NSF CAREER Award: Expert Expands Accessibilty for Dexterous Robotics Research

A Georgia Tech roboticist is making a complex and restrictive field of research more accessible by using frugal methods that do not require vast amounts of computer and data resources.

Harish Ravichandar, an assistant professor in the School of Interactive Computing, earned the National Science Foundation (NSF) CAREER Award for his proposal to develop dexterous robotic manipulation algorithms that use limited resources without sacrificing performance.

Ravichandar says dexterous robots are key to automating various industries, including manufacturing, agriculture, healthcare, and consumer robotics. Current methods to achieve dexterity require massive amounts of training data only available to well-resourced industry laboratories. 

That means many academic researchers and small businesses don’t have the privilege of developing dexterous robots on a scale that would impact the industry. But Ravichandar says combining classical tools relevant to robotics, physics, and math with the latest advances in machine learning can improve efficiency, self-sufficiency, and reliability.

“We present an alternative paradigm to approach these problems,” said Ravichandar, director of the Structured Techniques for Algorithmic Robots (STAR) Lab at Georgia Tech. The idea is for the robot to learn only what it needs to learn and avoid reinventing the wheel whenever possible.”

To Scale or Not to Scale

Ravichandar pushes back on the narrative that vast data resources are necessary to push the boundaries of robot dexterity. 

“Many institutions can’t take advantage of scaling due to limited resources, so let’s identify alternative ways to solve the problem,” he said.

“There is some sentiment in the field about what can be done at institutions with limited resources. There are valuable things we can do without scaling. Whether scaling is necessary to achieve robust robotic intelligence is an unsettled debate.”

Image

Ravichandar and his lab will spend the next five years developing new structured algorithms that learn dexterous manipulation skills by watching people through videos and motion capture. These algorithms will also test the learned skills in practical environments and actively learn from the experimentation — a stage he calls “learning by playing.”

Future Implications

Ravichandar plans to disseminate his findings through partnerships with HBCUs and minority-serving institutions as part of his education and outreach efforts.

Those plans include a guest seminar series and a student research exchange program.

He will also work with Georgia Tech’s Center for Education Integrating Science, Mathematics, and Computing (CEISMC) to introduce high school students to robotic learning research.

The Georgia Tech Research Institute (GTRI) is part of an ongoing project to automate chicken deboning through robots. The project is part of a $5 million grant from the U.S. Department of Agriculture’s National Institute of Food and Agriculture. GTRI received $2.1 million of that grant to develop the system with advanced robotics.

Ravichandar says GTRI has taken an interest in his work, and the new algorithms he develops could be used to automate deboning.  

“Georgia is one of the leaders in the poultry industry, but there’s a huge turnover in jobs because the job is difficult for people to perform for long hours,” he said. 

Ravichandar says a frugal learning approach that builds upon hard-earned insights about physical systems will help robots efficiently learn and reliably perform such tasks in the human world.