U.S. Navy Lt. Cmdr. Michael Tremel was in the cockpit of his Super Hornet fighter soaring 20,000 feet over war-torn Syria on June 18, 2017, when he detected a warplane heading toward a U.S.-backed militia force on the ground.
Tremel tried to shoo away the other pilot by firing off bright decoy flares. But the Syrian aviator ignored the warning and dropped bombs on the militia position. Tremel reacted in seconds, firing two air-to-air missiles that sliced off the other plane’s tail.
It was a dramatic display of American pilots’ traditional prowess in air-to-air combat. Since the Vietnam War, U.S. fighters have shot down 58 enemy planes. In that span of time, only one U.S. pilot has lost an airborne skirmish.
In other words, a highly-trained human being can still make a fearsome aerial warrior. But the scientists at the Defense Advanced Research Projects Agency (DARPA), the Pentagon’s fringe-science outfit, have long had a hunch that a drone might do even better. To test that theory, DARPA organized a computer-simulated war game between bots and one nervous human pilot.
That war game came to a close on Thursday with a dramatic (digital) head-on fight between an A.I. and a flesh-and-blood flier.
Modern military drones mostly fly surveillance and ground-attack missions. The U.S. military wants a new generation of fast drones to handle air-to-air missions, too. But dogfights can be quick and chaotic. It has long been conventional wisdom that it would take a very clever artificial intelligence to defeat a quick-thinking human pilot.
In early 2019, DAPRA recruited eight teams of coders to develop special A.I. for aerial warfare. The competitors in the AlphaDogfight program included major defense firms such as Lockheed Martin and Boeing, smaller companies such as Maryland-based Heron Systems, and scrappy squads from schools such as Georgia Tech.
In a series of trials beginning late last year, the agency pitted the eight A.I.s against each other as well as against a generic simulated enemy fighter developed by the Johns Hopkins Applied Physics Laboratory. The Johns Hopkins lab hosted the AlphaDogfight trials at its facility in Laurel, Maryland.
The contest “really feels like an e-sports event,” Tim Grayson, the director of DARPA’s Strategic Technology Office, which oversees AlphaDogfight, said at the time. The mock dogfights played out in a Flight Simulator-esque digital world. Red and blue icons representing bot-controlled F-16 fighters turned, climbed, dived, and shot at each other.
To P.W. Singer, an analyst at the New America Foundation in Washington, D.C. and author of the novel Burn-In, the AlphaDogfight war game was “a John Henry-versus-the-steam-engine moment.”
That is to say, it hinted at a massive shift in how war is practiced.
“There is a nobility to the human role, but it symbolically points to a future of more and more machines in more and more roles,” Singer told The Daily Beast.
Early results at the Johns Hopkins lab were mixed for the machines, however.
“The [A.I.] agents that were well-developed at this point—they were able to handle very well the adversaries that were somewhat predictable or operating the way they had trained against,” Air Force lieutenant colonel Justin Mock, an F-16 pilot who observed the initial A.I. trials, said in an April video. “But they struggled with those adversaries that did something even just a little different.”
Even so, the bots demonstrated some creative thinking. Some roared straight into the fight, hoping to surprise and overwhelm their opponents. Others moved carefully into an engagement, cautiously maneuvering for advantage. At least one A.I.-driven fighter flipped upside down before taking a shot with its simulated gun—a maneuver many human pilots can’t pull off.
Heron’s hyper-aggressive A.I., which favored head-on gun attacks, came out on top in Thursday’s trials, shooting down 213 computerized enemies while getting shot down just 16 times in return. The grand prize: a chance to go head-to-head with an actual F-16 pilot steering a simulated jet in a computer-game-style setup.
That bot-versus-human showdown was the final event in this phase of the AlphaDogfight initiative. The live F-16 pilot, identified only by his call-sign “Banger,” worried aloud about the A.I.’s aggressive tactics as he flew his simulated jet into battle.
For years, the Air Force has mulled adding a dogfighting A.I. to an array of high-tech new drones the flying branch is developing. But as recently as last year, DARPA was skeptical a bot would be up to the task. “No A.I. currently exists ... that can outduel a human strapped into a fighter jet in a high-speed, high-G dogfight,” the agency stated.
Attitudes slowly began changing a few years ago. In 2016 Psibernetix, a small software firm in Ohio, developed “Alpha,” an artificial intelligence specifically for controlling fast-flying, armed drones during an aerial dogfight.
The company tested its A.I. in a series of computer-simulated dogfights with manned enemy planes. Incredibly, the drones won by dodging enemy missiles before maneuvering for a good angle and firing missiles of their own. Where a human pilot needed a third of a second to react to an enemy’s move, Alpha could react in milliseconds.
“The speeds at which Alpha can intelligently operate serve as a distinct advantage within the context of air-to-air combat,” Psibernetix CEO Nicholas Ernest and his co-authors David Carroll, Corey Schumacher, Matthew Clark, Kelly Cohen, and Gene Lee wrote in a 2016 issue of Journal of Defense Management.
The U.S. military wasn’t the only one to grasp the potential of dogfighting A.I. In the past couple of years, the Japanese, Chinese, Russian, and Australian armed forces have also begun developing A.I. for air-to-air missions. “Everyone is working that angle,” Samuel Bendett, an analyst with the Washington, D.C.-based Center for a New American Security, told The Daily Beast.
But the Americans appear to be in the lead, thanks in large part to AlphaDogfight. And according to DARPA, the A.I.s didn’t even need to win all or even most of their dogfights for the program to be considered successful.
“Regardless of whether the human or machine wins the final dogfight, the AlphaDogfight trials is all about increasing trust in A.I.,” Col. Dan Javorsek, the DARPA program manager, said early in the trials. “If the champion A.I. earns the respect of an F-16 pilot, we’ll have come one step closer to achieving effective human-machine teaming in air combat.”
In parallel with AlphaDogfight, the Air Force is developing a drone A.I. it calls “Skyborg.” The plan is to add Skyborg to maneuverable, jet-powered drones that would fly in formation with human-piloted planes, adding new sensors and weapons to the battle.
"Skyborg is a vessel for A.I. technologies that could range from rather simple algorithms to fly the aircraft and control them in air space to the introduction of more complicated levels of A.I. to accomplish certain tasks or subtasks of the [combat] mission," Matt Duquette, an engineer with the Air Force Research Laboratory, explained to The Daily Beast.
For now, Skyborg is mostly for ground-attack missions. But the Air Force could take the code that comes out of AlphaDogfight and add it to Skyborg. That way, the Skyborg-controlled drones would be able to battle enemy planes, too.
Those drones are still years away. But dogfighting A.I. undoubtedly will continue to improve in the meantime, by way of AlphaDogfight and similar efforts. “The key is that this is not really a contest, but a learning exercise for both the military and the machine,” Singer said. “The military is learning what works or not, but so is the system itself.”
Back at the Johns Hopkins lab, F-16-pilot Banger was right to worry about his A.I. opponent in the final showdown of the AlphaDogfight trials.
Speeding hard toward the A.I. fighter in the digital air-space, Banger managed to avoid the bot’s first shot. But the A.I. kept up as Banger struggled to circle around. “The standard things we do as fighter pilots aren’t working,” Banger moaned.
In quick succession, the bot fired three times at Banger—and hit each time. It was a flawless victory for the A.I., and a possible hint of things to come as the Pentagon teaches bots how to dogfight.