While I see the obvious dystopian version of this, I'd still like to see this tech in the hands of search and rescue teams.
People get lost in the woods all the time. Image a swarm of 100 of these drones, 200, 500, all flying at 40km/h in a coordinated sweep, searching with multi-spectrum cameras and image recognition.
All of this tech can also be used for good. They're introducing that new Wifi standard because they claim it can help monitor elderly people's heart rates. Surveillance cameras are there against street crime and they scan our private data in the cloud to prevent child pornography. Or take a look at this pilot project in China: https://old.reddit.com/r/ABoringDystopia/comments/q32799/bec...
Somehow the idea of autonomous winged cameras flying around all over the place does not make me feel safe.
It's interesting how they sell this over there vs in the west. In the west you get very ominous, 1984 panopticon connotations whenever cameras are introduced. You never see this back and forth dialogue between your friendly overlords and the people being watched.
A network of tiny cameras paired with wifi, can identify everyone all the time. You also carry your phone on you and now with Apple’s latest improvements it won’t really be “off” when you turn it off.
They can hear through windows, lip read, and anything else. All that’s missing is a database that will cross compile all this information to reconstruct everywhere youwere and everything you’ve been doing, and your health characteristics at every moment. From there it’s pretty trivial AI to make sure to do precrime and nip “undesirable” gatherings or movements in the bud before they have a chance to take root.
Ironically, the one place you MAY have privacy would be end to end encryptednetworks online.
Every time I drive past this place, I get an uneasy feeling. An organization I thoroughly detest, setting up shop with one of the largest data centers, right in my home region.
> Image a swarm of 100 of these drones, 200, 500, all flying at 40km/h in a coordinated sweep, searching with multi-spectrum cameras and image recognition.
Now imagine those machines chasing down a thought criminal desperately trying to hide in the woods in order to avoid execution.
> Now imagine those machines chasing down a thought criminal desperately trying to hide in the woods in order to avoid execution.
If you're hiding in a forest and they know where you are within 30 minutes of flying time - you've already lost. If they don't have drones they can just use dogs.
The problems with this situation to solve, sorted by priority:
- don't have a totalitarian state
- don't let them know your thoughts
- escape to a different country while you can
- don't let them know in which forest you're hiding
- don't let them know where exactly
Drones only change the last point, and only marginally at that. Meanwhile we're collectively helping the internet overlords with the first 4 points by carrying smartphones everywhere and using google and facebook and whatsup etc.
Except that democracies can and do slip into totalitarian rule by autocrats who declare themselves the winner of elections and then obliterate their political opponents.
> - don't let them know your thoughts
Uh, whoops, they just know everything you ever searched for, watched, or bought online, who your friends and lovers are, where you live, what music you listen to, and what your genetic history is!
> - escape to a different country while you can
Except COVID and visas drying up, and anti-immigration sentiment rising everywhere.
> - don't let them know in which forest you're hiding
Better disable GPS on every device near me then! As well as disable every "CC" camera that's just for "security". Better also shoot down any surveillance drones or satellites!
The point is reasonable; but drones are quite different to dogs.
I imagine it is much cheaper maintain 100 drones, truck them to where they are needed and have them swarm with some level of facial or gait recognition than the equivalent operation would be with dogs.
Also this could be centrally coordinated with cameras a lot more cheaply over a wider area than at any point in the past, with fewer people and less requirement for locals to dob people in.
> ...don't have a totalitarian state...
After COVID, I'm not sure how feasible this step is. "People should be allowed to leave their homes without first having the required paperwork" does not even appear to be a consensus view. And speaking from Australia even that would be a step up from what we have had recently.
The same exact argument can be made against end-to-end encryption that is resilint against state actors … if you have to resort to sneaking around, your society has already lost politically. Fix your government.
The drones also solve the second to last point, if they make it viable to just sweep all forests within travel distance.
That's the important difference between dogs and drones: you can't substantially bring down the price of trained dogs with handlers. The costs of drones on the other hand are guaranteed to go down with scale, both in terms of capital costs and operating costs. So if your dystopian goal can't be achieved by a couple drones, just deploy more of them.
Reality is boring. The drones will be used for surveillance but the death will be delivered by regular firearms. The holocaust was "expensive" enough to justify shipping people into a central location where they are executed as efficiently as possible. What makes people think that a government would be willing to spend $5000 per killed civilian when it has to kill hundreds of millions?
I agree, with one correction: majority of Holocaust victims never seen a camp (neither concentration nor death camp). They were executed with regular firearms near the place they lived.
Camps were for undesirables that for whatever reason couldn't be killed immediately. Mostly because there were too many of them in one place.
And yet, Afghanistan exists today as an independent Taliban state. I think the English would still have trouble projecting power across the pond indefinitely.
Sometimes that's out of your hand. The people killed in Nazi Germany had no choice in being Jewish or disabled.
Governments can change for the worse pretty quickly, if you have one competent charismatic guy with the wrong ambitions. So we would better not give such people the tools to keep their power, if it ever comes to that.
But the problem to me with that vision is the existence of "thought crimes" and not the method by which such criminals are being chased.
The government isn't waiting to pass thought crime laws because they just don't quite yet have the ability to search a forest for people. That's not the big barrier here.
I’m imagining them in a snowy forest chasing a boy carrying a baby, while the boy uses memories of coldness to hide them from the heat-sensitive cameras on the drones.
The law enforcement is already equipped well enough to get you if you lead a normal life (addresses, phone etc.). The technology won't change that. It can be used for things they are not as good at though like finding people lost in the woods or chasing criminals. I mean, if they deem you a "a thought criminal" they can just knock on your door and take you to jail. They don't need a swarm of drones to achieve it.
According to my napkin math it would cost 1.5 trillion USD to kill every American with drones and I am already assuming very cheap and capable ones, that you deploy them near the target and that the target is incapable of defending itself. The cost of deployment itself isn't even included. Nuclear weapons are very cheap in comparison.
