back to article So long, Top Gun... AI software waxes US F-16 pilot's tail 5-0 during virtual dogfight drills

An AI bot defeated a human pilot in a series of virtual dogfights that unfolded in skies albeit within a flight simulator during a competition held by the US military research arm DARPA. The fighter pilot battling on behalf of us humans, a US Air Force instructor nicknamed “Banger,” struggled to fend off the AI system …

  1. Glen 1 Silver badge
    1. DavCrav Silver badge
  2. Anonymous Coward
    Anonymous Coward

    Well, duh.

    That new fancy F-35 fighter plane was obsolete before it could even take off. Everybody with half a brain cell knew it.

    1. Glen 1 Silver badge

      I dunno, autopilot is a thing. If they don't have macro hotkeys like "perform high-G blackout manoeuvre and don't crash/die until I come round" then are they even trying?

      I would have thought "aim bots" would have been in use since the 90's and possibly before then. Remember, they had laser guided bombs in 1968 (Vietnam). Not to mention radar assisted guns in the 40s.

      My point being that software like this would just be an extension of those systems, and thus usable on an F-35.

      There is an old saying, "If you're not cheating, you're not trying"

      1. Stork Silver badge

        I think Saab Viggen had radar aiming of the gun in the fighter version, ca. 1990

      2. I am the liquor

        Certainly they have "aimbots" that fire the gun automatically at the right moment, the pilot just has to point the plane in the right direction. It doesn't seem much of a leap to have the computer steer the aircraft as well does it.

        But arguably they've already been doing that since the 1950s: the computer in the air-to-air missile works out how to steer itself at the target. And it doesn't even need "AI" to do it.

      3. Anonymous Coward
        Anonymous Coward

        Glen 1 wrote:

        > There is an old saying, "If you're not cheating, you're not trying"

        I prefer - "If you find yourself in a fair fight, you planned it wrong!"

        (attributed to David Hackworth)

    2. Pete 2 Silver badge

      "Obsolete" is a rather harsh term. The americans (still) reckon that the F35 has a service life up until 2070.

      What they don't reveal is that the first 5 years of those 50 is as a promotional gimmick and the remaining 45 years will be as a target drone for testing AI-piloted next-generation fighters.

      1. Chris G Silver badge

        After the first five years the remaining 45 will be to operate 'No exit' maintenance contracts with the suckers that bought the thing from the States, a system pioneered by Xerox.

    3. Persona Silver badge

      Very obsolete if it gets down to dog fighting range. The gun on a fighter is fixed to the airframe so to aim it you point the plane. If the gun could be aimed over an additional arc the pilot would not be able to control it and fly at the same time. An AI pilot could both pilot the plane and independently aim the gun while pulling 9G.

      1. Anonymous Coward
        Anonymous Coward

        If the gun could be aimed over an additional arc

        I imagine that, whilst not necessarily impossible, doing this would be highly problematic, as a modern fighter's gun is a rotary cannon producing a huge amount of recoil when fired. For example, the recoil thrust produced by the GAU8 cannon in the A10 Thunderbolt is greater than that produced by one of its own engines (so it's a good job it has two), and while this is a slightly extreme example, you can see why having these sorts of forces off centre line would play havoc with aircraft control.

        1. Persona Silver badge

          you can see why having these sorts of forces off centre line would play havoc with aircraft control

          Just a degree or two would make a big difference to aiming accuracy and certainly help put a bit more lead on the passing target. As for the control impact, an AI pilot should be able to not only compensate for these highly predictable forces but also incorporate them into the attack "solution".

          1. Chris G Silver badge

            It's not only about the AI, it's also about very complicated engineering, to compensate for a directional change in stressing forces could require extensive redesign that could possibly be detrimental to the overall performance envelope.

            I have never worked on modern combat aircraft but I did spend a few years as an airframe engineer.

            The closest to combat aircraft was working on a WWII vintage T6 trainer.

            1. Glen 1 Silver badge

              So you're saying it would be easier to let the AI have control of the aircraft to provide extra "assistance" to to aim? Current off-bore targeting tends to be focused on missiles.

              Hell, having RCS thrusters / Viffing / Canards to point and shoot while stalled would make some high-G manoeuvres unnecessary. (Think B5 Starfighter) Altitude permitting to recover, of course.

              1. John Brown (no body) Silver badge

                Nor forgetting the manoeuvrers that Harrier pilots came up with. An F-35 under full computer control could probably pull some very interesting manoeuvrers against a human pilot. Although I suspect that when the first fully non-human piloted fighters are ready to go into service, they may not look like what we think of currently as fighter aircraft.

