In a groundbreaking development, researchers at ETH Zurich have made a significant leap in artificial intelligence, demonstrating that AI can now outperform humans in tasks requiring physical skills. This breakthrough was showcased through their AI robot, CyberRunner, which mastered the labyrinth marble game, a test of dexterity and precision, in a remarkably short time.
The labyrinth game, traditionally a test of human motor skills and spatial reasoning, involves guiding a marble through a maze-like board to reach a goal while avoiding pitfalls. This seemingly simple game demands considerable practice for humans to excel. However, CyberRunner, developed at ETH Zurich and detailed on its dedicated website, achieved this feat in an unprecedented manner.
Using advanced model-based reinforcement learning, CyberRunner demonstrates how AI can extend its prowess into the realm of physical interaction. This technique enables the AI to predict and plan actions by continuously learning from its environment. Equipped with a camera to observe the game and motors to control the board, the robot rapidly improved its gameplay through a process akin to human learning but at an accelerated pace.
Remarkably, CyberRunner completed its learning cycle in just over six hours, going through 1.2 million time steps at a control rate of 55 samples per second. This feat saw the AI surpass the record held by a highly skilled human player by an impressive margin of over 6%.
Interestingly, during its learning phase, CyberRunner even discovered shortcuts in the game, prompting the lead researchers, Thomas Bi and Prof. Raffaello D’Andrea, to intervene and guide the AI to avoid these paths.
This achievement by ETH Zurich researchers not only pushes the boundaries of AI in gaming but also signifies a major step forward in how AI can be applied to real-world physical tasks. The success of CyberRunner indicates a future where AI can undertake complex physical activities, potentially transforming various industries and everyday life.
This milestone in AI development marks a shift from virtual achievements, such as mastering chess or Go, to conquering physical challenges, blurring the lines between human and machine capabilities in the realm of physical skill and dexterity.
A preprint of the research paper is available on the project website. In addition, Bi and D’Andrea will open source the project and make it available on the website. Prof. Raffaello D’Andrea commented: “We believe that this is the ideal testbed for research in real-world machine learning and AI. Prior to CyberRunner, only organizations with large budgets and custom-made experimental infrastructure could perform research in this area. Now, for less than 200 dollars, anyone can engage in cutting-edge AI research. Furthermore, once thousands of CyberRunners are out in the real-world, it will be possible to engage in large-scale experiments, where learning happens in parallel, on a global scale. The ultimate in Citizen Science!”