Let's put that number in context. USD 1.5T is 2/3 (or 3/4) of the total cost US spent on their forever war in Afghanistan.[0,1]
Clearly doable for a dedicated nation, given that US lost the war, lost the 20 years, and with their botched withdrawal handed the country back to the very same political power they wanted to drive out in the first place.
In comparison, spending less than the cost of a lost war and having a guaranteed result of being able to walk into an empty territory sounds like a bargain.
Humanity has repeatedly proven that it is capable of extreme, systematic cruelty without high-tech gadgets (take the Holocaust or the Rwandan genocide, for example). Why do you think that autonomous drones would make this worse?
"Now imagine an angry mob chasing down a thought criminal desperately trying to hide in the woods in order to avoid execution."
The Holocaust is a pretty bad example for "cruelty without high-tech".
Nazis had census data on millions of people reaching back decades, to sort trough that data, and search it for "undesirables", they employed computation tech that was cutting edge for its time [0], just like the methods of killing millions of people saw quite some industrialized innovation.
As such it's a rather blatant example for the whole new levels of cruelty technology can enable humans to do.
The progress also didn't just stop there, what the Gestapo and Stasi did, is child's play compared to the amount and details of data that can nowadays be trivially collected on whole population scales [1]
Majority of Jews during WW2 died outside of death/concentration camps [1]. Army, police or SS rounded up Jews in some majority-Jewish town or village, took them to a nearby forest, forced them to dig their graves and then shot everybody. Rinse and repeat. Sometimes non-Jews helped, or even did it without Germans. Sometimes they helped Jews and were killed with them. Sometimes Germans just killed everybody. Sometimes they had list of people to kill for political reason (for example university professors, priests, politicians, army officers, teachers etc. were targeted no matter their ethnicity).
Vast majority of Jews before WW2 lived in small mostly-Jewish towns and villages in eastern Poland and western USSR. It would be inefficient to move them to death camps, also you could steal from them when you murdered them. Win-win.
There was no need for technology, nazis weren't checking the papers or bothering with recording anything. There are thousands of mass graves near formerly-Jewish towns. Non-Jewish population was mostly helping to point the Jews because they were competing for food and resources during harsh occupation conditions, and besides Germans were mass-murdering non-Jews too, just as a lower priority. And if you protested you were Jewish too, right?
According to Timothy Snyder the main difference between countries where high or low percentage of Jews survived nazi occupation wasn't antisemitism but how much pre-war institutions were preserved. In countries where the nazi occupation was negotiated and the rule of law was preserved - it took months to confirm these people were Jews, move them to camps, sort them, etc.
In places like eastern Europe - where Germans just destroyed everything and law was practically non-existant - why would you bother with all of that? It was basically 4 years of the Purge. 1/6th of the population disappeared in Poland between 1939 and 1945. Jews were about 10% of pre-war population.
The involvement of IBM was shameful, but it wasn't necessary. The only technology nazis needed was firearms. And they could probably do it with knives and sticks if they really wanted - like in Rwanda. Or like in USSR or Communist China for that matter.
>>There was no need for technology, nazis weren't checking the papers or bothering with recording anything.
Most of what you said is true, except for this part. Nazis did make actual recordings(as in - videos) of exterminations conducted in eastern Poland, especially in the area that is now Ukraine, conducted both by themselves as well as the local population(threatened with death otherwise of course), all to show back in Third Reich as proof fo barbarism in the countries they were conquering. I really recommend reading Stanislaw Lem's biography on that topic, since he lived in pre and in-war Lwów.
>>The involvement of IBM was shameful, but it wasn't necessary.
I mean, that's a bit of a weird argument to make. Yes, places like Auschwitz could have exterminated as many people as they did even without IBM's help, but the whole point is that Nazis liked efficiency and the tech allowed them to keep track of what they were doing(in the concentration camps, like you said outside of them Nazis didn't bother to track much of anything)
> There was no need for technology, nazis weren't checking the papers or bothering with recording anything.
The Nazis did census data to sort out all kinds of undesirables, one of the more infamous examples being the Pink List; Since the German Kaiserreich police made lists of homosexuals and those suspected of being homosexuals.
These lists made it trough the Weimarer Republic, and once the Nazis took over the Gestapo already had most of their work done for them, because these pink lists were only one of many lists Germans liked to keep about the population.
It's one of the reasons why Germany to this day tries to be somewhat careful about large collections of personal data because after the Nazis, there then was the GDR with the Stasi, who once again took it to a whole new level.
These are relevant and well known historical examples for data collections being abused, making the Nazis out as mere "Dudes with guns" is vastly underplaying the levels of sophistication and effort that were put into "sorting trough people" on population size scales.
You are talking about 3rd Reich only part of Holocaust. It was bad, but it was a drop in the sea compared to what Germans did in occupied territories during WW2. A few orders of magnitude more victims.
I'm not talking about "3rd Reich only part of Holocaust", I'm talking about how the Gestapo was responsible for collecting and siphoning trough census data even in occupied non-German territories [0]
In that context, whatever point you are trying to make about numbers of victims in territories is kinda besides the point.
The point being that even in occupied Poland everybody had to register for the census, Polish people that were found without their registration form were shot on the spot.
In addition to that the Nazis also got Polish population data from the Deutscher Volksverband in Poland.
And it wasn't just Poland where that happened, it happened in pretty much every territory occupied by the Nazis because that's what the Gestapo was created for.
On the IBM front - IBM helped Nazi Germany run concentration camps, the number tatooed on every prisoner was an IBM-managed inventory number, they had offices and workshops near concentration camps to help run, operate and maintain the machines used by the Nazis, as much as IBM would like to forget that part of history now. So yes, Nazis were definitely using cutting edge tech for its time to help with their extermination efforts.
Holocaust was so striking to contemporaries, despite genocide being fairly common in history, at least in part because of its technological efficiency.
