Chowdhary group uses AI to improve stability of autonomous aircraft
On an airplane, if something goes awry midflight, the pilot is there to take over and land the plane safely. When something happens to an aerial vehicle piloted purely by artificial intelligence (AI) technology, there is no fail-safe. CSL student Girish Joshi and his adviser, CSL Assistant Professor Girish Chowdhary, are working to improve the stability of autonomous aircraft in their newly funded project, “Fall-tolerant adaptive control for over actuated aircraft.”
“Autopilots have been around forever, but we still need a pilot in the airplane when crazy things happen,” said Joshi, an aerospace engineering student planning to graduate in May. “How do you create AI that can deal with these cases that only happen once in a blue moon, but when they happen, impact human lives?”
The duo and their research team have worked to create a drone that can keep flying even when there are issues mid-air.
“Faults can occur, airplanes can get hit, or the unpredictable thing can happen,” said Chowdhary, assistant professor of agricultural and biological engineering and aerospace engineering. “We’d like to create controllers using AI that can keep aircraft flying even in the presence of faults or uncertainties.”
Current advanced controllers operate through rule-based, state-feedback algorithms, without taking into account contextual information. This is problematic because flying in good conditions is not the same as flying at a high altitude, high Mach speeds, or in turbulent conditions. To include contextual information, the next iteration of AI-based controllers would use inputs from a variety of sensors other than the standard state measurements. These advanced sensors would not only provide information about vehicle health, but would also include visual sensors to access situational data, including the surrounding environment, for better decision making.
The group has had success with their strategies, as seen in this drone footage. In the video, shortly after the drone takes flight, one of the rotor blade tips breaks, but the drone manages to stay afloat.
“There are no current controllers that use Deep Neural Networks in the loop and yet guarantee this kind of stability,” said Joshi. “The type of systems we are developing are evolving, so there is a need for evolving controller architectures that can catch up with advancement in high-performance systems.”
There has undoubtedly been a lot of research dedicated to improving AI technology. However, Joshi and Chowdhary, who also has appointments in electrical and computer engineering and computer science, believe they are conducting the research that could “close the loop” and make AI-based advanced aircrafts a realistic possibility.
Other students involved in this project include Jasvir Virdi (Mechanical Science and Engineering) and Garrett Gowan (Aerospace Engineering). This project is funded by Sandia National Laboratory.