A robot that can screw the cap of a bottle, place a hanger on a bearing, or even manipulate geometric to a simple game of kindergarten, nothing really special today, so this is the robot that performs these tasks learned these gestures of itself, by trial and error.
A robot that learns in the same way one between humans, it is the turn of force directed by Berkeley researchers who have injected a “Deep Learning” algorithm in a venerable PR2 robot to give it the ability to learn. A breakthrough in artificial intelligence according to these researchers.
The ‘Deep Learning’ will bring a bit of adaptability to robots
Researchers in the Department of computer and electrical engineering from Berkeley will present the results of their work at the next conference of the ICRA (International Conference on Robotics and Automation) which will be held in Seattle on May 28. According to the release from the prestigious University of California what they have achieved is a major advance in the field of robotics. They have implemented “Deep Learning” algorithms in a robot. It is artificial Intelligence algorithms that are traditionally used in image analysis. It is through such algorithms that software is able to identify a cat on a photo or, in this case, a bottle, a hanger or form any. No need to learn and relearn the robot to open a bottle. It is able to apply its procedure for screwing on any bottle from the moment where he recognizes that the object facing him is a ‘bottle’, even if presented with bottles of shapes and sizes.
To achieve this “feat”, researchers have equipped their PR2 to a network of 92,000 software neurons that give this robot the ability to learn or rather the ability to name objects in its environment. It is what allows him to recognize a bottle, although its form varies with the first bottles researchers gave him to see.
The Personal Robot 2 (PR2) with his artificial brain was named after the sobriquet BRETT, for Berkeley Robot for the Elimination of Tedious Tasks. For the moment, with still limited intellectual capacity, it takes 3 hours to the robots to learn and perform a new task while it would take 10 minutes only for carrying it out in being programmed to do so. However researchers believe that in 5 to 10 years, this approach will allow robots to learn tasks becoming more complex and that the ‘Deep Learning’ will allow major advances in terms of learning for robots.
Translation : Bing Translator
Source : “New ‘deep learning’ technique enables robot mastery of skills via trial and error“, Berkeley News Center, May 21, 2015