Hacker News Reader

Comment by viraptor | Hacker News Reader

Completely different things. Self driving cars need to actually deal with vision and fuzzy real time response. Air traffic is a planning and scheduling task with known constraints and (in most cases) known minutes ahead. Comparing their risk is a complete apples and oranges situation.

Than again, ATC needs to deal with people talking on the radio, so the current system has a really long way to go to be completely automated.

View on HN

> Self driving cars need to actually deal with vision and fuzzy real time response. Air traffic is a planning and scheduling task with known constraints and (in most cases) known minutes ahead.

It's funny to read all of these confident comments claiming ATC is easily automated from people who obviously don't understand what ATC entails.

ATC isn't just planning and scheduling. There is a lot of quick thinking and communication with pilots. You might only be thinking of the everything-goes-perfectly-right case, but the real value of having trained ATC operators is handling all of the edge cases and making quick decisions under high pressure scenarios that may not have even been represented in the training set.

ATC is also partially a visual job. Did you ever notice that there's literally a tower at the airport for air traffic control people? The people in this tower will manage things like traffic on the ground and immediate airspace around the tower. Visual inputs and critical thinking skills are very necessary.

There are a lot of assumptions that people outside of aviation make - it reminds me of that “falsehoods programmers believe about dates and time” article that gets passed around from time to time. Off the top of my head, some easily believable falsehoods:

1. The system knows where every plane is going

2. Every plane is talking to ATC

3. Every plane that is currently taking to ATC will be reachable a minute from now

4. If you issue a plane an instruction, it will follow it

5. The planes want to go the most direct route to the destination (winds aloft can often mean direct is slower and more expensive than a more circuitous route)

6. If a plane has an emergency, they will declare an emergency.

7. Planes that are not currently talking to ATC will not fly into the regions where they are supposed to be talking to ATC

8. Planes that are not talking to ATC will not just show up and land at the airport. This happens for a variety of reasons.

9. All planes have working transponders

10. All planes are traveling from one airport and landing (once) at another.

It feels like a tractable problem from the outside, but the variety of issues ATC solves every day is staggering.

ATC here. One of my favorites is:

11. Planes have radios that can select all ten digits.

Someone's radio broke where they couldn't enter '2' into it, so we had to find frequencies along their path that they could use and where ATC could relay.

You are thinking about automating the existing system, but the current system is entirely defined by the constraint that it must be operated by humans on radios. When this constraint can be removed so are its specific edge cases. When your phone communicates with the cell tower a frequency also must be assigned, and no buttons have to be pressed to do it.

Opposing Bases a few weeks ago had feedback from someone who had a button on their transponder that didn’t work and needed a code without any 5’s in it. Good luck getting _that_ through to auto-ATC.

Can emit all bytes except for 00000101 isn't really the type of problem you see in a digital system. And even if it were, it's pretty simple.

plane 1 > assign code 4563

plane 2 > reject

plane 1 > assign code 0827

plane 2 > accept

Also assigning short codes like that isn't something likely to be necessary in an automated protocol like this. Why not just have every message sent between 2 planes include a sender_id: UUID header?

Because now we’re talking about putting deeply integrated equipment in every plane. It’s a certification and cost nightmare.

This is not a system where you get to do clean slate greenfield development. Whatever you do must work for the lowest common denominator. ATC is a fairly cheap societal expense compared to developing, certifying, installing, and maintaining systems with the level of integration you want in hundreds of thousands of diverse planes.

There are only 35,000 commercial planes on Earth. Even if installing costs $1 million per plane, that's only $35 billion.

The US has about 200,000 general aviation planes. You can’t ignore them, and you can’t just ban GA because that’s your pipeline for getting commercial pilots.

But a lot of these assumptions that are now incorrect could easily be made true if the system was automated.

Worth noting in your “if the system was automated”: There are aircraft permanently without electrical systems. There are aircraft temporarily without electrical systems.

This is no different than the current ATC system. A plane or tower can lose power too. It's not particularly hard for the software to detect a plane that isn't in communication with the rest of the swarm / not obeying commands, assign it highest priority and GTFO of its way. The key is to have the software running on all planes (which you can do with commercial aviation) rather than rely on a centralized system with a single point of failure.

> The key is to have the software running on all planes

Yeeeeah… we just went through the ADS-B mandate. It took a decade or more, cost pilots thousands and thousands of dollars, still doesn’t have 100% compliance, and does weird stuff sometimes. And this was orders of magnitude easier than any kind of two way system.

Respectfully, do you have any time in the front seat of an aircraft or a tower/TRACON position?

I have none. Did the engineers who developed autopilots start out as pilots? Did the people who invented email start out delivering mail for the post office? The point is that the new system doesn't have to look like the old system. Automating ATC isn't going to be the current ATC system, just done by computer. That makes about as much sense as designing driverless cars with a humanoid robot as driver.

nevermind misread it.

> read all of these confident comments claiming ATC is easily automated

You responded to the wrong comment then. I did not say in any place it would be easy. Just that they're very different class of problems. Nether did I say it's only planning and scheduling. Even the vision part is very different than cars. (Static in known environment vs dynamic in entirely random one)

You're arguing against others or a straw man here.

Air traffic also requires the use of visual skills - and it’s harder than driving because of the small target size and wide field of view.

“See and avoid” has a very high priority in the cockpit - not everything out there is on radar, not everything on radar is under ATC control.

Also remember that ATC is vital for emergency situations. "Your distress call is important to us, please continue screaming into the void and hopefully a miracle happens.

> Self driving cars need to actually deal with vision

Doesn't ATC also need to actually deal with vision?

Should've phrased it way better, that's true. It's a very different kind of vision when you do environment mapping and distance measuring -vs- when you do object tracking from a static location. Yes you need vision processing, but at a much higher resolution (sky is huge, planes are small) and much lower complexity. (moving objects between frames are easier to track) My point is that it's not comparable to what the cars use as vision.

We've known how to do identification and known object tracking for decades (for example https://www.academia.edu/122937237/Computer_vision_system_fo...)

Reply on news.ycombinator.com