To be fair, maintaining straight and level flight— i.e., not falling out of the sky— is the easiest part of flying an airplane. Computers have been good at that for a while; in fact, that’s how airline flights are controlled these days during basically all of cruise flight (by autopilot).

Takeoff and landing are the tricky parts.

(I say this as someone who’s firmly on the skeptical side of the spectrum regarding the near-term prospects for fully autonomous vehicle control.)

Garmin has already gained FAA approval for an emergency autoland system that integrates with most modern private aircraft. It is designed to be used in case of an emergency where the pilot has "stopped functioning". It finds the nearest appropriate airport, announces it's intentions to ATC and then flies to and lands at that airport.

https://www.garmin.com/en-US/autonomi/

There are landing procedures where the pilot simply follows the machine instructions for a long while already. And it works perfectly well where it's available.

Also, the take-off being easier, it's not conceivable that a machine couldn't do it. The reason there is no such machine today is exactly because it's easier, so pilots never need any help.

Instead, the hard part isn't flying. it's all the things an airplane does while landed. It's taxing, interacting with other planes, interacting with other vehicles (there are a lot of them in an airport), parking, etc. If the airport is entirely built for automation, I imagine that will become much easier.

That’s one thing that boggles my mind about airports: it seems like they’re all designed from the ground-up as if it’s the first airport ever built. Terminals too.

I get that the biggest airports have grown organically, but I’m surprised a city building a new terminal or airport (Berlin anyone?) can’t order a standard proven design like they’re building a McDonald’s or Holiday Inn.

Airports are built to fit the space they're allocated. And runways have to be sited accounting for approach patterns that can be many hundreds of miles. Given the variation in geography of the planet Earth there's zero one-size fits all construction that can be done.

Airports are standardized to the degree they can be standardized. There are entire sections of the Human Factors and Ergonomics Design Handbook devoted to airport terminals and other large transportation spaces. But these spaces have to be designed to fit the place that they are constructed in.

Berlin is disaster and running joke on so many levels.

If the city was competent about the project, they would have actually hired people who knew what they were doing. While new terminals and airports are not as standard as Holiday Inn, there are plenty of constraints in most cities [1], that do not allow a lot of room for radical change, and design usually come out not widely different.

[1] https://www.youtube.com/watch?v=5QZ3eozyQfU

> Also, the take-off being easier, it's not conceivable that a machine couldn't do it. The reason there is no such machine today is exactly because it's easier, so pilots never need any help.

I don't think it's just that take off is easier, but that it's also more dangerous than landing. If something happens during take-off and the computer needs help, it needs the pilot to step in quickly so you don't have time to do a computer -> person hand-off.

I don't think it would affect this proposal (no passengers on board, and package-only airports can be located far away from things to hit) but worth noting.

> If something happens during take-off and the computer needs help, it needs the pilot to step in quickly so you don't have time to do a computer -> person hand-off.

The same happens at landing, but the airplane is moving faster.

Yes, but faster is usually better in this case.

Over simplifying, there are a lot of situations during take-off and landing where stopping isn't an option. During landing, the extra speed helps you take-off and try again (go around) with minimal risk. A computer can pretty easily handle "something's weird, TOGO and have the pilot take over." Take-off doesn't have that safety margin.

> Also, the take-off being easier, it's not conceivable that a machine couldn't do it. The reason there is no such machine today is exactly because it's easier, so pilots never need any help.

It exists:

"Airbus demonstrates first fully automatic vision-based take-off"

https://www.airbus.com/newsroom/press-releases/en/2020/01/ai...

Takeoff is relatively straightforward: planes "want" to fly. Deciding if, when and how you should abort a takeoff is less straightforward.

Pilots have to learn the schematics of electrical wiring behind the dashboards. They are expected to do problem-solving in case of malfunction, and this happens really often. The first time asymetric thrust was used to pilot a plane (ailerons and elevators didn’t work), it was done without being in the books. I wonder how this can work with computers, given they have very fee sensors compared to humans (« Do you smell smoke coming from the vents? ») and they are not yet capable of reasoning. They need to be trained millions of time for each failure mode. Granted, thanks to 70 years of experience with MD, we have listed those failures quite extensively and this part is well documented.

