Gone are the days when taking classes was necessary to learn a new skill. With the whole Internet at our disposal, learning to knit, cook, apply makeup or repair appliances is just a Google search and YouTube click away.
As humans, we can watch a video, understand the steps and reproduce them (at least fairly well). Researchers at RoboWatch are teaching robots to do the same thing. Yes, you read that correctly. They are creating an algorithm that will allow a robot to watch a series of YouTube videos on a particular topic and acquire a new skill without prior knowledge.
Take making an omelet for example: the robot will begin with a YouTube search for “How to make an omelette?” and download many videos. It parses each video into steps that recur in each one (cracking eggs, whisking them, pouring them onto a pan, etc.), which creates a storyline.
The video below shows how it would work for baking a chicken breast and making Jello shots, ice cream, a milkshake, and an omelette. The omelette section begins at 1:45.
RoboWatch is different because it is employing unsupervised learning. Using their method, a computer could learn without the need for prior human classification, labelling and training. The research team, comprised of computer scientists from Stanford and Cornell, described their research in a paper on the new algorithm entitled, “Unsupervised Semantic Parsing of Video Collections.”
I guess we’ll have to be careful with what we put up on YouTube in case robots take over the world. I can think of some skills that I wouldn’t want them to learn!