Technology also allows for more leverage. Hitler need an entire modern nation state with millions of people, semi-autonomous robots may give this power to a group of thousands.
Not really, Genghis and the other khans killed more than Hitler using a fairly simple technique: divide the captives by the number of your soldiers and each soldier is tasked with killing the captives assigned to them.
They usually had around 100k soldiers, armed with bows and knives, and killed tens of millions.
My point is, you don't need technology to achieve bad things and it is the wrong thing we are focusing on when trying to prevent these things. Nukes can already annihilate billions yet here we are, not being annihilated by nukes.
You could program a ethereum smart-contract virus, which hires people to scam money, produce slaughterbots, ship them to a city and strategically hit the soft underbelly aka airports, roads and railways and communication.
If you stack the attack into waves and have the slaughterbot parcels piled into flats by gig workers, you can have a ideology free terror attack that kills millions.
Such ATTAC-CELLs without members, are part of my games lore.
Here is a picture of a standardized factory, made from standardized components, making "slaughterbots" made from standardized components.
PS: The main building block is a sort of hangrenadesized fuel-container, with a fuelcell useable fuel that can be chemically primed to be a explosive. A standardized battery that could be exploded would do too..
The difference is in speed and profile . A decision can be made for automated systems to depopulate a region using low profile weapons vs say, nukes. Within 24 hours the job is complete and not a single non-victim actor had to witness the events. Compared to conventional methods which may take years, and risk the losing support, drone swarm ethnic cleansing would be cheap, quick, and decisive. Impersonal and rapid like nukes but without the implications.
I can imagine cleanup effort required trying to recover 10-20% of those drones after crashes/battery depletion, resulting in littering forest with garbage.
Lots of people do die in the woods. It's a daily thing. What is much rarer is to know that someone is dying in the woods. The scenarios whereby these drones are lifesavers are only those where a search has been launch and the lost person is still alive but in peril of death. And even with 100+ drones you still need to know the area of the person. The woods can be very big.
The more efficient method, by far, is to have a device on the person from the start. Better coordination between search and cellular providers would do much in reducing search areas, at least when looking for people carrying cellphones.
Let's don't get ourselves lost in the maze of arguments here.
The big picture is, if you give the rescue team a good tool, their success rate will then increase, sometime by a big number.
There is a documentary about children missing in the woods called Missing 411 (https://www.youtube.com/watch?v=zEA9-mEOZtA). If you take look their search efforts, it is not hard to realize how many weight can be lifted by an automatic area scanning system.
In the documentary, the search and rescue team had to walk step by step doing almost blind searches. If you have an automatic system, you can just launch a scanning array of say 50 drones, very quickly the data will come in and provide indications on where the search effort should be prioritized.
Bottom line is, such system is better than the current (human manual search) one. If the technology can be properly developed and utilized, it can help save more lives.
Except that when you are looking for kids, and to a lesser extent elderly people, they become afraid and hide. Some rescued children talk of cyclops-like giants hunting them in the dark ... rescuers with headlamps. They hide from the people trying to find them, which makes friendly dogs a particularly powerful tool as nobody can really hide from a labrador. These people will probably hide from a swarm of drones too. I would.
> Except that when you are looking for kids, and to a lesser extent elderly people, they become afraid and hide.
Then maybe let each drone carry a Teddy Bear or/and a Rainbow Pony. I mean I would totally grab the wild drone for it's micro controllers.
Now, the truth is, we don't know how people would actually react to drones. Maybe a kid would hide even deeper, but that's an edge case here. The important thing is, this tool can generate real benefit for a rescue operation that no other tool can provide. How to use it properly is the responsibility of the rescue team.
Another thing is, with information provided by drones, the search and rescue operation itself can become safer. Which also enables a load of new possibilities.
Just because it can save more lives doesn't mean it overall 'better'. Because you're saving more lives at the cost of hugely increased ecological impact.
Like EPIRBs! The tech already exists and works anywhere on the planet with no dependence on cell towers. I've decided that if I ever go hiking in the woods, I'm going to get HAM certified and carry a GPS radio with a beacon, just in case.
Those only work for people that know that they need help and are able to activate the device. They don't work for people who do not know how bad a situation they are in and/or are injured to the point that they cannot activate the device.
Maybe we can have a drone swarm that cleans up the crashed drones. You should have a fairly good idea of where they are when they crash so you wouldn't need to do a full sweep of the location.
> As many as 2,000 elegant tern eggs were abandoned on a nesting island at Bolsa Chica Ecological Reserve in Huntington Beach after a drone crashed, scaring off the would-be parents.
The same is true for any human achievement: humans have used fire to burn other people as well as stay warm and metallurgy, knifes as tools or for war etc.
Problem is still that they might not fly for more than 20 minutes, especially if you add a lot of sensory and processing power.
I have an older gen drone and they are amazing dust cleaners when you launch them in your appartment. Only slightly irratates they eyes.
But honestly I think power is still the technical barrier and the reason we don't yet have drones for package delivery. They won't fail because of sensory or software.
SAR is a statistical numbers game. Each search method provides some degree of certainty of coverage. Aircraft can't see through trees and thermographic cameras have a tough time seeing hypothermic or dead people. Ground search is still necessary even after a helicopter or fixed wing aircraft have gone over a search zone.
A swarm of drones would allow a rapid search of an area with similar coverage statistics as a manned ground search. But a drone going 40 km/h could miss subtle clues like footprints or freshly broken branches.
The amount of energy, tech and brain power we're spending on ever lower impact / less rewarding topics is scary. It's equally dystopian to me to be honest
What I really want is a drone with a robust recovery system. I want it to be able to hit tree limbs, fall to the ground, but then right itself and take off again.
Dystopian movies have obviously gotten it wrong. The attack drones won't be slowly scanning the forest for escapees above the canopy but will instead rip through the forest in large packs at 40km/hr. How uber-dystopian.
