Researchers at UC Berkeley are working alongside those at other universities to build a large-scale online learning system called “Robo Brain” where robots can access a single, global database collected from the Internet to learn how to perform human tasks.
The effort is led by Ashutosh Saxena, assistant professor of computer science at Cornell University, who first presented the project at UC Berkeley’s Robotics: Science and Systems Conference in July. Saxena described Robo Brain as a “search engine for robots,” with researchers having already collected about one billion images, hundreds of thousands of videos and a million documents and manuals. This raw data is assembled to help robots identify and understand how humans and objects interact.
The ability of a robot to comprehend representations of the human world from multiple sources is challenging, according to Saxena. But Robo Brain aims to make this easier through the use of language, text, perception, visuals and planning. Once robots use Robo Brain to link text with visuals, they can ideally plan how to perform certain functions like a human brain does.
For example, Saxena described the case of a coffee mug: When robots learn about a mug through Robo Brain, they won’t just see what the object is but will learn its characteristics, such as how it’s grasped and how liquids such as coffee can be poured into and out of it.
Ken Goldberg, an electrical engineering and computer sciences professor and an advisor of the Robo Brain research at UC Berkeley, noted how Robo Brain is a peer project of the Cloud Robotics and Automation research — a 20-year effort at UC Berkeley to study robotics and the web.
“That’s what the Robo Brain is about — sharing data across different kinds of systems and platforms so they collectively get better,” Goldberg said.
Goldberg and EECS associate professor Pieter Abbeel oversee a group of robotics research students composed of graduate students, postdoctoral researchers and undergraduate students who work on Robo Brain alongside other projects.
Ben Kehoe, a graduate student in Goldberg’s research group, explained the advantage of cloud computing for robotics.
“There are many robots out there, and things that one robot does, all robots can learn from — and that’s where the cloud comes in,” Kehoe said.
One of the team’s postdoctoral researchers, Sachin Patil, expanded on the importance of big data within robotics and how there can be “potentially infinite storage in the cloud.”
In addition to Robo Brain and Cloud Robotics and Automation, Goldberg said that about a dozen faculty spanning various departments are conducting robotics research at UC Berkeley, covering areas such as surgical robots and exoskeletons.
The Robo Brain project is set to expand. Saxena said the first stage is a collaboration with four universities to get the system up and running, but that by the end of this year there will be 10 universities on board, including some international institutions.
“We are very actively involved in collaboration,” Goldberg said of robotics research happening at UC Berkeley. “It lets us build on the individual strengths of different research groups around the world.”