Every hole in the cheese slices, including operator error, is important. And operator error is an important kind of error in particular, not because we need to prosecute somebody but because many forms of operator error can be mitigated using better system design, but you won't figure that out if you blind yourself to operator errors.
Throwing up our hands about how hard a problem this is seems like a ridiculous position to stake. If we can design self-driving cars, that operate safely in far less controlled environments, we can figure out the linear offset of a large vehicle traveling along a physical rail running along a fixed pre-determined path.
Self-driving cars aren't supposed to drive quicker than they can look ahead, which makes the problem way easier because you don't have to figure out how to detect obstacles ahead of time and can simply rely on vehicle-local information to stop if necessary. And because cars can brake way more rapidly than trains, it means you can still drive reasonably quickly even within that limitation.
Plus in an emergency, cars have steering as an additional degree of freedom to attempt getting out of any nasty situation that may have developed.
(And of course the socially expected safety standards for cars are quite a bit lower than for public transport…)
Trains on the other hand have braking distances on the order of a kilometre or so, and high-speed trains at full speed several kilometres, so trying to get by on local data only is hopeless and the usual strategy is instead having central signalling system that can accurately track trains and coordinate their movements.
> If we can design self-driving cars
We can not design self-driving cars to the safety standard that is required for railways. They are way less safe than the current half-manual system for railways. What you suggest is using less automatic safety systems than what we currently do.
But there was a recent incident in Germany as well (https://en.wikipedia.org/wiki/Burgrain_train_derailment), which is somewhat similar to the earlier accident in Spain this year. Targeting high value projects, while not paying for proper maintenance is not something that only happens in a single country.
> but you can't say that it has ever happened in Japan.
That sounds very unlikely to me, and in fact Wikipedia indicates otherwise: https://en.wikipedia.org/wiki/Shigaraki_train_disaster
This you? Who's fault is it then? You said the tech doesn't exist (lets take that at face-value for the sake of your argument), so then if not tech, then what? You're just left with people. So are you saying it is their fault or not? Or is it just nobody's?
In addition: https://news.ycombinator.com/item?id=50297950 https://news.ycombinator.com/item?id=50293093
It's a mixed bag. That can also cause safety violations, they were suicidal incidents, because punctuality is placed over safety, in Japan. https://en.wikipedia.org/wiki/Amagasaki_derailment
Japan railway crew have stricter procedures when they are alone: https://en.wikipedia.org/wiki/Pointing_and_calling
Current operation of a train is basically self-driving and supervising driver. So if you suggest "just use self-driving", you make the system less safe, not more, because it means just getting rid of the human supervisor. This is not currently done, because it would be too insecure. It might be done in the future, but rather because less safety is going to be accepted in favor for getting rid of the costly human.
A self-driving train will happily cause a crash, if the target is "drive with high-speed in that segment currently occupied by another train". The issue is that that becomes a target, not that there is a human in the loop.
Sure, if you allow trains to ignore safety completely, then accidents will occur. That is a ridiculous statement.
I don't think that's relevant to trains. Cars can avoid a collision much more easily than trains can. Cars can swerve, and if they are at a distance at which line of sight tells them they will hit an obstacle, they are moving slow enough to stop before hitting that obstacle. Trains have to stay on tracks, and they need up to 4 kilometers to stop moving.
The constraints are very different, so the solutions and tradeoffs will be different too.
That is completely false. PTC does that, and we've had it for several decades. There are several self-driving cars operating in the world right now. I have never heard of a PTC system causing accidents; those occur in places where the system isn't required (like entering Grand Central).
You keep pretending that avoiding train collisions is too magical to achieve with current tech, but train traffic is easy to limit (but slightly harder to optimize), and systems exist to do that, because rail is unique amongst all forms of travel: if you drop speed to zero for any trains at risk of collision, they are safe. Planes will plummet; most cars are human-operated and you can get rear-ended, etc.
I think you are misunderstanding the sentence you quoted.
It says something about the safety of self-driving cars. That is they don’t meet the standards required for railways. In other words the expected safety for a self-driving CAR is lower than the expected safety of a RAILWAY. (Caps for emphasis, sorry about the shouting.)
Your comment reads as if the quote claims that we cannot make “self-driving trains”. It does not say that.
> You keep pretending that avoiding train collisions is too magical to achieve with current tech
I don’t think that they do that. Definietly not in the comment you responded to.
And they somehow work with radar, lidar and GPS from inside the train? No they work with classical block based systems and on signalboxes. So all the safety systems are outside of a train.
