chrisl

So basically the Borg.

It's certainly a doable thing - it's by no means a magic technology, and probably doable with public/private keys without even having to add another transceiver. The catch is that you have to implement it across all your drones, and the dronespace seems to be too diverse and moving too fast for it to get integrated without being in the way in Ukraine.

And about building up a model of what's around them and what not to run into. There will always be things outside the training space that the system has to realize it needs to avoid. And I was really just taking a potshot at Tesla's difficulty in detecting stationary emergency vehicles using its camera-only system. Sort of. Driving is also a much more defined and controlled environment than a battlefield (fully acknowledging that the driving environment is absolutely packed with unpredictable, uncontrolled things). At least you have a model of roads, standards for lanes and markings, standard signage and traffic controls, rules that you can expect other objects to mostly follow, etc. So anything that's a deviation from that model has to be deal with, but things that comply with the model you can "ignore", or at least treat as normal operating conditions. In the battlefield environment, the whole environment can be changing rapidly, and generating meaningful simulation data may be almost as hard as getting real training data. If you're training to hit big, expensive, slow to change things like tanks and trucks, you can train based on "find things like this and ignore other things". If you're training to find and hit things that only fill a pixel or two and don't stand out from the background (a drone at distance), it's trickier.

So you want to start a new US/Canadian defense company?

The Dalek approach to a defined area crossed my mind while I was writing, but I didn't address it. As you point out, it has some rather spectacular failure/spoof modes. If you don't solve the IFF problem, could end up with the "baby spider" situation, where they all just eat each other until there are one or two left, without regard for whose side they're on. (btw - I've seen *actual Daleks* patrolling the new Economy Parking garage at LAX. You have to be really careful getting to the shuttle)

To be found dead in a fall from a window in 3.....2.....

Sure. Some of those gas motors are even kind of cute.

Later versions even have a laser rangefinder in the submunitions. Enough pieces already exist to do lots of these things we're talking about, at least as far as the swarm of loitering platforms goes. For most of them it probably just comes down to someone in DOD getting the motivation to throw some money at doing it and making it a big enough pile of cash that the MIC frees up the right engineers to do it.

If one side has the Diamond Age swarm and the other doesn't, the swarm side is going to have an advantage no matter whether they're on attack or defense. So far they aren't quite overwhelming the defenders, but we saw the dystopian science fiction video a few pages ago, and that kind of quantity isn't all that far away. The limit on those swarms is going to be battery endurance.

That's always the hard part. Not just enough training data, but the right training data. If you're Google and just trying to squeeze another fractional percent out of your ad revenue, you can use half the internet to train and the other half as a test set. And you only need a tiny improvement in performance to make it worth the effort to get mediocre results. If you're Tesla it's trickier. You can put cameras on cars driven by people with excellent driving records and then correlate the video to the drivers' actions and hope you got it right. Sometimes they seem to drive into firetrucks, though. (Dude, where's my LIDAR?) At the low SNR, barely resolved limit, the tracking is easier than the detection - once you identify all the things, figuring out who is who from frame to frame isn't terrible if you've got a high enough frame rate. But at low SNR the noise and the signal look very similar in any frame, so you have to watch things (noise) move around, and then throw out things that only exist for a few frames in a row because they're noise. (or maybe they're drones popping up above the weeds to take a shot and then disappearing) All of that isn't so bad to do if you're looking for large targets, like vehicles, and trying to separate them from bald monkeys and buildings. Maybe even separating trucks from tanks from IFVs. It's a lot harder if you're Russia and don't have an internal semiconductor industry base. I wouldn't be all that surprised if they had well developed algorithms derived from open source, maybe even partially trained on video data, but not enough hardware to go past a few prototype units. But they'll have a terrible time doing it for drones - there are multiple pieces of silicon that you need to put in a row that are practically free in the west but will depend on some unreliable machinations in Russia. Which reminds me that I saw a spoof prototype of The Captain's wandering mines over on the site formerly known as twitter, but didn't notice it here or have time to link it before it disappeared in the ether. Someone did a little photo sequence pointing out that the Russian AT mines look like they fit neatly atop a Roomba, with the obvious extrapolation.

In a very narrow domain. Low SNR tracking of large numbers of objects at the resolution limit of my optical system using (deliberately) underpowered computers has been a headache of mine for the past decade. This is for targets entirely unrelated to anything of defense interest, but it's defense-adjacent, and there aren't good general solutions floating around out there.

