Self driving cars just feels like a bunch of engineers wanting to solve a really hard and really interesting problem. But is it really the right problem to solve? I don’t think so. Currently we have:

>other unpredictable human-driven vehicles on the road

>no sensors or passive/active guides in the physical infrastructure — guidance is completely determined onboard

>no signage designed for CV — everything written in plain English, so you have to waste compute on object detection and OCR. And, you have a constant and unending need to collect mapping data

>no vehicle-to-vehicle communication whatsoever

Design the roads and signage for self driving cars, enforce an industry-wide V2V standard with NHTSA, and start small in some test areas that only allow other vehicles that are part of the botnet (perhaps not necessarily self-driving, but at least part of the V2V network). That’s a reasonable problem space. The current one is unworkable — the models won’t converge and they’ll never make it out of “beta”.