Drastic wing failure, as shown in the above illustration, would probably result in irreparable damage to an unmanned aerial vehicle. Learning must start and finish within a single execution. But, the newly proposed vehicle would have more than just sparse data at its disposal. Physical information about its immediate environment could also be used as it learns "on the fly." Cockrell School of Engineering, The University of Texas at Austin
AUSTIN, Texas - Thanks to a Department of Defense grant, researchers are planning for a future when unmanned aerial vehicles (UAVs) have the ability to fly themselves in emergency situations.
A research team from the Cockrell School of Engineering and the College of Natural Sciences at The University of Texas at Austin has been selected by the DOD to lead a $7.5 million Multidisciplinary University Research Initiative (MURI) project aimed at developing artificial intelligence for UAVs.
While almost all artificial intelligence, or AI, technology is reliant on the availability of massive amounts of data, Texas Engineers have been charged with the task of developing machines that can learn "on the fly" in situations where there is little data to inform them. The interdisciplinary team hopes their combined efforts will assist the Department of Defense in the development of truly autonomous systems that can not only operate in challenging environments but also survive disruptions or recognize when they are fatal.
When faced with a tough decision and lacking the requisite understanding or experience to make an informed choice, human beings often rely on intuition, gut feeling or even figuring things out in the moment. This decision-making trait does not smoothly exist in current machine learning technology.
The MURI research team, therefore, faces some big challenges. But if they can provide new insights and technology, the implications for machine learning could be wide-reaching, applicable to everything from driverless car safety to unmanned aerial vehicles in emergency situations.
Working with the U.S. Air Force Office of Scientific Research, the team is being guided by Ufuk Topcu, an assistant professor in the Cockrell School’s Department of Aerospace Engineering and Engineering Mechanics who conducts his research within the UT Institute for Computational Engineering and Sciences (ICES), and includes researchers from Northeastern University and Princeton University, such as pioneering mathematician Charles Fefferman.
"Our approach to this unprecedented challenge embraces the fact that developing truly autonomous systems - in particular, control-oriented learning on the fly - is beyond the reach of any single discipline," Topcu said.
Topcu is bringing together engineers, scientists and mathematicians to tackle analysis, control theory, dynamical systems, learning theory, optimization, formal methods and other essential areas related to machine learning and AI.
Two core members of the team from UT’s College of Natural Sciences - Rachel Ward, associate professor of mathematics and member of ICES; and Arie Israel, assistant professor of mathematics - said they have their work cut out for them.
"The problems I’ve studied in the past involve situations where you have very limited or sparse data - and a wide range of uncertainty. So, it’s about discovering that uncertainty and fully leveraging the data available in order to make as many deductions as possible," Israel said. "But the algorithms I develop aren’t dynamic. They take a long time to process data and are inefficient. This is an exciting challenge."
Ward, on the other hand, specializes in machine learning and optimization and currently works with Facebook in this area.
"In particular, I look at how to make very efficient, fast algorithms based on a lot of data," she said. "I look forward to applying my experience to this project."
The two will work together with the rest of the research team to develop efficient algorithms using limited data.
Topcu will incorporate his expertise in formal methods, learning and control of autonomous systems by providing insight into how a machine deciphers the physical environment around it.
"The algorithms we develop will obey and leverage the laws of physics and contextual knowledge, adapt to unforeseen changes in the system and its environment and establish verifiable guarantees with respect to high-level safety and performance specifications," he said. "On-the-fly learning can be realized only by an integrative approach - precisely what we propose to develop."
The team’s ultimate objective is to develop advanced autonomous systems that will help reduce the number of UAVs lost in emergency situations caused by unforeseen system failures or damage.
This UT-led project is one of 24 total MURI awards given by the Department of Defense in 2018. The grants, which total $169 million over the next five years, have been awarded to academic institutions to perform multidisciplinary basic research.
In 2016, three Cockrell School professors received MURI grants totaling $22.8 million to help advance innovative technologies in energy, computing and nanoelectronics.