The future is not Star Wars style fights where lumbering ships can't seem to reliably hit each other despite clear lines of sight. The future is fast, precise, sterile and not something a human can react to. The future is drones with knives slicing your jugular before you know they are coming.
I'm an embedded SW programmer and the speed of computation still boggles my mind sometimes.
The thing those movies get wrong is the moments of suspense when the slaughterbot is chasing it's victim. They'll be dead before they even contemplate running.
It’s not inconceivable that they’ll follow their victims for a bit to make sure that the face recognition doesn’t have a glitch.
Apparently you can also recognize someone by the way they walk, so if you’ve recorded that before, getting them to run around for a bit provides an extra level of identification.
If you're a nation state maybe. The question here is how far down the cost of drones like these will go, and how easy it will be to contain the technology.
I remember the video with the swarm of tiny drones dropped from the airplane. It's very unrealistic in my opinion.
3 types of problems:
1) quadcopters battery is not great (yet)
2) AI is very specialized
3) You underestimate how fast and creative a human being is when survival is at stake.
1) quadcopters battery is not great (yet)
Quadcopters must balance weight, energy and payload.
Essentially drones don't fly very long and any payload you add will increase their weight and decrease their speed or flying time.
In the video the tiny drones are released from an airplane.
Already not sure if that would work, what about wind and altitude? The drones might spend a significant part of their battery just to reach the ground. They might be so light that a gentle breeze would swipe them hundreds of meters away.
Even if battery will be better there will always be a constraint on what it's possible to do and how much it cost to build it. In the video the announcer says 25 million USD for enough drones to wipe a city. Let's say there are 5 million in the city and half is 2.5 million people. Your budget than is 10USD per drone. Clearly that budget is just too small.
2) AI is very specialized
Just because an AI can beat the world champion at chess doesn't mean it can do anything else at all. Take the same AI and change the chess board size or another fixed parameter of chess. The AI will likely not work at all without being retrained while a human being will still manage to figure it out on the spot.
3) Underestimating humans
Alright somehow the tiny drones with explosives reach the room. It's not like in nature we don't have insects and birds flying at us. People have reflexes to fly or fight. Most people would duck or protect their face with hands. If you think the drone will be too fast remember it's a 10USD drone dropped from an airplane. Frankly it's a miracle even if manages to land on the ground.
It could but it's quite not sure it will.
In 2001 we were supposed to be doing a space odyssey on Mars and beyond.
And 5 years ago we were supposed already to have full autonomous vehicles.
Just because something looks close in sci-fi made-up movies doesn't mean it will turn reality. It's about the tech.
>How about you sign up for the "escape the AI slaughterbots challenge"
You are making a very different scenario. You are suggesting in 2030 will be possible to kill one person in a setup environment with very high budget for 1000 drones.
The scenario of the movie is to kill 2.5 million people with 2.5 million tiny drones with 25 million USD budget over a very large and not controlled environment like a city.
Depending on the capabilities of the drones that could be absurdly trivial. GPS jamming scales to an infinite number of drones.
For medium capability drones your budget is already in the millions so it is only fair to use a medium capability air defense system: https://youtu.be/1DXpPmpmcak
Edit: Before you say that I don't get to defend myself according to the threat level then a single person beating me to death would be enough to kill me. Not very interesting.
It’s naive to assume military research labs don’t have this in an advanced state of development right now, or even combat-ready weapons that just haven’t been used yet.
I mean, it’s obviously possible. That means militaries have at the very least developed it to a proof of concept by now. This kind of tech often exists for many years before becoming publically known.
Yes probably the US and Russia and China militaries have some advanced quadcopters. Do you think even to build one of them costed less 10 USD? You have to count in all the research and development put in.
Dropping does not require energy. A drone-mothership will obviously take wind in account when dropping the dronelets. The video was not realistic in that regard because the drones would probably be folded for initial descent.
I find USD 10 is not unrealistic. Already now a device with battery, chip, and motor can be manufactured for under that. It will get cheaper.
2) AI is very specialized
We've just seen how a drone navigates autonomously at high speed without crashing. They don't need much more to be killers.
3) Underestimating humans
You have to be prepared to react well to a drone attack like that. These things will not wait around for you to figure it out.
>Dropping does not require energy. A drone-mothership will obviously take wind in account when dropping the dronelets. The video was not realistic in that regard because the drones would probably be folded for initial descent.
That already exists. It's called GBU-39/B Small Diameter Bomb (SDB) and it has been obsoleted already by the GBU-53/B StormBreaker. Yet nobody is freaking out.
And yet last week a simple light show with quad-copters didn't manage to go as planned:
https://www.stuff.co.nz/travel/travel-troubles/126582371/its...
Now I don't think those drones costed 10USD each, they probably costed more and even having them doing a light show is very hard.
It's the same in the Culture series. Near-God AI motherships manufacture combat psychopath AI ships by the thousand, which fight lower tech ships by overwritting their CPU caches with EM effectors (very advanced versions of phased array emitters) in milliseconds, taking them over for their side. Rarely it gets to anti-matter and particle beams.
Or Stanislaw Lem from 1964's "The Invincible" where autonomous drone swarms take out all life and complex machinery. In the story, they evolve, but a lot of the swarm behavior is similar to killer bot drone swarms like from the clip above.
Yes, I've been saying this for years. Everyone seems to overlook the massive speed advantage they could have over us. Even if they're not trying to kill us, aside from for interacting with humans, there is no reason that they would be constrained to operating at human compatible speeds. It's possible that we could see them as operating on a completely different time scale than us, similar to how we see fungus.
Elon Musk got mocked when he talked about robots moving so fast you'd need a strobe light to see them but typically for Musk he was just a bit too early in his timeline. Definitely the way things are going; why wouldn't they?
In the future, the bad guys never miss. This is what 90% of pop sci-fi gets wrong.
Speaking of fast, the Russians now have hypersonic missiles that go roughly 14,000 miles per hour. That's more than 3 miles a second! If you have a missile defense on board a ship, that means the missile defense has to detect, target and kill that missile in just a few seconds maximum.
