DARPA Announces Breakthrough Advances in Aerospace Machine Learning
•
The Defense Advanced Research Projects Agency has been named a finalist for the 2023 Robert J. Collier Trophy, which formally recognizes recent breakthroughs that ushered in the era of machine learning in the aerospace industry.
The teams worked together to test groundbreaking implementations of artificial intelligence algorithms using the X-62A VISTA aircraft as part of DARPA's Air Combat Evolution (ACE) program.
“The possibility of autonomous air-to-air combat has been imaginable for decades, but reality has remained a distant dream until now. In 2023, the X-62A will be one of the most important obstacles in combat aviation. “This is a transformative moment, and it was all made possible by the groundbreaking work of the X-62A ACE team,” said Air Force Secretary Frank Kendall. Secretary Kendall will soon fly in the X-62A VISTA and witness the AI first-hand in a simulated combat environment during a test flight scheduled at Edwards Air Force Base, California.
In less than a calendar year, the team went from initial installation of a live AI agent on the X-62A's systems to demonstrating the first AI-to-human engagement within visual range, known as air combat. . The team made over 100,000 total lines of flight-critical software changes during his 21 test flights.
Air combat is a highly complex scenario that the X-62A utilized to prove that non-deterministic artificial intelligence can be used safely within aerospace. AI dogfight paired his X-62A VISTA with a manned F-16 aircraft over Edwards. Initial flight safety was first established using defensive maneuvers, then switched to offensive high-aspect nose-to-air combat, with the air combat aircraft traveling at 1,200 mph and approaching him as close as 2,000 feet.
The first use of machine learning-based autonomy in flight-critical systems will serve as the foundation for future aerospace AI advances that will be safer and more reliable for both commercial and defense applications.
“The X-62A is a great platform that can be used not only to advance the current state of research and testing, but also to prepare the next generation of test leaders. Industry can look to the results of what the X-62A ACE team has done as a paradigm shift when we demonstrate that this can be done safely and responsibly. And it fundamentally changed the conversation,” said Col. James Valpiani, Test Pilot School commander.
While traditional autonomy has been practiced for decades, machine learning has historically been prohibited due to its high risk and lack of independent control. The X-62A is flown by a safety pilot with an independent ability to disable the AI. However, test pilots were not required to activate the safety switch at any point during the dogfight over Edwards. “We have to be able to trust these algorithms in order to use them in real-world environments,” said Lt. Col. Ryan Heffron, his ACE program manager at DARPA.
Although air combat was the main test scenario, it was not the end goal.
“It’s easy to look at the X-62A ACE program and think it’s under autonomous control and capable of air combat, but that’s beside the point. “All the lessons we're learning apply to any task we can give an autonomous system,” said Bill Gray, the school's lead test pilot.
Advances in machine learning will continue as teams at both Test Pilot School and DARPA look to develop lessons learned into future programs of record. The X-62A VISTA will continue to contribute to research for a variety of customers and provide important academic lessons for the next generation of test leaders.
The ACE program is the result of strong collaboration between academia, government, and private industry. Government partners include the Air Force Test Center, Air Force Research Laboratory, DARPA, and Air Force Test Pilot School. Academic partners include Johns Hopkins University and his MIT Computer Science and Artificial Intelligence Laboratory. ACE's industry partners include Calspan Corporation, Cubic Corporation, EpiSci, Lockheed Martin Skunk Works, and Shield AI.
