Fleets of drones could aid searches for lost hikers
news.mit.edu
news.mit.edu
We seem to be running headfirst into the autonomous killer bot future without giving so much as a second look at the implications.
Slaughterbots [0] is a pretty good look at what can be achieved by refining and integrating technologies which all exist today.
Morality has kept nuclear weapons in check (so far anyways), I’m sure it will keep AI drones in check too.
Morality and logistics. The major nuclear powers have implemented a relatively successful program to make it very difficult for a non-nuclear power to develop nuclear weapons.
> I’m sure it will keep AI drones in check too.
There is simply no way to use logistics to curtail AI drones in the same manner as nuclear weapons; all that is needed to make a drone is a 3d printer, common industrial components and software.
That said, some cities are already so messed up, that cops won't respond to 911 calls in some areas. I would expect that to get worse should there ever be "thugs with drones".
This could only happen in a very distant future. As it stands, firing a gun from a 3D printed drone would make it fall out of the sky. Even if it didn’t, it would be laughably inaccurate. Even still, they could be easily defeated with a birdshot.
[Edit] Did I just create a DarkWeb business model?
Of course there is: you could attack only battlefield targets, or only strategic military targets, or only one of the enemy's cities. MAD requires two things: first, the technology for mass destruction, second, the geopolitical conditions where either side can credibly threaten complete annihilation if provoked. Nuclear weapons provide the first by themselves, but the second only incidentally: the nuclear powers have come to a credible mutual understanding that they won't use nukes in their various border skirmishes, and that any use of nukes represents an escalation which risks MAD.
World leaders could hypothetically come to the same understanding regarding drones: use them anywhere in the world and you will start a nuclear war.
Yes, except for one big difference that I see: individuals can't go purchasing a nuke like they can a drone. So while the MAD concept works with nukes, because only nation states have them, the same isn't true for drones, where individuals can have an unlimited number, in addition to governments. World leaders can agree re: drones and attacking each other, but that won't stop an intent individual. I think people might be quite surprised what you can do with a consumer grade drone purchased for a few hundred bucks hooked up to lidar, multiple gps sources, swarm technology so they communicate with each other and object recognition. Now just rig up an explosive device, chemical weapons or a gun. You can even drop items (bombs), or pick items up using retractable claws. You can do an awful lot with a consumer/hobby-grade drone hooked up to some cheap sensors using something even as basic as an arduino with open-source software like ArduPilot, which is so advanced (and cool) even Nasa uses it.
More people thinking about what they're doing wouldn't be "stopping progress", it would be progress.
Further, what pisses me off, is how idealistic-yet-naive-and-downright-corruptible many very smart people can be.
Take for example the genius nerd who is empowered and elated at the acceptance and praise they might receive for being able to hack a system which ultimately results in the death of anonymous-to-them-other-human.
This happens in every culture.
Mental exploitation is just as insidious as sexual or physical exploitation.
This is where there is one fundamental thing lacking in all of technology, aside from Asimov: What are our first principles?
A freaking POST-BIOS requirement for any system might ultimately have to be a checkin with a prime-directive which tells it that regardless of whatever code you are running, you can't do XYZ.
We are all guilty of "wouldn't it be cool if" syndrome in tech.
What you can do in 'cool' video games, is not 100% mappable to meat-space.
What are you referring to? The Three Laws? AFAIR they exist in order to explore how dumb the idea of them is in the first place. Such high-level rules don't map to reality at all.
A ruleset correctly encompassing human moral values is a complex and hard problem. It's so hard that we haven't even figured out the proper formalisms for it, not to mention that while all humans seem to share some basic moral values, it gets very muddy in the details.
That's a good thing. We're going to develop the search and destroy technology no matter what, so we might as well get something beneficial out of it along the way.
Haha, no, no it's not. (I'm very active in SAR -- headed out to a large search this weekend, even.)
We've had "search and destroy" for a decade now, but SAR still doesn't have effective use of drones. Quadcopters can navigate tricky terrain close enough to the ground to see a sufficient amount of detail, but they only get about 25 minutes of flight time.
Fixed-wing drones can get much better flight time, but we can't put them anywhere near close enough to the ground to see the average missing person.
Typically the next thing people suggest is IR. Yeah, IR's great, there's just one problem: there are a lot of IR sources on the ground, even in the wilderness, and people with clothing on aren't that great a source of IR by comparison. At a recent multi-agency training event, I reviewed footage from a repurposed Reaper drone flown by the national guard after the southern California mudslides. Making out the difference between a person on the ground and a streetlamp was impossible.
The only documented drone find I've heard of so far was a person sitting in an orchard. Probably there have been others, but the point is: SAR is a wildly different situation than people imagine.
There was just a study done by DJI and European agencies that found that drones were of negligible benefit in the missing person scenarios they tried: https://www.aopa.org/news-and-media/all-news/2018/october/01...
