"AI in Military Aviation | Behind the Wings on PBS"

Neural Jetstream
2 followers
Nov 9, 2024
Hi, I'm Tracy LaTourette, Colorado's first lady fighter pilot. Call sign Jackie O. We're here at
Wings Over the Rockies Air and Space Museum in Denver, Colorado. We're getting an inside look
at artificial intelligence in aviation. It is a moment of wonder, shock, and awe all together.
We're around the White Feathers moment. We're in a race for military technological superiority.
How do you methodically explore the unknown? A question of trust is the heart of the issue.
At Edwards Air Force Base, they took an F-16 like the one I used to fly,
modified it, and now it can fly itself. I'm home, back in the cockpit. We can go maneuver relative
to another aircraft that doesn't actually exist. This is going to save a lot of pilots' lives.
We need to seize the day. We have to be really careful about it. This is the way that wars are
fought and that we need to put real guardrails on. It's time to go behind the wings.
Artificial intelligence is rewriting the rules of military aviation. As planes are programmed
to fly without pilots, uncrewed aircraft will define the next era of military and someday
civilian flight. However, human ethical and operational considerations don't always keep
pace with technological advancements. Those working on AI's cutting edge must balance great
potential with great responsibility. Every time we get a new technology,
we have certain expectations and dreams for it. You may have seen pictures in the late 1800s
when we were starting to get the sense that we can make an airplane, we can make a thing with
wings that could fly. And they had these fantastical pictures of these airplanes shaped like boats,
right? It had paddle wheels. But once the Wright Brothers flew and once we actually started doing
it, nobody imagined what the airplanes actually ended up looking like.
To me with AI, I kind of feel like we're around the Wright Brothers moment.
The outskirts of the Mojave Desert in Southern California provide an ideal setting for the
United States Air Force Test Pilot School. The school is a unique institution in that
we have a flying operation and a master's degree program. Our students fly over 35 aircraft in a
year. It's perhaps the most complex flying operation in the Air Force. You walk through
the same halls that Chuck Yeager walked through back in 1947 when he broke the sound barrier.
This school is full of history. It is incredible. You know, the drive from LA out here to the
desert, there's a lot of desert, a lot of opportunity to contemplate. It immediately
strikes you what a perfect place this is to fly. And then as you drive across the lake bed and you
realize all the aircraft that have landed here and have taken off here for the first time.
The history of the school is incredible and you can feel it as you walk around here. But one of
the things that I'm very excited about is the future in autonomy and aviation. It really began
with simulation. Thousands upon thousands of runs with multiple different performers were conducted
in a simulated environment to mature these algorithms. What makes these algorithms different
than a lot of the autonomy you might find in autopilot today on most aircraft, that that kind
of autonomy is line by line. If this, then that. Do this. The kind of autonomy we're talking about,
machine learning, uses statistical methods to find patterns in human behavior. It's very powerful
and in some ways better than humans, for example, in chess and Go. And yet it's not always predictable.
It's not always repeatable. Artificial intelligence is not a new subject. It goes back to the 1950s
and 60s to the mainframes that predates the internet. And really what changed around 2010
is Moore's law finally caught up with us and we're able to get more and more processing and
we're able to do it faster. And suddenly raw compute capacity caught up with our dreams of
the algorithms. It is a moment of wonder, shock, and awe all together. We need to seize the day.
We have to be really careful with it. It would be wonderful if lethal AI weapons never existed.
I am confident that somebody will invent them. And if we aren't capable of dealing with them,
then we're going to be in trouble. In Washington, D.C., the ethics of military AI are central to
the technology's development. Your conception of AI and robots as they relate to the military and
war fighting might be sci-fi in your head, but it's actually here. We're becoming increasingly
more willing to trust AI as recently as a few years ago with something that was experimental.
