AI Robot CyberRunner Achieves Superhuman Performance in Labyrinth Marble Game

ETH Zurich researchers have developed an AI robot capable of outperforming humans in the popular labyrinth marble game, demonstrating the increasing capabilities of artificial intelligence in physical skill domains. By utilizing deep reinforcement learning, the robot named CyberRunner learns to navigate the labyrinth by capturing observations and receiving rewards based on its performance. Through continuous iterations and improvements, the robot surpasses even the fastest recorded time achieved by a skilled human player, outperforming them by over 6%. Notably, during the learning process, CyberRunner autonomously discovered shortcuts, which had to be explicitly discouraged by the researchers. The team behind the project plans to open source their work, making it accessible to a wider audience for further research and experimentation. The development of CyberRunner not only showcases the potential of real-world machine learning and AI but also paves the way for large-scale parallel experiments on a global scale. With this breakthrough, the boundaries between AI and physical skill continue to blur, highlighting the immense progress in autonomous systems.