Artificial Intelligence in the Cockpit
daedalean.ai
daedalean.ai
On the one hand, the author discusses what they have achieved so far - a machine-vision based system that can fly reasonably competently in good visibility and low traffic density, comparable, they say, to aviation 80 years ago (actually, as I mentioned in another comment, aviation in 1941 had already advanced significantly beyond that.)
A little later, they write this;
There is no reason to believe computers will always be worse at that than you are. There is no reason the machine can’t reliably make the call to land in the Hudson when all engines are out and to do so in adversarial conditions safely.
True enough, as far as it goes, but there is also no reason to suppose that the technologies the author is discussing here will deliver that level of performance. The good judgement demonstrated by Sullenberger that day (and by many pilots in many other dire situations) depended on an extensive understanding of how the world works, and to reason about outcomes outside of the rules of the game, so to speak (for example, short-cutting the checklist in order to ensure the aircraft continued to have auxiliary power.) Current machine-vision systems, on the other hand, lack the ability to reason about how things ought or might be, and so can make utterly bizarre-seeming judgements about what they are "seeing."
Personally, I believe fully-automated aviation will become both feasible and acceptable, but with arguments like the one quoted above, this article is glossing over the challenges that remain.
Tree bad, river pretty?
I would guess that there are probably enough things on there that are constantly monitored theses days so the plane could just give you a big green "all systems nominal" light?
Maybe I am just naive :-)
One example that comes to mind is visual and audible alerts when you haven't put the landing gear down and the system thinks you're going to land (speed low, flaps extended, descending). Yet gear-up landings are still common in single-pilot general aviation aircraft.
How about an "assistant" that calmly tells you "it seems like we are landing, but your gears are still up, wanna do something about that?".
I guess acceptance for something like that will be low.
The inverse of your "all systems nominal" light exists, in many if not all commercial airplanes of any size, in the form of the master caution and master warning lights, but you also need to know what, specifically, is going wrong.
You have pilots for the situations the available automation doesn't work.
Everybody else has to get out of the way, though, when it declares an emergency. It can't really communicate with ATC. So it's strictly an emergency system, for now.
This is really just integrated control of the existing avionics. The existing systems are good enough that you can input waypoints and have them followed, and do an automatic instrument landing on a designated runway. This mostly sets up a flight path. It can't deal with traffic. There are no new sensors. The additional hardware just lets it lower the landing gear, apply the wheel brakes, and shut down the aircraft after landing. After which you're blocking the runway until someone comes out and moves the aircraft. Again, emergency use only.
It's intended for the "sick pilot, healthy airplane" case - the pilot is out of action, but the hardware is fine. It's not helpful in making hard decisions when the aircraft is having problems.
Real world flying with the 10^7 or 10^9 kind of scale for safety isn't in need for a better autopilot. That would be like making a more precise cruise control while you want a self driving car. What it needs is better decision making and problem solving skills. And those are very much lacking from these current systems.
I have an instructor rating. We can teach an average human in 5 to 10 hours how to control a plane. Then we spend an additional 40 to 60 hours on how to make the right decisions, solve problems, deal with weather, traffic etc. And humans are really good at that.
What is needed if you want autonomous flight is really good AI decision making, not aircraft control.
That said, the reason pilots are paid to fly dangerous whirlybird machines is primarily for take-off and landing. Takeoff and landing are operations that require the most concentration, coordination with air-traffic controllers and other aircraft and regardless of weather carry the most risk. Takeoff and landing are likely to be some of the last functions to be automated, I'd argue this even further with take-off. Another important aspect here is the pilot should be able to act independently of multiple systems failing - as the moniker goes... "fly the plane until its on the ground or stopped".
The coolest pilot safety automation tool to improve safety I've seen thus far is an iPad app developed by the OG developer of the X-Plane flight sim Austin Meyer[0] (definitely check out his blog) called Xavion [1]. Basically, it's an app that with decent GPS will calculate a glide plane to the nearest airport in seconds. Austin is an avid pilot and clearly a brilliant guy - I'm eager to see if he starts commenting on autopilot ai initiatives like that of Deadalean.
The better analogy, IMO, is how people are being pushed into a computing model and adopting computing attributes as characteristics. We are becoming like "semi-autonomous computers". And worse, it is a client-server model, with corporations and governance taking the executive decisions!
