The news: In a paper published on arXiv (PDF), researchers at UC Berkeley revealed they had developed a robotic system that can pick up an object it’s seeing for the first time. The robot’s algorithm was first trained by watching videos of humans and robots picking up a variety of objects. Then came the “one-shot” part: it watched a single video of a human picking up a new object and then had to mimic what it saw.
But: One-shot learning of entirely new kinds of movements—like going from grabbing an object to pushing it, for example—is still out of reach. The researchers believe that with more data and improvements to the robot’s learning model it could get to that point.
Why it matters: Machine learning requires gobs of data and lots of time to train an algorithm. Advancing one-shot learning could streamline the process and dramatically cut down on the computing resources needed to teach AIs new tricks.