I'd think the Russians training waterfowl strapped with mines (two birds with one mine?) would probably be more effective than anything "autonomous" that they can field today with their current resources. Automatic? Sure, that's Shahed. Automatic with a basic IR sensor for picking out warm targets? Probably, but not widely available. Truly autonomous and using some kind of AI to pick targets? Bring in the trained ducks. Air defense against the Diamond Age swarm isn't going to be any single system - it's going to have to be many layers dispersed over a large volume, and insensitive to any particular node or 20 being knocked out. Basically attack of the killer bees on both sides. Ground hugging drones is a nice approach for anything that can be a lightweight payload. Lots of background to make them hard to pick out, but without the mobility issues of a UGV. Arguably a dumptruck full of ground hugging drones will be more effective than most UGV concepts.

Yeah, I'm very aware of that, and I suppose that's a cue for getting a little into why we don't have anti-drone drones. There are two parts to the anti-drone drone: detection and attack. The detection is the hard part. Destruction is easy - we already have no end of systems that can very accurately destroy anything that you give them coordinates of. We can accurately fire projectiles, exploding projectiles, exploding projectiles full of razor sharp hoops, high energy beams of photons, rings with chains on them, rings with strings on them, giant wads of gooey stuff, or anything you want to take out a drone. But you have to detect it. For an anti-drone drone, there are sort of two categories of drone you're targeting: open loop (no comm back to the sender) and closed loop (some comm back to the sender, whether full two-way control, occasional updates, or whatever). Detection of the first type (no comm), which includes Shaheds, is tricky - unlike the F-35, these *start* with the radar cross section of a goose* and then you can make that even smaller. These things are all small on visual and radar cross sections because you can paint them and they don't have a lot of metal. You're going to track them with frustrating "visual" algorithms, where "visual" can mean different things in the optical vs. radar wavelengths, but you're still trying to pick out changes in the scene to decide where the thing is. I'm not going to spend much time on it, other than to say that unless you have really high signal to noise and high resolution (both of which the target is trying to reduce), it's a lot harder than you think, and in general you're not going to get there with simple image differencing. And this problem exists for commless drones whether you're using another drone, a gun, or a death ray to take them down. Shaheds at least have a very characteristic sound that you can probably use for detection and targeting once they're within audible range. Detection of the second type (active comm) is easy. It's transmitting, and transmitting enough to get clear signal back to its operator, who is farther away than you are if it's attacking you. Triangulation is old technology. Piece of cake: you lock onto the frequency, have some kind of sensor so you know your own orientation relative to the sensor, and just maneuver in a way to make the signal from the drone stronger until you hit it and destroy it with whatever mechanism you prefer. Or have a few sensors that are networked to give you the position (helloooo MLAT) and shoot it with your favorite method of action-at-a-distance. Except for one problem: whose drone did you just destroy? In the Ukraine environment, IFF is the hard part of doing radio based anti-drone systems. There are tons of things flying around, as evidenced by the daily releases of yet another view of every bit of ground combat we ever see. It's not quite Diamond Age concentrations of them, but they're working on it. And they're all sorts of random drones, including commercial drones, custom drones made with commercial off the shelf parts, custom drones with a mix of commercial and special mil parts, totally custom mil drones, and who knows what else. And they're all using similar frequencies, because the combination of physics and the atmosphere force you to the same frequencies if you want a particular range and data rate at powers that you can reasonably supply to both the ground operator and drone with batteries. If you don't sort out the IFF thing and you set an autonomous anti-radiation based anti-drone system loose, it's just as likely to attack its allied drones as the enemy drones, because it has no way to tell them apart. That means you have to have your complete drone ecosystem integrated (ring that cash register over at Lockheed/Northrop Grumman/Raytheon!!) or you're just going to be attacking your own stuff. And part of why we aren't seeing even rudimentary versions of it in Ukraine is that it's not a function that people were already spending much effort on for commercial/hobbyist drones. You can't just pop over to Robotshop.com or Alibaba and order tunable RF sensor kits (or a few thousand of them) the way you can other types of sensor, or actuators for operating your 3D printed grenade dropper. It's possible to get relatively inexpensive software-defined radio modules that are small (that's what feeds ADSBExchange so you can see who's flying around Ukraine), but the environment is so variable, along with the need to confirm what drone you're attacking, that at least for now you're going to need a human in the loop, even if you can semi-automate your remote control drone sensor. And even with a human in the loop, nobody is painting national flags on their drones, so unless you know "this is one that our side makes" after you get up close to it (assuming you're doing that, rather than sending a death ray at it from 5 km), you really don't know who you're shooting down. So the basic tech isn't all that hard, but because it's not just point and shoot or point and drop, it's a lot more dependent on integration of the whole system to be usable. *geese, like all waterfowl, are incredibly mean and probably deserve to die. That's why there's a book entitled "Ducks and how to make them pay". If we can do an autonomous system for drones, it should probably be immediately applied to geese and ducks.