Rather than describing some software as intelligent, we would be better describing the change that this idea (and the idea of collectivisation - as opposed to individuation) has had on people. It is people that are changing, machines are still inanimate.
I found reading 'A Collection of Definitions of Intelligence' quite interesting myself https://arxiv.org/abs/0706.3639
I would prefer we use "automation" instead of "AI". That way, we are forced to speak specifically ("automation of what?; it's always specific) and say things like "machine-automated aviation" because that's what this is. It's clear, faithful to the truth, and obvious what we're talking about and what's happening at a general level, and we aren't reifying some vague bullshit fantastical term like "artificial intelligence" which only leads to romanticized mystification and projection. There is no artificial intelligence. Machines doing so-called AI are the same kinds of machines we use to write email and make phone calls. We've just configured them in a way that makes them useful in different situations in different ways for specific ends, even if the applicability appears general; we've done the generalization which is then represented in concrete ways which are not themselves general.
(N.b. central to intelligence is the ability to abstract (not the lambda calc/CS meaning) from particulars. I have the concept Circularity which is not just an image (there are potentially an infinite number of circular objects), but based on experience of particular circular things, my intellect has abstracted from this experience the universal concept of Circularity. I understand Circularity, apart from any given circle, and yet what is true of it is true of all circles. I can analyze the concept to infer that any circle's circumference is twice its radius times pi or that its area is the square of its radius times pi. Computers don't do this and cannot do this even in principle because all physical things are always concrete. There is no physical Circularity, only physical, concrete circles. And abstraction is not regression. Indeed, all the abstract values in your computer aren't really abstract except in the mind of the observer. Those in your computer are representations only, devoid of denotation except what's in the programmer's head.)
* https://www.electronics-tutorials.ws/systems/negative-feedba...
* https://en.wikipedia.org/wiki/Feedback
There are no heuristics involved:
* https://en.wikipedia.org/wiki/Heuristic_evaluation
* https://en.wikipedia.org/wiki/Heuristic_(computer_science)
although I think when autopilot was invented, the concept of "AI" didn't exist, so maybe not a perfect example
A feedback loop — namely between a system and its measure of its own performance— is central to the idea of AI. At least according to Peter Norvig, the director of research at Google, who defines intelligence as ‘the ability to select an action that is expected to maximize a performance measure’ (Russell & Norvig, 2016, p. 37).
(I do go along with author Charles Stross, occasionally seen here on HN: that Corporations are Old Slow AIs, in which case we've already had AIs around for centuries.)
Personally I'm happy to grant a degree of intelligence to plants (and probably even bacteria, if I squint hard enough) though it's of a quite different nature to our own. Certainly feedback loops are central to the idea of intelligence, but there's a whole lot more wanted/needed than merely feedback. And so, I find Norvig's definition a tad wide of the sort of intentionality and sentience we'd want flying a plane, and certainly far, far short of the sort of thing we'd call AGI.
Is it useful to view corporations as old, slow AI? I certainly think so. Otherwise we get really confused about AI. Look at Zillow. That was a deep misunderstanding of AI— the product isn’t done until we take the humans out of the equation. No. What is intelligent is to have a system that uses its own measures of success to improve. This, by the way, is why cybernetics is so critical to understand in the context of AI design.
I have experience in adaptive education where massively expensive teams of engineers missed the point that the “smartness” of the system needs to be based on improving outcome measures (namely, learning outcomes) and instead focused on massive, complex modeling initiatives with no feedback loops to indicate whether the models were doing anything useful.
If more people understood why a steam governor or a 1914 autopilot or a corporate bylaw were primitive forms of AI, they wouldn’t be looking for magic. “If I can understand it, it must not be AI”
By the same argument, a person could say "AI is like a steam governor. I understand completely how steam governors work, and I know they cannot possibly translate from one language to another or recognize faces, so any claim that AI can do so is nonsense." This, of course, would a completely fallacious argument - and where it goes wrong is precisely with the assumption that AI is anything like a governor, except in the broadest possible way that gives precedence to a commonplace resemblance over all the substantive ways in which they are almost completely different.
I understand your desire to persuade people to not regard AI as magic, but I do not think this is helping.
I am also curious as to how you reward a steam governor - does it get more sex, with better governors? You might reward the inventor of a particularly effective governor with orders, but that isn't rewarding the governor.
1. Venus fly traps detect if an insect is on the trap and close the trap. So far so known, but less people know that the trap will open again after a while if there is no movement detected (ie if a stone fell in). Likewise digestion will only start if there is movement detected for a while.
