I am a consultant who developed the companion computer/autopilot and software environment for DRL drones used at the AIRR race sponsored by Lockheed in 2019. UZH was the team to beat and I met Davide and Elia there. They almost pulled it off but sports being what they are had a bit of a heartbreak at the end, and the team from Delft ended up pulling through and winning the $1mm prize. Delft then raced against Gab from DRL and iirc came behind by only 6 seconds.
So glad to see this team from UZH continued pushing the envelope and are now beating human champions. If you saw the team and what they managed in under a year, it was clear they were highly talented and human racers had their work cut out for them to stay ahead.
What is the general outline of going from a model of a craft (drone, sailboat, etc) to building a sim that can do reinforcement learning over a physical object interaction with its env?
I want to start playing with models, sims and collected data for sailboat racing- I know the RL/data science stuff, and I assume a good model of your craft takes time to build, and can be improved with collected data. What are some areas to explore when chaining model -> sim -> RL for performance?
I realize this is an extremely complex topic, with several PhDs worth of potential input- if you had to explain to someone technical what it looks like and where to keep digging, what would you say?
Models for drones are actually pretty simple. The races are indoors, so we can assume 0 wind. It’s really basic physics and aerodynamics. We had some sessions in a mocap lab where contestants were able to perform some flights and get some ground truth to tweak their sim -> actual drone dev loop. The simulator itself was based on the DRL simulator, but not the same version on the Steam store.
I did not develop the sim itself but did develop the hardware-in-the-loop portion of it along with things like real-time debugging, and output to the hitl. We had the sim rendering cameras which we output from the workstation to custom HDMI bridges over MIPI that we could treat as real cameras on the NVIDIA Xavier AGX. There was a data channel over Ethernet for IMU data.
I made a custom version of Eclipse that interfaced w GDB for debugging, which also was modified to stay in sync w the sim using PTP, w rewind capability.
As for sailboat modeling, yes it’s more complicated because of the effects of both wind and fluid dynamics. If I were approaching this, I’d probably try and find a physics simulator to start with. Getting ground truth will be difficult, but I imagine you would start w the IMU and GPS data off the boat, but having time synced ground truth for the waves and wind will probably be the hardest part.
Numerical simulation of aircraft is pretty straightforward. The hard part is going from the inertial frame into the aircraft's frame of reference, and back again. Once that part is down, the main procedure is to find the gravity vector in the body frame, calculate thrust, lift and drag vectors in the body frame, calculate any torques, sum everything, and then convert back into the inertial frame to do the numerical integration step. From memory, torques are the hardest part, since you need to work with either the time derivative of the quaternion, or the time derivative of the rotation matrix. But if you stick to standard coordinate frame conventions, it's just applying a formula.
The only difference between fixed wing and multirotors is the different models for forces in the body frame. A fixed wing aircraft would use an aerodynamic model of whatever complexity you like, and a model for the prop. A multirotor would use multiple prop models, and a drag model for the body. Most models for these things are pretty well defined, but you can choose just about anything that "fits" so long as you can justify it. And you can use the same basic setup for any aircraft (and indeed any robot) with a bit of forethought. The data part is mostly specifying things like, say, damping coefficients that you would then want to measure.
The dynamics are not the difficult part of simulating aircraft for RL, it's actually the rest of the environment. E.g. navigation tasks require you to build a navigable environment around your physics model, and specify things like reward functions.
I don't have any expertise in this field, but I expect sailboat training simulations would be much harder than aerial drones because of the underlying physics involved.
I'm sure flight simulations have to deal with some amount of fluid mechanics and turbulence to get an accurate simulation, but I suspect it's fairly simple and you can mostly model it accurately enough using Newtonian physics.
But for sailboats, the entire system relies profoundly on the interaction of two separate fluids with a single body and turbulence and viscosity are deeply interwined in making the boat go. Not to mention that the sail itself is a flexible deformable surface.
As a sailor, you're giving sailors a bit too much credit. Those things all matter, but to say that the average sail boat racer is even optimizing for all those things efficiently at once is a huge stretch. A lot of sail boat racing is strategy and properly performing the basically in a hostile environment. If you could ensure great strategy based on learned data, and immaculate execution based on properly tuned controls, I think you'd win easily.
I don't think GP is saying that sailors take all of those effects into account constantly, so much as he's saying a useful simulation would be monstrously difficult due to the underlying physics.
I've made robot models in the past, but never for anything involving water (although I actually have done physics simulations with water but nothing with robots). I don't know anything about sailing and I'm definitely not an expert at building simulations, but I can grok what GP meant and I agree with his assessment.
To give a rough analogy - it would be like trying to learn how to be a racecar driver on a videogame with simplified physics. There are professional drivers that use simulators and iRacing (or Assetto Corsa or a handful of other games) but none of them are training on Need for Speed, and it's because the difference is so stark you're actually handicapping yourself instead of learning how to drive. You need the simulation to be close enough to reality before it starts to become useful.
A simulation doesn’t necessarily need to be remotely physically accurate to be useful, as long as the simulated behaviour is qualitatively similar to the real thing. For sailing a small boat, I’d imagine the important things would be the shape of the sail and the heel of the boat for a given apparent wind, presented to whatever sensors your sailbot is using. For the water’s surface, random semi-regular perturbation (sum of sine waves) would be fine, and something similar (or maybe a random walk) for wind direction and strength.
For this kind of real-world robotics, you can’t optimise in any academic sense so you just have to go for effective and robust.
