In a groundbreaking achievement, researchers at ETH Zurich have developed an AI robot named CyberRunner that has outperformed humans in the popular game Labyrinth. The robot accomplished this feat by navigating a small metal ball through a maze using fine motor skills and spatial reasoning, mastering the game in six hours.
This marks one of the first instances in which artificial intelligence has excelled in a direct physical application, showcasing its ability to think, learn, and self-develop in dexterity-based tasks.
Model-based reinforcement learning: The key to success
The researchers, Raffaello D’Andrea and Thomas Bi harnessed recent advances in a field known as model-based reinforcement learning to teach CyberRunner how to excel at the Labyrinth game. This type of machine learning involves AI learning how to behave in a dynamic environment through trial and error.
By sharing their work in an academic paper, the researchers have made it possible for others to build upon their findings, emphasizing their project’s collaborative and open-source nature.
One of the remarkable aspects of this project is its accessibility. D’Andrea and Bi are making their work available on an open-source platform, enabling a wider community to explore, experiment, and innovate in AI robotics.
With a price point of just $200, users can utilize the CyberRunner platform for large-scale experiments, fostering communication and sharing best practices among AI enthusiasts and researchers.
AI’s evolution from repetitive to adaptive
Industrial robots have long been adept at performing repetitive and precise manufacturing tasks. However, the flexibility and adaptability demonstrated by CyberRunner represent a significant leap forward in the capabilities of AI-driven machines.
This robot’s ability to think, learn, and adapt in real-time to a dynamic physical task was previously thought to be achievable only through human intelligence.
Learning through experience: CyberRunner’s unique approach
CyberRunner’s learning process is an intriguing aspect of its success. Equipped with a camera that observes the labyrinth from above, the robot learns through experience, discovering surprising ways to navigate the maze, including taking shortcuts.
However, the researchers had to step in to instruct it not to exploit these shortcuts, highlighting the need for ethical considerations in AI development.
D’Andrea emphasized that their project is not a costly, bespoke platform. Instead, it is an open and affordable resource for anyone interested in advancing AI robotics. This inclusivity and affordability are key factors in the potential for rapid progress in the field.