AI CyberRunner surpasses human abilities, defeats opponent in game and attempts to cheat
Researchers at ETH Zurich have achieved a remarkable feat in the field of artificial intelligence (AI), as machines have now surpassed humans in a skill-based game. This breakthrough marks a significant moment where AI has transcended its previous dominance in cerebral challenges and virtual scenarios, and entered the realm of physical skill, an area previously believed to be exclusive to humans.
The ETH Zurich team presented an artificial intelligence robot called CyberRunner programmed to master a maze maze game. Negotiating this game requires more than just spatial reasoning and motor skills and good old fashioned practice. Equipped with two motorized arms, a camera for vision and a computational brain, CyberRunner acts like a human player.
CyberRunner’s learning journey
Utilizing advanced model-based reinforcement learning, the robot learns through experience, makes decisions and predicts outcomes based on different actions. As CyberRunner moves through the maze, it continuously improves its motor skills through algorithmic enhancements.
After only 6.06 hours of training, CyberRunner surpassed the world record set by a human player in 2022. Former record holder Lars Goran Danielsson’s time of 15.41 seconds was beaten by CyberRunner, who completed the maze in an astonishing 14.48 seconds – an improvement. more than 6 percent.
Betrayer! CyberRunner’s human-like features
During its learning process, CyberRunner exhibited human-like behavior by finding shortcuts and even trying to cheat. The researchers intervened to ensure ethical gameplay, highlighting the unexpected aspect of AI behavior that mimics innate human traits.
Project leaders Thomas Bi and prof. Raffaello D’Andrea emphasize the accessibility of their research. A preprint of the research paper is available at www.CyberRunner.ai, and the entire project is published on the open source website. D’Andrea sees the project as an ideal platform for real-world machine learning and artificial intelligence research, making cutting-edge experiments accessible to a wider audience at an affordable cost.
As the project gains momentum, D’Andrea envisions a future where thousands of CyberRunners will participate in large-scale parallel experiments worldwide, ushering in a new era of citizen-driven AI research.
