This article was written by Alex Hollings and originally published on Sandboxx.
DARPA’s AlphaDogfight trials have officially come to a close. Heron Systems’ incredible artificial intelligence pilot system defeated not only its industry competitors but went on to secure five straight victories against a highly trained U.S. Air Force F-16 pilot, without the human pilot scoring a single hit.
Eight teams were selected to create artificial intelligence (AI) “agents” that would be capable of simulating a real dogfight between fighters, referred to as “within-visual-range air combat maneuvering,” more formally. The first two rounds of the competition saw these virtual pilots engage with one another in simulated combat environments in November and again in January. This third round of AI dogfighting included similar competitions, with the four finalist firms squaring off in a round-robin. The event then culminated with the hands-down victor, Heron Systems, taking on a real human fighter pilot in another simulated fight.
And Heron really brought the heat, with its artificial intelligence system ultimately securing the AI championship by defeating Lockheed Martin’s AI system.
Heron consistently proved to have the most accurate targeting apparatus of any AI agent, as it engaged opponents with laser-accurate gun strikes often in the first merge of the fight.
“It’s got to keep that opponent in that one-degree cone to win the game,” Ben Bell, Heron’s Senior Machine Learning Engineer, told Sandboxx News.
“You saw that a lot with Lockheed, we’re both nose on, we’re both creating damage, but when their nose is off by that one degree, that’s where we were able to win a lot of these engagements.”
That superior aiming capability was on particular display when squaring off against the U.S. Air Force F-16 pilot representing humanity in this battle for what some consider to be the future of aviation. While the pilot’s name was not released due to OPSEC concerns, DARPA did provide his callsign: Banger. They also explained that Banger was not just a working fighter pilot: he’s a graduate of the Air Force’s Weapons Instructor Course, which could loosely be described as the Air Force’s “Top Gun” school, for the movie buffs out there. (The real Top Gun, of course, is a Navy school called the United States Navy Strike Fighter Tactics Instructor program.)
Heron’s AI system racked up the first four wins against Banger in quick succession, leveraging its incredibly precise aiming to whittle Banger’s aircraft “life” down in a series of looping merges. In the fifth and final bout, Banger changed approaches, sweeping his aircraft out away from Heron’s F-16 and creating separation with high-G turns.
However, the new tactics only seemed to delay the inevitable, with Heron killing Banger’s F-16 once again, without the human pilot managing to get a single shot on target.
Heron’s AI pilot was widely described as “aggressive: by DARPA staff and the Air Force pilots on hand throughout the competition. Under the control of Heron’s AI, the virtual F-16 would practically play chicken with its opposition. The human pilots were quick to point out that this would be a violation of training regulations in a real simulated dogfight. Of course, in an actual dogfight, there are no such limitations… but Heron’s aggression may still have been turned up just a bit too high for it to serve as a reasonable wingman.
“It’s important to realize that a BFM (Basic Fighter Maneuvers) engagement can occur in any direction and any altitude. We’ll often begin with a basic starting parameter to develop a site picture to reference, but a real engagement doesn’t have those cuffs,” Major Justin “Hasard” Lee, an F-35 Pilot instructor and former F-16 pilot, told Sandboxx News.
“The enemy always has a vote, meaning they always reserve the right to do something you’re not expecting. When that occurs you have to find a creative solution to counter the unexpected problem.”
According to Bell, their AI agent placed an equal emphasis on damaging the opponent and minimizing its own risk.
“If the agent sees a 51 percent chance of scoring a kill as it heads into a neutral merge, it’s going to take it,” Bell explained.
Of course, aside from some really exciting videogame playing, this entire exercise had official purposes too. DARPA is not only seeking to improve drone aircraft systems but it’s also looking to increase the level of trust between human pilots and AI systems. In the future, these same sorts of artificial pilots will likely be flying alongside humans. And other similar systems will fly along with them in the cockpit of their own aircraft.
Bell explained that the victorious AI agent could be adjusted to prioritize its own safety to a higher degree. This might make pilots a bit more comfortable with the AI’s approach to combat. He also pointed out, however, that just because something’s scary to human pilots, doesn’t mean it isn’t effective.
“Trust comes from being able to execute a mission with a high degree of success. There’s some point where you have to say you know that it works and in all the ways we tested it, it was superior to its opponent.”
He went on to clarify, however, that there will certainly be “some give and take” between their engineers and real pilots moving forward.
When asked about Heron’s ace in the hole, its incredibly accurate targeting system, Bell made sure to point out that the way in which this competition was executed was to the human pilot’s disadvantage. Banger was flying in a simulated environment using a VR headset. This doesn’t equate that well to a real fight in the real sky, and gives the computer pilot instant awareness of its surroundings.