Computer vision and pattern recognition represent a challenge in robotics because we need to be surrounded by robots able to understand images, process high-dimensional data from the real world and able to learn everything about anything. It took just a few decades for computers to evolve from a full-size room to a credit card sized device and a vastly more processing power. It took just a few decades for robots to evolve from a big machinery able to move its hands to a robot able to walk, dance, talk, recognize humans, or learn and adapt to different situations. In the same basket can be added the evolution of programs and algorithms in robotics and not only so.
A truly social robot should have the ability to learn while interact with the physical and social environment, an instinct that people are born. This is the idea of the LEVAN program written in the labs of the University of Washington and the Allen Institute for Artificial Intelligence in Seattle. LEVAN is a program engineered to learn how to differentiate an apple by an egg, and learn how to handle different objects such as a sharp object that should not be moved away from people.
The LEVAN program is able to search on the Web in millions of books and images to discover the associations between textual and visual data. This exploration of the Web creates a set of possibilities to link a set of phrases with pixels from an image. Identifying the pattern of an object, the program learns all the terms relevant to this object, information grabbed from the Web.
How it works
Even in testing phase, the program still needs an input from a human subject. The user has to browse among the existing concepts already available in the program. A concept is a keyword such as “innovation,” “car,” or “robot.” If your concept doesn’t exist in the library, all you have to do is to define a new one with your keyword and wait the program to filter the words searched on Google Books and in relevant images from the Web and then learn about your subject.
This program can jump the robots vision and pattern recognition into a new era where any autonomous machinery can learn about objects and humans using the information from books and images from the Web.
The current version of the program is limited in speed and capabilities, but with few improvements the time consumption to a result and the processing power will drop dramatically.
Probably you think that this program can fill some important gaps about computer vision and pattern recognition in robotics. You have right!