Scientists Create One Robot Brain to Rule Them All

Robo Brain is a cloud-based repository that enables robots to learn from each other

Researchers are creating a single, massive repository of robot knowledge so machines around the world can learn from each other.

Dubbed Robo Brain, the repository, which robots can access over the Internet, is designed to let the machines draw on its more than 10 terabytes of data whenever they need it. The knowledge store resides on Amazon Web Service's public cloud.

"I'm really looking forward to building this brain, with all this information that robots  need," said Ashutosh Saxena, an assistant professor in the computer science department at Cornell University. "Instead of teaching each piece of knowledge to each robot, when a robot goes out in the real world, it can query the brain and learn how to do things."

Robo Brain is a collective project, with scientists from Cornell, Stanford, Brown and the University of California, Berkeley, collaborating on it.

Part of what can be laborious and time consuming about building efficient robots is that each machine has to learn so many individual tasks, Saxena explained to Computerworld. Every robot has to be taught the same thing, such as how to open a carton a milk, time after time.

Each robot also needs to understand how the world works – what is a table, what is a living creature and how should it be treated – while also understanding its place in the world.

With Robo Brain, individual robots, whether it's a robotic arm working on a factory floor, an autonomous car or a robot assistant helping an elderly person at home, can draw on this store of information and learn from what other robots have already learned.

"Our laptops and cell phones have access to all the information we want," Saxena said. "If a robot encounters a situation it hasn't seen before, it can query Robo Brain in the cloud."

In July, the team of scientists began uploading about 1 billion images, 120,000 YouTube videos and 100 million how-to documents and appliance manuals. They also uploaded all of the training and information they already had given the different robots created in their own individual laboratories.

When a robot learns something, it should feed that experience back into the repository. That way the brain will grow and hold more and better knowledge and experiences to pass on to other robots.

Robo Brain is designed to process the images uploaded into it and pick out the pertinent objects in them. It is learning to recognize objects and how they are used, along with human language and behavior.

That way if a robot needs information, such as how to put away dishes, Robo Brain will be able to provide videos and images related to that particular task.

A robot could learn not only that a cup is a container designed to hold liquids, and that the liquids can be poured into and out of it. The machine also would learn that people use cups to drink liquids, such as coffee or tea.

Cornell also noted that the system employs what computer scientists call structured deep learning, or information stored in many levels of abstraction. An easy chair, for example, is a member of the class of chairs, and going up another level, chairs are furniture. Sitting is something people do on a chair, but a human also can sit on a stool, a bench or the lawn.

"The Robo Brain will look like a gigantic, branching graph with abilities for multidimensional queries," said Aditya Jami, a visiting researcher at Cornell, in a statement. It might look something like a chart of relationships between Facebook friends but more on the scale of the Milky Way."

Only the four institutions involved in the work have access to Robo Brain, though Saxena said he hopes that within six months, that number should grow to 10. In two years, he hopes 100 institutions and companies will have access to it.

This story, "Scientists Create One Robot Brain to Rule Them All" was originally published by Computerworld.

Insider Resume Makeover: How (and When) to Break the Rules
Join the discussion
Be the first to comment on this article. Our Commenting Policies