Well, for the drones, we did base things off ground truth captured in a mocap lab, and this helped the models significantly. There always will be a jump from simulation to real world, but the smaller you make that jump the quicker you’ll have a functioning vehicle. Without that, you can spend a ton of time (wasted) in simulation when the real world is substantially different. I’m sure a sum of sine waves would get you somewhere, but only so far. To me this is a difficult problem and I think it might be best to do most development in a real world environment.
I was just talking about sailboats specifically, where the external forces are highly unpredictable and similar in magnitude (albeit hopefully a fair bit smaller!) than your control authority over the boat.
The work you did on drones sounds very cool and in that scenario (high powered motors, crazy nonlinear dynamics, relatively small external perturbation, at least from how you described it?) I don't doubt that ground truthing your simulation using mocap data made a huge difference.
> as long as the simulated behaviour is qualitatively similar to the real thing.
My point is that I suspect that unless you have a very sophisticated simulation, it won't be.
> I’d imagine the important things would be the shape of the sail
The shape of a sail is itself a very complex thing to model and simulate. I'm not even a real sailor and I'm already thinking about camber, twist, and luffing. When sailing upwind, the sail is functioning like a wing using lift. When sailing downwind, it's relying on drag. Points in between use a mixture. This implies that turbulence and stall must be simulated at some level of fidelity.
If you've got a mainsail and a jib, then you have to worry about the slot effect and one sail blocking the wind of the other.
> and the heel of the boat for a given apparent wind
Not just heeling, but weather helm, leeway, drag, displacement, and how those interact.
Agree, I imagine just maxing VMG on the fly as conditions change is something a computer could already outperform humans easily.
I've wanted to make a simple sailing game for a while now but found it very hard to actually get something that feels right and compared to other activities there doesn't seem to be great resources on all the different parts of the physical system (in terms of Math).
> If you could ensure great strategy based on learned data
That's my point. If your learned data is coming from a simulation that doesn't model actual sailboat fluid dynamics, you've learned little.
A robot that learned to sail by studying a hollow cube floating in a bathtub full of marbles is not going to be able to make an actual sailboat do anything useful out on real water.
Yes getting timestamped and synchronized ground truth to build/verify your model will be quite challenging for something like this. For the drones we just took them to a mocap lab, and races are inside so we had the luxury of assuming 0 wind. The only remotely tricky part is handling ground effect from the props.
That's a very interest project. As a amateur sailor I have played with some models for sailing path optimizatio, wich can be somehow related to your problem. I started with a very simple model using sailing polar graphs, these map wind speed and course to maximum speed and usually based on empirical data, I found some DBs with this data, I can check if you are interest. Nevertheless for a better simulation you will at least need a simple model that gives you the forces at least especially if you want to simulate the boat acceleration, heeling and leeway. Simulating the sail foil behaviour would be very computational intesive as the airflow around it can get very complicated especially as the sail changes it's shape with it.
Reminds of me of my time 7 years back when i took the robotics courses from Davide. What a legendary professor alongwith Bernstein for AI, Sven for Computational-Econ and Boehlen for DBs.
It is, yes, but also this was done purely by VIO - courses were not mapped ahead of time. Contestants had very limited time on the actual drone, almost all their work had to be done via HITL simulation. The entire stack was done in maybe 6-8 months, then contestants produced their first code for the drone in a couple more months, and were flying shortly after. It was very clear that given more time, 6 seconds was going to be surpassed very quickly. All of this was live events too - contestants were tweaking code up until cameras rolled and spectators were in the stands.
It was mostly VIO. The drone had 4 cameras, two front facing, two aimed more at the sides. There were two IMUs on the companion computer (and more on the flight controller of course), as well as a downward facing rangefinder.
It was a software competition, whoever developed the best AI was the whole point. And we did NOT give them full course data ahead of time - they had to do this based off VIO only (visual inertial odometry).
Exciting! I do a lot of FPV drone racing and it always felt ripe for an AI control to take over. Because the video system is so low quality the course is intentionally very high contrast and almost perfect for a vision system. Along with the relatively limited and constant inputs available.
I will say, flying indoors on a relatively simple track like theirs is a lot easier than flying outdoors on more real world tracks.
It's a bummer that they didn't fly the standardized 2023 multiGP global qualifier track. Then we could rank their AI against every pilot more objectively.
You can see that track design here: https://www.multigp.com/global-qualifier/
It may well be because it simply can't complete that track yet. As soon as it can complete it at all it will probably be near the top performers. But if the situation changes considerably from the training conditions I would still expect it to fail, as well as with relatively minor (but unexpected) changes to the track.
I think that the idea of using the residual model (trained on high-resolution outside-in pose estimation) to correct the inside-out model is particularly interesting because it might actually be possible to generalize; if the kinematics residuals (basically the delta between simulation training and the real world aerodynamic + inertial behavior of the aircraft) are able to be made general enough to reflect the flight dynamics of the drone, rather than the specific actions on a specific course, this is a promising approach for general purpose flight.
It's a really fascinating approach compared to most "AI-guided" drones which use models for vision and pathfinding but a traditional IMU+PID loop for kinematic control.
It reminds me of the way the Mario Brothers game changed overnight: for the longest time it was considered impossible and then suddenly it was unbeatable. But both benefit from the static course, if it generalizes it will be a game changer, otherwise, I'm not that impressed. But there may well be some gold to be found here and I applaud them for making it work this far. Maybe they could purposefully improve their performance in real world situations by making this one harder, for instance by changing the lighting from one run to another, introducing or removing obstacles or moving goals around. That might force the model to come out more general. We've seen similar strategies used with good results in image classification problems. In fact I used them myself when building the lego sorter, as long as everything was always lined up perfect it worked a lot worse then when introducing various complications. During the real world runs those would show up all by themselves anyway and where before they were classified wrong or ended up in the recycling bin for another shot they were suddenly classified right.
