How many times have you turned to YouTube to figure out how to do something? The site is chock full of instructional videos on just about anything you can imagine. Soon, robots might do the same thing, thanks to a new project working on parsing YouTube videos.
Called the Robowatch Project, the effort aims to generate a set of instructions through automated study of video content. The algorithm starts by searching YouTube for a single query, such as “how to make a milkshake.” From there it downloads the videos and begins to parse them one by one, using both visual and speech cues.
The resulting output comes in the form of a storyline, complete with pictures of the actual steps to complete the task and automatically generated text descriptions of each step. While it seems like a complex process, researchers found that the videos in general contained identical steps, allowing the algorithm to differentiate essential steps from inconsequential ones quite easily.
How could this technology be put to practical use? It’s not too much of a stretch to think that this technology could be the base for something bigger. For example, your home robot would just need to connect to YouTube to learn new skills rather than rely on complex programming. It’s a pretty exciting concept.
This isn’t the first time YouTube has been used in artificial intelligence, either. Three years ago, Google itself used YouTube in an experiment where a neural network built in its secretive X labs division was able to find cat videos with surprising success.
Obviously here, the skills learned by these computers would have far more practical use. But just looking at how far artificial intelligence has come in just those three short years shows you how close we really are to AI with a practical use.