What's always been amusing to me is that trains should be self-driving, but they always have a human in charge. I mean it's not like they have a steering wheel.
They don't always. There are self-driving trains operating full time in London and Paris as well as many other cities around the world. But they are low-speed trains operating at only around 30mph. High-speed trains run in excess of 180mph and heavy goods trains are often a mile long and weigh more than 10,000 tons. While much of the signalling is automatic these days, I think it will be a long time before we could think about running something so dangerous without a skilled human in the front ready to override the controls if necessary.
If there's between 100 and 1000 people in a train, getting rid of the driver saves between 1% and 0.1% of an hourly salary by ticket. The economic pressure to do so is very low, compared to the difficulty and legal liabilities.
Automating out a taxi/Über driver saves 50× to 1000× more money per traveler. And the infrastructure for a car ride cost way less than train + track + catenary lines + stations + antiquated ticketing system etc., so salaries represent a higher percentage of total costs.
I believe the main reason is the long tail of unexpected events. For instance: I regularly use the tram. Every few months or so, the driver has to get out to manually operate a railroad switch that does not react to the wireless control signal. If you had a driverless tram, the tram would have to wait for personnel to arrive at the switch for the manual override. Extrapolating from my experience manual overrides are probably necessary a few times a day across the entire city's tram network, so we'd be looking at massive interruptions multiple times a day that cause cascading delays throughout the whole system.
So unless you put in a lot of engineering effort to get rid of these unlikely-but-still-common problems, running a driverless operation is going to be more disruptive and less cost-efficient than paying the driver salaries.
It's also a cautionary tale for self-driving car advocates. Running a driverless train on segregated track with signalling is far easier than running a driverless car on a mixed street with lidar.
Yet it still has enough edge cases to mean it's not worthwhile