Of course a setup where you can gather your training data with thousands of images per hour has some advantages over one where if you get it wrong you have to rebuild your drone...
A changing course is one of the biggest impacts of flying outside. You have the wind directly acting on your drone, and you have the wind acting on the gates and flags. Flags will spin around in the wind, double gates will lean, sometimes quite a lot in the wind.
There's a whole skill to feeling the wind on your body and anticipating how the drone will behave. When I feel a big gust of wind I'm going to slow down out on the course to get my bearings.
Navigating between gates is usually done by flying what you can see and reacting real-time to conditions.
When you get to a gate, or perform a turn or a split-s, these manoeuvres can be executed faster than you can see and process so Instead you go faster than your brain by memorising the timing of your inputs and execute them from memory rather than flying what you can see.
For example, if i want to do a barrel roll, i don’t watch the horizon. Instead i rely on my memory of how fast it rotates at full stick deflection.
Sometimes these set pieces go wrong - you clip a gate for example - it’s usually my hearing processes that a fraction of a second before my vision does.
It’s quite common to fly without vision for short periods, the analogue video feed is not reliable.
I'd say it's a combination of both. Just like in auto racing, some people are really good at flying "seat of the pants" or "reading the course" and adapting to a new layout gate-to-gate, while other pilots are better at a course with practice and memorization. At the top, it's about both together. Most races are conducted with analog video at a very low output power to reduce interference between racers, so the visuals are pretty weak and most pilots do rely on a fair amount of memorization. But, there's also a not-inconsequential conditions angle that comes in when flying outside, so pilots need to be adaptable.
In terms of "fairness" to the computer, I think this approach isn't up to snuff with a human pilot until it can fly an outdoor course in changing conditions. Still, I find this particular approach very interesting since it's inside-out (self-contained) flying once it's trained, with guided learning to start. I found the earlier purely outside-in guidance approaches to be rather ho-hum as they weren't very practical and basically skipped the "hard parts."
In auto racing (other than rally), pro drivers typically know the track incredibly well, down to every bump that could possibly affect traction. Often times there are blind crests where they have to know the track perfectly to be able to go through a corner at maximum attack. F1 drivers know the tracks like the back of their hand except for the times where they go to new tracks, and even then they can practice in the simulator.
Their world-view (and knowledge) of disaster response was shaped by steady diet of 90’s disaster films — Volcano, Dante’s Peak, Armageddon, Twister, etc.
I'm guessing faster Street view updates could be next. Park the car, release bunch of drones (permission from the town is no problem, it's good for local businesses), have all nearby streets scanned in a few minutes. Eventually have a base station in each city and do daily (hourly?) updates.
That may be true in a small number of very pro-technology cities in the US, but I think it's far from the case around the world in general.
Btw, I think the point has been missed here a bit. Autonomous drone technology has been available for quite a while, and is used for use cases like the ones you describe (but usually on private property). The new thing here is drone racing - very high speed, high precision flying.
Gotta love the cheeky euphemisms of these companies. In this case, disaster response isn't responding to a disaster, it's a response that causes a disaster. Preferably for people you do not like.
I have a client who does drone flyovers for oil and gas industry monitoring. They also do some disaster response work, and it can indeed be delicate flying, especially given the value of the drones and their sensor packages. Companies like theirs will absolutely use tech like this to replace expensive human pilots with AI once it’s widely available.
Although, I have a wild fantasy that someone will build fleets of drones that’ll take to a battlefield with the sole purpose of stopping a conflict.
Buzzing around, disabling vehicles, grabbing weapons and ammo, plugging up artillery barrels, flashing lights in combatants eyes, softly playing contemporary adult jazz… basically making it impossible for the two sides to fight.
Everytime I see the Slaughterbots video posted, I roll my eyes. There are currently many weapons capable of VAST destruction, and suddenly _this_ is where we should start drawing the line? If military drones were to become real and as capable as in the video, they would be hard to aquire, maintain, operate and keep hidden.
Besides that, yes, I do tend to believe that autonomous flying drones will play a large role in future conflicts. This dude I follow on Reddit thinks this based on what he is seeing going on in Ukraine[1]. I got sold on it as a possibility.
You should watch more combat footage on Reddit. Artillery is woefully ineffective against even lightly protected troops. It takes about a dozen expensive 155mm shells to get one kill against troops out in the open.
It still takes about two of the much more expensive cluster munitions to get one kill. The submunitions in those things are about as expensive as a “murder bot” would be.
An actual person-seeking drone might have a failure rate lower than 50%.
A single shell with 20 mini drones could take out a dozen troops, even if they’re in a trench, or spread out, or lying on the ground to avoid shrapnel.
It would be the end of infantry warfare, forever.
Squishy meat troops can’t evade tiny bots with 1000 fps cameras and inhuman reaction times.
Do you have any information on electrically-powered drone boats? Or information that they are electrically powered at all? Evidence I'm finding is that they'r e based on combustion motors.
I'm going to assume that those are low-speed at best. Water transport is efficient at scale, but has tremendous drag, and electric boats (of a non-nuclear sort) will tend to be quite slow.
WWII-era diesel-electric submarines typically have a top speed of less than 10 knots submerged on battery power, as compared with 20 kt surfaced (or at snorkel depth), and in excess of 25 kt for nuclear powered submarines.
