth, the objective is to guide a marble through a maze without it falling into any holes. This entails the player manipulating the ball’s movement by rotating two dials, which in turn tilt the game board.

What distinguishes CyberRunner is its ability to learn and evolve through experience. Equipped with a camera that observes the game, the robot’s algorithm comprehends and refines its game-playing strategies with each attempt. As a result, CyberRunner showed exponential improvement, ultimately achieving a remarkable feat by conquering Labyrinth in a mere 14.5 seconds, a record-breaking speed previously unattainable by any human.

Emphasizing its adaptability, CyberRunner even discovered unintentional shortcuts within the game, prompting researchers to intervene and guide the AI to adhere to the intended path. This successful demonstration marks a significant leap in AI capabilities, particularly in physical skill-based games, which had not been widely explored by AI until now.

Furthermore, this accomplishment aligns with the trend of AI surpassing human capabilities in strategic games such as chess, Go, and Dota 2. Notably, the researchers have made CyberRunner open-source, paving the way for collaborative efforts to enhance its capabilities across various applications. The potential for AI to outperform humans in activities necessitating physical prowess expands the scope of AI’s impact and presents opportunities for its integration into domains requiring physical skills.