> You keep pretending that avoiding train collisions is too magical to achieve with current tech
Huh?? Railway interlocking exists for more than a century. It's very safe and I didn't wrote anything contradicting that.
If I can have reasonably accurate gps readings of my phone while I'm on the train, there is zero reason for train companies to not have systems that are at least as accurate.
Aviation also doesn't exclusively rely on GNSS positioning, so why should train protection, given that it's also extremely safety critical?
My car's GPS does, using dead reckoning.
Trains can also use dead reckoning.
/me checks notes, I've heard that they move on inflexible things called "rails" that almost never change the location.
Also the hard part is knowing the position of the others and their trajectory, which changes in between moving towards them and is done by a third party. The normal passing distance is smaller than the width of your vehicle and you can only change your own direction every few kilometers. That means that the others trajectory towards you can only become known after you are already past your last opportunity to change your trajectory and even within your breaking distance. Good luck!
[1] Friction for cars on the road is 0.010-0.015, for trains 0.001-0.002; train speed 120 km/h * 7.5 = 900 km/h equivalent for car; that's not even a high-speed train.
"Reasonably accurate GPS" trying to determine which track you're on at 150 mph is not as simple as you make it sound. Yeah, it can determine how far along the track you are. But when you get two trains moving toward each other, each at 150 MPH, do you really trust that GPS to tell you that they're on different tracks?
Keeping track of which parallel track a train was on automatically was a solved problem a hundred years ago, using relays and lights and hard-wired logic. Adding computers and GPS greatly increases the precision with which dispatch knows position and speed along the track, but knowing which track has not changed in the last century.
What we're talking about here is some sort of TCAS-equivalent for the case where the central signalling is down – while it's easy to accurately localise trains with the aid of a central signalling system that authoritatively knows about the infrastructure layout, it's quite more difficult when you're trying to do the same thing in a decentralised, train-local fashion.
Single track means that at least outside of stations, collisions can be reliably identified without much danger of false positives (and for station areas you can just institute a blanket speed limit to compensate for the gap in RCAS coverage, which might not even be any lower than the speed limit required for traversing diverging/spring-operated/hand-operated points there anyway), and no already existing full-scale signalling system means that RCAS actually provides a significant safety benefit and is therefore likely worthwhile from a cost-benefit-analysis point of view, too, if you're actually regularly operating multiple trains simultaneously on that kind of infrastructure.
Whereas on multi-track mainline railways avoiding false positives is much harder and normally you should already have proper signalling and train protection systems in place anyway, so the additional safety benefits of RCAS would be much more marginal.
Oh, they can, but it's usually a very traumatic event.
It's not impossible. But handwaving GPS placement of trains at the correct location, direction, and track deserves more than a handwave as trivial.
Presumably switches already report their position to somewhere.
It's absolutely unacceptable for this problem exist in today's world.
Yeah and this somewhere already knows whether a crash is going to happen. In which case your solution brings no benefit. You can't rely on the correctness of a system, which you want to make an alternative to, to help in the case the original system fails.
This discussion focuses on a non-problem (granted partly due to an imprecise comment of mine). The system knows where all the trains are and also that a crash will happen, if the decision is made, that then indeed was made.
What is needed is that the system needs to convince the operator there is still a problem, after it proved itself unreliable. My suggestion would be a log of the constraints the prover encountered while determining whether there is a conflict. This certainly already exists somewhere in the background, but I am not aware whether that is currently shown to operators in deployed systems. It does exist somewhat in the research systems we use in university.
The thing is that in modern automated dispatching you only have the same world model as the automatic system, so when there is an error you have the same world model as the system, but that just detected, that that does not equal the actual state of the world. And often the error is in fact a false positive.
But now, I actually think that this accident was caused by several systems being in effect: https://news.ycombinator.com/item?id=50298796
When Google Maps gets wrong what lane I'm in, nobody dies. It can be right 99% of the time, that's good enough. It's not good enough for a train system, though - at least, I suspect it's not. I'd want a lot more than "it works in my car" before I trusted it.
On the other hand this is a complete non-issue and solved for more than a century on the network level. So there is zero need to invent satellite or vision based train detection at worst even using statistical systems, when you can use deterministic systems, for example using the fact that a train is out of metal, so it's detectable by the train closing a electric circuit between the rails. Or using the sensors that are installed at every signal, switch or other positions for precisely this question.
As a passenger, I'm very glad software engineers haven't (fully) taken over aviation and railway safety yet.
