Advertisement

It took approximately 100 years of virtual computing time for Elon Musk’s non-profit artificial intelligence lab, OpenAI, to teach a robotic hand how to juggle a cube, according to AP News.

The team paid Google $3,500 to run its software on thousands of different computers, which turned the actual computing time into 48 hours. Once the robot hand, named Dactyl, was trained in a virtual environment, the researchers put it to test in the real world.

The researchers wanted to use AI to teach the robot to handle objects in a way that reflected the way humans move. OpenAI set the task of teaching a robot hand to manipulate a six-sided cube, according to The Verge. Dactyl had to move the block from one side of the cube to a specific side facing up, but with one alternative: ever-changing simulations.

First, the researchers added random visual noise. Then they changed the colors of the virtual hand and made the cube look different. They would also randomized the cube by changing how slippery the surface of the cube was and how heavy it was. The even changed the simulation’s gravity. This was all to give the AI better tools to adapt to real-life conditions. Although the simulation was not a complete accurate depiction of real life, its variations allowed the system to deal with the unexpected.

The end results allowed Dactyl to become fairly trained. The hand was able to move the cube from one position to another up to 50 times in a row without dropping it, and by moving the cube around it developed human-like attributes.

“This shows that what we humans do for manipulation is very optimized,” said Matthias Plappert, OpenAI researcher. “It’s a very interesting moment when you look at a robot trying to solve a problem and you think ‘Oh, hey, that’s how I would do that, too.’”

Advertisement
Advertisement