Yep, we might face a world where gun drones have perfect accuracy from hundreds of meters away as well as split second kill on sight decision making. You'd be dead before you'd see it coming.
While I have the same feeling as you, let's not forget that those things don't have unlimited range and are vulnerable to simple tools like a shotgun.
Anyway if they have infrared cameras, they'd be more efficient by scanning the forest high in the sky and if they just rely on normal cameras, good old camouflage still works.
I could be wrong but IMHO we're entering a period of rarefaction of resources. I don't think in the future they will waste precious resources like energy and drones to hunt common people, it will be reserved for high profile cases.
And then, once finding the target, an explosive charge blows up in the drone to fire a single high speed projectile straight to the target, punching a hole clean through the forehead for a quick and humane death. You know you’re finished once you’ve locked eyes with a drone, maybe next time don’t break the law.
Keep in mind that Skydio currently manufactures and sells affordable drones ($1000 USD) that can fly autonomously through pretty much any environment, including forests.
Disclaimer: I don't work for them nor affiliated - just been following their progress from a few years back and can't wait until they start selling to consumers in Australia.
I watched the second video and was a bit disappointed. It lost him twice in 6 minutes even though he didn’t leave its frame. Maybe version 3 will crack it
One thing that stood out to me is that they train it entirely on simulated data and deploy it without needing any fine-tuning on real-world data.
I took one of Andrew Ng's deep learning courses a few years ago and remember him saying that this approach typically doesn't work. It sounds like that's no longer the case which is really cool because it's relatively cheap/easy to generate simulated data.
Is this a more common approach these days? Are there any other interesting examples to point to?
This depends almost entirely on the accuracy of the simulated environment relative to the real world. If we have very accurate simulators then sim2real isn't too bad. If we have poor simulators (or simulators which don't fully capture that <0.1% probability action that happens to have REALLY bad consequences...) then sim2real doesn't work.
Flying a drone through a static forrest in no-wind (no moving branches, etc.) seems like a reasonable sim2real environment, in part because... well... video games have had forrests and trees don't tend to move very fast unless there's wind.
But sim2real for self driving cars is a lot harder given the huge diversity in streets and freak events that happen (people, bicycles, shopping carts, piano-delivery-trucks)
I'll add that real world sensors (including cameras) are noisy in ways that are difficult to model 100% accurately in simulation. Same thing with aerodynamic effects, especially when the drone gets close to the ground or other objects, or y'know...it's windy outside. Dealing with these types of noise and unpredictably and building in resilience to them is a big part of testing and validating systems like these, and can only happen in the real world.
I'm pretty confident the drone linked in this article doesn't have high reliability. I'd be shocked if they managed to get >70% critical without significant real-world iteration.
I think it basically helps that quadcopters are already computer stabilized and highly manoeuvrable so the problem is more identifying spaces to move into and motion planning through them. What’s impressive since the last time this kind of thing was posted is the increase in speed and at least in the provided video the ability to avoid really quite small branches.
It looks from the video someone linked to above that these things don't use cameras, just depth sensors -- maybe they're easier to simulate? It looks like they do introduce noise into their simulated depth-sensor data, but maybe that's more straightforward than creating realistic visual noise.
The simulation is a bit of a red herring, in this case. What operated in
simulation was not the neural net but a different algorithm. The article calls
it a "simulated expert" and my guess is that it was most likely a classical
planner. This "simulated expert" was given all the information and, crucially,
all the _time_ it needed to fly through a simulated environment (or possibly
many more than one). The neural net was then trained on examples of the
"simulated expert's" behaviour. Then the trained model was loaded onto the drone
and was used to pilot the drone, essentially predicting the path the "simulated
expert" would have taken in a similar situation.
In short the "trick" is that they generated a plan offline and then learned to
approximate it to avoid having to re-generate it for every flight, which is very
time-consuming and wouldn't allow fast flight.
It's clever and as the authors are quoted to say in the article, it doesn't rely
on a very precise simulation, since the neural net doesn't have to be trained in
a simulator and a classical planner doesn't need very granular detail (it's just
slow). On the other hand, it's difficult to know how well it works in practice
because "flying through a forest" is not really an objective measure, insofar as
there is no ISO standard for what "forest" means. Sparse forest? Dense forest?
Conifers? Other...? The news item above says that the approach was tested on
"previously unseen" environments, but that, too, is a very fuzzy expression. So
while this is very impressive news (without having read the original work) keep
an open mind about whether this will really work in arbitrary environments in
the real world.
The sim2real approach can work pretty well as long as you are very in tune with where your simulator falls short relative to the real world and take steps to circumvent those shortcomings.
We were able to train the robot to climb stairs completely by feel/proprioception without any sort of vision. We trained it in simulation, and then transferred it to the real world without issue.
Yup! There are a ton of examples of this in robot learning. For instance: https://arxiv.org/abs/1809.10790 by NVIDIA uses photorealistic images and something we call domain randomization to efficiently bridge the sim2real gap for robot grasping.
A core issue seems to be that the better the simulation, the harder it is to generate (compute wise) but simple tweaks to the generative process also go a long way.
This is reminiscent of intense wingsuit videos. While the drones are small and the speeds are not quite there yet, being able to change course on a whim may prove sufficiently appealing. I wonder if zipping through obstacle areas could become a notable application area.
There's so-called "nuclear breakout time" for how quickly a non-nuclear nation could create nuclear weapons.[1]
Time for a similar "slaughterbot breakout time."[2] However unlike with nuclear weapons, this is something a small team or even individual could soon achieve and then hold as a deterrent or use for sociopolitical aims.[3]
It's easy to imagine an arms race between drones and the tiny point defense turrets (likely lasers, maybe kinetics or even EMPs) trying to disable them, both systems operating far faster than humans can react. In a strange way, wars fought like this may involve less loss of life, with soldiers surrendering to a cloud of slaughterbots that destroyed the automated defenses in seconds.