> Do you smell smoke coming from the vents?

This is not the most common way to detect a fire in an aircraft (unless GA). The most common way is "FIRE" being indicated in the annunciator panel.

> Pilots have to learn the schematics of electrical wiring behind the dashboards.

No they are definitely do not. They will only know the wiring if they are also mechanics.

They need to know which fuses control what, but that's a much higher level.

> They are expected to do problem-solving in case of malfunction

Also no. The first thing to be done is triggering the memory items. Those are memorized. Zero thought here. Pilots have to (and generally will) fall back to what they have been instructed to do.

Next there's the checklists. They will execute the checklists line item by line item. No thinking happens here either, nor it is supposed to.

If all those fail... usually you'll read the resolution in some NTSB report.

> The first time asymetric thrust was used to pilot a plane (ailerons and elevators didn’t work), it was done without being in the books

Not sure if that was the first time, maybe first time in an _airliner_. Also in that particular incident, they had an experience commander, first officer and an instructor for that specific model of aircraft working together (the instructor was handling the thrust levers).

That was also a very extraordinary incident.

> flight crew performance was highly commendable and greatly exceeded reasonable expectations

> I wonder how this can work with computers

During the Columbia incident, the computers didn't know the wing was melting. But they detected a deviation from the flight path, which they tried to correct with control surfaces. When that failed, they added RCS thrusters too. The 'wing melting condition' wasn't programmed, but still they managed to use all available resources.

For more on how we could allow craft to 'intelligently' respond to failures, take a look at the Remote Agent, from NASA's DS-1 mission.

If a plane is unmanned, then, in case of malfunction, just safely smash it into the ground.

Handling mechanical failures and other in-flight emergencies are the tricky parts.

That depends on your definition of "tricky". If you mean something that requires human intelligence, I would say it's the complete opposite. More or less the first thing pilots start doing during an emergency is to follow lists, it's all very procedural and most failures have been predicted in advance, as well as what would need to be done to resolve them, or at least eliminate the immediate threat -- and if these procedures do not help, you are in a very bad place anyway. There are exceptions, of course.[1] I would say that the "tricky parts", as always, are very generally the situations that we did not predict and did not program for.

[1]: https://www.youtube.com/watch?v=099cHWSbAL8

There is one benefit of having people in the loop in that they come with a completely independent set of sensors for flying the aircraft. Look at the recent issues where the single sensor on the Boeing Maxs crash the aircraft. They're far from perfect for sure though(see the times people did the opposite and ignored the correct sensors causing a crash) but there's a definite benefit to having two completely different systems monitoring the data. It's unusual to see that level of safety differentiation, not only having multiple copies of a sensor to take measurements but multiple different designs and types of sensors measuring the same thing to avoid correlated failures.

> the single sensor

And this is the problem. We wouldn't have autonomous aircraft without sensor redundancies. Even pilots have multiple redundant instruments.

Having multiple instances of the same type of sensor doesn't necessary help a lot because they tend to have highly correlated failure modes. A human pilot contains entirely different sensor hardware which, although unreliable, tends to fail in different ways.

The tricky part is not decision making, but sensory awareness.

Takeoff and landing in rough conditions are tricky.

Takeoff and landing when the weather is fine is already a solved problem. So is taxiing.

https://newatlas.com/aircraft/airbus-attol-autonomous-airlin...

What's the advantages of a human compared to a computer for landing in rough weather? I would think it's a fairly logical problem. Wind is rotating the plane left, need to move the stick right. Wind is pulling the plane up, need to slow down or push the stick forward, etc. Need to make adjustments past safe levels or can't keep the plane level enough means it's too rough and the landing needs to be aborted. In my mind, it seems like the thing that's perfect for a computer. It can instantly take into consideration hundreds of precise measurements, sensors, and history, and make corrections for the precise amounts. A human has a much slower reaction time and can't make adjustments with the same degree of precision. I feel like a computer could correct a quick gust of wind before a person even has time to move their hand.

Compared to a self-driving car where you have all sorts of situations that require complex analysis, like trying to determine if that's a person lying on the highway or an empty garbage bag, this seems like it should be a walk in the park.