Though on writing this I'm noting that the Swedish Gotland-class submarine reports 20 kt submerged on batteries, though only 5 kt on air-independent propulsion (AIP), a Stirling-cycle engine design.
I'm finding very little on UA drone ships / unmanned surface vehicles (USV), but this CNN report shows a still photo of one generating a considerable wake, which I'd estimate shows at least a 15-20 kt speed and possibly greater. The copy reads "faster than anything on the Black Sea". I still suspect that stealth and battery limitations would make lower-speed operations far more common, particularly where the hull could hide from radar or other detection in wave chop.
The accompanying video seems to indicate engine noise, which would indicate combustion-based propulsion. Given that the explosive payload is 300 kg of explosive, and in the Kerch Bridge attack, 1 tonne of high explosive (HE), as well as the extended range of the missions (Odessa to the Sea of Azov), I'd expect that any battery capacity would be limited at best. Fuel, on the other hand, would help enhance explosions and seems far more likely.
Though that does raise some interesting relationships between insolation and combat situations. Night-time, winter (short daylight cycles and low solar angle), and overcast / stormy weather might favour infantry over solar-powered / assisted drones.
You don't need to worry about the drones themselves. The drones are useless without the explosive munitions. Good luck acquiring that without raising any flags. Not to mention somehow acquiring all the training data and ML modals to actually make them useful. Then there is the compute power that would be needed and coms infrastructure.
it puts it firmly out of reach for anyone that isn't state sponsored.
Not saying autonomous weapons won't/couldn't be an issue. just pointing out that drone swarms won't be.
>>explosive munitions. Good luck acquiring that without raising any flags
Not if you were a bit precocious in HC chemistry; I know my friends and I were not even close to alone in knowing all kinds of interesting fireworks recipes, all of which were well tested, and could be configured for anything from a small 'pop' to far more power than a slaughterbot. All from basic chemicals that can be easily be acquired without raising an eyebrow (could alternatively just repurpose a bullet/ shotgun shell, very easily obtained). The ability to put it right on target means that very little is needed.
>>acquiring all the training data and ML modals to actually make them useful.
Sure, if you are looking for fully ai-controlled and integrated with a city-wide CCTV system, that's a big task. But drone control and swarming is already out of the bag. Even a basic open-source Pixhawk does obstacle avoidance, and so can likely be reprogrammed for obstacle-targeting. Multiple off-the-shelf drones have done "follow-me" for years. I'd be astonished if a competent team with five-figure funding couldn't make a working system that could be deployed near a target, ID, it and home in.
We're way past needing state funding. The fact that it is not here suggests that while the scenario is scary, it is not actually that useful? Perhaps that's hoping too much.
I agree with all your points. I've been saying to my friends that I think one of the wealthy billionaires could possibly become a serious power. I think we can all agree Elon Musk could make this happen easily with his existing SpaceX people, and then build a bajillion of them with his Tesla people.
I wonder how far EM could go if he went full evil, full throttle and tried to take over something [the world? a country?] with a drone fleet?
Charging stations seems easily solved, so the drones never need direct human interaction.
I don't. He's the obvious choice for this thought scenario for several reasons:
1) He's the richest man on the planet
2) He has competence (arguably, but that's another debate)
3) He already has all the infrastructure, experience, and people to do close to this exact thing
4) He's already widely considered evil or at least hated
If you sub in Bezos, I'm guessing he's got a lead on EM in the software, but he'll be way behind in other areas.
If you sub in Buffet or Gates, then I have no idea how to even have this flippant conversation, because I have no idea how to rate their ability to get the job done.
I'm confident 3 of the 4 above could solve the problems of "make autopiloting, auto-targeting AI drones, and produce millions of them." I'm sure countless people could solve those problems and have the means to get it done. Have your choice of billionaires and motives, or nation-states and motives, or aliens and motives.
Maybe I should have chosen Santa Claus. I do really have a hard-on for him.
They do, but I don't think this is an example of that — but Musk's position w/Tesla and SpaceX makes him uniquely positioned to develop & deploy such tech.
That said, with all the advantages Musk would have if he "went full Evil", if he challenged something that got the US military involved, I'd bet on the US Military. In a straight-up fight, they're the most lethal force that ever existed. You might get away with 'sucker-punching' them a few times, but not for long.
> Good luck acquiring that without raising any flags.
How is the explosive used in the slaughterbots scenario different from what you can buy by the box in cartridges of bullets? Those aren't exactly highly protected even in the USA.
I imagine e.g. just tying a bullet to a drone, without the structure of a gun around, it isn't exactly going to be effective. But how far off would it be?
It’s a decently large difference. A couple oz of C4 is amazingly effective, no casing required - but an additional couple oz of ball bearings or a copper cone will make something truly scary. [https://youtu.be/AwyniA5ryhY?si=a562j-T8pyQlwaH4]
1/4 lb total for a devastating precision or area weapon.
The casing required for any nitrocellulose based device to do anything dangerous at all will exceed the weight of both of the above and be barely a firecracker. Fragments will be limited in velocity too.
Because if you look around at sophisticated military hardware (which these drones, when they will exist, will be) it's all like that.
I feel that the point our stances diverge is you feel these kind of weapons systems would become something a skilled hacker could make in a shed out of AliExpress parts for a 10$ bom, whereas I view it as being a complex weapon system that would be sophisticated and relies on battlefield infrastructure and complex supply chains to be feasible to properly deploy.
I think if you look at cutting edge weapons, my belief is no hacker could build them in their garage, no matter how sophisticated. Hell, take a look at the pathetic state of "3d printed" weapons to see how I imagine DIY "slaughterbots" would look like.