Forests are irrelevant; rail right-of-way contains no trees, and is wide enough to assure clear "view" of satellites.
Absolute nonsense. Do you have any idea how long some tunnels are?
But yes there does not need to be a limitation on there only being a single train in a tunnel.
No, what I describe has nothing to do with the companies, that's the research state. It's just an indictment of you making bold statement about a subject, you don't now enough about.
> I can have reasonably accurate gps readings of my phone
Would you dare to cross on a rope over a deep valley, blind using only GPS readings of your phone? Because that's kind what you are suggesting, expect you also move a few hundred other people.
Hang on, this is nothing at all like that - trains run on fixed tracks. The only bit of data we are interested in for collision avoidance is the linear offset along the track. Even the shittiest GPS fix of all-time would give a good-enough intersection with a fixed path to avoid crashing into a second stationary train in the open country-side (do you see any satellite line-of-sight blockers in the crash footage?)
So that is naive, but assuming that would be all that you need, with GPS you would also need the positions of all other trains. But why would you transmit all this information to every train, instead of just computing the collisions in a central node? (That's what is done now.) A better fit for what you suggest would be radar, and this does exist. [1] But that's where your original assumption is naive. You regularly pass other trains within <2m at high speeds, so linear offset is not all you need.
[1] https://www.dlr.de/en/kn/research-transfer/projects/next-generation-train/rcas-railway-collision-avoidance-systemBecause central nodes can fail (as they did here, both human and machine), and providing this redundant data allows both the computer and the human-in-the-loop on the fast-moving chunk of steel to independently validate the decisions coming from the central node.
> You regularly pass other trains within <2m at high speeds, so linear offset is not all you need.
In this collision, both trains were on the same track. So while we obviously care about adjacency issues in the general case, this specific failure mode can be detected purely via a linear offset.
> In this collision, both trains were on the same track.
Yeah, but how do you distinguish that? So when I understand you correctly, what you want is: Find another moving chunk of metal in the radius of 2km, which moves toward you, but not in a direct or ballistic trajectory, but in a arbitrary complicated trajectory, but which can lookup in a database. Also you are operating in the real world, so there are potentially houses, trees and mountains in between. I guess that would be possible, but it would be really expensive. Your best bet would be to adapt a missile detection and targeting system from the military. A system which does exist and is somewhat similar is RCAS (https://www.dlr.de/en/kn/research-transfer/projects/next-gen...).
By enumerating the tracks?
I feel like you are intentionally trying to turn this into a spherical cow situation by insisting on solutions that work in an information vacuum. All of this sort of problem in the real world is solved by sensor fusion from multiple data sources.
Trains are captive to a particular track - they can only change tracks at specific pre-determined junctions. With a pretty minimal sensor feed from each junction we can statically determine which track segment a given train is on. With GPS we can shrink that segment length down to a few kilometres. By sharing data persistently every single train can validate that central command-and-control is not about to run it into another train from behind...
There cases which are genuinely challenging engineering problems (i.e. derailments that impinge on adjacent tracks), but this trivial case of bumper cars is not one of them.
> By enumerating the tracks?
How do you distinguish that the <2km away detected moving train runs on your track on not on the one directly besides you?
> Trains are captive to a particular track - they can only change tracks at specific pre-determined junctions.
Yeah, but you do not know which one it runs on. Your entire point is that you do not trust the information the infrastructure knows. If you do allow to use information from the infrastructure, then why not use the sensors that already exist along the whole train network? It's also very unlikely that there is something electrically connecting the two rails of a track, that you would like to run over. So why wouldn't you trust that information, but some vague GPS signal?
> we can statically determine
But the current state of the art is, that you deterministically and provable know where a train is. Why would you use a more expensive, less safe system? The current focus to reduce errors in computing train conflicts is to prevent random bit flips due to radiation at the same time in three different computers running different software.
> I feel like you are intentionally trying to turn this into a spherical cow situation
The situation where you proposal makes sense, but the current system does not, is a spherical cow situation, that's my point.
You are the one who framed this conversation as it not being solved how to deterministically know where a train is. Everyone else in here has been kindly suggest that if perfect information is unobtainable, then we should work with what we have
> Trains are still not able to assess properly where they are, even using satellites and inside train sensors. Of course there is research on that, but it's still an unsolved problem.