Helicopters and civil air patrol are superior to drones in almost every way. I've hung out the side of a UH-60 and a couple of Lakotas during exercises looking for brightly colored objects on the ground. At around 100 feet above ground level our success rate was around 60%.
Heck, we had a missing small aircraft, in the snow, a year and a half ago. There were hundreds of searchers on the ground, multiple CAP and helo overhead. Guess who found it? Some folks on snowmobiles a few weeks later.
We use a number called "POD" -- Probability of Detection -- in our reports after every assignment. It's an ass-pull guesstimate at what our odds were of locating a responsive subject, an unresponsive subject, or a clue during the assignment. High POD for a close grid search pattern on the ground in light forest for an unresponsive subject is about 70%. Average POD for an unresponsive subject by a fleet of quadcopters in the same environment would be like 5%. To effectively rule the area out, any decent search manager would deploy a ground team anyway.
So, in the first place, this MIT experiment -- while very cool -- still suffers from a lot of the same issues that this technology has in a SAR context. And in the second place, shooting down efforts to develop this technology for civilian use is silly because, well, I hate to break it to ya, but we've been killing people with (mostly autonomous) drones for a while now. That horse is already outta the barn.
I've done SAR at sea and it is frightening how close you can be to who/what you are looking for and not see them/it because of the waves, or reflections, or low cloud/fog, or rain, and then there's the constant movement of everything due to currents and wind, the inaccuracy of GPS and shooting resections... And you can't just have guys out looking, there are only a small number of RIBs or other assets, and they have limited endurance, and the sea is very very big.
SAR is a hard, hard problem as you say. When I dive for fun now, I take an excessive amount of gear incase I ever need to be found.
Obviously drones are noisy, so maybe fixed wing. Or maybe you could have a microphone suspended on a cable fifty feet below a quad, with an apparatus that shields noise from above. But there seem to be promising characteristics too for this situation: 1.) human voice is very different than propeller hum, 2.) wildernesses have very minimal background noise. Filter and classify outliers.
I say this as someone whose only SAR situation was getting himself out of the Vermont woods at dusk by hiking roughly a mile towards the sound of a barking dog.
I think search organizations are certainly interested in UAVs, but at the moment nobody's sure how to use them effectively. It kinda strikes me as being like blockchain: cool technology, everybody wants to use it, but there aren't really that many real-world use cases where something else isn't just as good or better.
Probably the highest priority should be fixing their flight time. 25 minutes is just not enough time in the air to be useful, and fixed-wing is too tricky to navigate in too many kinds of terrain. If someone figures out how to start approaching 60 minutes in the air with a payload -- speaker & microphone, IR, comms -- that would probably start getting more attention.
Another thing that would be really cool is an airborne minicell. Just about everyone's carrying a smartphone now, they're just going into areas where there's no signal. If something could go out there offering signal, and report back when a device tries to connect to it, it wouldn't necessarily need to get close to the ground and that could really narrow down some of our searches.
For all of our flaws I'd still rather that our military was on an even footing with theirs. And I can't see how autonomous killer robots won't simply be better at warfare than non-autonomous machines.
I am a SAR volunteer, and our team has used drones on real searches. What I can tell you is that the problem is not navigation (even through forested areas), but object detection. Even for a human being, detecting humans (responsive or unresponsive) or signs of human presence (trash, abandoned camps, etc) from drone camera footage is extremely difficult.
Autonomous flying drone navigation is hard but identifying signs of human presence from onboard sensor information will be much, much harder.
The scenario is Joe is lost and incapacitated in the Australian bush. "Joe" is a human sized dummy with a (very) small infra red source under his clothing. He needs to be found and then have something done. He is located within an area of about 50 sq km.
The requirement has always been the aircraft must be totally autonomous from take-off through to landing - and that includes finding Joe. In the previous challenge, once Joe was fond the aircraft had to land beside him, and take off again - in an area it had never seen before.
About 6 years ago Andrew Tridgell (original author of Samba and Rsync) joined Canberra UAV (an amateur group of UAV enthusiasts), and started working on the Ardupilot AUV software Canberra UAV used in future attempts at the competition. At the time nobody had completed the challenge. After Andrew joined Canberra UAV, Canberra UAV completed the challenge every time (and indeed until 2018 was the only team to complete it) forcing the CSIRO to up the difficulty for the next competition to be held.
The location algorithm has been pretty simple in all cases - this is a group of amateurs who could not use expensive LIDAR's. From memory it's multi pass. The first pass happens from relatively high, and then promising signatures (which are barely more than a pixel) and then investigated more closely.
Seems like a place where pattern recognition could be an asset, even if it just supplements rather than replaces a spotter. Or heck, send a Google Maps-style image to land, and let volunteers flag points of interest.
