You might claim I am naïve for saying this, but I would have the right to counter with: why have you not done it, then?
The concept of autonomous vehicles has come a long way: from cartoons and dreams to BYD, Waymo, Google and a whole host of vehicle manufacturers making (potential) people-killers and sticking them on the road. These manufacturers have had a hard time integrating them into existing traffic because human beings, animals, weather and "climate".
The technology they are using to try and make vehicles work autonomously is thousands of underpaid slaves in Kenya, India and other countries who try to "teach" those things to see objects and situations and make good decisions like a perfect, perfectly well-trained and rational human driver. That is why they are failing like Chinese rebar.
How would I do it?
There are billions of smartphones, aircraft, ships and other devices which use inertial navigation systems in partial or complete form to find their locations in space. These systems integrate one or more of the following devices: barometers, compasses (or) magnetometers, gyroscopes, accelerometers, star maps and/or others to find location without having to consult the American, European, Russian, Japanese or Chinese navigation satellites.
These items are very cheap due to mass production, so even at the scale of cars, their inclusion will not affect vehicle prices much if at all.
The next (again, obvious) step to take here would be to create a giant database of all the coordinates of all the vehicle-navigable routes and places worldwide which includes road and terrain characteristics, side of the roads to keep, speed limits, lane and road boundaries among others, with the most important in the list being the coordinates and road boundaries.
This database can be broken down into fragments of different countries to reduce computational power required by onboard computers with these inertial navigation systems. All the driver would need to do as a result would be to tap on "Fucking" on the screen and off goes the car.
Any vibration-induced inaccuracies have solutions for them. I will not consider them here.
You might also need to compensate for vehicle width.
Road furniture such as wireless beacons can be embedded or planted at intervals along every road that the vehicles can communicate with and use as guides, but that is impractical because of the cost, and some random humans will pluck them off for any number of reasons.
Do you need millions of strong GPUs? No.
Do you need millions of "AI" slaves? No too. hehehe
Can this technology work on I.C.E vehicles? Absolutely.
You know the problem, then, do you not?
I do.
Money.
As demonsrated by Google with their Street View feature, it is entirely possible to map all these routes manually and in a cost-effective manner. All you will need is a few men willing to do the mindless task of driving around all day every day until the mapping is complete and to occassionally update those maps with new or closed routes. You can also do it like Open Street Maps and get a team of volunteers from across the planet to do it for you. Hell, you can even pull the data from Open Street Map and othe mapping service providers because why not?