And acquiring the real weapons would not be easy at all. Heck, even now, if you want to acquire for your projects certain camera or heat sensors you have tons of hoops to jump through to get your hands on them.
Besides, even IF it would be something a DIY-er could do, someone wanting to kill civilians would probably use the most effective tools, least resource intensive, least effort tool, least conspicuous to gather resources and this would not be it. Today any maniac can use a car to plow down civilians or get a gun and go on a rampage. That is low effort high impact. This would be a lot higher effort. Probably kind of the same impact.
> Because if you look around at sophisticated military hardware (which these drones, when they will exist, will be) it's all like that.
I agree that it will be much like sophisticated military hardware is today - but I say that as someone within 20’ of his own third-generation night vision, body armor, and weapons that are approximately on par with that of a contemporary US soldier.
> Heck, even now, if you want to acquire for your projects certain camera or heat sensors you have tons of hoops to jump through to get your hands on them
Nonsense. You can order very usable thermal sensors from China, shipped to your door. You can order even more capable sensors that are covered by ITAR and prohibited for export without so much as signing something. A notice on the website saying that they’re restricted to the US is the extent of it.
> Today any maniac can use a car to plow down civilians or get a gun and go on a rampage. That is low effort high impact. This would be a lot higher effort. Probably kind of the same impact.
I totally agree with you here. Firearms aren’t the most effective way to cause harm to a large number of people even today; drones certainly won’t be for the foreseeable future.
FPV killer drones used in Ukraine now are really slapdash designs: often just motors on a flat fixture with a charge rubber-banded to it. They are used to devastating effect.
The only part of this technology that is remotely possible to tackle on is proliferation of explosives.
Fully agree. I mentioned in my original post I believe them to be very capable on a battlefield. But from there to slaughterbots like targeter drones there is a looong way. I believe my points still remain valid.
These drones won't be sophisticated. They'll be cheaper versions of off the shelf drones. Quantity > Quality.
Imagine a drone that flies towards a cellphone signal and when it gets within range does some basic shape recognition and adjusts it's course to fly towards a the nearest neck until it detonates a crude fragementation grenade.
How much would a drone that can do that cost today? They don't need to be durable as they are meant for single use, and they don't need to be particularly fast, just faster than a person can run.
Consider that despite grenades being dropped from drones onto Russian soldiers in Ukraine on a daily basis, it happens to American civilians never despite America's extremely high homicide rate. This is due to a lack of grenades, not a lack of drones.
The main thing gating slaughterbots from reality is the proliferation of the shaped explosive, not of the delivery mechanism.
Americans kill each other with guns because guns are cheap and plentiful, not because grenades are scarce. If guns were as scarce as frag grenades then it would be a competition between which one is more effective at killing and the availability wouldn't factor into the weapon choice at all.
What you're telling me is that slaughterbots will be armed with guns that shoot bullets, not fragementation grenades.
Well, it's a bit more than that. People kill each other with whatever they have close at hand when they get into a fight; handguns are easily carried into a bar or on the street, which is why more people are killed with handguns than long rifles in America despite more of the latter existing. People aren't going to start carrying drones with propellers around in their pockets.
It could be used by some people who want to do targeted killings, e.g. in gang violence, where it could be better than other methods maybe. But on the other hand, RC helicopters able to carry and shoot a gun have existed since at least 2006: https://www.youtube.com/watch?v=bZCH1492CzA
The number of people who actually want to kill lots of people without being physically present is extremely small; so small that in a country the size of the US, one only crops up every couple decades, e.g. Timothy McVeigh.
I would make a solid bet that proliferation of cheap drones does not impact the homicide or terrorism rate in the US in either direction.
Grenades are low precision weapons that are relatively safe to use. Guns are high precision but very difficult to use in a firefight.
If you watch Ukraine footage, soldiers barely even look where they’re shooting. They’re just trying to establish suppression over an area so that they can toss more grenades to kill people.
That’s because both sides have guns. If you aren’t fighting an armed combatant to the death. The calculus changes a lot.
(Video shows airsoft players using a standard FPV drone with an actuator to press a trigger, and the guy flies around the field getting "kills" and it looks incredibly effective (partly due to no recoil, but .22LR exists))
The ordinance is controlled, and without that the system loses a lot of potency. Not to mention that without decent onboard AI, jamming is a pretty good countermeasure since target selection and much of the control input isn't actually done onboard the vehicle.
Those seem like pretty large hurdles imo.
Edit: and, as others mentioned, goods even tangentially related to defense are routinely controlled. In the extreme case, countries like the US would onshore manufacturing of certain components, acquisition would require licensing, and competing products from overseas would simply be banned.
What is so difficult about creating a small explosive that could take a chunk out of someones neck so they bleed out?
> In the extreme case, countries like the US would onshore manufacturing of certain components, acquisition would require licensing, and competing products from overseas would simply be banned.
Which parts exactly? the PCBs? The motors? the microcontrollers? the batteries? How would that work exactly? How do you regulate any of this without per component DRM? What would that look like exactly? Would I need a license to buy a child's toy that had a microcontroller in it? Would they nerf microcontrollers for devices like appliances and toys so taht they weren't fast enough to handle the needs of a drone?
Materials for making decent explosives are usually controlled. If you're talking homemade explosives, you probably aren't getting that on a quadrotor, or if you are, you're seriously compromising its performance. Bullets are controlled through licensing in most first world countries.
