Autonomous drone navigates independently through the forest at 40 km/h
media.uzh.ch
media.uzh.ch
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.
Somehow the idea of autonomous winged cameras flying around all over the place does not make me feel safe.
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.
Did you forget about this[1]?
https://www.theregister.com/2021/03/31/wifi_devices_monitori...
The more you look into it the more messed up it gets. Complete with a Huawei guy chairing the group developing it.
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.
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.
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.
https://en.wikipedia.org/wiki/Emergency_position-indicating_...
"Mr Burns, we've lost 20% of our drones"
"Well then... Send more drones!"
But honestly, if the drones are operating as a swarm then a downed drone could report it's last known location to it's peers for easier recovery.
https://www.latimes.com/california/story/2021-06-07/thousand...
> 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.
Now imagine those machines chasing down a thought criminal desperately trying to hide in the woods in order to avoid execution.
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.
0: https://theconversation.com/calculating-the-costs-of-the-afg...
"Now imagine an angry mob chasing down a thought criminal desperately trying to hide in the woods in order to avoid execution."
This isn't any less frightning, in my opinion.
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.
To control them you need one- or even none.
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.
https://springrts.com/phpbb/download/file.php?mode=view&id=1...
https://springrts.com/phpbb/download/file.php?mode=view&id=1...
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..
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.
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]
[0] https://en.wikipedia.org/wiki/IBM_and_the_Holocaust
[1] https://opendatacity.github.io/stasi-vs-nsa/english.html
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.
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)
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.
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.
https://books.google.co.uk/books?id=ICISEAAAQBAJ&pg=PT298&lp...
If a G7 state were to change in that way again, they would almost certainly achieve their eradication goals
Because it reduces the costs of doing it.
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.
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.
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.
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.
https://m.youtube.com/watch?v=HipTO_7mUOw
With this:
Camps were for undesirables that for whatever reason couldn't be killed immediately. Mostly because there were too many of them in one place.
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!
> - don't let them know where exactly
The solution is almost never technology, it's usually politics and legislation.
It's much easier not to ban something than to ban it, especially when the state is hostile.
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.
But people are still lost in the woods.
Or in other words, should we slow or stop technological advances because they may get into the wrong hands?
And, for part 2, who decides what's wrong?
2: Everyone. It's sort of a planetary wide, group decision.
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.
autonomous drones can clearly be very useful.
With good equipment you can scan large areas using only one helicopter.
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 trick is to get out soon enough to find people alive.
That said, if hypothermia has gotten so far that one is hard to spot on thermal camera I am afraid there isn't much hope left anyway.
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.
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.
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.
How about you sign up for the "escape the AI slaughterbots challenge" in 2030. Win $10,000 if you survive the 1,000 Slaughterbots trying to kill you.
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.
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 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.
> 1) quadcopters battery is not great
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.
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.
Ohh and even a nice fictional animation: https://youtu.be/n8u0l_c2_NA
From the same Iain Banks, "Against a Dark Background" features space battles where rules of engagement ban the use of AI-controlled weapons.
...or engage in wrongthink. The Supreme Commander must never be questioned even in private, peasant.
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.
In reality, automated killing happens without the target receiving previous warning.
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.
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.
[1]: https://www.youtube.com/results?search_query=skydio+2+forest
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?
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)
- Twitter: @iamtrask
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.
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.
Here's some work my lab did on sim2real for a roughly human-scale bipedal robot (Cassie): https://www.youtube.com/watch?v=MPhEmC6b6XU
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.
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.
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)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.
The clear outcome of this is the current ptb enslaving the world without needing so much as a police force.
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.
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.
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.
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.
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 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.
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.
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.
Slaughterbots could carry a small explosive, just like in the video. Doesn't have to be melee.
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.
edit: formatting
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.
https://shield.ai/products/nova-class
https://www.wired.com/story/shield-ai-quadcopter-military-dr...
Newer models (Nova 2+) are much faster and explore in multi-level buildings.
> I guess I was expecting something more like this goshawk flying through a forest, but maybe it's too much to ask for :) https://www.youtube.com/watch?v=p-_RHRAzUHM&start=100
Now 6 years later, the drone seems almost on par with the capabilities of the goshawk. Pretty amazing!
(Though that POV shot in the forest looked straight out of Evil Dead II.)
Also, the cost of mistake is vastly different. Drone crashing into a tree? $1000 mistake. Car crashing into a group of schoolchildren?
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.
The applications are not limited to quadrotors. The researchers explain that the same approach could be useful for improving the performance of autonomous cars ...
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.
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.
Fleeing into the woods will no longer be an option...
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)
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.
(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.
Fascinating!
On the other hand, I really don't these things in the woods for anything other than emergencies.
We humans are noisy, and our toys are no small part of that. Guess that's what's growing old is like. Complaining about noise and other people
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.
As an American, I can guarantee you the latter is not required for the former.
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.
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.
https://en.wikipedia.org/wiki/Trail_of_Tears
https://en.wikipedia.org/wiki/Timur
https://en.wikipedia.org/wiki/2020_Nagorno-Karabakh_war#Retu...
This is called a shotgun wielded by the best AI ever devised: a human.
The Talibans proved that low tech isn't dead.
(Missiles are the original drones. Cruise missile = single use drone.)
1) Make a simulation of the world for monte carlo
2) Do complex processing for obstacle avoidance
3) Make the drones “learn” from only the internal states what to do
And voila, you teach a new robot old tricks
my guess is that politicians will get scared when it becomes a reality
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.
Article is at: http://rpg.ifi.uzh.ch/docs/Loquercio21_Science.pdf
Wanna bet where the money comes from?
Great to see good tech deployed in the wild, but I'm not terribly impressed by the machine learning side of this.