Science

AI Robot Masters Marble Maze Game Surpassing Human Players

Published December 20, 2023

Scientists at ETH Zurich have made remarkable strides in artificial intelligence with the development of an AI robot named CyberRunner. Crafted by Thomas Bi and Raffaello D’Andrea, this robot is designed to navigate the intricate paths of the classic physical marble game Labyrinth with an efficiency that surpasses human capabilities.

Understanding the Challenge of Labyrinth

Labyrinth is a game that tests players’ dexterity and precision. The objective is to move a marble through a maze by avoiding pitfalls, which is accomplished by adjusting the incline of the playing surface through two dials. This dynamic requires not just mental strategy, but also careful physical manipulation.

The Rise of CyberRunner

The innovative robot, CyberRunner, is equipped with a model-based reinforcement learning system. This system enables CyberRunner to improve its performance through practice. Employing a camera that monitors the game's progress, CyberRunner's algorithm digests each round of play, tuning its approach as it discerns more effective techniques and maneuvers.

After roughly six hours of practice—a duration also allotted to human players—CyberRunner achieved mastery. It completed the game in a record-breaking time of approximately 14.5 seconds, a feat the creators tout as unmatched by any known human attempt. Moreover, CyberRunner became so skilled it discovered shortcuts within the game, prompting the developers to adjust its programming to follow the intended maze pathway.

This development is significant as CyberRunner's expertise embodies a fusion of AI with a physical skill set, a domain where AI has seen fewer triumphs compared to strategy games like chess, Go, and Dota 2.

Open-Sourcing the CyberRunner

In an effort to foster further advancements, the creators of CyberRunner are sharing its source code with the public. This gesture enables enthusiasts and researchers to potentially enhance the robot's abilities or adapt its learning framework to other physical skill games.

AI, robotics, technology