If you wanted to control components going into homebuilt drones, you would probably target the guidance, since even basic waypoint navigation requires something to tell you where you are in relation to the target. To an extent this is already done with GPS (the true precision was classified last I checked, civilian uses are downgraded), but controlling manufacture and distribution of the stack would shut any kind of home robotics down while leaving a lot of channels untouched. If people wanted PCBs, fine, they just wouldn't be able to get certain sensors on them without a license.
You don't need global DRM to control exports. It's routine for some industries.
Of course, much of this is besides the point, because if someone really wants to hurt other people, they'll find a way. But the components to build a cruise missile in your backyard not only exist, they're mostly available COTS, but the critical parts like the payload are controlled. In some hypothetical future where AI drones became a threat, you can bet that the critical paths would be identified fairly quickly and locked down.
> If military drones were to become real and as capable as in the video, they would be hard to aquire, maintain, operate and keep hidden.
Couldn't a terrorist load an explosive on a drone and drove it into a music festival or some other open-air large gathering? This is what concerns me personally.
Although long-term (~10 years) I wouldn't bet on more capable weaponised drones being hard to aquire, maintain, or operate.
Someone tried to assassinate Nicolas Maduro with drones carrying explosives back in 2018.
Protections against this sort of thing usually exist in the form of radio jammers; disrupting the control transmission from the pilot or the video transmission from the drone. If the control system is onboard the drone it's a lot harder to disrupt.
As far as I know the systems designed to counter this sort of threat look like shipping container sized microwave emitters that interfere with the electronics of the UAV.
Any of the drones used in the Ukraine right now would be able to do a lot of damage. And those things are relatively low tech.
The reason these horror scenarios aren't so common is that most people simply are not suicidal murderers. It's always been easy to target crowds of people. But it's not that common for people to actually do that. When it does happen, there usually isn't a happy end for anyone involved with the attack.
Any determined idiot armed with a knife, baseball bat, truck, improvised explosives, etc. would be able to do quite a bit of damage. And in the US you can of course add all the light military gear that lots of people there buy for fun (assault rifles, etc.). That sometimes goes horribly wrong of course but you don't need drones to do a lot of damage.
Semiautomatic rifles were legal in the US for a time before someone shot up the first high school. But now it's a concept in the public consciousness and schools need to have shooter drills due the frequency.
The entire premise of the video is not what is possible right now, but what could be possible in the not to distant future. I.e. that these drones do become available to anyone. Easy to acquire, maintain, operate and keep hidden.
It's also not about warring states, but instead very targeted assassinations, where the perpetrators doesn't have to be anywhere near the target, allowing very little risk of punishment, retaliation, or even identification.
Aren't assassinations targeted by definition? And surely this kind of thing is already doable if you're motivated enough, e.g. every smart phone is both an identity mechanism and a signal (thanks to all of the tracking/location systems) that could probably be used to guide a weapon system to an intended target. Hell, there was a kiwi engineer who built a cruise missile in his garage almost 20 years ago. I don't know that a killbot system offers much of an improvement over existing methods for the difficulty of execution, with the probable exception of, say, air superiority where the capability of autonomous systems will far surpass human ability. But even there we're probably two generations away from fully autonomous systems; the next generation seems to be robotic wingmen.
What really stops the average person from building these systems isn't necessarily the difficulty of the system itself, but rather the impossibility of getting the components. The drone is just a delivery mechanism for the ordinance, and good luck getting your hands on that.
Also, for said killbots to be anything other than a trivial threat (just jam the signal) they would need pretty decent onboard AI to function independently, which I guess is the point they're trying to make. But there are almost certainly far easier and more cost effective ways to get the same effect that don't rely on currently non-existent technology.
Sure. I should have said assassinations and targetted attacks. The video shows the ability to target specific races/genders/demographics.
> The drone is just a delivery mechanism for the ordinance, and good luck getting your hands on that.
I don't know why a shotgun cartridge could not be used in place of a shaped charge. Just pop down to your local Walmart.
You seem to be arguing that this work speculative fiction isn't interesting because we already have a huge number of ways to kill people, and that it's not currently possible. The point of this sort of fiction is to get us to imagine a world where these things are possible, where it is achievable.
A shotgun cartridge would be vastly less effective, and in many countries, even getting your hands on ammunition is difficult.
My objection is that it's mostly fright porn for drumming up mindshare. I don't think it's particularly compelling because a) imagining such systems is trivial, and b) there are already ways for us to achieve the same goal, so it's not even a novel threat.
But of course, that's just my opinion. Others obviously feel differently.
But that's beside the point. There are folks on YouTube making all sorts of dangerous things in their garden sheds on a shoestring budget. If Explosions&Fire can do it, so can a dedicated attacker.
>If military drones were to become real and as capable as in the video, they would be hard to aquire, maintain, operate and keep hidden.
Really? Because as of right now everyone in HN and other tech communities thinks government regulation of AI is "market capture" and should be blocked entirely. Many will not be convinced until the horse is already long gone.
The parts to make an 8 inch kill drone (except the explosive, but you could use a blade or other simple weapon) from scratch are sold at microcenter. The only missing piece is the software.
"I can only imagine what would happen for example if the US still had a relatively large contingent in Afghan, and the militants got a supply of even few hundred high precision aerial IEDs per month."
[Nah] Being afraid of automated, weaponized drones
[!!!] Being afraid of social media enabling something like automated, weaponized drones
This boogieman from the future is kind of overshadowing the fact that social media is already weaponized and used to harm people, and can be used in increasingly effective ways without the need for shaped charges or even physical objects.
