Air Force, DARPA

The U.S. military is claiming a world first after conducting a live aerial “dogfight” exercise which pitted a human pilot against an AI algorithm embedded in a modified F-16 test aircraft, the X-62A Vista fighter jet.

Development of AI-controlled fighter jets could pave the way for a future of fully autonomous air defense. However, a safety pilot was seated in the aircraft running the algorithm during the exercise.

Dogfighting refers to close-range engagements between two or more fighter aircraft, typically involving maneuvers to gain a tactical advantage over the opponent.

The AI-piloted X-62A which successfully navigated complex dogfight scenarios is the latest of a series of test runs focused on autonomous combat maneuvers.

The U.S. Air Force Test Pilot School (TPS) at Edwards Air Force Base in California began the tests in 2023. The exercise was conducted as part of the Defense Advanced Research Projects Agency’s (DARPA) Air Combat Evolution (ACE) program at the TPS.

Flight safety was initially established through defensive maneuvers, before progressing to offensive engagements with the aircraft flying as close as 2,000 feet at speeds of 1,200 miles per hour.

A video of the dogfight was published on YouTube, with Secretary of the Air Force Frank Kendall calling the dogfight a “transformational moment” and a breakthrough accomplishment in the history of aviation.

“The potential for autonomous air-to-air combat has been imaginable for decades, but the reality has remained a distant dream up until now,” he said in a statement.

The program highlights the potential of autonomous AI systems in addressing complex tasks like dogfighting, with lessons learned applicable to various autonomous tasks.

Moving forward, the advancements in machine learning will be applied to future programs by teams from the TPS and DARPA.

The X-62A VISTA will continue serving research purposes, providing insights for the training of future test leaders.

During an online roundtable discussion with reporters about the successful dogfight, Lt. Col. Ryan Hefron, ACE program manager for DARPA and Colonel James Valpiani, commandant of the Test Pilot School, explained the tests are not necessarily geared towards removing the human element completely, but rather augmenting pilot capabilities the way autonomous systems have been doing for decades.

“It’s important to highlight that dogfighting is a challenge problem from a DARPA and a TPS perspective,” Valpiani said. “It is a challenge problem for understanding how to develop machine learning and how to employ it safely.”

He said the dogfight problem was chosen precisely because it is dangerous and difficult, but also because it’s well characterized.

“You know what success and failure looks like in a dogfighting combat set, know what the rules of engagement are and what the training rules are,” he said. That gives us a set of benchmarks we can measure the machine learning against.”

Lt. Col. Ryan Hefron, ACE program manager for DARPA, explained the program leverages a modeling and simulation framework through a technique called deep reinforcement learning.

“It takes an AI agent that has controls for the aircraft–stick inputs, rudder inputs and throttle inputs–and allows the agent you’re training, to try different things,” he said. “As it interacts with the environment, over millions and in some cases, billions of runs, it learns how to control the vehicle.”

The program directions can also specify “rewards” for the algorithm when it achieves desired outcomes, which come in the form of positive assessments of performance that encourage repetition of successful maneuvers.

“The danger, of course, is in anthropomorphizing these things–they are essentially blocks of software,” Hefron said. “What we’re using is a very artful combination of complex mathematics and optimization techniques.”

Valpiani noted the test team’s primary contribution here is not so much that a set of AI agents can perform dogfighting, but broader questions of developing the technology and a workforce that understands this capability and can safely test it, as it is expanded to other capabilities.

Those could include sensor management, beyond visual range sites, or other kinds of applications – for example, delivering cargo.

“The potential for machine learning in in aviation, whether military or civil, is enormous,” Valpiani said. “If we’re focused too much on dog fighting, we’re missing the point of the program which is really how we apply machine learning to combat autonomy.”

With the Pentagon’s anticipated $26 billion expenditure on military simulation and training by 2028, AI is poised to play a pivotal role in various military operations, including the U.S. Army’s “synthetic training environment” capable of simulating joint operations and incorporating cyber, space and maritime domains.