This forthcoming ability to intimidate and harm precisely, anonymously, robotically, and at scale makes autocracy more likely and more durable.
[1]: https://en.wikipedia.org/wiki/Nuclear_proliferation#Breakout_capability
[2]: "Slaughterbots TED talk." https://youtu.be/9CO6M2HsoIA
[3]: "Hated in the Nation" Black Mirror Season 3 Episode 6 (Netflix)
The problem with the Slaughterbots narrative is it's not actionable.
We can choose not to build nuclear weapons: they're weapons with lots of weapons specific research associated.
But drones aren't. The technology is general use: it's a natural consequence of a couple hundred incremental improvements, and the only difference between "weapon" and harmless is a software parameter and a payload.
"Find and kill this person" is indistinguishable to a search and rescue AI, just with "on found and centre camera" triggering a bullet or a spike or a bomb and not text notification.
> Guns are also "easy" to manufacture, but are successfully restricted
Hahaha. This is a privileged take. 3d printing guns has been a thing for years now, and the FGC-9 (Fuck Gun Control 9mm) is a carbine that uses zero restricted firearms parts.
Gun control is in serious danger of becoming completely obsolete.
Edit: I'd like to add that the FGC-9 is also durable, none of that single-shot disposable .22 pistol stuff.
The Halle shooter built his own weapons out of metal but he didn't have access to real ammunition. He couldn't forcibly open the gate to the synagogue with his guns. Gun control works just fine.
The recently departed Jstark actually commented on the ammunition. Smokeless powder isn't hard to make, he simply didn't know how. The information has since been released.
On an international scale guns aren't restricted at all in any practical sense.
Nuclear weapons can have international conventions and treaties because they're difficult and obvious to build. Mass producing cheap drones and downloading an AI from the internet to run on a repurposed Android phone is not only normal, it's already happening now.
You also don't need a proper gun. A gun is a big advantage if you can't safely get close enough to someone without them getting a chance to react.
But if you can fly a drone or a number of drones up to someone too fast for them to react, there are any number of other ways to kill them with easily obtainable materials.
Totally agree. Each drone already has 4 sharp blades. Easy to replace them to be slightly bigger, sharper and from e.g. titanium instead of plastic and covering with poison/venom.
Enough to slit someone throat/wrist and wait until bleed to death or detect someone eye and stab with nail.
Wow, sharp blades now that sounds dangerous. If your drone is basically a flying lawnmower I can see how that will kill or heavily insure hundreds of people with a single drone.
But that isn't very surprising though. An Apache helicopter can kill hundreds of people as well...
What's surprising to me is that someone suggests that we do away with the toy drone swarm idea and to increase the capabilities of the drone without requiring an additional payload.
> wars fought like this may involve less loss of life
Western audiences forget why wars were often fought to the last man. Mass murder could often be the merciful option. Vae victis.
A military superpower unwilling to impose its will through sheer brutality is a huge historical outlier. As Pax Americana disintegrates, new powers will arise in its wake. ISIS was a brief glimpse into the future.
The trick is going to be improved friend/foe detection. Historically, wars simply ended when the victor would kill all remaining enemy soldiers and enslaved their families and other people inside the areas they were defending. Brutal but effective.
The last century since basically WWII and especially the Vietnam war, we've seen increasingly less brutal wars where harming innocents is seen as a negative (in the case of Vietnam, that was what eroded the will to fight). Lately, we're seeing a type of asynchronous war fare where huge armies are effectively powerless because an insignificant opponent is hiding among civilians delivering very minor attacks with no strategic relevance. There are a whole range of stale mate situations across especially the middle east where the attacking side is simply hard to root out without lots of undesirable casualties among civilians. Just ask the Israeli's. They live tens of kilometers away from people willing to die to risk launching rockets at them, despite the fact these rockets rarely hit anything. It's been the status quo for the last decades. Their military is big enough that it keeps foreign countries from trying to invade but domestically, it's clearly not able to end the conflict.
Stalin would have ended that conflict in no time. There would be no-one remaining to figut, the local language would be Russian, and there might be a new Palestine presence somewhere deep in Siberia. His successor Putin has fared less well in Chechnya mainly because he would suffer internationally if he did it the Stalin way.
Guerilla warfare is dependent on the other side being humane enough or unwilling to not want to cause unnecessary casualties. As long as attackers manage to hide, they can't be hunted down without large amounts of innocents dying. WWII style leveling of entire cities is very much frowned upon now. I live in such a city (Berlin).
However, the ability for guerilla fighters to hide after an attack is time sensitive and critical to their survival. Drones swarming an area would make that a lot harder. Urban warfare particularly is dreaded by armies because of high casualty rates related to simply not knowing where the enemy is and the need to fight door to door to find them. This could fix that. It could be over in minutes. It won't end all wars but it will make particularly guerilla warfare a lot harder.
ISIS could be a thing of the past. The next engagements with the likes of those might be short, brutal, and effective. Minimum (but not less) loss of life, maximum effectiveness. It would follow months of covert mapping the territory, inventorying its inhabitants, and tracking all their movements followed by a few minutes of mayhem. Some people find that scary but they forget what decades long of conflict or a real war looks like.
IDK. First, I don't believe there will be any kind of surgical precision with autonomous weapons systems. Human drone operators have taken out way too many weddings for my liking, and the German army obliterated a lot of people gathering around a stuck tanker.
Second, with a delay of say 10 years, swarm autonomous drones will be in the hand of insurgents. We've seen how ISIS took out tanks and troop by dropping tank shells from drones. They are not dumb and will happily adopt drone swarms. The science and IT is out there.
It is true that armies have a harder time carpet-bombing cities, but at the same time, what modern armies hate even more is losing their own guys. Even a semi-capable killer drone swarm could make it real hard to take and hold cities if seemingly out of the blue your buddies drop dead. Insurgents typically will not care that much if there's dead civilians if they can terrorize invasion forces.