          2. TeeCee Gold badge

            You are assuming that an AI controlled aircraft needs to fly in a straight line and can, therefore, only fire straight ahead.

            I'd have thought that slewing the airframe to aim would be simpler, using stick and opposite rudder, although the reaction times required to hit anything by doing that are rather beyond human capabilities.

        2. iainr

          The thunderbird's cannon is optimised for ground support/anti tank operations. the gun in most modern fighters is lighter and the F35s is somewhere between.

      2. Why Not?

        Why would you risk a pilot & multi (tens of) million or even a billion plus aircraft on dogfights?

        Surely a fleet of high altitude cheapish (<$15 million) drones with remote humans targeting the weapons would be the obvious solution. In a nice underground bunker with a Starbucks of course!

        I sort of assumed they had already got intelligent targeting for their missiles, the drone operator lights it up once and the missile follows the target regardless of the pilots attempts to evade, so maybe this AI is best for the hellfire missiles.

        A core mounted gun with a more agile drone could do close up stuff.

      3. John Brown (no body) Silver badge

        "while pulling 9G."

        That's the deal breaker when it comes to a meatbag pilot against AI assisted remote pilot. Suddenly the flight envelope is only limited by the strength of the airframe and power of the engines. And you no longer need the weight of the pilot plus gear/life support and the heavy ejector seat.

      4. Pascal Monett Silver badge

        They do that on attack choppers all the time. The Apache had a forward gun controlled by where the pilot (or gunner - don't remember specifically which) was looking, not where the chopper was pointed.

        But a helicopter doesn't go supersonic, and dogfight tradition is you point the plane. It was rather successful in both world wars.

      5. iainr

        The pilot aiming the cannon is probably not that difficult, the chain gun on the apache helicopter is aimed via an ocular sight so will shoot anywhere the pilot can shoot. I imagine the difficulty is designing a steeerable gun pod that can cope with supersonic stresses.

        1. hoola Silver badge

          In modern aerial warfare I thought that the traditional dogfight was pretty much dead as it was cheaper & easier to fire various missiles at your opponents. If you reach the gun stage then you had lost. I don't know what ammunition the guns have but invariably it is very limited.

          Assuming a missile is still cheaper than a drone or UAV in military parlance one would have thought that pilots/technicians playing shoot-em-up fighter games is a long way off.

          I suppose the only advantage comes in that you no longer need an expensive pilot in the air so if they do get shot down you no longer have to worry about recovery.

          1. CrazyOldCatMan

            raditional dogfight was pretty much dead as it was cheaper & easier to fire various missiles at your opponents

            I remember a sci-fi short story called "A hawk among the pidgeons". The basic plot was that a modern (for the 1980s) fighter[1] somehow gets thrown back through time to WW1.

            "Aha!" thinks the pilot - at a stroke I will kill the German biplanes - how hard can it be?

            The answer is - very hard. Not enough metal or heat for the missiles to lock onto and the speed difference is just too great for the gun to be useful. The most he can do is try and tear them apart with his slipstream and even that doesn't work.

            Just as he's about out of fuel, the time-rift opens again and he's back in his time.

            [1] And F104 Starfighter if I remember correctly. Surprised that the pilot survived..

  3. Andy Tunnah


    I know the usual thing is to make skynet jokes and be all gloomy about the rise of the machines, but I can't. This is just INCREDIBLE.

    How far AI is coming and mow well we're doing in automation just puts a big smile on my face. Technology feels like it moves along at a snail's pace nowadays, considering we've mastered everything; there's no more revolutions, just evolution. But AI feels like a new frontier where we can make massive impressive strides and I am LOVING it.

    1. Grunchy

      Re: Awesome

      Nice try, Hal

  4. Blackjack Silver badge

    Do it for real or scram

    Was it really that expensive to have both the AI and the human remote control a plane or drone each?

    A good programmed AI is always gonna be better than humans in a simulation but real World conditions won't be do accommodating.

    1. HildyJ Silver badge

      Re: Do it for real or scram

      I'd like to see it done for real too, but not in drones. A pilot in the plane can't maneuver as fast because of g forces. The AI will still win.Your suggesting drones for the test just points out that planes with pilots and pilot support systems are becoming obsolete.

      1. EricM

        There are differences between a simulation and reality

        Not necessarily

        While an AI can undoubtly pull G's up to the frame's limits and has faster per-se reaction speed it also will have to deal with real-world, unclear and contradictionary radar, microwave and optical sensor inputs to keep its situational awareness in an environment where electronic countermeasures are actively used and expected to evolve fast.

        That awareness is currently delivered by the simulation for free.

        So all this simulation currently shows is that the AI can play an elaborated videogame better than a human.

