News By Jan. 6, 2015 5:25 pm
This robot learned how to make dinner from watching YouTube videos | News | Geek.com

Researchers at the University of Maryland and NICTA, Australia have created a robot capable of complex actions like cracking an egg and using a vacuum. And if you’re wondering who taught it how to do that stuff, it learned it by watching You… Tube videos. In a paper describing their development process, the researchers explain how they taught their robot to process information by “watching unconstrained videos from the World Wide Web.”

Teaching robots to imitate human actions has already been done plenty of times, but for researchers Yezhou Yang, Yi Li, Cornelia Fermüller, and Yiannis Aloimonos, the goal was to create an advanced language and grammar that let automated robots learn on their own with little context. In their words, “beyond simple learned schemas, we need computational tools that allow us to automatically interpret and represent human actions.”

This robot learned how to make dinner from watching YouTube videos | News | Geek.com

The system they developed uses a Convolutional Neural Network, or CNN, deep-learning framework along with object recognition and grasping type recognition. Simply put, by watching a how-to video, the robot can see what the person is holding and how they are holding it to break down and repeat the action.

While similar recent training programs like Robobrain rely on product manuals and the limited information that comes from how an object looks in a flat video, this system uses knowledge of the objects themselves to deduce how to better manipulate them. Flipping a steak on the grill? Use a small power grip on the tongs. What about whisking batter in a bowl? Use a small power grasp with the right hand to vigorously stir the whisk and a large precision grasp with the left hand to keep the bowl in place. Given the plethora of cooking videos on the Internet, no wonder the robot was stuffed with food data for these trials.

This robot learned how to make dinner from watching YouTube videos | News | Geek.com

Overall the action hierarchies the robot chose for each task were mostly accurate and the experiment was a success. Logically the next step is to just add more categories of objects and grasp types to the robot’s vocabulary so its knowledge can expand. The team also hopes create a model for multi-step actions. So like concerned parents in a PSA we may actually need to worry about our young robots learning how to use drugs by watching us. Or even worse, they might learn how to build self-learning robots by watching us build self-learning robots and cut humanity out of the picture entirely.

Here’s a video from a few months ago about Robobrain:

Now read: NASA is training Swarmie robots to prepare alien worlds for our arrival

This robot learned how to make dinner from watching YouTube videos | News | Geek.com

发表评论

电子邮件地址不会被公开。 必填项已用*标注