2) minimum size before it becomes a viable strategy
3) vibrations affect any cameras, requiring extra damping
There's the problem of false positives for sure (e.g. deer), but it gives you more than enough starting points.
This research sounds genuinely really fascinating, and I think they are making huge progress on a difficult problem (specifically, navigating unknown 3D space). I'm just not sure that this use case is the most important one.
[0]: https://www.radionl.com/2018/03/11/3624/
[1]: https://fosterelli.co/file/talk/tracking-cattle-with-infrare...
And I'm calling BS on IR being able to see footprints under fresh snow, unless that fresh snow is a light dusting about 1 cm thick -- which just about any tracker could follow anyway.
IR requires a steep difference in thermal gradients between the subject and the environment. In cold environments, subjects tend to wear insulating clothing (and if they've shed that, then they've gone into the paradoxical undressing stage of hypothermia, and unfortunately they're done for). The cold environment also acts as a massive heat sink, quickly masking any heat signatures.
Warm environments suffer from reflectivity. Rocks bake in the sunlight and then re-emit that radiation all night long.
There are some environments where IR might help, but many more where it doesn't.
> I'd love to hear more about cases in which IR has made a difference in a search, because that hasn't been my experience or what I've heard about it.
Here's another link where IR was used to find seven missing people, same trial program: https://bc.ctvnews.ca/a-b-c-first-heat-seeking-drone-used-in...
While looking more into this though, they mentioned that one of the things the program found is that the drones are not good at seeing through tree canopies. This is surprising for me, and I suppose answers my question about why this MIT project has some benefit! I stand corrected in that regard.
> There are some environments where IR might help, but many more where it doesn't.
Absolutely. There's no magic solution for finding people. The project here is going to have pros and cons too, you need experienced people to know what tool makes sense where.
> *...one of the things the program found is that the drones are not good at seeing through tree canopies. This is surprising for me...
Yeah, if you dig around a bit you can find infrared and near-infrared images of trees and forests, and overhead IR shots of forest canopies. Trees give off a surprising amount of IR.
1) A generally lowered body temperature from being wet/cold
2) Clothing or shelter designed to keep heat in or limit exposure to the elements (and therefore also cameras)
3) Trapped under rubble, landslides, snow, etc. that could obscure their IR signature.
Granted, not in all cases, but I'd suggest that it's not quite as simple as just slapping an IR camera on a drone and flying it around.
It seems that the "no GPS required" spec and the under-the-canopy flying and mapping they showed makes this considerably more likely.
Small drones on rescue missions need to find the lost person, identify their location and transmit it back. The search/find part would be best done above the canopy with appropriate sensors, covering as much ground as quickly as possible. Flying under the canopy, and mapping it the whole way would dramatically slow this process -- probably quicker to have a human horde search on foot.
Once found, the GPS coordinates would be the most essential data point to transmit back to the rescuers. The relative positions of all the local trees would be kind of irrelevant, but that could be quite useful in targeting various kinds of attack.
IOW, it seems the other Search & Destroy comments have got it right.
I've posted a few criticisms of IR already in this thread. There are a narrow range of environments where it's helpful, and there's a tradeoff between image quality and flight time -- you can have good flight time and terrible image quality, or good image quality and terrible flight time.
Also, and this may sound just totally perverse, when i go into a wilderness area I really do want to be alone and out of range of the world.
Edit: really? Downvotes? Hiker gear is half neon colors and colored in reflective tabs. Easy visibility for SAR purposes is a design criteria for most hiking gear.
Photographers, hunters and people who's taste in clothes doesn't run to the garish?
I’ve never noticed a correlation between photographers and clothing colors. I tend not to be in my own photos, so it’s not something I select gear based on.
Obviously not every bag is bright, although my wife’s bag is a nice metallic teal, but every hiking bag I’ve ever seen has reflector tabs built in. Ditto with our tent and rain flies, and quite a few of our tops. They help in a SAR situation, and improve safety if you ever need to come close to a road at night.
Taste is subjective, but outdoor gear trends towards the bright for safety reasons.
I have never seen a hiker wear camo, all of them have reflector tabs, and a large percentage have tops, bags, and hats in bright colors. And that’s not even getting into inclement weather gear, which trends towards the super garish. There’s a reason why Everest has a “rainbow valley”, and it’s not because heavy coats are designed to be subtle.
I pointed out that hunting isn’t hiking (opinion, but they do have different names), and that regardless hunters regularly wear gear that is anti-camouflage by design, making any distinction there moot.
I then proceeded to point out that hiking gear has reflective tabs and bright colors for similar reasons. The more extreme the gear, the more intense the visibility features on average.
I then made an observation about what I’ve seen.
At no point did I say that any of those groups aren’t “real” hikers. Therefore, not a true Scotsman argument.
Not to cheapen the work the authors have done, but I dont have a lot of faith that this tech will be used for "good".