Vast destruction is almost never used because of the consequences and collateral damage. Surgical, cheap attacks would be used on an everyday basis. This is what makes it terrifying. It doesn’t even have to be murder, even surveillance would be the end of a way of life.
There is a vast difference between the 'surgical' precision footage pushed through regime media such as CNN, and the reality of undiscriminate carnage in the actual warzone.
Still need to solve the power problem. When drones can fly for 8+ hours straight or charge themselves using the electrical infrastructure then I'll worry about "Stabby the Robot".
AI isn't really the limiting factor as far as I can tell. We have good enough pattern recognition and flight software already.
There are much higher density energy sources than lithium batteries which militaries can use, I’m not sure that is at all insurmountable.
In other words, time to worry :)
https://newatlas.com/nuclear-uav/22041/
If you can accurately strobe it with a laser, why not use the same targeting system to shoot it down? But good systems like that with sufficient capacity won't be cheap. And they'd be prime targets for artillery during an offensive operation.
Then again, from watching the publicly available videos, it seems like these drones have been most useful for defensive operations on the battlefield.
> If you can accurately strobe it with a laser, why not use the same targeting system to shoot it down?
There will always be a chance that the targeting system misses so strobbing a laser (or bidirectional jamming) will minimize the risk of damaging the surrounding area (people, animals, etc.) vs shooting it down with bullets/nets.
Also a laser that emits enough light to blind a sensor may not be classified as a munition, whereas something that emits projectiles is almost always classified as a munition.
Dazzle camo and decoys seem like the obvious answer, and anything flying is still gonna have to contend with AA weapons platforms. It doesn't seem like it would be the hardest problem in the world to honeypot an autonomous drone into a kill box.
Those sound like captchas for drones to me. We've already started to see that bots can defeat some of those better than humans can.
Decoys are likely the best answer, but not necessarily practical for offensive operations. And given the low cost of drones, likely limited in usefulness anyway.
I'm not worried about battlefields in the "stabby the robot" scenario. I'm worried about unprepared civilians being exsanguinated en masse by a fleet of drones.
The scary scenario as far as I'm concerned, is mass genocide based on some characteristic. 10k drones wiping out 100k civilians who are identifiable based on location, skin color, clothing or some other measure. It's a scenario that has seemed vaguely plausible to me for several years now.
I refer to it as "stabby the robot" because drones wielding blades or other means of exsanguination do not run out of ammo and have little limitations on how much they can achieve their goals(other than powering themselves).
It’s made by a team from Zürich. If the Swiss army get hold of this technology then the applications will cover everything from whittling wooden sticks, to removing screws, to opening wine bottles. Advanced models might even have a toothpick or a pair of nail scissors.
People don’t understand what SWARMS of bots can do.
Online reputational attacks overnight.
Astroturfing a revolution on any topic. Disinformation at massive scale. Amassing karma points and deploying sleeper accounts with morals totally off.
Destroying forums and overwhelming every human fitting a filter.
Getting people into trouble easily in 800 different ways (think Trump lawsuits)
And this is before we ever get to the physical world.
People are still stuck wondering how a poker bot will do at a table with 9 humans. The issue is what will humans do when surrounded by self-organizing swarms of 10000 bots for every human?
> Its integrated inertial measurement unit measures acceleration and speed
If that’s fed into the neural net, I don’t this is an actual matchup. An accomplishment for sure! But the computer didn’t beat the human.
I’d consider this a win when the computer has to send its inputs over the same wireless link that humans use, fed vision inputs with the same latency and quality.
Edit: I want to add this is a criticism of the headline, not a criticism of the developers and project managers
I agree with your sentiment. The computer should also fly without any extra components.
But I think it's pretty clear that
a) this poses a bigger challenge for which the computer isn't ready yet - it barely beat the humans with its advantage.
b) we could also reach equality by feeding the humans the measurements from the intertial measurement unit and allowing humans to ride the drones. Whether they can make use of it is their problem. Showing that "equal conditions" aren't really possible.
c) The current matchup: "Computer with extra sensor data, humans with lag" presents how the battle will be in real-world applications, like warfare.
It would be trivial to hook accelerometer output to audio signal generator and transmit it together with video - pilots could then learn to interpret audio cues as direct force feedback.
The proposed use cases in the paper like "forest monitoring," etc. seem like uses cases in highly dynamic environments. It seems like the applications will lend themselves more towards human pilots for the time being.
Fun experiment and test, though! Very interesting.
When’s the next contest when each team of paintball players each get a single AI Drone armed with a paintball gun? Let’s not kid ourselves, that’s a more accurate simulation of where this research is going.
The world champion level results were achieved when using external motion tracking, not onboard camera. So, not as impressive as the title made it look.
> Until very recently, autonomous drones took twice as long as those piloted by humans to fly through a racetrack, unless they relied on an external position-tracking system to precisely control their trajectories. Swift, however, reacts in real time to the data collected by an onboard camera, like the one used by human racers. Its integrated inertial measurement unit measures acceleration and speed while an artificial neural network uses data from the camera to localize the drone in space and detect the gates along the racetrack. This information is fed to a control unit, also based on a deep neural network that chooses the best action to finish the circuit as fast as possible.
Well, they've muddied the waters well enough. The b-roll in the last segment of the video you linked shows motion tracked drones and the 2022 event that used motion tracking (the balls used for motion tracking are visible on the drone arms). The voiceover tells us the system won against human racers. So, even if they have a vision only run made on the same track that beat human times, that run wasn't shown (but you couldn't tell that if you just watched that video).
And from BMSThomas' video, it's obvious their vision only system wasn't ready for prime time in 2022, but they were still working on it.