I would take my chances against a slaughterbot everytime compared to more mundane things like mustard gas.
A slaughterbot obviously needs melee weapons and needs to be athletic. Ballistic platforms are probably far more effective, but it doesn't send the right message.
1) current small racing drones have speed >80mph, that's 35 meters per 1 second. You won't even have time to hear them and react while under attack.
2) current drones already have 4 tiny knives (blades). Enough to make it slightly larger/sharper and/or from titanium instead of plastic. That's enough to cut your throat or wrist and let you bleed to death ...
3) ... cover those blades with some venom or poison for "extra efficiency"
4) ... they probably don't even need some bullet - enough to detect your eye and stab you with a nail covered with venom/poison
I doubt you would have any chance with just a dozen of slaughterbots.
Probably not, but I also have noch chance against mustard gas and bleeding out seems at least preferable.
Drones are extremely loud, especially those that fly that fast and small. If you have finely mashed net or large piece of cloth, drone already get into trouble.
Not sure if it's a fair comparison when the links you posted do not include any real evidence of flight agility or video of the actual drone. It's probably because of classified nature of this project, but still we don't know how similar these capabilities are.
This is great. About ~6 years ago there was a story on HN from MIT titled "Self-flying drone dips, darts and dives through trees at 30 mph"[0]. It was a fixed wing drone, rather cautiously flying near a few trees. At the time my comment was:
In Romania, Rangers and activists are often threatened or even beaten by wood thefts. Drones can be a highly effective method of stopping deforestation.
People get overconfident with self-driving cars because they consider the problem as navigating from A to B while driving along the road within the law. That's the easy par The hard part is the long tail of human behaviour and the unexpected you have to deal with.
Just consider an object on the road. Is it roadkill? Trash? Will it get out of the way? Is it a person under a blanket? In some places it may well be a person to get you to stop so they can rob you or worse.
It's said that humans are motivated by fear and greed when it comes down to it. Let's talk about fear.
A lot of what keeps people inline is implicitly or explicitly fear of what people might do. They're an unknown. It's why a burglar may not care about an alarm but a dog is a deterrent. Will that dog bite? Fear of those unknowns makes the burglar look for easier prey. Not always but statistically often. It's one reason why people get dogs.
It's the whole basis for having doormen and the like in buildings over automated systems. That doorman is a much bigger unknown and can react in ways machines can't.
It's the same on the road. There's only so rude or inconsiderate many drivers will be because they fear the consequences of road rage and such.
So what happens when such a driver sees a car that is self-driving? It is absolutely naive to assume that drive won't change their behaviour.
When flying through a forest, you're basically just dodging trees. Wildlife will tend to avoid you. Just identifying larger wildlife and steering clear of them is likely sufficient. Driving is just much more difficult.
I'd say self-driving cars in the sense where the car isn't even designed to have a human driver anymore is at least 20+ years away (and, much like AGI, probably will remain that way for some time).
Instead you'll have what we already have to some degree: driver assistance for the simpler parts of driving like staying in a lane on a highway and avoiding hitting the car in front of you. That's the easy part. Over time, more and more of those features will be automated. But it's still going to be a giant leap to autonomous driving.
Almost static scene.
Drones are also not trying to avoid each others, I'd be curious to see them trying to fly as a dense flock in the forest, or at human height in a crowd.
These drones need to be able to do one thing:
- do not hit the tree (or ground)
And the things not to not hit are basically static. Year, Wind might move parts of them here and there, but you can just try to make sure to avoid them far enough. And we do not see if they can do event that.
Self driving cars on the other hand...
- avoid static objects (like trees, or houses)
- stay in lane
- detect and act on traffic signs and lights
- detect other moving object and predict where they will be in the near future
- watch for suddenly appearing lifeforms all around you, do not hit them!
- if you leave your lane, do so orderly
- do all that at speeds far over 40km/h
I think the "static objects" parts is not the hardest part on that list.RADAR/LIDAR are good at that. I imagine the "everything and everyone is moving in random directions" part makes things harder.
Maybe this approach will help, as suggested in the article:
The applications are not limited to quadrotors. The researchers explain that the same approach could be useful for improving the performance of autonomous cars ...
the drones are doing 3D from stereo. My understanding those autonomous cars we see around SV don't do it (except for some Ford sedans - probably Ford venture - who have what looks like stereo for the front view). Google and some others build 3D from lidar - it is almost as good as stereo though much less resolution. The low lidar resolution is close to being good enough for the street driving while it wouldn't work for the forest at meaningful speeds (even if somebody managed to place a lidar on drone)
Don’t be distracted by the apparent application. This is academic work. The achievement is the state estimation and control sufficient to fly fast around a semi-structured environment without a map.
No one really cares very much about flying fast through trees, though it’s cool as hell. The point is that they are pushing the performance envelope so that more applications fall inside it.
Davide’s team has made great contributions in sensing and controls for UAVs over the years. Great stuff.
Not that I don't also find this dystopian, but thermal cameras have been a thing for years and they're already pretty good at catching people who flee into the woods.
You also have many animals in the woods that will show up on thermal camera. It's a different level of scale, sending humans with thermal cameras, walking slowly and identifying if someone is human or animal versus dozen drones doing the same things autonomously and much faster than human.
It isn't clear why it needed AI. If there is enough compute power to run a neural net, surely there is enough computer power to parse data from sonar or compute near-term avoidance vectors from bifocal imagery parallax/depth-queuing. Not clear how the AI obviates this.
It's not about AI. They trained a neural net to approximate an offline
calculation that required more time and processing power than what's available
on a drone, then they used the trained neural net to navigate the drone by
trying to predict the approximated calculation. Thus, saving on time and
processing power and allow the neural net to pilot the drone at the vertiginous
speeds of 40 km/h :P
(but through trees and all)
The article calls the algorithm that performed the offline calculation a
"simulated expert", probably because it operated in a simulation. My guess is it
was a classical planner. Planning works very well but it can be slow and
deliberate because it usually has to solve some kind of NP-complete SAT problem
and so on. By approximating it with a neural net they avoid the time-consuming,
intractable calculation. The question is how broadly it can generalise.