        I'm not sure how this would translate into a real-world battle, where the AI "pilots" , their capabilities, tactics and reactions are known to the adversary.

    2. This post has been deleted by its author

    3. LDS Silver badge

      Re: Do it for real or scram

      For example, how the AI got the other plane data (speed, heading, etc.)? From the sim data directly or having to parse simulated sensor data (radar, cams, FLIR,LIDAR, etc.)? Did they have to simulate a way for AI to track a plane outside sensors field of view?

      I'm quite sure that a computer fed with the proper data can be faster to find the right manoeuvre, the problem, like in self-driving cars, is to obtain and map properly those data.

      1. vtcodger Silver badge

        Re: Do it for real or scram

        NOT like self-driving cars. With cars the system must fail rarely and always fail safe. Were it not for the rarely and safe requirement, autonomous vehicles would already be flooding the roadways. With weapons systems, you can tolerate a few lethal (to the WS) mistakes.

    4. RPF

      Re: Do it for real or scram

      Having done it for real, I think the result would actually be even worse: the g-forces, the noise, the adrenaline all conspire against the pilot, which is why you train hard to condition yourself against these effects as much as possible; but they\re still a factor against top performance.

      I thought the AI planes were bloody good and Heron terrifyingly good at head-on gunnery (which is usually a complete no-no for obvious reasons).

  5. Anonymous Coward
    Anonymous Coward

    Anyone know where I can get a big EMP?

    1. Grunchy

      EMP hoax!

      It’s a book somebody wrote!

      1. Klimt's Beast Would

        Re: EMP hoax!

        What about ELP, or would Motörhead be more effective?

        1. Aussie Doc Bronze badge

          Re: EMP hoax!

          I'm old. ELO anyone?

          1. Anonymous Coward
            Anonymous Coward

            Re: EMP hoax!

            ELO ... well that's certainly blue sky thinking!

            1. Steve K Silver badge

              Re: EMP hoax!

              Or Confusion?

    2. Ken Hagan Gold badge

      You are probably better off learning about computer security, so that you can takeover the enemy's assets mid-war. Given the "state of the art" in that, I'd have thought that the outcome of WW3 was completely up in the air. (The side that *builds* the best drones may not end up being very important.)

  6. Pete 2 Silver badge


    > The fighter pilot battling on behalf of us humans ... losing 5-0 in one-on-one virtual combat.

    So the death delivery business is going through the same phases that chess went through, apropos artificial intelligence.

    First there is the scoffing "it will never beat a human"

    Then there is the grudging acknowledgement "well it might have some limited application ... in special circumstances

    After that comes denial "Ahhh, but it's too unreliable / expensive / slow / limited"

    Followed by the reality check "AI beats the world's best human"

    And the inevitable excuses "Yebbut it isn't really playing, it's just searching for past moves - played by humans"

    Finally the turnaround "Major contractor buys-out fledgling researchers, adopts technology and declares THIS IS THE FUTURE"

    1. Boris the Cockroach Silver badge

      Re: Chess


      "Finally the turnaround "Major contractor buys-out fledgling researchers, adopts technology and declares THIS IS THE FUTURE" "

      Shortly followed by

      "All air combat craft now fly with a 100% safety record"


      "The system became self aware at 2.37am PDT"

      <<currently stocking up his bunker with 40w phased pulse plasma rifles

  7. Julz Silver badge


    the video. The biggest difference seemed to be the 'AI' could shoot better. On each merge, the 'AI' did way more damage than the meatbag. Perhaps the lesson is to let an 'AI' be in charge of the gun.

    1. Chris G Silver badge

      Re: Watching

      I would be more impressed if the software succeeded multiple times against multiple pilots on different platforms.

      As it is there is little to show its not a one trick pony, although it does stand as a proof of concept.

      1. jdiebdhidbsusbvwbsidnsoskebid Bronze badge

        Re: Watching

        Agreed. It's probably more complicated than this but for me, this story headline is merely "human loses to CPU opponent in video game". As someone who has often lost to the computer on Mario kart, this is just so much "meh".

    2. RPF

      Re: Watching

      This is true and was mentioned in the commentary

      But also IRL the enemy wreckage that's closing head-on at 1000+knots doesn't just miraculously disappear....

  8. chivo243 Silver badge

    How long until Ender's Game?

    Virtual war now, how long until the soldier's don't know their part of it?

    1. amanfromMars 1 Silver badge

      Re: How long until Ender's Game?

      Virtual war now, how long until the soldier's don't know their part of it? ..... chivo243

      Are you suggesting that soldiers know they are part of virtual wars, chivo243?

