Swiss researchers achieve the world’s first direct physical application of AI in which it beats humans

Maze

In a groundbreaking demonstration of artificial intelligence’s potential to master physical skills, researchers at the prestigious ETH Zurich University have created a robot that learned to excel at a popular maze game in just over 6 hours, surpassing even the fastest human players.

The robot, nicknamed “CyberRunner,” was designed by a team led by Thomas Bi, a doctoral candidate, and Prof. Raffaello D’Andrea, a pioneer in the fields of robotics and autonomous systems. Its goal was to navigate a marble through a labyrinth game by tilting the board, requiring fine motor skills and spatial reasoning.

While AI has achieved superhuman performance in games of strategy like chess or Go, mastering the physics of the real world had long remained an exclusively human domain. CyberRunner represents a breakthrough in this arena through its use of an advanced machine learning technique called deep reinforcement learning.

Much like a human, the robot learned by experience, observing the maze through a camera and receiving feedback on its performance. Over 1.2 million iterations completed at a rate of 55 per second, its algorithm built an understanding of the game’s dynamics, allowing it to hone its strategy.

Within 6 hours and 6 minutes of continuous gameplay, CyberRunner was navigating the labyrinth with dexterity exceeding even expert human players. In addition to mastering the intended route, it discovered shortcuts that allowed it to skip sections of the maze.

“We had to explicitly program it not to take those shortcuts,” explained Prof. D’Andrea. “It was an intriguing demonstration of how AI can discover unintended solutions.”

The research, described in a paper available on the project’s website, cyberrunner.ai, marks an important milestone in AI’s transition from the virtual realm to applications in the physical world. D’Andrea believes it will open up new possibilities for machine learning research.

“For less than $200, anyone can now engage with cutting-edge AI using this accessible testbed,” he said. “As more CyberRunners are deployed, large-scale experiments with parallel learning across the globe become possible, bringing us closer to the ultimate vision of citizen science.”

D’Andrea’s Institute for Dynamic Systems and Control at ETH Zurich, founded 15 years ago, has been a powerhouse of robotics innovation. Past creations include the pioneering Flying Machine Arena for aerial robot swarms and the Cubli, a cube-shaped robot that can jump up and balance on its corners.

The CyberRunner project will be fully open-sourced, with details available online to replicate the maze-solving robot. This will allow the machine learning community to build upon the advancements made in Zurich and usher in a new era of physical AI.

Explore more topics
Author
Prolific News
Related news stories
Advertisement