I imagine for well-plied, fixed routes (interstates between two warehouses) there could be training runs (with a driver) to check edge cases along routes befor

I imagine for well-plied, fixed routes (interstates between two warehouses) there could be training runs (with a driver) to check edge cases along routes before going to prod. There could also be periodic retraining with drivers, as well as offline model retraining with the data that the trucks will be collecting continuously. This simplifies the problem significantly (even though it doesn't solve it completely -- there's always the 1% off case). I don't think one would just plop a general system on any route without human-supervised customization -- at least we're not there yet.