*Trains* on there own! This IS a hard problem, how to do that with any other infrastructure. It IS current research and not yet solved. That is for example relevant, if you want to be able to run trains at full speed <1m tolerance, so that you can have more than one train stopping at a station <60s. All while still not paying for infrastructure and maintaining backward compatibility with a century old trains. (I mean in the future you could use the same to also couple trains at full speed, but that's still sci-fi.)
...is that something you'd ever want? In general, it is, but the way it's handled is by having those trains stop on separate tracks.
Running multiple trains through the same station track in under 60 seconds means that all passengers must embark and disembark in under 30 seconds. That's very tight. I would venture to say it's too tight and meeting this goal would make the trains worse.
(And if passengers aren't getting on or off the train, then the train isn't stopping at a station.)
> Safety critical systems need to work 100% of the time
Railway safety must be safe 100% of the time, but not available 100% of the time. Also the point of a signal box is to make decisions about the order of train assignment to track. You absolutely do not want multiple points of truths for that.
There are optimizations possible like make segments smaller etc but the original algorithm is trivial. If we can't figure things out, we make segments longer and we slow down the trains until we can. I am positive for a sufficiently large track segment and a sufficiently slow train, we can always answer the question with a no.
>Even GPS + Starlink + 1 monitor (is that 50k€?) could have prevented this.
My reading on that sentence was that even a simple setup would have been sufficient to prevent this. Not that it should be the only thing.
- I don't think X would actually work, without actually eventually recreating part of the current railway control system, and the current rail control system alone already works better.
- I don't think another less reliable more costly safety system helps, when the current one already stated that there is a problem, and it was ignored.
X on its own would have solved this particular crash, yes.
X + existing setup would solve this crash + all the other crashed the existing setup solves.
There are also other solutions, e.g. LZB uses a phase change of the signal every 100 meters as an additional precision signal.
Isn't the ability to do so the basis of operation for e.g. the German LZB train protection system, which has been around since the 1960s?
If you mean that train protection systems don't (exclusively) rely on self-reported positions, your statement makes a bit more sense, but even that is changing with e.g. newer ETCS levels which do rely on self-positioning more and more.
I did not say "train systems are ...", I said "trains are ...". This is referring to e.g. that odometry based on wheels results in a drift. GNSS is to imprecise and unreliable. I'm aware that you simply use balises, that's what I meant with "markers in the track". However that's based on the infrastructure again, so it's again not trains on their own.
https://news.ycombinator.com/item?id=50293093
It was simply supposed to be an example how GPS sounds like it should easily solve a problem, but does not without additional measures, that then do not need GPS anymore.
And some problems are just _open_ to outside looks. My assistant tells me that the room for error in trains on one dimension ('left - right') is mere 10s of centimeters. I understand that problem. But what track you are on, isn't that obvious to the human in the cabin? (Again, simplifying, not flaming. Production ready safety is _hard_.) So could a vision based solution help count tracks from left to right? (Assistant: "US PTC, on about 54,000 route-miles, uses GPS for the lead locomotive plus crew-entered track information and switch position".) So you use the dimension that matters most in crashes (forwards / backwards) from satellite, augmented with vision or human based left-right positioning. Some trains already have sensors to sense things like leaves on the track to adjust speeding and braking... So the sensor based approach is already there.
Perhaps I am read as suggesting that the problem is easy, and that is not my intention. But for cars this was solved decades ago for routing and years ago for self-driving. I use that daily. Flying is organised with a very clear 6 sigma approach. Crashes are nearly always situations where events / risks stack. Trains (in Spain)...?
Tl;dr - Safety is a function of (human; process; technology) and there is clearly something lacking for trains, when compared to other types of transport.
> Would you mind sharing some more of the why it is still an open problem?
Because for train sensors, there is a small drift which accumulates too much, hence it is synchronized by magnetic position markers in the tracks.
> I understand satellite positioning isn't definitive
Yes, and the point is it might be some solution if we would have the safety systems from 1860. But we do not and hence it needs to be more accurate than the current system, and it is not and in addition also more expensive (and subject to remote influence and reliance on potentially adversary nation states).
> Sorry for the bleak introduction of economics
No, economics is basically the primary constraint for safety. We could pay million people to make sure a single person does not stumble if we want total safety, but that's completely pointless, and also these million people would starve, because no one is producing food anymore.
> But even as a tertiary or lower risk mitigating factor
But for that it's too expensive and still unreliable. There is RCAS[1], which I already linked elsewhere in the thread. That's basically cutting edge and too expensive to introduce everywhere. Also previous safety systems are already not rolled out enough, which is causing these accidents, so rolling out yet another expensive system is not going to improve much.