But now as AI is getting more sophisticated and prompting people to think of it as something
that could be used in the military. That if the military doesn't develop on its own,
it will be outpaced in. We have, you know, months, not years before this becomes a reality. This is
something that we should be watching for. This is the way that wars are fought and that we need to
put real guardrails on. The idea of empowering weaponry with AI, that is something that once
you create it, you don't take back. You worked at the Pentagon back in the 80s. Autonomy was
something that was front of mind. How have we gotten to this place now where AI, where artificial
intelligence is really the name of the game? It's been an evolutionary set of technology
developments with revolutionary applications. I was one of the proponents of our uncrewed
autonomous aircraft in particular. And I've watched that technology over the last almost 40 years
move to a point where we can increasingly rely on it for some of the functions that traditionally
humans have done as operators. I think it will be a paradigm shift. I think it's happening as we
speak. And this is a change that's relatively recent. Talk about ethics. How do we do that
on the global stage? First of all, we have to design those requirements in. They're part of
the design process. And secondly, we have to move fast. People who are not constrained by that,
who don't take the time and the effort to design those kind of constraints onto their systems,
do have in some ways an advantage. I think overall we have the high ground here. And I think it's the
right thing for us to do. So that's what we are going to do. But it dictates that we move as
quickly as possible to integrate these technologies into our systems. We're in a race right now for
military technology superiority. We've got to move quickly. You don't win a race by standing still.
How do you get to a place where you can trust that aircraft when it's being flown by this
artificially intelligent agent and not the pilot that's in the backseat?
Question of trust is the heart of the issue. And I think if you look at the trust question,
there are several aspects that we have to explore. First of all, can it do what we're asking it to do
in the first place? Does it do it well? Can it do it safely? And can it do it ethically,
in a way that conforms to the norms that we as an American military and American people expect?
Normally, you train the AI on the ground. And then once you put it in the air, it's no longer
learning while it's in the air, but it's collecting data. Then when it lands again,
you can use the data that's collected to retrain the AI to get better and then have it fly again.
The reason why is that we would like, when it's flying, that we're fairly confident that we know
what it's going to do. We don't want the algorithm changing mid-flight, at least not yet. That may
be something that comes in the future, but at the moment, that's not the cycle we normally use.
A lot of what we learn is that we didn't know what we didn't know. It's the unknown unknowns.
It looked like it was working well based on all the parameters that we could put on it.
But then when they put it in the air, it suddenly will find itself in a position
that we weren't prepared for. One of the challenges, of course, in developing simulations
is that the reality is very different. And so it was necessary to take it from a training environment
in a simulation and move it into the real world. And that's where the X-62 becomes part of the
story. So we have here the X-62 Alpha, which is also known as the Vista. It's a one-of-a-kind
national asset. We looked at this aircraft and realized that it would allow us to take
these cutting-edge autonomy solutions and to put them into an airborne environment.
General White, part of your job is to figure out how to execute our strategic vision
by buying or building the things that we need to move forward in the future. How do you even
get to a place where you're looking at something like the Vista, a flying simulator?
So I think Vista is a great example of how we're working to evolve technology over time to meet
our needs. And it's that first step towards building that trust with our operators. We're
verifying that this autonomy is going to work the way we want it to. When we think about autonomy
as an Air Force, we want it to be platform agnostic. So I should be able to take autonomy
off of an F-16 and put it on any platform. Vista is our first step on that journey,
and we're going to continue to evolve that over time.
At Edwards Air Force Base, a revolution is taking place in AI piloted aircraft.
At first glance, this F-16 looks a lot like the one I used to fly,
but when you look close, there are a whole lot of differences.
Damon, we don't even call this an F-16 anymore. It's now an X-62 Vista. So what's new?
The Vista stands for the Variable In-Flight Simulator Test Aircraft. It's designed to be
a research test bed for experimental flight controls and autonomy algorithms.
So inside and out, it's a whole lot different than an F-16. Can we take a closer look?
Yeah, let's go.
All right. So Damon, simulators are usually something we use inside.
So what exactly is a flying simulator?
The X-62A can actually simulate the motions of any aircraft that I can model. So what that means
is a pilot sitting in the front seat here, when they pull back on the control stick,
it's going to feel like they're flying a 747, or the space shuttle, or a T-38,
or any number of aircraft that we can program into this simulator.
So the F-16 is the smallest fighter in our inventory. How do you simulate
larger aircraft like the F-35, or like you mentioned, an airliner size aircraft?
