- author, Michael Dempsey
- role, Technology Reporter
I'm flying a Typhoon fighter jet over the Irish Sea and I'm experiencing major problems.
There's an enemy jet behind me, and no matter how much I push the joystick, no matter how much I manipulate the throttle, the enemy is still there.
The threat appears as a black triangle on the computer screen in front of me, along with a view of the Typhoon's cockpit, and I engage in three consecutive dogfights in the simulator, with poor results.
That annoying little image never left the six o'clock position, and we lost 3-0 to the ominous black triangle.
The ace I'll be facing off against is an AI dogfighter developed by Turkish aerospace engineers on behalf of defense giant BAE Systems.
They are working on developing an AI co-pilot that could rescue real-life pilots in dogfights, which will be deployed in advanced combat simulators.
One of the engineers, Emre Sardiran, a student at Britain's Cranfield University, which has strong ties to aerospace, explains how the AI co-pilot learned combat tactics through a process of trial and error: “We feed the AI more data into the dogfight simulator, which helps it learn better.”
One of his goals is to combat the information overload fighter pilots endure. His colleague Mevlϋt Uzun asserts that AI will need to learn a lot to beat humans: “It's made millions of mistakes. Teaching an AI is like tutoring a child.”
But Uzun says that once trained, AI can offer valuable advice.
“The AI can tell the pilot to slow down or speed up. It can also assess an upcoming dogfight and warn the pilot that there is a 70% chance he will lose if he enters the fight.”
There, the AI would warn pilots of situations that could shoot down the jet, and make the decision within milliseconds, though the design team makes no grand claims that the AI will replace pilots.
“It's just code, so you can run it on your phone,” Uzun said. Currently, their program runs on an ordinary laptop.
The US Air Force will unveil its own, more sophisticated AI dogfighter in 2023, demonstrating it flying F-16 jets in combat exercises.
The flight was the culmination of years of research aimed at creating an AI that could beat living pilots.
In 2020, eight US AI companies went head-to-head in a three-day competition called the Alpha Dogfight Trials event, which pitted competing AI programs against experienced US Air Force fighter pilots in simulated online dogfights.
The winning program beat pilots multiple times, and the three-person team that developed it also included Brett Darcy from US defence company SHIELD AI.
He remembers the Alpha Dogfight event vividly: “Participants were [defence giant] Lockheed Martin came to us.”
They started by pitting the AI pilot against targets flying straight and level — “easy targets,” Darcy said.
They would be pitted against other AI pilots, forcing the AI to come up with tactics, with certain rules set for the length of each dogfight (usually five minutes) and the maximum speed that could be achieved.
But there was no need to follow Air Force doctrine: “Our AI used a head-on collision with a target as an opportunity to fire,” he says.
This novel tactic went against common air combat doctrine, and the AI had learned to reject the rules if it found a better move.
Points are awarded for each battle and the AI evolves based on successful outcomes – this evolution spawns multiple copies of the AI as competing AI pilots respond to each other's changing tactics.
In these qualifiers, Darcy's group was pitted against experienced US Air Force fighter pilots who were fitted with VR headsets to feel like they were in the cockpit of an F-16.
Thanks to their victory over the human pilots, Darcy's small team was invited to join the government's Defense Advanced Research Projects Agency (DARPA), which develops technology for the U.S. Department of Defense.
Specifically, they participated in DARPA's Air Combat Evolution (ACE) project.
When DARPA's AI-driven F-16 flew, control during combat was provided by a distant descendant of software Darcy's team created in 2020.
AI is evolving at an astonishing rate. Darcy says this is “the many-generational grandchild of Alpha Dogfighter AI.”
Despite the incredible pace of development, AI still has a long way to go. The ACE jet has a safety pilot on board during takeoff and landing, and the AI can be switched off at any time.
To be effective, AI pilots must gain a lot of trust and be able to integrate into the broader forces around them.
Interestingly, Darcy says a big question is how an AI pilot would be able to “account for itself on the ground” and report back to human controllers about its actions and motivations.
The UK's AI dogfighters are much more energy efficient than their US counterparts. “They're teaching it to fly the plane,” Dr Uzun says. “We don't have to.”
Shrinking the project by focusing only on the combat movements allowed the Cranfield team to work quickly: “What they did in weeks, we did in two days.”
One man who has spent his career on the rise of digital defence tools is Michael Hull, now a principal engineer at BAE Systems in Wharton, Lancashire, who joined the company as an apprentice electronics engineer in 1990.
Among the dramatic changes he has witnessed is that innovation that once came from within defense companies is now moving in the opposite direction: “We're taking technologies like AI from the public sphere and into defense.”
So the humble tradition of AI dogfighters includes air-to-air combat tactics downloaded from Wikipedia, no consideration of classified information, and contributes to the pace at which the project moves forward.
How did the quickly assembled British AI dogfighter fare against the real Top Gun?
Ben Westby-Brooks flew Typhoons for the Royal Air Force and now works for BAE Systems. He has fought and defeated AI dogfighters.
AI dogfighters are no substitute for thousands of hours flying fast jets in extremely challenging conditions, but they could help with realistic online combat exercises and reduce pilot overload in real cockpits.