This paper is based on the performance at the 2022 event, but it took this long to write it up and get it published.
The tracking system was used during all of the races for data collection of both the autonomous and human-piloted drones, which is why the reflective markers are visible.
They did do some demonstrations of the drones controlled with the tracking cameras, and they were significantly faster, but the vision-based drones were definitely able to fly faster than the human pilots in some races.
My read is they were only using the motion tracked system to train the residual network, but the actual competitive flights were done without input from the external tracking but they still captured the data, so they could feed it to their training for future runs.
“Now a group of researchers from the University of Zurich and Intel has set a new milestone with the first autonomous system capable of beating human champions at a physical sport: drone racing.”
This is somewhat impressive but let’s be honest here: drone racing is about as much of a physical sport as riding a Peloton.
Might be a mistranslation but I think what they meant was sports situated in the real world as opposed to the digital world. That requires a lot more feature learning and engineering than let’s say chess and go where board state can be more easily encoded.
It seems to require training on a static environment before it can actually execute and drive the race track, is that true? If so, wonder how long time / how much resources the training would require.
I first was under the impression that the entire thing was executed in real-time with onboard sensors and computer, but seems that's not entirely true if I understand it correctly.
Oddly, I would think the "limitation" of only using the onboard system for control is an advantage? In that not having the latency of transmission back to a controller has to be measurable?
I probably have very bad intuitions on how long it takes to transmit video and such?
I’m not familiar with the specifics latencies for drone racing but hierarchical planning/control is very very common in robotics.
It’s natural to say plan a longer term trajectory at 1Hz, a short-term trajectory at 10Hz, and perform control at the 100Hz, obviously with those rates varying based on the system.
And since (generally speaking) higher levels of planning require more compute resources, it might make sense to partition the compute and take the latency penalty.
Makes sense in general. In the specific case of these racers, it seems you wouldn't be losing out on much by keeping everything local, though?
That said, I'm still mostly curious on how long it would take you to get the sensory data off the drone. My gut is that the video transmission done would blow through most of the time budgets you listed. I'd love to see a good rundown on the relevant latency values involved.
Good high-resolution specialized video-codec based (H.264 / H.265) digital FPV video systems like Walksnail or DJI hover at around 30ms of latency glass-to-glass.
Analog systems used in racing are at around 15ms glass-to-glass, with most latency coming from the camera's image processing, but would be a bit non-standard to train or build a system around without transforming back into the framebuffer/pixel-value domain, which would introduce some slight degree of extra latency.
There's a rather unique uncompressed transmission system called HDZero which roughly matches analog latency.
Purpose-built low latency uplinks add 5-10ms to get data turned around and back into the flight controller, albeit at fairly low bitrates as they're mostly serial based.
So, for realtime control (kinematics), which runs at the multi-kHz rate on racing drones, onboard control is pretty much a must, but for both short-term and long-term planning, offboard control becomes more practical.
Most non-FPV drones are already architected this way, with gyro-to-motor PID control performed on a microcontroller running an RTOS, short-term planning information coming in asynchronously from a larger SoC running Linux and ML-type stuff, and long-term control information coming down over the air.
Thanks! This is really cool stuff. I'll start digging online looking at some of this stuff. Curious what the resolution constraints are for this video. I'm assuming for cinema purposes, they have separate control video to keep things fast?
And I now have to convince myself that i don't, in fact, need to physically play with one of these things. :D
Right. DJI's FPV video link has around 60mbit capacity, 50mbit used for video in stock form. I think Walksnail is similar.
For FPV/"freestyle" cinema and even FPV YouTube videos, the onboard cinema camera will be totally separate from the video link. In the case of FPV, this will be "cinelifter" FPV drones which lift RED/Arrai/BlackMagic cameras, or "freestyle" FPV drones with a GoPro.
For "camera" style drones, a separate gimbal camera from the "main" flight camera will be employed. These often have real-time capability, but with downscaling and a higher-latency / lower-resolution video downlink, like DJI's Zenmuse series found on the Matrice and Inspire drones. On these drones, you basically pick and choose what's sent to your display given the downlink limitations, but you also have a much higher latency tolerance as they can fly themselves.
Slightly OT but I don't think that Deep blue's achievement against Kasparov is the milestone they make it to be, as the moves of the computer were still vetted by other grandmasters.
There's no doubt that if deep blue would've not beaten Kasparov another computer would've just few years later (I also think algos have made insane improvements since then, it's not just a matter of raw speed) but still that game is not the kind of achievement that it's touted imho.
Great video! Noise was added to the AI decision making. Reminds me the the saying: "Silence among silence is not true silence, silence among noise is true silence" Next challenge: moving gates!
I suppose they mean you won't need a drone pilot, which takes training and takes a soldier off the battlefield in terms of them needing to have on a headset and can't fire a conventional weapon at the same time. I don't think it will come to the war so quickly though, it's the least of the problems at the moment for Ukraine that badly needs long range missiles, mine clearing equipment and F16s.
FPV pilots need special training & jamming could interrupt the control link.
A fully autonomous drone is a nightmares inducing option that just does not have these downsides. For extra inhuman points you could have it sleep in a box & launch on trigger or timer.
Like the hunter seeker from Dune, just without a pilot, lightning fast & with a RPG7 round stripped to it.
Oh those rose colored visions of future utopia people had in the past. ;-)
So glad to see this team from UZH continued pushing the envelope and are now beating human champions. If you saw the team and what they managed in under a year, it was clear they were highly talented and human racers had their work cut out for them to stay ahead.