Autoland systems and humans have different limitations. Humans are limited by visibility, because we look out the window and control the final parts of the landing by what we see. The autopilot is limited by wind restrictions, because the control input rates of an airport are limited.

That limitation is there because the sensor inputs aren't as good as you think. Looking out the window reacts much faster to wind gusts than receiving a radio signal to determine runway line-up.

If you want to have a computer execute a full autoland in varying and gusting crosswind you'll probably need it to use cameras and computer vision, but then it becomes limited by the same visibility issues humans have.

You would think we wouldn’t have an endless string of 0-day exploits for operating systems as well but here we are, for the last 25 years. It should be a simple and logical problem as well.

An operating system has to be able to handle arbitrary data, arbitrary network traffic, and arbitrary user input, with the goal of enabling its operator to perform arbitrary computation.

That might be a slightly less tractable problem than controlling a fixed set of flight systems to maneuver an aircraft with a fixed set of degrees of freedom.

Definitely not to understate the complexity of flight systems software at all, but the analogy to operating systems seems unhelpful.

I think one of the key things computers aren't as good as people are is communication. There is a lot of info that a human can convey with a short sentence that a computer would have no chance of doing. The ability for a human to effectively communicate with ground, tower and other aircraft in inclement weather is key.

Computer systems can obviously also communicate over radio, and almost certainly at a higher information density than a human operator.

> To be fair, maintaining straight and level flight— i.e., not falling out of the sky— is the easiest part of flying an airplane.

Sure. But if it does happen, the self-driving airplane needs to have a landing plan that isn't "fall on a population center." It's not insurmountable, but does need to be baked in.

>But if it does happen, the self-driving airplane needs to have a landing plan that isn't "FALL* on a population center."*

Planes don't simply "fall" though, or more specifically it's not a limiting factor for autonomous flight because if they do Fall (due to total structural breakup) there is absolutely nothing a human pilot can do about it either. The more likely worst problem that is still within the realm of "anything can be done about it" is total engine failure (and with it the potential for failure of systems that depend upon engine power as well). Commercial jet aircraft tend to have glide ratios (how far they can go horizontally for each unit they drop vertically) of high single digits to low teens, though some go farther (the 747 for example is 17:1). So from 35000' say they'll typically be able to travel at least 60-110 miles (with some bonus for their initial velocity).

The power problem has also needed to be considered long since due to all the fly-by-wire systems, though for an autonomous cargo aircraft this may be another area worth a second look since some weight savings from normal life support might be well spent on additional redundancy.

As other posters have pointed out a fully autonomous aircraft just needs to aim for the most sparsely populated area of land or body of water. No need to worry about human life on the aircraft dramatically opens up potential ok ground areas. So a failure at cruising altitude that a piloted aircraft might survive anyway should leave a sizeable radius for an unpiloted aircraft to find a ditch point with low risk to human life. So ultimately while yeah sure, of course this should all be formalized, it shouldn't be a real blocker for this application.

So your aircraft is stricken and finds the nearest local field (putting aside commercial desires to protect the cargo too). The aircraft data tells it the field is empty but actually it’s filled with cattle/machinery/a circus/state fair/Trump boat flotilla or otherwise.

Isn't turbulence hard to predict? Could something cause an issue. I always assume pilot take over or at least monitor things more aggressively when the plane starts shaking a lot.

Turbulence doesn't mean loss of control. It's just rough air for a bit, and computers can manage through it. Modern passenger/cargo planes are designed to be naturally stable (positive static stability) in flight so that disturbances are dampened out.

Bigger issues like bad weather and other traffic would likely be programmed well in advance rather than leave that to real-time decisions.

Correct me if I'm wrong, but I was under the impression that most pilot intervention for turbulence was for passenger comfort. If your cargo is secured properly, you should be able to fly through rougher stuff and navigate around the real bad weather that's easier to detect.

Avoiding icing situations (i.e. thunderstorms) is also important, so route planning and revision is too.

You wouldn't think twice about someone driving that exact same package 3000 km though right?

Planes simply don't fall on people even in worst case scenarios.

Losing power is actually what every meat pilot trains for.

In general you're right, planes falling on people is rare, but IIRC sadly one person on the ground died when the Airbus flight PK8303 flight crashed this year in Karachi; some others were injured and most people on the flight were killed.