One of the big problems with robotics – especially limbs and other humanlike features -is dexterity and making them move with fluid movement like us humans.
Now Elon Musk-founded robotics non-profit OpenAI has moved one step closer to introducing robots that move just like us real beings.
OpenAI’s project is powered by AI to understand how humans move, manipulating objects just as we would. Rather than working on the actual hardware like many other companies are doing, OpenAI is focusing on the teaching methods that on’t rely on the usual trial and error techniques.
The company wanted to test its method by trying to teach a robot hand to manipulate a six-sided cube, turning it over and over so different sides took turns to be face-up. These manipulations included pivoting the cube around its fingers, sliding it across digits and “finger gaiting”.
Although it sounds pretty straightforward, the researchers wanted to make this more difficult by changing things. For example, they changed the size and shape of the cube, they made it more slippery and changed colours on the sides, breaking the habit they’d come to learn.
This taught the robot how things can change and they must develop an understanding of this, with different factors changing frequently to confuse them.
It took around 100 years for the AI system to get to grips with the changing scenarios and unsurprisingly, it took a lot of processing power. Dactyl was powered by 6,144 CPUs and eight powerful powerful Nvidia V100 GPUs to process the data – more than the average research institution let alone a home computer.
“This shows that what we humans do for manipulation is very optimized,” says Plappert. “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.’”