      I imagine if they knew that they would be absolutely furious .... and extremely dangerous.

  9. KBeee Bronze badge

    Bit of a "Dreadnaught" moment..

    ..when you realise all your hyper expensive fleets of State-of-the-Art planes are rendered useless against Next Generation pilotless planes that are available to anyone willing to pay or capable of building. And don't think "Only We!" are capable of building these things.

    1. Anonymous Coward
      Anonymous Coward

      Re: Bit of a "Dreadnaught" moment..

      Problem is, big militaries don't recognize the dreadnought moment until they're getting defeated in actual battle, and the generals are forced to admit that they're still trying to fight the last war.

      1. ciaran

        Re: Bit of a "Dreadnaught" moment..

        In the next big war, all the aircraft carriers, AWACS and aerial refuelling tankers will be out of action at the end of the first day. GPS will be inoperative, and battlefield wireless communications will be jammed.

        Compare that to all the shiny stuff the militaries want...

  10. Jan 0 Silver badge

    Rockets and Jets!

    Why were they fighting with bullets in the 21st century? Don't 'planes carry autonomous, manoeuvrable, high velocity air-to-air missiles anymore (and decoy countermeasures)? I'd still expect the AIs to win most dogfights in identically equipped combat aircraft.

    1. ectel

      Re: Rockets and Jets!

      Yes they have guided missiles, but they have envelopes of engagement. 2 aircraft fighting will go with missiles first, but as they get closer together they will be out side the missiles envelope ( a missile finds to go straight off the plane and there is a trade off between speed and maneuverability the)

      So quickly, when the planes are close the missiles don't work, then you are down to guns.

      The Phantom went to Vietnam with out a gun, big mistake, every fighter since has had one

  11. J27 Silver badge

    This sounds good, but the situation favours the AI as much as possible. A 1 vs 1 encounter wiith the same model aircraft is incrediblly unlikely in reality. Real battles are much more complex, with more factors to take into account.

  12. Sanguma

    Butlerian Jihad time, again?!?

    Sheesh, just when it'd gotten quiet and I was taking a nap ... Snoopy'll have to keep his eyes peeled this time - it's not the Red Baron he's got to worry about, let alone Natural Stupidity but Artificial Intelligence!!!

  13. Michael Hoffmann

    Youtube comments

    Only watched the video briefly because it was WAY more fun to skip straight to the YT comments.

    All those video game jocks claiming how they would have totally beat that AI and that the human pilot just sucked.

    Great entertainment!

    1. RPF

      Re: Youtube comments

      Yes, what tits they are.

  14. amanfromMars 1 Silver badge

    A "'Curiouser and curiouser!' cried Alice" Moment/Event Horizon:-)

    What you might not know about Heron Systems, but which might surprise and please you .....

    1. RPF

      Re: A "'Curiouser and curiouser!' cried Alice" Moment/Event Horizon:-)


      What's a "small disadvantaged business" though?

      1. Tom 7 Silver badge

        Re: A "'Curiouser and curiouser!' cried Alice" Moment/Event Horizon:-)

  15. Justthefacts

    Generals fighting the last war again.....

    Putting an AI drone up against a human pilot is like all those Hollywood movies where the minions politely agree to attack the hero one at a time. Unfortunately, real enemies are unlikely to be so compliant.

    The thing to remember is that while dogfighting happens occasionally, almost all combat is Beyond Visual Range. This fatally compromises the case for having a human pilot.

    The whole point about a drone is that you can afford to lose it. 99% of the cost of a fighter aircraft is in making it capable enough of protecting its human pilot. For the price of one F35, make a hundred dumb drones, and arm each of them with a single mIssile. The first drone gets shot down in seconds. So does the second and the third. Thing is, how many missiles was the F35 carrying again? After the F35 has shot off its rack, now it’s eighty drones against one F35. The F35 has only short-range gunnery remaining, although it could still take on any three drones in a dogfight. But the drones will just sit fifty miles away, and snipe away with their eighty missiles. Again, the F35 will take out the first ten or twenty with countermeasures. But after that, it’s still facing sixty missiles incoming simultaneously, having exhausted its countermeasures. Then it’s toast, Top Gun or not.

    1. Julz Silver badge


      expensive. A BVR missile like a meteor costs around £2,000,000. So by your numbers, a hundred dumb drones carrying a missile each would cost £200,000,000 to arm. The single F35 would cost the same to build (your assumption) plus it's missile load. Lets assume it's in a non-stealth configuration and is carrying six meteor missiles. This adds up to £12,000,000 for its armaments, a cool £188,000,000 less than the drones.

      Yes I know I'm ignoring a load of other stuff, but this is just order of magnitude, fag packet (is that still a thing), arithmetic.

