Actually, there are plenty of autonomous trains: https://en.wikipedia.org/wiki/List_of_automated_train_system...

The problem is automatizing the existing infrastructure, 'cause it requires great effort (so, money and inconveniences).

I don't think however that the lessons there are applicable to cars as well, and while I think we could and should invest a lot on train systems, the final goal for cars should be reducing them, not replacing them with autonomous systems.

It's been done for trains. But they're already down to one person for a huge train, so the payoff isn't huge.

And that one driver now carries all the responsibility. To automate them out, you need 100% accuracy which closes in on impossible to achieve. Any lack of perfection would be down to the automator. So it's a very scary piece to take on when there's no human in the loop left to hold the bag when something goes wrong.

Computers are more accurate than humans in a number of the things that matter for trains.

As the other poster said, the main need for humans is to fix mechanical problems. If the door jams, unjam it (why can't passengers use a different door or figure out how to unjam it>) If someone is having a heart attack do CPR until medical crew arrives, then stop the train until the medical crew is off (central dispatch can teach anyone CPR over the phone and stop the train). If someone is attacked - stop the the attack (unless you are a trained police office you will probably make the situation worse trying to stop an attack)

In short there actually isn't that much that a human can do that a computer cannot do better.

> In short there actually isn't that much that a human can do that a computer cannot do better.

Passengers fixing their own doors and responding to their own emergency alarms doesn't fit most people's definitions of "better". Hence why even though train automation has been possible for decades, most trains and especially most trains travelling long distances aren't.

> To automate them out, you need 100% accuracy which closes in on impossible to achieve.

I don't follow. As Animats indicated, there are already plenty of autonomous trains that work fine.

Those are usually "closed" systems, that don't have to deal with unpredictable events that often.

And in non closed systems the set of problems that prevent fully autonomous cars is the same set of problems that prevens fully autonomous trains, ie, recognizing shit getting into the lane of travel but not panic braking for random trash.

If there is something on the track the only option is hitting it. You cannot switch track (even if there is a switch on the tracks, it can't safely operate in time) You can hit the brakes, but your braking distance is so long that you won't be able to slow down much before you hit it.

I have talked to somebody in the SBB R&D department recently about this and the problem is actually that there has to be at least one person on the train.

Imagine a door gets jammed or there are some problematic passengers. If there is someone on board they can fix the problem in a couple minutes. But on a completely automated train there is no one on board to fix the problem. One solution would be that you send out crews to arrive at smaller stations and fix the issue, and larger stations could have dedicated people, but that is already more investment and less efficient than just having someone on board.

Retrofitting automation for a train network in a wealthy country like Switzerland wouldn’t be a big problem, but it just makes more sense to leave the driver in there. Accordingly, SBB is investing much more into driver augmentation to help the drivers drive as efficiently, safely, and consistently as possible.

> It's been done for trains

Has it been done on anything more complex than 100% underground subway line with the train just going back and forth?

Yes, the DLR in London has goes overground and has run without drivers since the 80s.

Railway is a different problem space than road vehicles. Trains run on fixed tracks, has long braking distance, are much heavier, and accidents can be much more devastating. While self-driving cars may use cameras, lidar and radar for detection and brake on sight, trains need to know exactly how far they are allowed to go and what speed restrictions are ahead in order to brake in time.

Traditional lineside signals and signs do give enough information to drive a train by a computer, but the rail industry has already chosen decades ago to integrate signalling and automation with the infrastructure, which is a completely different approach than today's self-driving cars. As a result, heavy rail and metro trains will not be "self-driving" in the same way as self-driving cars.

Now, if you say trams that run on streets, then possibly. Trams are often driven by sight and has stronger braking so it is possible to apply the same principles from self-driving cars. In fact, Siemens has been developing and testing one such tram. [1] Though I think the problem space is very similar to self-driving cars. There is no steering, but it still has to handle pedestrians and other road traffic.

[1]: https://www.mobility.siemens.com/global/en/portfolio/rail/ro...

They exist, many new lines or expansions are highly automated. Retrofits are hard for two sets of reasons:

Political: union strikes (meaning if you want to automate a line you may have to accept it being closed for the duration) and public outcry over a reduction in jobs.

Cost: You often only get the benefit of the simpler domain if you can implement the whole system from scratch, if you have to deal with the existing signalling, engines, (and human-driven trains) that's a complex problem, with massive capital outlay.

> Political: union strikes (meaning if you want to automate a line you may have to accept it being closed for the duration) and public outcry over a reduction in jobs.

For local public transit subway (ny mta, path ny nj), I think the cost is worth it if it means fewer delays.

Talk to your local politicians and demand it. This means changes in state laws as well to be less union friendly in cases where automation is replacing their jobs (I'm not saying that all union friendly laws need to go, but some specific ones)

The Sydney Metro has 22 driver-less trains operating across 13 stations:

https://www.republicworld.com/world-news/australia/australia...

The problem of automating trains is easier, but train labor requirements are already very low compared to truck requirements, so there isn't much of a win there. A typical freight train has a 2-person crew carrying 200+ TEU containers.

There are self-driving trains:

https://en.wikipedia.org/wiki/List_of_automated_train_system...

We have them, but on a smaller scale, on airports like Munchen or Singapore

Even if it is technically solvable, don't underestimate the work to be done to comply with regulations. Some might even not be available yet, especially when talking about automatization with AI.

Complex non digital signals. My neighbor is a train conductor and he had an accident because he missed a signal. He almost died and had to be cut out of the cabin. Some passengers were also hurt. Imagine the outcry when a ai train would do that.

Other than that isn't there in London a autonomous train? I remember riding that in 2005 or so

The DLR is semi-autonomous. It has the ability to act autonomously, but it must be started with a button and has an attendant on-board that also fulfills other duties and can be there in the case of an incident.

Indeed, it has a grade of automation of 3 (out of 4), so it is driverless, but not unattended (as that is level 4)

According to this article (https://railinsider.co.uk/2020/10/27/making-london-undergrou...), it occasionally has to even fall back to effectively level 2 operation (i.e. operator at the front of the train) because increased passenger usage makes level 3 operation as originally built (no platform edge doors or obstacle detection systems) unsafe at certain stations during peak times.

Automated trains wouldn't do that because we can design signals so that they cannot be missed (that is if you don't have positive go/no-go you stop in time)

Automated trains do have difficulties, but they are very different from the ones you see with human drivers.

The DLR (Docklands Light Railway) is autonomous, but it's a pretty slow light rail system.

I think that automation is possible because it was built de novo and doesn't interface with the rest of the rail network.