Robots programmed to acquire knowledge like children do
The Robotics Institute at Carnegie Mellon has been known to carry out some of the campus’s most futuristic and cutting-edge research. This week’s Research Spotlight focuses on experimentation conducted by Assistant Professor of Robotics Abhinav Gupta and his team. Gupta’s research aims to improve the coordination of robots via the use of machine learning and visual learning. The research team and their project will be aided by a three-year $1.5 million award courtesy of Google.
Robots are being used in a variety of fields, and one overarching requirement across the board is that the robots must have dexterity when interacting with objects. This research shows that such nimble and precise movement cannot be achieved directly by hard-coding it into the robot’s central processing unit, but rather by allowing it to heuristically learn how to behave. Just as a child learns to hold different objects with time, so too do the robots Gupta experimented with.
Whereas research groups of yesteryear have suffered largely from data shortage and very expensive robots, Gupta and his team have benefited from the economic advantages of the robotics industry today. “The cost of robots has come down significantly in recent years, enabling us to unleash lots of robots to collect an unprecedented amount of data on physical interactions,” commented Gupta in a Robotics Institute article. With more and more data, combined with mostly isolated visual learning breakthroughs, robots can now use sensory information to understand the ever-changing physical world they occupy.
Excitingly, if one of these many robots learns something new, it can share that knowledge with the others through an adaptable network. This greatly reduces the learning curve and lessens the man hours such experimentation requires. Lerrel Pinto, Ph.D. candidate and member of Gupta’s team, discussed the benefit of this faster data accumulation technique: “If you can get the data faster, you can try more things — different software frameworks, different algorithms.”
Robots that are able to learn and adapt like children, albeit within a slower time frame, are a huge step in the direction toward a more precise artificial intelligence. However, it is this research team’s method of data collection that has the ability to benefit fields far beyond robotics.