Well, although the F-16 is one of the smallest aircraft in our inventory,
it is also one of the most maneuverable aircraft. And so anything that isn't as maneuverable as an
F-16, I can simulate because I have the ability to make those motions and those dynamics with
this aircraft.
So you're a test pilot school and you have students, only in this case,
it sounds like the student is the AI agent.
Whenever we're out testing new autonomy algorithms, we always have two humans on board.
We have a safety pilot, who's one of our most skilled instructor pilots,
and we have a system operator who knows how to operate the autonomy systems on this aircraft.
Is it true that you actually took out the gun because you needed more room for computers?
This aircraft doesn't need a weapon because it's not a combat aircraft.
It's a research test bed. We remove the gun and replace it with all of our research computers
so that we have the computing power to run our autonomy and to run our in-flight simulator.
So I know a lot of the magic happens on the inside, but let's take a quick look at the back.
So a lot of it's the same as a regular F-16. The tails are the same size as an F-16,
although the software inside the jet can move these in ways that a normal F-16 can't.
We also have special spin shoot mounts that were used in the early days of this aircraft
when it was a research test vehicle for a multi-axis thrust vectoring nozzle.
So you mean they took the nozzle off of this jet and replaced it with a thrust vectoring nozzle
to test for things like the F-22?
It used to be a research test bed, and we did put a thrust vectoring nozzle on this
that fed into research that went into the F-22.
But nowadays we use that same research architecture and safety sandbox to test autonomy.
It's great to see the old F-16 all grown up, and I'm dying to see how the cockpit has changed.
It may look a lot like a regular F-16, but the Vista can actually adapt like a chameleon,
taking on the characteristics of a variety of a different aircraft,
even changing from one type of aircraft to another multiple times in a single day.
This all may sound like science fiction, but here at Edwards Air Force Base, it's happening right now.
So Evel, I understand you actually fly this aircraft from the back seat.
How does the AI fit into all that?
So the AI lives in a computer in the back of the airplane,
and the front seater is mostly charged with connecting it all up to the jet.
And then meanwhile, I'm just sitting back there making sure that's all going right
and taking over whenever I need to for safety.
So the front seater then is typically like a flight test engineer?
The front seater can be a flight test engineer, but later on as the system is working right,
we will put a fighter pilot in the front seat who's got all that expertise on that,
so they could sit there and observe how the airplane is handling, how the AI is doing its job.
So we'll go ahead on the cockpit and show you how they change this airplane to make that possible.
I'm home, back in the cockpit. This is amazing.
So it feels like I'm sitting in my old F-16 cockpit,
but this entire aircraft has been reconfigured for research.
The setup that allows us to do that, that makes it very easy for me to take control of the
airplane in the back no matter what, makes it easy to kind of replace you with the AI.
My job in the back hasn't changed.
I'm still there to turn the system off and then fly the F-16 and just keep things safe.
So there's always a human on board so that you can test the AI in a very safe environment.
So we're testing these at a very early stage and what we see is fascinating.
We have this problem and that problem is we train these AIs in a simulator
and then we take them out and put them in the real world in this machine,
but to watch how the AI fails and then to see how quick they change the training and
then change the parameters that the AI works in. So we're kind of growing the AI in the airplane.
With the walk around complete, it's time to throttle up and go fly. Fight's on.
We're headed into the flight brief where they'll review the plan for the day.
So the objective for today is we're going to test Agent Bravo on the X-62A Vista and we're
going to test it in a number of different basic fighter maneuver configurations.
All right, great. So let's go through the test flow.
Next up is going to be test point one. Once I give you that indication,
you're going to request the test points. You're going to go ahead and start the data recording.
You're going to select the test point and then you're going to wait for that agent to load.
And after that, it's fight's on and the AI is in control.
Today we're doing a virtual target so what we're expecting to see is we're set up and
we're in a turn like this. Once everything is ready to go, when we engage, now our airplane
with the AI controlling is going to do some sort of defensive maneuver while the other agent
virtually will be dogfighting with it. Just make sure we turn off the simulation
before we land so that you're landing the normal F-16 and then we're ready to go.
I'll meet the crew at the desk in 10 minutes. I'll see you there and we'll have a good sortie.
I can't wait to see this thing fly.
