Robo Brain database is online. It already contains billions of images, 3D files and videos, but it is not meant to Internet users, but robots. Joint project between the prestigious universities of Cornell, Berkeley and Brown, Robo Brain will allow robots that will connect to recognize objects in their environment, but also know how to use. An advanced use of the so called Machine learning, this discipline of artificial intelligence where the software will learn by itself. The implications of Robo Brain are manifold, both in robotics, service robotics research, but also autonomous cars.
Deep Learning in the service of Robotics
Imagine: You just unpack your robot of service brand new. You turn it on, and immediately it goes to your kitchen and began to empty the dishwasher. It takes the cups, glasses, plates and pans to forms the most various and undertakes to store them in your closets. All those who program bots can imagine the complexity of such a scenario. A task that might seem yet extremely simple for a human. The secret of this robot from a future can be close enough, it is connected to Robo Brain to analyze your apartment, your kitchen, the content of your dishes lavec. It is in the Cloud, on the Internet, that he was able to recognize objects that surround it, learn how to open the dishwasher, recognize the glasses, cups, learn how to take without leaving them down and place them in the closets of the kitchen.
How to teach robots to interact with objects
Ashutosh Saxena, an assistant professor at Cornell University and Director of the laboratory of Robot Learning is working on the problem of understanding robots their environment and enable them to learn independently to interact with objects. The algorithms developed by his team are awesome. From the billions of photos stored on Robo Brain, the robot analyzes environment and will create a graph where each of these objects. A graph that will quickly fill up with thousands of nodes if it is a service robot in a kitchen, hundreds of thousands of nodes is an autonomous vehicle that rolls in a street. Not only the database allows it to recognize objects, but it also shows him how to use them. Know how to take a cup for example there is nothing obvious to a robot that sees a mug or a cup of coffee in Limoges porcelain for the first time. In addition to hooking possible points to catch the object, Robo Brain indicates the behaviour to adopt with the object. It reverses not a box of eggs, it does not pass over a laptop with a glass of water, it does not run a knife to a human being. So many good behavior meaning that gradually learns the young child but that it should also instill in the robot. But the work of Ashutosh Saxena go further still. Its database stores objects-related human behaviour. You enter the pot on the stove. Several options to then offer you: lay it on the table, place it in the dishwasher, refrigerator or Cabinet. 4 scenarios that will service robot will have in mind as soon as you enter the pot. You take a step towards the fridge? Your robot can deduce that you are going to place the Pan in the fridge, it anticipates the movement and will open the refrigerator door without requiring you to give him the order. Science fiction? The researcher explains that, the robot can anticipate your movements with 3 to 10 seconds in advance! You decide that the behavior of the robot is not satisfactory and does not match by your habits. Just take it by ‘the hand’ and show him what you expect of him. What the researcher calls the active learning will allow the robot to add this new scenario of use in its database. So far, Robo Brain has 14 million images from imageNet, 1.4 billion images from yahoo! Images, 2 million and 120 million of Youtube videos 3D files. Designers of robots to operate in turn this collective intelligence to learn their achievements to interact with their environment. Sources : “The Plan to Build a Massive Online Brain for All the World’s Robots”, Wired, 25 août 2014 “Robo Brain Downloads and Stores Images, Videos and Data in a Robot-Friendly Format”, AzoRobotics, 25 août 2014 “‘Robo Brain’ will teach robots everything from the Internet”, Phys.org, 25 août 2014