2. Mast years (https://en.wikipedia.org/wiki/Mast_(botany)) ... somehow trees communicate when to produce seeds on mass ... from what I gathered we have no idea how they do that.
Alternately, I have also read of trees exuding pheremones via their leaves as a warning to other trees in the vicinity when predators (antelope) come around to munch on the leaves, resulting in surrounding trees rapidly increasing tannin content in their own leaves to make them unpalatable to the browsers.
There's a whole lot of shit going on out there that we're scarcely aware of...
A feedback loop may be necessary for AI (or even "natural" intelligence, and sentience (nevermind sapience)) to happen, but is simply having it sufficient?
see also the last few chapters here; http://lavalle.pl/planning/book.html
i would say relative to classical control, fast gpus have replaced the need to have certain closed form solutions that are easily analyzable.
Alexa, Tesla autopilot, Google search, speech to text, recommendations— any practical example of AI— these are tools. Not human replacers.
Is it your position that commercial flying can never be automated, or only that, if it is automated, it would not, by definition, be AI?
My position is that “fully automated” is a logically inconsistent human objective. It means that there is no possibility of human control, because if there is human control / supervision, the system is not fully automated. And, so long as there is human control, you are designing tools for people to use. No one wants anything to be fully automated. It’s the biggest fallacy of AI in design.
If your basis for so believing is the same as you expressed a couple of posts back - that "fully automated" means that there is no possibility of human control, because if there is human control / supervision, the system is not fully automated - then you are simply avoiding the issue by using a pedantic definition of automation. It would seem that, to you, full automation of anything is a logical impossibility as there is no such thing as automation (in the usual sense) in your dictionary.
If this is your argument then, by your definition, even the task of governing the speed of a steam engine cannot be fully automated, as someone sets the desired speed. This sort of reasoning is not insightful; it merely turns your participation in a discussion into a statement of your private lexicon, and avoids engaging any substantive issue. Private lexicons are not useful in communication!
Furthermore, I can respond to your pedanticism by rephrasing the question: do you think it is impossible that commercial aviation could ever operate without human pilots either aboard the airplanes or controlling their flight remotely, beyond just setting target routes and schedules?
I think it is fair to say that, in the context of the article we are all responding to, the relevant sector of commercial aviation is that which currently uses airplanes flown by human pilots.
We should not remove humans from the loop until the system passes the Turing test and solves the Trolley program concurrent to acceptable ethical standards.
A "good" three axis autopilot consists of 3 quite independent systems. One controls pitch/altitude using a feedback loop very similar to an old-style car cruise control (no traffic sensing here :). A second system controls heading, similarly simplistic. And a third controls yaw, even simpler. There might be an auto-throttle, almost identical in function to conventional cruise control. There might be some useful variations (eg VS and FLC modes) - all very useful but there are NO SMARTS HERE.
Modern systems have edge-case handling (triggered by too-slow, excess load, or anything going outside the normal operational envelope). The AP handles such cases by going "beep" and turning itself off - just handing over to the pilot!
see where you are without GPS or radio or inertial navigation
see where you can fly without ADS-B or RADAR or ATC
see where you can land without ILS or PAPI
A big part of safety critical systems is having redundancy and often voting systems, or sensor fusion. So having a single source of information means that you can potentially go down with a single-fault condition.
(e.g. If your vision system is in the visible spectrum, things like sun glare, snow, bird poops on your lense/windshield, one of your cameras fails and you lose stereo vision etc)
The choice to focus on purely visual systems seems an odd one, given that aviation did not really get going until it advanced beyond that stage. Even by 80 years ago, instrument flying with radio navigation in conditions that did not permit visual flying was a routine practice, and this, together with the increasing number of aircraft in the sky, necessitated the development of air traffic control.
While one can very plausibly argue that we are approaching the point where ATC as we know it could be largely dismantled (not, by the way, through the use of AI), that will not be achieved by reverting to purely visual methods, which can only function safely in fair weather with good visibility, and also only where the traffic density and speed are low.
We have better technology than that now.
Strictly speaking that wouldn't be vision only though, so who knows.
Limitations: https://www.garmin.com/en-US/legal/ALuse/
There isn't anything in the article that suggests that the system does not have sectional charts in a database...
Once you know where you are in the world in VFR mode, it's a fairly simple affair (eg landmark detection) to register that positons on a sectional map, however invisible it may be in the real world.
