Robot Can Read, Learn Like a Human : Discovery News
"Marge" can figure out misspelled words. She can even read the New York Times and learn from her reading.
Marge, a robot on wheels, can read and understand human signs.
While reading for a literate human is pretty easy, it is quite difficult for robots.
Later versions of Marge's software could be installed on everything from cars to cell phones.
For her age, Marge is pretty smart. She can find and correct misspelled words. She taught herself that Barclays is a bank, Strada is a restaurant, and that both are in the UK. She is even smart enough to read the New York Times and the BBC Online. And soon she will take her first steps, or more accurately, her first rotations, into the real world.
Even though she is young and has a lot to learn, Marge isn't a child. She's a robot, developed by UK scientists, that can read and learn in the real world.
Her creators, Ingmar Posner and Paul Newman at the University of Oxford, along with their collaborator Peter Corke at Queensland University of Technology, hope that Marge, and future versions of Marge, could navigate through the real world using the same words and phrases that humans use.
"Text spotting is hard because text is a such a variable thing," said Newman. "It appears in so many guises in so many places, in so many sizes, and of course the real world is full of reflections, occlusions, etc."
To a literate human, reading is a simple matter. If a word changes size, or the lighting in the room changes, you don't instantly become illiterate. That's because the human brain can make intuitive leaps of logic, said Edward Grant, Director of the Center for Robotics and Intelligent Machines at North Carolina State University.
"The brain says, 'I've done something kinda like this before, so I can adapt to this new activity that has been presented to me,'" said Grant.
A robot, or more specifically, the mathematical algorithms installed on a robot that are its brain, can't make that intuitive leap. For a robot, a change in the angle, lighting or size means they have to learn to read all over again, said Grant.
Because of this limitation, a reading robot ("or 'robot literacy,' as we have come to think of it," said Posner) is usually found in the lab, where the light, the angle, and other variables remain constant and unchanging, so the robot doesn't get confused.
But in a move both simple and brilliant, the Oxford scientists installed text recognition software (technically called Optical Character Recognition, or OCR), complete with spell-checker and dictionary, onto Marge, a small robot on wheels. By using a few new tricks to separate text from, say, sticks or trash, and correcting the image based on a simple spell-check and the word's meaning in the dictionary, Marge can bridge the gap in intuition.
That means she can read newspapers like the New York Times and learn about banks like Barclays and that restaurants like Strada are good places to eat.
"On the one hand it makes perfect sense," said Gregory Dudek, a scientist at the Centre for Intelligent Machines at McGill University. "The environment has markings for people to use, and this exploits those human markings for use in a robot."
Robots come in a wide variety of shapes and sizes, from Google's self-driving cars to toys for tots. This advance could enable car navigation systems that don't rely exclusively on GPS (a big help when in long tunnels or underground structures) or a robot that can lead you to that perfect gift in the mall. The Oxford group already has newer version of Marge, dubbed Lisa, that they are also testing.
Since Marge's intelligence comes from software, not hardware, that means that it could be used in devices besides robots as well, said Dudek. Cell phones and even eyewear (equipped with some kind of head-mounted display) could eventually read signs or menus and instantly provide information to the user.
There is still a lot of work to do before Marge's technological children, including Lisa, could live on your phone. But when they do, the bigger question will be what can't a literate robot do, rather than what can they do.