Hacker News Reader

So, NASA, is in the process of working on some of this. They’re trying to come up with a new ATC model which will work better for drones / eVTOL air taxis / etc.

Some of the ideas are around “carving” out corridors, where you would be able to fly, and this corridor could be reasonably large (across multiple towers today). The other interesting thing is hierarchical ATC automation.

So, normal ATC would still exist, but they would interact with a tier of controllers who are responsible for them interacting with UAVs.

A lot of this apparently came from the idea that eVTOL is the future, and we will all be commuting to work in an electric airplane.

See: https://www.nasa.gov/simlabs/atc

I await our future where nobody will go outdoors anymore because all you will hear is vuvuzelas / drones.

It's not the mistakes that matter, it's the proportion. Pareto says most cases should be ordinary, so looks like we got that covered. The question is given x amount of flight hours under likely ranges of events, who does better and by how much.

If the human can perform better in niche situations but not better in ordinary then that may lead to an overall decrease in crashes or loss of cargo (the metrics we actually care about). If the human only performs slightly better, well it makes sense to switch to autonomous because humans are expensive, especially when things go wrong.

The point is that no one expects the machine to be perfect. You can get caught up in the "well it can't handle x well" all day, but that's missing the forest for the trees. A robot isn't going to get frustrated if you only fly it on VFR days and ground it for 20 hours in Little Rock because weather changed.

Along these lines, would an autopilot ever think of landing in the Hudson? (Particularly if it hadn't been done before?)

One side benefit of using robotic planes (in theory) is that the fleet can be OTA'd to include emergency water landings, once the models are updated.

Devil's advocate: Would a human pilot be skilled and lucky enough to pull this off a second time?

IMO, the critical thing the human crew did in US1549 was the decision to take the Hudson; once they decided that, most crews could pull off the execution of that decision. (Then the presence/absence of boat traffic comes down to luck.)

> once they decided that, most crews could pull off the execution of that decision

That's not correct. Ditching is very hard. Mess up your angle or sink rate or roll even slightly and you'll be in pieces. Same if the water doesn't cooperate.

One interesting thing is that it happened with an Airbus. The flight computers actually modified Sully's inputs on pitch by about 2 degrees, which ensured a softer splashdown than what Sully was commanding.

But that was a very subtle adjustment. Sully was an experienced pilot, with Glider training too.

Yes. The aviation system produced the human pilot who was able to transfer their skills and experience from all previous landings to this new unusual one. I do believe another equally as well trained pilot would have been able to pull it off as well. What would a computer program have done?

Perhaps pilots of previous generations could do this but, having trained numerous pilots for international airlines, as well as pilots for "fee-for-departure" (regional) airlines in the US, I wonder how many crews are up there that could competently handle most catastrophic events. Though I didn't train him directly, the pilot who crashed the Atlas 767 was pushed through the training program I worked for. Described as a "train wreck" in articles, his ability level was characteristic of others I trained (which I regret). Major US carriers do well screening their pilots but, at lower tiers of US aviation employment, I wonder if more automation would be the safer choice.

Pilots train for water landings. Why wouldn't an automated system get the same training?

Human pilots are trained in how to ditch, but the decision to ditch in that scenario was, as far as I'm aware, an instance of creative problem solving.

Before Hudson, the idea of a water landing looked like this: https://youtu.be/RM7zMW8WVYU?t=19

It was not considered a serious option.

Taking a step back, do we want autopilots flying over a dense population center like NYC?

There are endless "what if" type scenarios where the answer boils down to simply detecting that this is one of those "what if" scenarios and then initiating the appropriate emergency procedure so it can be figured out properly. E.g. an emergency landing is always an option.

That's basically what pilots do as well. They spent most of their time assessing whether everything is OK via checklists and when it is somehow not OK there's not enough time to think and they reach for some emergency procedure instead. Mostly this never happens and there are whole sets of procedures that most pilots never even handle outside a simulator. E.g. landing a 737 with all engines out is a thing but obviously not something most 737 pilots would ever have to do in a real plane.

ATC obviously needs to be automated too. Automating routine interactions like exchanging altimeter settings, agreeing on what the current weather report is, what the active runway is, etc, would eliminate a lot of need for humans to talk to each other and pass the information on to dedicated computer systems on both ends (which is largely what pilots and controllers do). These systems could, and should, communicate directly. Any autonomous plane would have to be integrated this way. Situational awareness is currently bottle necked on the ability to pass information around, which happens to be something computers are good at.

A stop gap solution is putting a remote human pilot in between to enter information relayed by human atc. The main challenge is not just replacing the pilots but also ATC and standardizing the interactions between them.

This is just as true for Human pilots as well, there are plenty of judgement calls and scenarios average pilots also will fail .

Are the autonomous systems better than our best, most experienced pilots? not yet. They are however likely already better than the average pilot. Updating the autonomous software to factor in new scenario is easier than grounding pilots and getting them retrained on the simulator

The expense and availability of type certification and simulators is one of the reasons the 737 MAX was designed to handle like the older model, although it was fundamentally different. After the fixes, if the FAA require separate type rating it will be very expensive and slow process for all airlines.