> that will not be achieved by reverting to purely visual methods
Another extrapolation on your part: they never said "purely", did they?
The visual part of the system is nowhere listed as being the "only" way for the plane to know where it is.
It's very likely their system with gulp in whatever signal they can get their hands on (GPS, etc...) and integrate it into a final solution.
The final system will be much more robust than either of the existing components use on an airplane to "know where you are".
While the author may not have written "purely visual", there is this quote, immediately preceding the claims quoted at the root of this thread:
"At Daedalean, we have built a system that uses nothing but visual input, just like the human pilot in that almost-uninstrumented aircraft from 80 years ago, to..." [my emphasis.]
So, while the author does not use "purely", he states something equivalent to doing so. Furthermore, the entire tenor here is that they are not using any of the modern technologies that are currently central to commercial aviation (why he wishes to stress this fact is another issue altogether.)
My point here is a bit more subtle that pointing out the obvious: that automating VFR flight will not cut it (with or without GPS and other aids, for that matter.) It is, rather, that there remains a huge gap between what they have achieved so far (or could be achieved just by integrating GPS and similar aids) and what will be required to achieve the level of performance blithely suggested in the comment about landing on the Hudson river.
UPDATE: I made that last point in a different comment; what I was wondering here was why the author seems determined to point out that their current system does not use GPS etc. - perhaps because any of that would make it look less like AI?
It wasn't Dadedalus [sic] who flew towards the sun, but his son Icarus:
> In Greek mythology, Daedalus (/ˈdɛdələs ˈdiːdələs ˈdeɪdələs/; Greek: Δαίδαλος; Latin: Daedalus; Etruscan: Taitale) was a skillful architect and craftsman, seen as a symbol of wisdom, knowledge and power. He is the father of Icarus […]
That said, this is still pretty cool, and I could see something like it being one component of a much larger fully automated flight management system.
Edit: link should be fixed. If I'm still crushing it, the title of the lecture is "Children of the Magenta Line."
I find it very surprising that there isn't at least non-AI software to monitor what the pilots are doing.
Plenty of accidents follow similar patters: pilots forgot to enable some warning system, or ignored a warning because it came up in the wrong context, or where overwhelmed by the amount of checklists they had to do or by multiple alarms coming up at the same time.
There is no monitoring system that is aware of the full context, including the state of the plane, the history of previous warnings or malfunctions and the intentions of the pilots.
Its called the pilot. We even include a second system, fully programmed and built entirely by a separate team, to double-check and challenge the primary system: the copilot.
> Plenty of accidents follow similar patters: pilots forgot to enable some warning system, or ignored a warning because it came up in the wrong context, or where overwhelmed by the amount of checklists they had to do or by multiple alarms coming up at the same time.
And yet, commercial aviation is by far safer than driving. General aviation, with fewer of these systems is roughly the same as driving.
You also day plenty of accidents, but I'd be willing to bet that just as many if not more were situations beyond the pilots control.
You're also not comparing it to the number of times that pilots got out of sticky situations by following checklists and/or not having warnings suppressed.
This is much less safe than driving a car.
Can you please spare the patronizing tone?
> And yet, commercial aviation is by far safer than driving.
So what? It could be safer than what it already is.
> but I'd be willing to bet that just as many if not more were situations beyond the pilots control.
Then you should review a good number of accidents.
> You're also not comparing it to the number of times that pilots got out of sticky situations by following checklists and/or not having warnings suppressed
That's an incorrect comparison. I never said that the checklists should not be followed or anything like that.
Avionics: the original microservice infrastructure
For a multi-crew aircraft, I believe this is the entire point of Crew resource management [1].
> "Specifically, CRM aims to foster a climate or culture where authority may be respectfully questioned. It recognizes that a discrepancy between what is happening and what should be happening is often the first indicator that an error is occurring"
For commercial aviation, I'd rather take my chances with the crew with 60+ years of combined experience vs. an AI model.
Not at all, CRM is about cognitive and interpersonal skills.
It's completely orthogonal to having some software that monitors the plane for (subtle) faults and unexpected behaviors and provides good contextual information.
> I'd rather take my chances with the crew with 60+ years of combined experience vs. an AI model.
That's a false dichotomy and also I did not talk about AI.
I'd rather take my chances with the crew with 60+ years of combined experience TOGETHER with a all-seeing monitoring system.