Hacker News Reader

Comment by petermcneeley | Hacker News Reader

There shouldnt even be pilots much less air traffic controllers. This is the same as the people that think subways need human drivers.

View on HN

The topic of automation always comes up when discussing ATC or aviation incidents. One says automation, another says it can’t be done. And both parties end up talking past each other.

Let’s go back in time. Long ago pilots would give position updates. Now that is done with radar and transponders. This is a form of automation.

There seems to be an uptick in ATC mistakes. Recently as this week where a tower cleared a takeoff and also cleared four aircraft to cross the same runway. So a form of automation I would like to see is something to communicate to everyone that a runway is not clear.

Air Traffic Controller here. Tour an ATC facility and talk with them. You'll see why it can't be completely automated.

Yes but 99% probably could be. Go to Vancouver sometime. All the subways are automated. But in rare cases they need to be driven remotely and every few years they need to be driven by a human.

ATC is a unique thing - if you’re interested, suggest listening to the https://www.opposingbases.com/ podcast. It’s eye opening how much complexity they deal with, and how frequent the edge cases really are.

It's the same issue that self-driving cars have. 99% can be automated. The remaining 1% is a bitch, and life-threatening.

I've heard piloting described as 99% moments of boredom, followed by 1% moments of sheer terror. If you actually could program a computer to anticipate all of the possible scenarios in that 1%, it'd be good to take the sheer terror out of the equation. But it's frightening because those are the moments when something has gone wrong and normalcy no longer applies, and you need to apply collected knowledge, wisdom, and experience to save your life and the lives of your passengers.

It is actually not like self driving at all. In the self driving case the nominal situation requires hard AI.

In aircraft the nominal situation does not require hard AI, only the emergency situation. If aircraft did not have mechanical failures the nominal situation is really well defined.

The ATC job is even more nominal since they are not actually dealing with the mechanics of the plane.

I would guess that you could make a video game that covers 99.999% of all ATC jobs and you could undoubtly with enough effort also program an AI to cover this.

With self-driving cars the nominal situation does not require hard AI. I remember sitting in a TGIF back in 2012, when Waymo was still called Google Chauffeur, and they projected what the self-driving car sees as it drives down a road. It's just boxes of potential hazards, detected largely by LIDAR but backed up by cameras and traditional computer-vision approaches.

The non-exceptional case for self-driving was a solved problem in 2012. My boss rode in a self-driving car at the time; they were available for beta testing by Googlers in Mountain View then. Heck, this is what L3 self-driving is, and is offered on the market now by Audi, BMW, Mercedes, Kia, and others.

Take all that you think and say about self driving cars and apply it to an ATC. The ATC job is probably an order of magnitude more automatable.

Again I think something that would resolve this question quite easily would be a game that covers the domain. You can actually get a game that covers ~98% of what cars have to deal with. This would be something like GTA V

You can get a game that covers 98% of ATCs have to deal with:

https://en.wikipedia.org/wiki/Flight_Control_(video_game)

It's the missing 2% that's hard.

This game looks very low budget and simplistic. Perhaps with higher budget one could get to 99.9% and then one would have a proper simulation of the jobs of ATCs. Then you hook in your AI prove that it can perform. Then you replace 99.9% of the human ATCs with this AI.

simple as.

When one read the acronym "AI", should be replaced with the synonym "statistics", which shows how contradictory about redundancy and accuracy can be the people who ate the marketing about these algorithms, and are thinking about to replace real trained intelligence, humans in this case, with it.

I dont mean AI here as in NN or LLM I mean like the traditional AI like you would find in an RTS.

But the proof is always in the pudding. I highly suspect that someone could write an AI that could play "Flight Controls" and play it flawlessly. So what does this prove? Well it proves that 98% of the work of the ATC can be automated.

> It is actually not like self driving at all. In the self driving case the nominal situation requires hard AI.

I agree with that. One could easily automate aviation like subways. However, you need to push the humans out, at least the controllers, possibly the pilots. This is far from doable and that’s why we end with the driving situation. In order to solve it, you need a lot of AI in order to deal with all the human-machine interactions.

First step in this would be to install a high-bandwidth digital comms between ground and planes. Currently they use analog AM, which is a century old! CPDLP is from the 80s, so much like SMS in phones. Nobody should even dare to market it as new.

Then again there simply isn’t anyone or any organization that is seriously pushing new technologies on planes and on the ground simultaneously. The plane-ATC interface has not gotten any update for decades and there is nobody aiming for redoing it. ATC and airlines, together with their respective manufacturers and suppliers are trapped inside their silos, unable to see trees from the forest.

Reply on news.ycombinator.com