Columbia University is part of a national effort to teach artificial intelligence to do something today’s most advanced models still struggle with — act safely and reliably in the physical world. The National Science Foundation has awarded a new $20 million grant to the NSF AI Institute in Dynamic Systems, a multi-university research center in which Columbia Engineering plays a leading role.
The award is the second tranche of a $40 million total investment. It will fund fundamental AI research in aerospace, robotics and manufacturing, as well as training for a new generation of engineers who can bring AI onto the factory floor.
Five researchers, four universities
The institute is led by five researchers from four institutions: Steven Brunton and Krithika Manohar of the University of Washington in Seattle, Hod Lipson of Columbia University in New York City, Na Li of Harvard University and Ricardo Vinuesa of the University of Michigan in Ann Arbor.
Lipson, the James and Sally Scapa Professor of Innovation and chair of Columbia Engineering’s Department of Mechanical Engineering, leads the Columbia portion of the institute.
The next big phase of AI
The institute is focused on what many experts see as the next major frontier for the technology — physical AI. “Robotics, embodied AI and physical AI are all synonyms for the technology that allows AI to gain agency in the physical world — when it can touch, move, and poke in the real world,” Lipson explained.
That shift is far harder than it sounds. “Today’s state of the art AI models can see, hear, think and talk about the world, but they have a hard time physically interacting with it,” Lipson said. The physical world is more complex than the virtual one, demands more energy and unfolds in real time. Errors also carry a much higher price. “Losing a game of chess is different than physically breaking something,” he noted. As a result, AI needs new ways to learn and train before it can safely operate around people and property.
Physical AI is not limited to humanoid robots. Some of its simpler forms, such as driverless cars, are already on the road and are already eight times safer than human drivers per mile, according to Columbia Engineering. But as AI systems move into factories, fields, streets and homes, new technologies and safeguards will be needed to make that transition smooth, safe and effective.
Aerospace as a proving ground
The institute will concentrate in particular on learning the dynamics of complex robotic manufacturing systems, with a special emphasis on aerospace. The industry — spanning everything from transportation to defense — builds highly complex systems in relatively small quantities, which makes traditional mass production difficult and automation especially valuable.
Researchers see many ways AI can speed up that automation, from self-learning digital twins and advanced world models to physical systems that can repair themselves and adapt in real time. As these innovations mature, they are expected to spread into other industries.
Robotic foosball at 60 mph
At Columbia, one of the tools used to train engineers is a robotic foosball table that can play the ball at speeds of up to 60 miles per hour. Engineers use it to develop fully autonomous, fast, dynamic and visually guided control systems, and they compete by submitting their AI players to the Robotic-Foosball world tournament.
Alongside the research, the institute will focus on preparing engineers who are well versed in applying AI to manufacturing, robotics and aerospace — the workforce that will be needed as physical AI moves from the lab into everyday life.
Source: Columbia Engineering