I don’t want to denigrate the admirable skill of air traffic controllers, but it seems to me that this work is highly suitable for automation by computers. The movements of planes with known bearing, altitude and speed are highly predictable, and the minimum horizontal and vertical distances that must be maintained between planes are classic examples of a constraint that a computer could be programmed to enforce. You wouldn’t even need machine learning for that, though ML could, of course, be brought into the game if needed. I can see that people might have a feeling of unease knowing that the safety of their lives is in the hands of a machine; but we’re also automating driving, and then, humans aren’t infallible either. Of course you’d want to build in several layers of redundancy into an automated ATC system, but that would not be an insurmountable problem.
This, even leaving out the possibility that someone might deliberately sabotage an automated system and cause multiple mid-air crashes. However, I think systems could be built to aid humans and avoid bad things happening because an air traffic controller gets tired or loses track of a plane, and I assume there is already already some kind of system to do much of the safety verification you are suggesting.
There’s a lot more to it than bringing a plane down safely to the tarmac. NATS in the UK (despite the recent problem) is moving towards a more flexible approach to routing, especially in the highly congested skies above London.
Much of this will be managed by AI, but I have no idea how this will affect the last mile. I think the ability to react to emergencies, poor piloting, and weather surprises will keep humans in the control tower for some time to come.
It seems to me that there would be an easy first step to help controllers with their workload. I don’t know whether something like that has been tried already.
A lot of what controllers are doing is telling pilots which speeds, bearings and altitudes to fly. Instead of telling them directly, they could instead enter that into a control system. That system would then tell the pilots (literally, by doing a radio call with a computer generated voice message).
Benefits: The controllers have to talk less. The pilots receive directions in a clearly pronounced, unhurried, standard format. Much better for anybody that isn’t a native speaker. The controller always can see what the last instruction they gave was.
The pilots then do a normal read back, that the controller verifies.
Benefit: Easier to check whether someone is repeating the words written on your screen than remembering what you said a minute ago. The controller can tick a checkbox that the read back was correct.
Benefit: The control system now knows what the plane should be doing, and issue an alert if the plane is doing something else. The system also knows what all the planes should be doing, and issue an alert if that will lead to planes getting to close at some point in the next minutes. Same if one plane is currently doing something it shouldn’t that will lead to conflict with the orders for another plane.
Obviously, the system would need to be adjusted for the amount of leeway to give pilots in following the orders.
Crucially however, it would not require you to use AI or simply trust the system in another way.
Part of the difficulty of updating ATC is the need to “build the airplane in flight”. Meaning, the system you’re replacing needs to function not just up to the changeover, but beyond it. The new system has to have time to get settled in.
So you can phase in incremental changes, but a wholesale change of the entire system is a serious undertaking.
I like the idea, but I also think that, in the event of any emergency/problem, humans immediately become involved in that particular situation. Computers are great at interpreting and acting on data, but sometimes the problem is that the data is incorrect or conflicting. Experienced people are needed.
Plus/minus a bunch of details, much of enroute control is already as you say: computer messages between ATC & the pilots.
It’s not quick enough for arrival, departure, or tower control.
One power or network failure in the computer network could kill a lot of people in a big hurry.
The chaos in the UK NATS system this week doesn’t inspire confidence in the idea of handing it over to AI. At least with the existing system they could identify and fix the problem.
A power failure in a manual ATC system would do the same. It all relies on electronics.
Even with increased automation–which would have to be added in a manner that really benefits controllers–the need for humans-in-the-loop seems clear, as ATC is a vital function that ensures the safety of millions (billions?) of passengers and crew per year.
Where, exactly, would we see the benefit? Increased traffic flow? More air traffic? Fewer humans to employ?
Well, humans are not infallible, and that includes human ATCs. An example of an accident caused by an error on the part of an ATC is the 2002 Überlingen collision. So perhaps an advanced automated system, with in-built redundancies, would be safer than a human one. I’m not sure it would; that’s what I’m asking.
a) a lot of airports need all of those built in redundancies & that will cost a lot of money. For example, the March ARFF truck crash at EWR may have been prevented if the apparatus had a transponder in it; because it didn’t, it was invisible to the automated systems.
b) 98+% of the time it may be better but where humans still are better than AI are those relatively rare & unusual situations - emergencies.
This is the only justification that matters to the governments which supply ATC.
It’s 99% labor cost cutting & 1% traffic flow improvement.
Which one percent is itself cost cutting: trying to increase airport & airspace capacity wo spending money to physically make more of either.
This is a critical point. Automation requires sensors that aren’t installed universally all over the entire ATC-verse. That kind of upgrades is a megaproject of possibly decades to accomplish before humans can be taken out of towers and centers.
Whereas the current system works with humans conversing over radio and aerodrome controllers using their Mark I eyeballs.
There are manual backup methods for when the computers go down. They still need power for the radios, though.
The “radios” are scattered around the countryside connected to controllers 500 miles away. All that relies on a dense computer network to move the digitized audio between the humans and the transmitter / receiver.
ATC ops also consists of a dense network of virtual phonelines that work as intercoms between all the various controller desks that need to routinely talk to each other. All that is a VOIP network with scads of computers involved.
Replacing the human in the ATC seat w a bot-in-a-box would be only a trivial increment in redundancy & network complexity.
All the hard part is making the ATC AI bot.
Maybe this is where the next automation step could be tried?
For example, the approach controller only telling the system which approach a plane should fly, without specifying each step separately?
That’s also already routine where / when traffic is light.
They can still hand-massage more airplanes into too little space by active voice control. So that’s what happens at big busy airports.
I think you are overoptimistic about the skills of the software engineers doing these systems. Just today I was reading:
The air traffic control outage that caused thousands of flights to be grounded across the UK earlier in the week was caused by a military aircraft entering “spurious” flight data into the system, according to a report. Hundreds of thousands of passengers were affected after more than 2,000 flights were cancelled on Tuesday when the system used by the National Air Traffic Services (Nats) shut down for four hours. According to the Financial Times, four people briefed on the incident claimed that the shutdown was caused by a flight plan submitted by a UK military aircraft, which caused a meltdown in air traffic control systems used by Nats which remained unfixed for several hours.
You or I would expect that spurious data entered by one airplane would either be rejected by the system or have a very isolated effect.