Using software agents is inherently revolutionary and risky, but I certainly would not use the word “never” and I also would not assume control from a local aircraft was the default for long.
It all depends on what your usage philosophy is. AI won't ever be able to fully replace a human pilot until one day we "crack our own code", but AI/Autonomous systems can very well do the more menial tasks - like fly itself, navigate, lob missiles etc.
In terms of command and control, I'll qualify this a little more minutely. There are different levels of controls with regards to any UAVs. You can have broad strategic control (i.e patrol this area), overall tactical control (i.e coordinate an attack with other aircraft from this aspect") and you can have low level tactical control (i.e engagment altitude, speed, distance etc). I expect all UAVs / CCAs right now to be able to do low level tactical control fully autonomously. None of them are going to be used to do the strategic control section, but overall tactical control can vary depending on how exquisite your software for that is. Which brings us to ...
Aircraft AI are not LLMs;
I agree that aircraft are not machine learning models, but in many ways, CCAs and today's autonomous systems rely on ML as indispensible organs within the system (also LLMs are probably not the best kind of ML model that goes into a system like this). It's just not the only or even the single most important part of the autonomous system.
The entire system includes ML usage where appropriate, with the inappropriate situations being defined in large part still by written program logic. The two go hand in hand and for any modern self respecting autonomous system, one can't do without the other.
On the ML side, the main usages for it is different levels of decision making, and signals processing. The signals processing part is fairly trivial and not within scope of this. To me, it would be the decision making that forms the core innovation of the whole CCA concept compared to other autonomous systems from before. At the broader tactical control level, say in an air to air engagement, manned aircraft won't be operating the low tactical level of a CCA. At the broad tactical level, it has to account for the rest of it's immediate system and decide on how to act. In practice, it'll be an entire frame work of multiple ML models in addition to rules based systems in which each model has a specific job or is integrating the outputs of lower level ML models to from a bigger picture before making a decision output. Ofcourse it should be stressed that this is not your end result at all, but it is a rather central part of having a robust yet safe decision process.
The rest of the meat and bones is the autonomous systems software that wraps ML models and subjects ML model's output to the rules determined by man. From the high level - such as defining guard rails for engagement behavior, down to the low level - such as flying complex maneuver - most of the work I've seen in this area largely doesn't use ML models or uses them rather modestly, usually integrated with certain sensors (such as the recognition models in your Tesla that tells the car to stop if it identifies immediate obstacles). Quite frankly, we don't need ML to fly a plane autonomously. Without this added layer of safety and injection of human determined rules, no amount of ML will ever get anything done.
Without ML, robustness is certainly still possible, but compared to ML augmented decision making, your robustness is limited. Most early UCAVs would have probably been using rules based logic for autonomy. With CCA's you aren't just expecting them to follow written rules. When you've got a hundred different units in theater, commanded by manned fighters at a roughly 6:1 ratio, you'll have to have greater robustness for independent decision making and taking action within a system. That's ML augmented layer is probably what sets CCAs apart from older UCAVs like the early predators. However, whether it's from a safety or a effectiveness perspective, no ML model can just complete these tasks as a standalone thing either. ML usage in autonomous systems can be thought of as a sensor. As with various sensors, you always have to account for noise and inaccurate signals coming from any of your sensors. Yet you also can't do without it. That brings us to ...
they do not have to be skilled at interacting or random situations - they have a very specific role with a purified dataset.
I'd argue that they
do have to be skilled at interacting with random situations not in the least because randomness is a central and sometimes even decisive part of warfare. Randomness and noise are also the direct result of degraded and even denied operating conditions, so in order to have resilient systems, they must be able to account for randomness as much as possible. Even within their own specific roles where whatever models are used in the system is trained on a purified dataset, there's randomness that you cannot possibly account for no matter how pure your dataset is. If there aren't any ML models to worry about, then you'd arguably have an even bigger problem of randomness to account for. The more exquisite a system, the more it must account for randomness.
Keep in mind that even a fully autonomous system is still acting inside a broader system formed by other UAS and manned elements that may or may not be fully connected with the current autonomous system. That's not even accounting for opposition elements, jamming, poor weather, unexpected actions from manned elements or other degraded conditions yet. In a shooting war, even if all your unmanned elements perform as expected, they still have to interact with the sometimes irrational, unnormal actions of manned elements. The permutations of randomness for an aircraft of even a single role is already overwhelming let alone pitting it against an opposing system. Having a baseline ability to account for random situations is absolutely necessary and the more exquisite a UAS, the more randomness it needs to account for within its mission set and within the broader system in which it operates.
And quite honestly I think the USAF is lying through its teeth when it says this is all about local control: IMO, it makes much more sense for control to come via LEO low latency satellites and control to come from an overseer AI on the ground with humans on the loop.
Agreed. Pretty sure I saw some article recently about the Chinese working on laser datalinking from satellites. Kendall at the time also talked about laser datalinks though in the context of control between manned and unmanned elements.
The efficiency of looking at the entire picture at once with machine learning is likely to be extreme - what if you could count every time an opponent launches or measure their performance at the theater level to determine tactics and aid in ID of platform types? The fact that the first render of the YFQ-44 had an antenna farm convinces me this is where things are truly going, and that local human control is a crawl before you run/fallback control scenario.
Note that in this scenario, the underlying satellite network and the algorithms that control all the NGAD automated platforms are far more important than the individual platforms. Note that automation of the manned platform could also be in play.
Well yes you are right, but that's situation specific. There's no single AI that can "do it all", but there are AI applications that can do things much better than humans can. And as you've noted, how well such an AI can perform is still dependent on the input from the sensor nodes. Fooling an AI then in part becomes a problem of electronic and kinetic warfare. It's also not impossible to have AI fool AI either. Still AI will be an invaluable tool in such applications and that's obviously where things are going, but the human in the loop is at best going to stay where it is and at worst going to still be making decisions when AI gets confused and the broader system (in your example) is sufficiently degraded.
As I've said previously - we need to think of AI as a sensor, or rather a sensor integrator: you have curate the input, validate the output and process the output. No output, not even from a general AI that is just as intelligent as a human, is error prone. And every AI is only as good as the data it can access and the data it's been given. As it applies to CCAs, I don't think control will ever be completely autonomous no matter the controlling medium. At the strategic level, no one is going to offload actual command of a theater engagement off to AI entirely. At the broad tactical level, the CCA will be able to do 90 - 95% of the reasoning autonomously, but when it's about to shoot down an airliner thinking it's a P-8 or something, someone somewhere needs a kill switch. At the low tactical level is the only place where a CCA can be fully autonomous but only if its equipped sufficiently (such as collision avoidance sensors, which apparently some initial CCA test articles didn't have).
Type C* is thought to be a CCA or “unmanned wingman”. It features jagged swept wings, canted tail fins, a dorsal engine intake and a single engine. Type C appears similar to the earlier FH-97.
Type E is yet another CCA or “unmanned wingman”. It features diamond wings, canted tail fins, a dorsal engine intake and a single engine. Type E is believed to be in service with PLAAF stationed in Southwest China (S/N 53x3x).
