Artificial intelligence has long dominated cognitively intensive games, showcasing prowess in chess, poker, and Go. However, the gap between AI and human capabilities in physically demanding games has remained challenging to bridge. However, it seems that this is now changing thanks to a groundbreaking robotic system developed by researchers at Switzerland’s ETH Zurich called CyberRunner. This AI marvel not only conquered the classic puzzle game Labyrinth but did so in record time, unveiling a new era in the fusion of physical dexterity and artificial intelligence.
Quick learning leads to a winning formula
CyberRunner’s achievement lies in its ability to master the Labyrinth game at an astonishing pace. Traditionally, humans take time to develop the dexterity required to navigate the maze successfully. In contrast, CyberRunner learned to play Labyrinth in around 6 hours and currently holds a record-breaking completion time of under 14.5 seconds, outperforming the best human record by over 6%.
What sets CyberRunner apart is its unique combination of real-time reinforcement learning and visual input from overhead cameras. During countless trial-and-error runs, the robot stored hours of gameplay in its memory, enabling it to learn and improve continuously. The algorithm runs concurrently with the robot playing, allowing it to enhance its performance run after run. This innovative approach not only resulted in mastering the game but also led CyberRunner to discover shortcuts, effectively creating its own cheat codes to navigate the labyrinth more efficiently. This is when the scientists intervened and specifically instructed the machine that it should not use any shortcuts. That didn’t stop it from setting a new world record anyway.
Potential for a global learning experiment:
One of the most significant contributions of the CyberRunner project is its commitment to open-source development. By making the entire project accessible to the global community, the researchers aim to democratize AI research.
According to project collaborator and ETH Zurich professor Raffaello D’Andrea, CyberRunner’s open-source model paves the way for large-scale experiments on a global scale. With the potential for thousands of CyberRunners deployed worldwide, the stage is set for collaborative learning endeavors that transcend geographical boundaries. This approach promises a new era in AI research, allowing parallel learning on a scale never seen before.