US/NATO almost certainly aren't giving a direct data pipeline, but Ukraine probably wouldn't be able to process it as well or as quickly as they're likely getting pre-processed from US/NATO. The US has had electro-optical/digital (non-film based) satellites since the mid 1970s and has a lot of infrastructure built up for the processing/evaluation/distribution that would take a long time and a lot of trial and error for Ukraine to replicate. Commercial sources actually have very rapid revisit times these days - that's the big selling point for Planet Labs - moderate (0.5 to 3 m) resolution at very high revisit rates. If you look at the whole Maxar fleet, they can revisit anyplace every few hours, and it looks like you can probably get commercial SAR at similar return rates.

Yes, it's a "difference detection" problem, but if you're looking for something that's near your limit of resolution it gets harder. And all measurement systems have noise. For most human optical applications you're used to signal levels that are *way* above the noise, but someone trying to hide from a satellite is going to work hard to keep the changes that you see down close to the observer's noise level. That makes life much harder for the automated system and its trainer. And aside from the changes they're looking for, there are lots of other changes going on above threshold because stuff just moves around on earth. That's noise, too, but of a different sort, and part of the developer's goal is to be able to discriminate benign activity from targetable activity when they both show up in the signal.

Not just good quality data. Good quality data that has been accurately labeled, and often in massive volumes and of high diversity. Depending on your system, that can be very difficult to get - I do some optical systems where we're almost guaranteed to have low signal to noise (always working at the limit of our resolution) and are likely to run into things that are outside the training set, and we spend huge amounts of effort on both improving the repeatability of the acquisition (relatively easy), and automating the development of training data, which sometimes involves a lot of bootstrapping. You think watching the scanners at TSA checkpoints is painful? Have I got a job for you...

What makes you think that isn't how Bayraktar found a lot of high value stuff early on, or that that's not how we got all the river bridging massacres? You can be sure that anybody who owns satellites and computers is spending a lot of effort on that. Whether it works for artillery hides will depend on a lot of things - the satellite resolution and wavelengths: optical are generally higher res, but easier to camoflage against; SAR doesn't really care what color you are, but does care about your shape and what you're made of. That's also why artillery units train to be able to hang out behind CB range, or just drive around, and then just like a panda: park, shoot, and leave in 5 to 10 minutes.

And my point was that the original estimate missed enough key features about how DE works that it wasn't terribly meaningful, and the discussion of energy source is a distraction.

As already alluded to - high power lasers are typically pulsed. They might be powered by a system that’s in the kW range and hit GW in pulses that are a few ns long. And the short pulses typically are putting all the power in a small area virtually instantly, so you get tremendous local heating and ablation with every pulse - you neglected thermal diffusivity in your bierdeckel calculation. For example, you can power a laser that can engrave rock using a 9V consumer battery.

Kinda like a camera flash…

Anything that is radiating energy is telling you where to find it. Anything radiating structured energy (active ISR, comms, jammers) is telling you that you want to find it first. You either have to have the biggest, fastest network of active systems to beat a nominal peer, or you want maximum autonomy with very infrequent active radiation that is somewhere else unpredictable every time it radiates a signal. More when I’m not typing on a phone…

16 grams for a Cox 0.010 engine - a little over half an ounce. The relatively common 0.049 is a couple ounces.

Why bother with the $100K ghillie suit with a person inside? Just wrap it around an autonomous robot.

And arm them with what? T-34s pulled from monuments? Mosin Nagants? Maybe there are a few IS-2s still floating around on a back lot. I know they’re not quite down to that, but they’ve burned through decades of Soviet and post-Soviet production and are nowhere near being a manufacturing powerhouse to build back up. They don’t have a huge base of non-military manufacturing that they can convert. China is the only real source they have for manufacturing resources, and they’re not being particularly generous. And Russia has been suffering brain drain for three decades. And who will train the newly mobilized on the higher tech equipment? Maybe it doesn’t matter, because it will all have been destroyed or fallen out of the sky by the time the next wave shows up.

I'm also a little curious as to the mud situation. Ideally, Russia will be struggling to shuffle all those forces to Kherson on trucks in mud. And if mud makes it harder for Russia to bring up supplies and reinforcements to the left bank, so much the better.