A robot with artificial intelligence beat people at a game of Labyrinth marbles. This is the first time a machine has shown it can learn and be physically skilled at the same time.
You have to move a marble around a set of obstacles and keep the ball from falling into any of the holes. This game takes skill, patience, and a light touch.
The record time to get through the maze is 15.41 seconds, set by Lars Goran Danielsson of Sweden, who has played the game for 35 years.
But CyberRunner, a robot made by experts at ETH Zurich in Switzerland, ran a time of 14.48 seconds after only six hours of practice.
Prof. Raffaello D’Andream of Dynamic Systems and Control at ETH Zurich said, “It’s not just about beating people at games.” You have to be very physically skilled and quick to beat people at a game.
“It was important that the first model let a machine beat a person at a game. It’s not just a change of task that we’re making here. We’re adding measurements and more lines that need to be broken.
Labyrinth is a board game that was first made by the British company Brio in 1946. The marble moves on tipped or tiled boards that are handled by two handles.
Even though the game is pretty simple, you need to be good at fine motor skills and spatial thinking to get good at it, and it takes a lot of practice to get good at it.
The robot moves the handles with two motors, and a camera on top of the board lets it see how it’s doing.
CyberRunner learns by doing, just like people do. It figures out which behaviors and techniques work best by doing them over and over again. This is called model-based reinforcement learning.
Computer scientists were surprised to learn that CyberRunner’s first instinct was to cheat, skipping parts of the maze to get through it faster.
A doctoral student at ETH Zurich named Thomas Bi said, “It’s interesting that CyberRunner found shortcuts naturally as it learned.” Some parts of the puzzle were skipped so it could “cheat.”
“We had to step in and tell it very clearly not to use any of those short-cuts.”
AI has been beating people for a long time. In 1997, it beat grandmaster Garry Kasparov in chess, winning its first competition.
AlphaGo, an AI system from Google DeepMind, beat Lee Sedol, the world champion at Go, in 2016. It did this by coming up with new moves that no one had seen before.
Experts say that this new discovery is the first time that AI has been used physically to beat a person.
Bi also said, “The AI robot beats the previous fastest recorded time, which was set by a very skilled human player, by more than 6%.”
The team is going to soon share the CyberRunner’s code and hardware information.
The work was written up in a paper that was put on the arxiv.org preprint website.