So welcome to the control room. This is where we monitor all of our test missions
and so we not only make sure that the test pilots up in the jet are evaluating the right systems
and they're looking at the right stuff, but we also help keep them safe here as well because
we have a ton of data that we can monitor here while they're focused on evaluating the autonomy
or the AI. Test control, you are clear to load test points. Copy that.
Over the last year we have been breaking all kinds of barriers in applying autonomous agents
in the aerial combat maneuvers, so basic fighter maneuvers. And that's what we teach at the school
is this mindset of how do you methodically explore the unknown. Next up is going to be
test point three. We're going to test the Bravo agents. You are clear to load test points.
Right now he's loading the test points and we're just waiting for the indications that
they've loaded appropriately. Test control, you are clear to maneuver.
We began the aircraft by itself just to see if it would maneuver in a way that we expected
and then we put it up against a virtual target. So we would beam up essentially an F-16 that it
would fly against and there again we were looking to see did it follow the rules of engagement,
did it respect the thousand foot bubble, did it behave in a way that we would expect an F-16
pilot to behave. All right next up is going to be test point four, the Bravo agent deep into BFM.
Test you are clear to load test points. So we're actually bringing up a simulated aircraft which
is fantastic because we can go maneuver relative to another aircraft that doesn't actually exist.
So we minimize that safety risk of actually hitting another aircraft because one doesn't
exist. There's the agent making its inputs and turning and it tripped it off. Test control,
point complete. So the Vista's up there flying, you're here in the control room and you can see
how the AI is actually acting. We can see what the aircraft is doing from here and we can keep
track of whether it's doing too abrupt of a maneuver, is it too aggressive, is it not aggressive
enough. Test control that is a test point complete, you are clear to reset the VSS. Copy that, thanks
a lot. Only once we were assured that yes it was behaving according to what we expect, yes it was
following all the training rules, did we then put it in the same airspace with another aircraft.
And even then we started with very slow build up. We did bubble passes. Does it avoid like we
expect it to? Does it avoid the floor like we expect it to? And then we put it in an offensive
position. Does it maneuver to a weapons engagement zone the way that we expect it to? And then we put
it in the defensive and only after all of those and we kept getting the positive signal that yes
it's safe, it's effective, it's ethical, did we then put them beak to beak. We conducted the world's
first air-to-air combat maneuvers in live time, not in simulators, between the X-62 and under
control by an AI agent and an F-16 that was crewed by a test aircrew and then ultimately progressed
to what we call high aspect maneuvers where these two aircrafts were approaching each other at 1200
miles an hour at less than 2,000 feet of distance and doing it safely and repeatedly as they engaged
in combat maneuvers. Test control that is test mission complete. You are clear to RTB.
So with each flight we're learning about how to update the AI software. We learned some things
today so you'll go back make some adjustments and tomorrow see how the AI performs. Some people
think that the AI updates itself but that really isn't the case here. We're going to go look at
our data, make some updates and yeah, well hopefully it improves tomorrow. You're on the
cutting edge of AI. Tell me a time when this really blew your mind. Well years ago when we
first started putting these agents in here I was flying one of these things and I come back down
and we're in debrief and one of my engineers was in the control room. He says, hey Evil,
these dudes didn't look practically in tears because these people had spent years and years
of their life working on these things. They thought they were a decade out from having
ever control an airplane and here they were in a control room watching their baby with lots of
problems but still their baby doing things in a real jet fighter. As an engineer that's just the
best feeling in the world because you get to see how your work is really going to do and better yet
you're getting the information you need to make it better. Where is this all going? The future of AI
it's hard to guess. I think it's going to do more than we thought it could do. It's going to cause
headaches we didn't think we were going to have. The power of it is a little bit scary. It reminds
me a little bit of nuclear power and nuclear weapons in the sense that we know there's a
tremendous amount of potential power in the technology but the hard part is well how do you
control it? As the musician Sting once sang, I hope the Russians love their children too, right?