      1. ChrisC Silver badge

        Re: Still

        Indeed. Also, even without having a detailed costed BOM for the F35 (or indeed any other recent-ish fighter) in front of me, I'm going to stick my neck out and suggest that apportioning 99% of the per-unit cost to those parts of the design which could be done away with in an unmanned drone is quite some distance off the mark. Closer to the mark if the drone is *also* designed to be as cheap as chips, with just basic subsonic loitering capability, limited maneoverability, little or no stealth (other than what would be provided by the smaller airframe), less capable sensor packages etc. etc. etc...

        However, it's worth bearing in mind that this "manned fighter vs swarms of missiles" scenario isn't a million miles away from what we already have if you bring SAMs into the equation, and air forces the world over have spent decades coming up with tactics and new aircraft/countermeasure designs to reduce that threat. Replacing the SAM launchers with airborne equivalents will change the threat somewhat, but it's far from a certainty that simply having a giant swarm of BVR-capable drones will be the killer blow to manned aircraft.

        Also let's not forget that, in order for these BVR drones to have any hope of downing the incoming F35, it needs to first find it and achieve a sensor lock... That in itself isn't guaranteed, and on top of that if the drones really are pared down to a minimum for cost purposes, then they're likely going to need support from other sensors - AWACS, ground radar stations etc - which won't be as plentiful as the drones. Knock out those sensor feeds (either directly via an AA kill on AWACS or Wild Weasel hit on a ground station, or indirectly by jamming the uplink), and you'd significantly hinder the ability of the drones to do anything effective. Stick enough sensor kit onto each drone (or even onto a "Queen Bee" drone attached to each swarm or sub-swarm) to reduce/eliminate the dependencies on more easily attacked supporting units, and you're increasing the price per drone to the point where you're going to struggle to field enough of them to have the advantage of numbers.

        Not saying drones can't find a place in air-air warfare, just that there's far more to it than simply assuming that "drone = cheap" and building a case against manned aircraft from that.

      2. Peter Gathercole Silver badge

        Re: Still

        Unfortunately, fag packets are now covered by so much graphic horror and warnings that there's no space to scribble on them any more (or, it seems to put enough text to identify the brand easily!)

  16. Mike 137 Silver badge

    One day maybe...

    Eventually, all the fighting might be done by machines, and the only human losses would be "collateral damage". But there's also another alternative. There's an episode of Start Trek ("A taste of Armageddon") in which computers fight a theoretical war with theoretical casualties, but the identified "casualties" have to report for extermination.

    What is war for?

    1. Julz Silver badge

      Re: One day maybe...

      Absolutely nothing.

      War, huh, yeah.

      1. hoola Silver badge

        Re: One day maybe...

        And then it all falls apart when all the tech fails, reboots gets a but or just simply has someone do something basic, with very little tech that takes the drone out.

        This then give rise to the unknown, if a technically superior power cannot be beaten by the military through established warfare, just go the nuclear or biological route in some large centre of population.

        The overall collateral damage could actually go up, potentially resulting in a catastrophic exchange. Generally some form of peace is achieved when the different parties do not have significant advantages over each other.

  17. Bearshark

    Making a correction to this article. Heron Systems is based out of 'California, Maryland'. Not the state of California. (Hollywood MD is only 5 miles north :-)


  18. Cynic_999 Silver badge

    Oh good

    We will soon have a situation where all warfair is conducted by pilotless aircraft, guided missiles and perhaps remote-controlled tanks etc. So the only casualties will be innocent civilians who are killed by bits of falling aircraft, crossfire and off-course missiles. Still, so long as the soldiers are safe ...

  19. Anonymous Coward
    Anonymous Coward

    AI war vs relaity

    The problem is the pilot is fighting an AI plane and not massive holes in the AI when it comes to edge cases. They need to be fighting the data set used to train the neural network. Paint the plane sky blue, cover it in clouds, make it polished and reconfigure the radar profile to match that of a commercial airlines (Use Pyotr Yakovlevich Ufimtsev 1960's maths, but not for stealth this time). Only fight at dusk or dawn, because most of the training dataset will be from night or daytime. Basically generate as much conflicting information to the sensors as possible to force the AI deep into the valleys of uncertainty of the dataset used to train it.

    Modify the real world input data to generate uncertainty, is how you win against an AI. Unless they train the drones with rules, like if it moves shoot it, then we are all screwed.

POST COMMENT House rules

Not a member of The Register? Create a new account here.

  • Enter your comment

  • Add an icon

Anonymous cowards cannot choose their icon

Biting the hand that feeds IT © 1998–2021