When an AI trashed a human fighter pilot 5- 0 at the culmination of DARPA’s AlphaDogfight event last month, skeptics were quick to point out that simulated combat is not the same as a real thing. Two new rounds of events, one military and one from the gaming world, will go further towards settling the issue.
In the DARPA event the human pilot followed his training and established Air Force doctrine. This meant being wary of maneuvers that took him to dangerously low altitude, and not engaging in high-risk tactics like head-on clashes with his opponent. By contrast, throughout the competition the AI pilots were bloodthirsty risk-takers that acted without any thought for their own survival. Falco, the winning AI produced by Heron Systems, was one of the most aggressive.
In other words, the AIs played like it was a game, whereas the human treated the simulation like reality and would have ‘crashed and burned’ in the real world. A better contest would pit the AI against an experienced gamer used to virtual combat rather than an experienced pilot.
The AI makers have risen to the challenge.
“You asked. You talked trash. We listened, nodded, and reached out to Mover. Let’s do this!“ was Heron’s response.
Folds of Honor, a charity supporting military families, stages an annual online dogfight contest on Digital Combat Simulator software. This year the winner will get the opportunity to take on Heron’s Falco AI. Details of how to sign up are here. Will a new champion regain humanity’s honor? The gaming world, among others, will be watching.
But however good an AI is at gaming, it is not the same as the real world, and the software’s perfect knowledge of its simulated environment gave it an unfair advantage over its human opponent. The only true test will come when the combatants are flying real aircraft – and the U.S. Air Force is setting up just such an encounter.
“These simulations will culminate in a real-world competition involving full-scale tactical aircraft in 2024,” U.S. Secretary of Defense Mark Esper told the DoD Artificial Intelligence Symposium and Exposition this week.
The virtual combat involved simulated F-16s, which are not the Air Force’s first-line aircraft, but their details are declassified (and so could be provided to the software developers) and they do not have the complications of stealth like the more modern F-22 and F-35. The Air Force has recently modified some of its F-16s into unmanned QF-16 drones that are flown remotely as targets for air-to-air combat training and missile tests. These would be an obvious choice for the next stage, as they are already set up for operation without a anyone in the pilot seat.
However, it might be argued that putting an AI in a plane designed for a human loses one of its biggest advantages in dogfighting: its ability to withstand maneuvers that would cause a human to black out. Fighter planes are designed around the limitations of the pilot. A highly agile dogfighting drone developed in 1971, flown by remote control, easily beat two human combat pilots. (It was promptly shelved. Nobody likes an uppity robot).
There were even more ambitious plans for dogfighting drones. Under a program codenamed Red Baron, the Air Force modified Ryan Model 147G drones to the XQM-103 configuration. These had a more powerful engine, reinforced wings, and a new digital flight-control computer, and could pull 10G turns. The acceleration on a rollercoaster might feel powerful, but they only briefly produce 3G-4G, the F-16 can pull 9G, meaning the pilot experiences a force nine times greater than gravity. F-16 pilots must not only wear G-suits but practice the Anti-G Straining Maneuver, a combination of breathing and muscle tensing just to stay conscious.
“Under peak G, you’re spending the majority of your effort pancaked into your seat, trying not to pass out,” notes one pilot.
Electronics can tolerate much higher G-forces, so an AI pilot can turn continuously at high rates for extended periods while maintaining complete focus on shooting down its opponents.
It is unlikely that any human vs. AI dogfight with cannon that the contests simulate would take place in a real war. Air-to-air combat starts with radar-guided missiles fired at far beyond visual range. At shorter ranges, modern heat-seeking missiles can be fired ‘off axis,’ meaning the shooter does not need to get lined up with the enemy, let alone get on their tail. The latest AIM-9X can reportedly hit targets at angles of more than 90 degrees, so little maneuvering is necessary.
However, drone versus drone dogfights are another matter. The U.S. Navy is already looking at combats with swarms of hundreds or thousands of small drones on each side. Such drones are likely to be armed with cheap unguided weapons rather than missiles, to the result will be a mass dogfight. In this situation, a force with superior algorithms could win even if hugely outnumbered, if their drones can win every individual one-to-one combat.
Our pride may be at stake in the new round of contests, but in the longer run, machines beating mere humans may no longer be the major issue. Their real goal will be beating other machines.