> But what track you are on, isn't that obvious to the human in the cabin?
Yes, but if it is already known, why use GPS to calculate the same information with more error?
> So could a vision based solution help count tracks from left to right?
If the train driver or dispatcher makes use of vision to assess the train position, the error will be higher than the current sensor/electric-resistance based system.
> Assistant: "US PTC [...]
BTW, that seems to be a verbatim quote from https://www.railwayage.com/cs/the-way-forward/
> US PTC
According to Wikipedia, ASFA being introduced in 1975 is rolled out in the entire Spain network. ETCS which is the European standardization seems to be rolled out on all high-speed lines or currently being rolled out.
> So you use the dimension that matters most in crashes (forwards / backwards) from satellite, augmented with vision or human based left-right positioning.
Duh, that just sounds like trains would be operating blindly in space currently and need to assess their approximate position.
> Perhaps I am read as suggesting that the problem is easy, and that is not my intention.
No, the issue is that you are suggesting far more complicated solutions to already solved problems, with simpler systems in operation for decades. The current system is that nobody in the train even has a say on where the train goes, so the problem you are trying to prevent doesn't even exist. This IS why it's way more safer than a car. There are movement authorities, that define from where you are allowed (and have to) drive with your train. The problem is not, that the train driver isn't able to know where he is and hence drives wrong or too far (that would be prevented by the on-board safety system), but that the movement authority itself is faulty.
On an up-to-date signaling box (withing the past 6 decades), even the train dispatcher is no say in safety questions. You could put a monkey in charge of it. There would be chaos, but they would not cause a crash. The current job of a dispatcher in normal operation is to assign resources to prevent deadlocks and to implement vaguely defined priorities. (Basically the equivalent of the public thinking a programmer is a code monkey, while the actual point is clarifying stakeholder interests.)
The actual issue here is, that while these safety system have a very small false positive rate, they do have a too large false negative rate for there to be an acceptable availability. Hence in the latter case, the human needs to take over and then you are back to a much higher false positive rate. The issue would be alleviated by relaxing the constraints on cost per availability. Either by allowing for less availability or by reintroducing more humans, so that the system situation can be assessed by humans directly.
These accidents are caused as follows: the automatic system fails, because the internal state of the system does not match the actual state of the world. Hence the human needs to take over. But the human also needs to rely on some system to know the world state. Which systems can he trust, given that the whole situation is that the systems report that he can't trust them? So he needs to make some decision and in most cases, that decision is still correct. But there is still an error rate to that, and then something like what we are discussing happens.
> But for cars this was solved decades ago for routing and years ago for self-driving.
This is absolutely not solved for the cars, and to the sense that it is, it was solved and rolled out from the 1870s onward for railways using mechanical computing systems, the switch to using electricity happened in the 1910s.
> Flying is organised with a very clear 6 sigma approach.
You are correct, that aircraft safety systems need to be designed to a higher reliability than railways, because on a rail you can always stop in case of an error, while in the air such a choice doesn't lead to a safe state. But if you compare these safety systems, basically aircraft is operated like railways used to be operated in the 1920s. (My peers actually had a task to compare them and find ideas from one that can be applied to the other. While there is a lot of automation to be taken from rail that can be applied to the aerial industry, there is not really much that can be applied the other way.) It is also operated with the amount of personal than 1920s railways. That's because there is a lot of money in aircraft, while railways are mass-transit and need to be cheap. If you would operate today's railway with the amount of human oversight than in the past, it would be a lot safer than aircraft.
> there is clearly something lacking for trains, when compared to other types of transport.
What are you talking about? Railways are so much safer than cars. https://news.ycombinator.com/item?id=50288518 Yes, there are less accidents in planes than in trains, but this is due to what I already explained.
[1] https://www.dlr.de/en/kn/research-transfer/projects/next-generation-train/rcas-railway-collision-avoidance-systemThat's a research topic, because the infrastructure companies would like to not pay as much for maintenance and still have a high sequence of trains. Also with track based train control systems you can only have very high resolution in specific places, where you place your track-side infrastructure, but it would be nice to have that resolution anywhere you like.
I guess a much less dense and largely overground network, mainly used by fairly slow moving freight trains as opposed to highspeed passenger connections, changes requirements somewhat.
I mean they are systems that rely on GNSS in Europe, but no safety systems, that would be simply to unreliable. For example I believe the railway announcements in Germany to depend on GNSS. They are also unreliable.