Ah. I see. It's using a neural net in the sense that it is literally a "generic function approximation" to make decisions on pathfinding based on depth maps (see 'johnthewise' response below). That's interesting: it foregoes the complex pathfinding algorithm for a trained "guess" based on how expert drone pilots handle the same stereoscopic depth map.
I understand this drone's obstacle avoidance as it sensing the terrain ahead, feeding it into a NN and deciding what the least-dangerous route to take is.
No doubt that score can be calculated with an explicit algorithm that says "score a tree as -100 and an empty space as +10" etc.
I liken it to a game developer designing an NPC's choice of path. Sure, Doom had enemies attacking you without neural networks, but maybe it's a better way to do it today.
(this is all just my understanding rather than actually having worked on this)
The thing is, I've worked extensively with object detection and segmentation neural nets for the past three years: they would barf all over the place in this scenario. You can't train for enough of the types of terrain obstacles encountered in such a wide-open range of outdoor spaces, and you certainly couldn't do it fast enough with any degree of accuracy, even with a 3080 GPU, let alone an embedded Xavier-class NVidia module. A dinky int8 quantized "tree/no-tree" wouldn't stand a chance.
This is why I asked for more details. They seem to indicate training, which today means neural net and not expert systems of yore. But that one be one hell of an object detector!! My question still stands.
Well, my understanding is that they are not training object detectors or segmentation models. They wouldn't be very useful anyway, you still need to have a 3d understanding of the scene and going from 2d->3d mapping wouldn't cut it. What they are doing instead is using stereo cameras, they are estimating the depth map of the field. Their model takes two rgb images, and produces a depth map. They combine this depth information with sensors on the drone such as acceleration etc and try to predict what an expert drone agent trained on perfect information in simulation would do.
They train in simulation, restrict some of the information to the student agent and have it rely on only stereo cameras and sensors like in the world to mimic the 'privileged drone agent'. Computation wise, if you can run the depth estimation network on the hardware, remaining steps(given the depth map and sensor information, predict privileged drone agent path/vector), should be trivial and most likely are shallow networks.
While reading the comments I shared most oft the dystopian visions of people being hunted down by scary drone swarms.
But armor (especially armored vehicles) is still a thing and for something to be able to be fast and manouverable it also needs to be designed within size and weight constraints, which automatically limits possible payload and armor.
So if there ia a technology to race around a forrest and autonomously find and engage slower targets, there must also exist the technology to engage flying objects with something like the Phalanx system, while not facing the same weight constraints as the moving attacker.
It probably boils down to target saturation in symmetrical warfare and the end of assymetrical warfare fantasys akin to Wolverines.
Armored vehicles are obsolete. See Azerbaidjan campaign in Nagorno Karabakh. More than 200 tanks & armored vehicles obliterated over a few weeks by drones raining hell from above.
The assumption behind your question is that the victors somehow owe something to the defeated, akin to US occupation of Germany and Japan post WW2. US as a superpower is a huge outlier. Most of history developed along the lines of the Trail of Tears, if not downright Tamerlane.
Use drones to create an air kill zone and keep the enemy military from stepping in said territory. Expel potentially hostile civilians. Reward your cousins with free land in the freshly depopulated lands.
The challenge will be cost. If you hide your expensive slaughter bots in a swarm of cheap toys that only shares an outward appearance, your opponent need to take out all of your drones, while you only need to get lucky once.
Couple that with good enough autonomous navigation and imagine an adversary suddenly releasing thousands of drones in between people and traffic in a city where they don't care about collateral damage, but you do.
I'm an American and radically pro-gun. I'm also radically pro-responsibility. If we want to renormalize firearm ownership (As in, you can mention it in Whole Foods without getting dirty looks), we need to reshape our culture of safety and ownership of our own actions.
This feeds into my dystopian believes. I a couple of years those drones will be affordable to every individual. This individual will then savely engage in a drone strike against other individuals. Hence I predict an end of the celebrity cult :)
You know, I'm honestly amazed we haven't seen any serious terror attacks use drones yet. I guess there is still a bit of a barrier in terms of cost and technical know-how for your average mass-murderer, but there are already drones which could do a huge amount of damage in an open space if equipped with the right weaponry.
There are people [0] who claim that air traffic will be the first target of this development. It is very fragile, has a huge attack surface and efficient counter strategies are non existent. I am really amazed why "nothing" we know of has happened yet.
How long before a small explosive is strapped to the tip and aimed at humans with low social credit score? Would make for a Running Man sequel, with some future Caligula.
Yes, but the innovation could be individual precise targeting. You could potentially keep a battery of these drones at every intersection. Track every human using phones, then deploy at will, at some threshold event, like a failed covid test, or mean post on fb against the dear leader.
I'm not sure if you've seen the original Mortal Combat movie. There is a scene if I recall correctly where he's running away through a forest chased by such a knife.
"Our work replaces the traditional components of sensing, mapping, and planning with a single function that is represented by a neural network. This increases the system’s robustness against sensor noise and reduces the processing latency."
I get the feeling a bit that when they compared their depth pixel in => control out NN with the other two baseline algos , a lot of the difference is attributed to latency (one of the other algos did more temporal averaging explicitly for example). The depth input already provides a pretty good source of data for planning.
The NN doesn't do anything you can't do already with the other algos, but it gets tuned automatically to a very good compromise between latency and performance, something that would be very tedious work for the designed algorithm systems and this is a strong point for building stuff like this as NNs.
People get lost in the woods all the time. Image a swarm of 100 of these drones, 200, 500, all flying at 40km/h in a coordinated sweep, searching with multi-spectrum cameras and image recognition.