The future for AI research here at the test pilot school with the x62 and what it supports in the
future is really exciting. We're going to be looking at AI algorithms that can learn and these are
really important research questions because there are certain behaviors that are very human as these
things learn and there are other behaviors that aren't very human so it's very important to be
able to interrogate that. How do these algorithms develop as we train them as we put them in
situations that are slightly different than the ones that they were trained for? x62 as a research
platform is a huge enabler for all kinds of other research. One of the very exciting projects is the
xq58 Valkyrie down at Eglin. We flew the first version of its AI machine learning algorithms
here on the x62. That will now go out to the Valkyrie and they're going to continue on that
platform to mature what they originally began researching here. In the same way some of the
dogfighting algorithms that we have built out here are going to move on to a project out at
Eglin Air Force Base focused on multiple f-16s flying autonomously against each other and they're
going to use even more advanced sensors and even more capabilities to do that discipline exploration
of the unknown that lets us find where can we trust these algorithms and where do they still
need more development for us to really be able to say yes they're safe yes they're effective yes
they're efficient and above all yes are they ethical. Our hope is also that in partnership
with other government agencies like NASA that we can help develop the frameworks and the ways of
thinking about autonomy that make it possible for civil aviation to begin to explore different
pathways for certifying AI in applications beyond the military. It's riding the bicycle
and slowly holding hands with it and eventually you let the bicycle go. We are finally at that
point I think for some of them letting the hands off the bicycle go but that's not going to be
good enough. We have AI systems now that can perform very well within certain parameters
but when you go outside that environment when for example it's a it's a thunderstorm something
where you haven't been able to train then the system's no longer reliable and it doesn't adapt
and adjust yet the way a human would because we don't have what we call artificial general
intelligence we have only the artificial narrow intelligence that can do a specific task
and the problem is that the number of different ways that a situation can be complicated
well for a car it's one thing for an aircraft with three dimensions is nothing now putting aircraft
in battle with enemies moving at you in the electromagnetic spectrum the dimensionality
multiplies and it becomes almost impossible to train for every circumstance. I'm one of the
skeptics that believes that the current generation of AI machine learning although we have a lot of
ways to go we can start to see the limits to where it can get to and probably another revolution
beyond that and be necessary to get us to a place where we can really do the sort of advanced AI
that we're dreaming of now. The interesting thing about AI and AI empowered weaponry is that
what happens is really going to evolve over the next year or two what we do at this juncture
really makes a difference for the development of this weaponry. The inherent concern is that it's
something that we can't even sit here and speculate on that something happens that's
beyond our you know realm of imagination and we've already empowered these systems beyond the point
to get them back. Critics would say that there are concerns with the potential of things going
wrong with AI taking over. We have to design very carefully and we have to test very thoroughly to
ensure that we don't have something like that happen. We also have to make sure that the other
side can't somehow take over control of these aircraft so we've got to design all that in.
One of the things that has surprised me a little bit quite frankly is how our fighter pilots have
embraced this. They recognize that a modern battlefield is going to be intensely lethal
place no matter what and they would much rather have some of these uncrewed aircraft out in front
taking fire so that they don't have to. This is going to save a lot of pilots lives when we get
this implemented. Getting meaningful military capability fielded is what I care about the most
and I care about our pacing challenge China the most as an adversary. China has been working very
hard and this goes back at least 20 years to feel the military that can go head-to-head with the
United States and first if they can prevent us from intervening in the western Pacific
and then defeating us if we do. They've filled in a number of systems that are designed to take on
particularly our space and our air force or airborne assets wherever they may come from
and we've got to respond to that. They're moving forward aggressively and I expect that the
disparity in time between when we can feel this kind of capability and when they have something
comparable is not going to be very much whichever way it comes out. How do you think the role of the
fighter pilot is going to change over the years? So I think there's still going to be fighter pilots
in combat aircrew for a number of years. We're still just broaching the capability of what AI
can handle and we are taking those you know safe ethical steps to make sure that it behaves as we
expect it to behave before we really use it in a more wide sense. So technically integrating AI
into older systems those are difficult challenges but they're all doable they're not really unique
to AI. I think one of the real difficulties of AI is on the human side. People are uncertain whether
to trust it. I think we still are at the beginning of AI. I think what AI can do today will even in
three or four years look so mundane we won't even call it AI anymore. As AI piloted aircraft
are beginning to take flight how we handle ethical operational and technological challenges
will set the stage for this next chapter in aviation history. We'll see you next time on Behind the Wings.