The autonomous systems will behave consistently under a given scenario, human responses and errors may vary. Fixing the former is easier than quality checking the later.

Planes already fly a lot on their own, they also crash on their own already too, due to software bugs or faulty sensors, 737 MAX is software bug from a faulty sensor reading on the autonomous system after all. It is a blurry line between what we have and a fully autonomous system. The risks are already there, might as well embrace it and test the software lot more rigorously than depend on the human to intervene and override as we do now.

Tech for automating ATC has been available for decades. The biggest resistance has come from the major US airlines, and associated regulatory capture. Automating ATC would significantly increase efficiency and landing slots. Increasing landing slots at major US airlines would create more competition for the majors. Monopolizing landing slots helps majors suppress competition on key routes and hubs. Wonder if these dynamics are moot now however post covid...

>Autonomous systems are great when things are going well, but at the end of the day they fare very poorly in edge cases sometimes with catastrophic results

That describes humans as well.

Depending on the circumstances, a robotic system might offer a better response than the humans: it does not have emotions.

If anything, robotic systems might be more predictable.

Automated cars have decidedly NOT been predictable in their failure modes

I think the best case scenario is to have an autonomous system that only steps in when the human pilot is doing something wrong. Then the pilots stay well-trained, while getting the benefits of autopilot. You could then manually give autopilot control for short periods in times when it is too dangerous for human pilots, ex. windy dark landings.

This has sort of been tried before for some time with occasional disastrous results. When the pilot was trying to do one thing to save the aircraft and the autonomous system was trying to do another, crashes have resulted. It's likely that either one would have been ok on their own, but the combination can be bad.

I know the rebuttal to this is "override the pilot" but it's really not so simple. Sometimes the pilot has switched off part but not all of an autopilot system, and it makes sense to maintain that capability.

It's a very difficult problem to solve.

I believe this is speculated to have happened with an Airbus doing a low pass or touch-and-go at an airshow. The aircraft seemed to believe it was landing, the captain was requesting TOGA thrust.

Airbus auto throttle is a nice idea when it works (once you understand it), but there's certainly something uncanny about the idea of not having direct control of the primary and secondary flight controls.

If I was coming in with one engine out, the last thing I'd want is the aircraft trying to out-think me and tell me what's right, or decide I'm at risk of stalling (2 feet above the runway), and apply power (which with only one engine available would risk skewing the landing).

This reads like 737-Max MCAS. Who/what decides that 'the pilot is doing something wrong'?

Well there’s an airplane manufacturer in Toulouse, France, named Airbus...

Well this is where I think autonomous planes would be superior. First off they could make determinations on their own, the old have three or more computers making decisions and voting, as well as be constantly updating and receiving from ground stations. This can be vital in a situation where the computer decides it no longer trusts its own inputs.

It all comes down to this, it won't be solved quickly, but with proper flight routing and use of specialized airports it can all be resolved without endangering the public. For the most part planes are not falling out of the sky on a regular basis but it would be interesting to see statistics on the number of registered events per miles flown.

Human ATC could still be involved, the words used for events can be very specific and the system can know how to respond to all of this. Automated flights would have special call signs and even may end up handled by assigned ATC to keep them separate.

>Autonomous systems are great when things are going well, but at the end of the day they fare very poorly in edge cases sometimes with catastrophic results. [...] All in all, while technologically I feel this could have been pulled off yesterday, I doubt we will see fully autonomous aircraft operating long routes in controlled airspace anytime soon.

You're correct that edge cases are tricky, but you've made the mistake of forgetting the other possible half of dealing with them that full robotics makes possible: simply sufficiently reduce the cost of edge failure. An aircraft with humans on board crashing is a tragedy, even if it's in the middle of nowhere. In general a cargo aircraft carrying mass produced goods crashing in the middle of nowhere is merely a financial statistic. By far the trickiest part of flying is takeoff/landing and navigation on the ground. But players with shipping volume at the level of FedEx, UPS, USPS or probably even Amazon/Walmart alone can flat out afford their own dedicated airports. With no existing path dependency, they can easily build them quite a ways from population.

This completely changes the calculus and possible responses. If something goes badly wrong on takeoff/landing or at the automated airport itself, it just costs money. There is financial incentive for the players to figure it out, but it doesn't involve anyone else. And once at flight given modern aircraft glide ratios, there is no place on the planet that doesn't have ocean/empty land within glide range. For human flights, ditching the plane as safely as possible, or hopefully making it all the way to an airport, are critical considerations. For a robot plane though experiencing an all-engine out just specifically ordering it to crash somewhere without anyone around is mostly okay (there will be some environmental pollution, but it'd be trivial on humanity's list of environmental damage, and the operator could be fined for cleanup). Robot aircraft could even flat out be equipped with self-destruct systems like the range safety systems rockets have.

So for the specific domain of dedicated cargo craft it seems reasonable that fully autonomous control could exist far ahead of human transport simply by virtue of being more expendable and having infrastructure purpose built exclusively for it in ways that are impractical elsewhere.

Autopilots also rely on the instruments working perfectly. A human can usually safely land a plane even when all the instruments (and even the engines) fail.