Dec 21, 2023 12:30:00

AI ``CyberRunner'' wins against humans in a game that requires physical skills, inadvertently devising a cheat and panicking scientists

AI has already overwhelmed humans in games that rely on intelligence, such as chess, Go, and

strategy games , but humans still have a day ahead of humans in terms of physical skills. However, it was announced that an AI that shows super play that humans cannot match even in the ``maze game'', which is played by physically tilting the game board with your hands.

CyberRunner
https://www.cyberrunner.ai/

AI beats humans for the first time in physical skill game
https://thenextweb.com/news/ai-beats-humans-first-time-physical-skill-game

This time, a robot equipped with reinforcement learning AI developed by Thomas Bee and Raffaello D'Andrea, researchers at the Zurich Institute of Technology in Switzerland, has defeated the human record holder in a game that requires physical skills. The system is 'CyberRunner'.

By playing the movie below, you can actually see CyberRunner brilliantly playing the maze game.

AI physical boundaries breaks: CyberRunner, the superhuman AI robot - YouTube


The maze game is a game where you move the knobs on the box to tilt the maze and roll the marbles to guide them to the goal. Although the mechanics are simple, it requires fine motor skills and spatial reasoning skills, so it takes many hours of practice for humans to be able to successfully play this game.

CyberRunner is equipped with a motor that moves knobs instead of a human hand, a camera that acts as an eye, and an AI that acts as a brain.

When CyberRunner first started learning, it would just randomly tilt the board, causing marbles to fall into holes.

The true power of AI lies in its ability to continually optimize technology through model-based reinforcement learning. After training CyberRunner on this maze game, it was able to complete the game in just 6 hours and in 14.48 seconds, 6% less than the fastest human record.

Surprisingly, CyberRunner discovered shortcuts during training and even came up with a 'cheat' to skip the route, so the research team had to intervene in the training and instruct it not to cheat. I was told that there wasn't.

Researchers at ETH Zurich have published a paper summarizing the results of this research

on the preprint server arXiv , and plan to open source the research project in the near future.

D'Andrea said, ``In the past, research in the field of AI was only possible for organizations with large budgets and custom-made experiment infrastructure, but now anyone can do research without spending $200 (approximately 28,000 yen). I can now participate in cutting-edge AI research.'