Artifical Intelegence in Railroading...

The tl/dr: There will always be human oversight in railroading operations until AI's become self-aware and human themselves, and even then...

Now that we got the impatient riff-raff out of the room, we can get to the real meat of my argument.

Lets start with the current state of AI in transportation, which is in car technology. Right now, it's still problematic:

• It's not real AI, but more like synthetic intelligence of limited capacity. Not there yet! AI is AI when it becomes self-aware.
• And yes, very limited capacity. Half the systems don't even have the right equipment to do the job at all!
• Tesla's getting sued because of two "AutoPilot" related decapitations and many other issues. They can't detect tractor trailers, road work zones, people...
• Uber Autopilot's getting stuck by having a safety cone on the car hood... and don't forget about the times where they crossed into an active fire fighting area.

Now take all those problems and apply them to the rails. Yes, you could have them run according to PTC and fail-safe to full emergency stop. But what if you strike a deer? A moose? A car at a grade crossing? Another train (that could even be derailing!)?

The tech's not there yet. And I think it won't be for many decades. However, sensor tech can be refined over those decades and a train engine has plenty of room. If it can detect an obstruction over a mile away, that gives it time to slow down enough that emergency braking WOULD "stop it on a dime." It would not take away an engineer's job, but it will definitely automate a lot of jobs an engineer would need to do and provide a bit more oversight.

Would we have AI's running the rails? Well, only if they were to the same level as science fiction depicts them, and they were able to sit down at the controls to tell the limited "SI" what to do.

I don't think we're talking about self-aware AI; that's for the sci fi novels. But Machine Learning (ML) systems are getting really good and can improve safety and productivity on something as straightforward as a railroad. Today it will probably be a hybrid system with a human ready to hit an override button, but soon ML will prove to be good enough to handle the whole job.

That decapitation incident was almost 5 years ago; you can be sure that the systems, and training data sets, and computational power have leapt a few orders of magnitude since then. As we know a computer is far less likely to be distracted (Chatsworth), stoned (Chase), or lose situational awareness (N. Philly, Washington).

Railroad operations should be easier than self driving -- no steering involved, less weather interactions, fewer bikes/pedestrians/construction detours, etc.

Further, I think railroad dispatching is the exact type of mathematical puzzle that ML systems should be able to handle extremely efficiently, even today.

AI has no place in the cab of a locomotive beyond what PTC already does.

The real place you'll see it is in operations management. Most dispatch centers (be it railroads, airlines, or OTR trucking) already have real-time decision making and adhoc predictive modeling tools for decades.

I don't think we'll ever get to automated dispatching, but I can definitely see a place for more complex "run train 123 at X speed until Y and then run at Z speed" type solutioning being implemented.

I think AI assistants are already in cabs overseas and here's an overview of how BNSF is using AI.

Not happening anytime soon. Look at the Tesla mess – can't even handle safety cones. Trains hitting deer? Nightmare. But, hey, sensors might step up the game.

>> AI is AI when it becomes self-aware

It's possible you're thinking of a theoretical concept called "AGI" or artificial general intelligence. Applied use of AI today is more about data-driven automation, predictive analytics, the use of models to effectively scale our ability to process a lot of data, making it useful for our purposes. I see it as a tool to be used by people, feeding people with information faster and more accurately than was previously possible, and not a technology that takes over for people.

Even then, it hasn't even been sorted out theoretically whether fully implemented AGI would create something that is "self-aware". Today's applied uses of AI are not even in the same ballpark as, say, the fictional character HAL from 2001. And even with that character, it's not clear whether it's self-aware.

I use AI for counting rail cars from video, sensing whether at-grade crossings are blocked and by what, turning Lidar point clouds into categorized features for mapping, models that can provide inspection results from drone photography, recording near misses from front facing cameras, stuff like that. And for those things AI (and within it, machine learning and neural network-driven deep learning) is very effective.

Trains hitting a deer is a red spot on the front of the engine.

Trains hitting a car would be more of a challenge but that could be detected and the train stopped. There just wouldn't necessarily be anybody to immediately investigate.

Trains hitting other trains can be avoided to the extent they can be avoided today.

There have been trains run by AI, with zero human intervention, for decades in various passenger and transit operations.

Generally what AI does today is fuel and train handling management (not making any judgement on how well it does those things, just that that's what the focus of the AI is.) There is limited AI in the network management. Various systems have been under development to automate the dispatching functions with minimal success. You can teach AI to move trains, teaching it to move trains efficiently as or more efficiently than a human dispatcher is very, very tough. By the time you get a system trying to optimize a chunk of a multi-track railroad several dispatcher territories in size, it the system often takes longer to calculate the next move than it takes for a trains to get to the next decision point. Back when I was minimally involved in that effort, the estimate was that if you gave a big desktop computer a railroad the size of a typical class one, and all the positions of the trains now, then told it to figure out how it should line the switches and signals to minimize delay across the railroad, it would take approximately 100 years for that single computer to figure out the next move. To be effective the computer would have to optimize the system about every 15-30 secs.
Millions of dollars and decades of development has been spent trying to do that and I haven't seen anybody come up with a fully automated system.