A developer built an AI system that learned to play the physical Labyrinth Game using only a camera and two motors. It now completes the maze in 13.8 seconds, beating the human record of 15.1 seconds, down from needing 5-6 hours of training to just 1-1.5 hours.
About 13.8, if I remember correctly.
Are the humans upset that you've built this? What is the game called?
The Labyrinth Game? Yes. You can buy them on Amazon, most of the toy stores. It's very widely available.
So that's computer vision?
Yes. It's really very simple. We just have the camera and we have two motors, and that's kind of the system itself. And then we just have this third component of this reload mechanism so it can learn autonomously.
So the first time You set it on there, it knows nothing.
It knows nothing, and the learning is very gradual. So it learns how to go past this section, and once it knows how to do this, it kind of proceeds to the next one, and then it really learns in a very incremental fashion. When you just watch it learn, you really see how it gets better and better.
How long did it take to get good?
At the start, it was about 5 to 6 hours until it gets to like this proficiency. We got it down to about 1 to 1.5 hours now. So it only needs to practice 1.5 hours without knowing anything, and then it kind of gets to here.
And the human record is 15.1 seconds? I think so. And this has gone—
About 13.8, if I remember correctly.
Are the humans upset that you've built this?
Um, I don't know. So on one of the videos, there's a comment saying, "AI does it better." I guess it just shows how far these algorithms have come.
It's not going to take too long until those robots really can do all the tasks that humans can do.
