AI Beats Human In Mock Dogfights


An artificial intelligence algorithm beat an “experienced” F-16 pilot in five straight simulated dogfights in a project designed by the Defense Advanced Research Projects Agency on Aug. 20. The pilot was up against an AI system named Heron and the two duked it out in a Second World War-style forward firing gun battle. The human didn’t score a single hit. Heron, which was developed by Heron Systems, competed against five other AI algorithms for the right to fight the human in a computer simulation and the result of the final battle neither surprised nor concerned observers.

While significant, the milestone is an early one in DARPA’s goal to develop a machine-learning enhancement to human performance with the ultimate goal of “a future in which AI handles the split-second maneuvering during within-visual-range dogfights, keeping pilots safer and more effective as they orchestrate large numbers of unmanned systems into a web of overwhelming combat effects.” It’s also not the only AI/human research being undertaken by the military. As we reported earlier, a human pilot in a fighter is expected to go up against an AI-controlled drone in 2021 although it’s not clear what aircraft will be used.

Russ Niles
Russ Niles is Editor-in-Chief of AVweb. He has been a pilot for 30 years and joined AVweb 22 years ago. He and his wife Marni live in southern British Columbia where they also operate a small winery.

Other AVwebflash Articles


  1. No one? NO ONE watched Star Trek? The M-5 unit worked ok in demo but was ultimately flawed. Putting planet killing arms in the hands of “programs” is the ultimate flaw. Dr. Strangelove anyone?

  2. Yahh, yahh, better than humans, infallible, only a matter of time, etc. etc. I’ll give credit when credit is due, when AI flies a glider task on a marginal day and then orchestrates the retrieve when it lands out.

    • Arthur: A better reference (among a myriad) is “War Games”. So they put a miniaturized WOPR in an F-16; what could possibly go wrong? It’s only a matter of time.

      Chris: The only reason we haven’t seen AI technology applied to un-powered aircraft is that they make poor delivery systems for ordnance. But they make great platforms for surveillance and intel-gathering: practically undetectable (thus difficult to hit), easily re-positionable, and offering a generous loitering time at high altitudes. And a lot cheaper than a U-2/SR-71/whatever black budget bird they are using now. It IS only a matter of time.

      As for “Star Trek”, if I owned an Alexa, Siri, et al. I would reprogram it to respond to “Computer”. Aside from the somewhat creepy “always knowing where you are” factor, how is “Star Trek” technology not already ever-present?

      As for cinematic references, “Star Wars” (and many others, going back to “The Jetsons”) portrayed a future in which we blithely travel the skies in our personal self-flying pods, which deliver us to our destination while adroitly avoiding the swarm of everyone else’s pods. It’s only a matter of time.

      • The only reason AI is not seen in unpowered aircraft is that it cannot experience flight for flight’s sake.

  3. AI (“artificial intelligence”) employs a heavy dose of “machine learning.”
    This is shorthand for experience-based “code branching,” which purposefully causes “BEHAVIOR branching.” The designer WANTS the machine to “learn.”

    (Think back to your dog-eared copy of The Fundamentald of Instruction, and this axiom: “Learning is a change of behavior that occurs as a result of experience.”)

    Unpredictable behavior can be very useful in a dogfight. It is utterly counterproductive in ordinary aviation applications. Which is why it is the WRONG way to design an autonomous civil aircraft control system.

    • Bear in mind that whilst a human could take several seconds to decipher “unpredictable behavior”, the computer may see what’s happening in real time, and react more quickly than a human could (if you have a 12-G airframe, the computer could hit that number when required with zed chance of blackout.) It simply wouldn’t be “unpredictable” to a fast processor.

    • It’s not so much that, but rather this was a simulated combat. The human pilot didn’t have any of the normal cues they would have in the actual aircraft, and apparently the AI had “perfect state information”. That is, it knew the heading, altitude, airspeed, G-load of the enemy aircraft, while the human did not know the state of the AI aircraft.

  4. In my early aviation career we used to joke that they would replace the three man crew with one pilot and a dog. The pilot’s job was to feed the dog and the dog’s job was to bite the hand of the pilot if he tried to touch anything. Now we are going to skip this phase altogether.

  5. “a future in which AI handles the split-second maneuvering during within-visual-range dogfights, keeping pilots safer and more effective as they orchestrate large numbers of unmanned systems into a web of overwhelming combat effects.”. Wonderful DoD talk to keep pilot recruiting and retention up until they get their walking papers in about 10 years or so. Hey, war is the ultimate human failing anyway.

  6. If it works better, do it. We got enough examples of people choosing poorly when they know better. Don’t need our military losing because there was too much pride. That’s been done before to good outcome.

  7. Some interesting things about this one. 1. the pilot didn’t have much time in a simulator of this type, 2. the simulation didn’t have to simulate the full physics of a gun shot it just had to get the opposing aircraft in a cone in front of its aircraft and 3 since the simulator didn’t have the ability to simulate sensors radar, cameras etc. the AI was provided with the exact xyz position and velocity information of the opposing aircraft at all times allowing it to calculate exact positions for a gun shot this is why it always tended to go for head on shots. The AI is getting better but it is still not there yet. If I was that pilot I wouldn’t worry about my job yet.