In recent years we have seen a multitude of artificial intelligences that can end up surpassing human beings in games as intensive as chess or Go, but there has not been much progress in regards to competing against humans in games that require physical dexterity.
But this could be ending, because researchers from the ETH Zurich in Switzerland have presented CyberRunneran artificial intelligence system that, based on visual learning, training reinforcement and precise physical controls, has even been able to surpass the world record enjoying Labyrinth.
If you haven’t played it yet, you’ve probably seen people enjoy it. Labyrinth. It is a box covered with a wooden plane that tilts on an axis. On top of the board there is a labyrinth with numerous holes. The objective is to move the marble or a metal ball from start to finish without it falling into one of the holes.
It is one of the most frustrating board-type games you can enjoy, requiring a huge pulse and therefore great physical dexterity.
With many hours of practice, it is likely that a human being can end up mastering the game in less than a minute, but not before getting frustrated on a few occasions.
However CyberRunner You can master the process in just five hoursand now you can complete the maze in just under 14.5 secondsand that’s 6% faster than the world record.
The key to CyberRunner It is a combination of reinforcement learning and visual input from cameras.
On the other hand, success and error tests are stored in the memory of CyberRunner allowing you to learn step by step to know how to move the marble properly around the board.
“It is important to highlight that the robot does not stop playing to learn; The algorithm runs simultaneously with the playing robot. As a result, the robot continues to improve, run after run,” say its creators.
Not only did the system learn the quick way to finish the maze, but on top of that he was able to find bugs and in this way discover new shortcuts that a human being had never even considered before.
