DARPA Robots Learn YouTube Cooking

For modern computer systems, it is not a problem to recognize a physical object or to calculate some parameters of such an object. In addition, computers are very easy to “learn” - just enter a new program, and the system already “knows kung fu”.

However, self-learning computers are not yet so common, and empirical training in computer systems is rare. The agency DARPA recently introduced a new project , the purpose of which is to teach robots something new by demonstrating how-to videos from YouTube. Looking ahead, let's say that the developers succeeded.

So, at DARPA they were able to create robots that were able to learn the culinary videos on YouTube, then using the information received to prepare new dishes. More specifically, the project was run by specialists from the University of Maryland, and DARPA funded this program. The program itself was called Mathematics of Sensing, Exploitation and Execution (MSEE) .

As for the results of the project, they are very interesting. So, watching the video, the robots were able to recognize the kitchen “accessory” from the video, select it from a number of tools laid out on the table, and use this tool. At the same time, the robots used the kitchen "tools" correctly - they did exactly what was shown in the video.

Another achievement of the project is the ability of robots to accumulate information obtained during training, as well as the ability to share this information with other systems. According to the developers, the achievements of the project allow us to say with confidence that it is possible to train computer systems not only by creating new programs, there are other methods.