To prompt some discussion, say someone tried to build "self-driving" cars in the following way: put a big brain full of rat neurons in a vat, hook it up to a car, and then train it to navigate.

I don't think that would solve any of the problems of driverless cars. The problems of driverless cars are about handling a lot of edge cases that don't come up often that a human can solve by having a decent mental model of the situation he's facing.

Why should we expect those rat-brain to have good mental models?

According to Musk self-driving cars while need to have a 10x reduction in traffic accidents to be viable. That won't work if you basically train mouse brains to do the same thing that humans do.