Sigh. Stop rationalizing your pet Google peeves.
> The software/AI is secondary to the ability of the solution to work in hundreds of millions of cars in hundreds of countries (all with varying regulations) under an infinite array of conditions.
No it doesn't. It's perfectly fine if it works even with one car model if that spurs demand and further development. It's perfectly fine if the self-driving system initially only works in "good" conditions, since a driver should always be prepared to take over. It's perfectly fine if the system detects ice or snow and sets off a shit-ton of alarms warning the driver to take over. No initial offering of a brand new technology ever deals with every imaginable scenario.
> The car companies have decades of experience at this.
The car companies have decades of experience designing mechanical and electromechanical systems. Their software expertise sucks. Other than the Tesla, have you seen the dismal state of modern car software? Even BMW uses a system of rotary dials, and "apps" are just now becoming commonplace.
Robust self-driving cars involve an unimaginable amount of software complexity. There are machine vision algorithms, sensor fusion algorithms, estimation and planning algorithms, learning algorithms, and probably a bevy of common-sense heuristics. There needs to be extensive testing of these algorithms, and careful bounds on their latency and reliability. These are computer science problems, not automotive problems.
> Google has none.
Google has cash. They can acquire people and expertise. See [1]
> And even worse they have a penchant for beta software and a concerning inability to ship products properly.
May I be bold enough to presume that the vast array of smart people they've hired might have realized that Gmail and self-driving cars require different approaches to software development....
[1] http://www.wired.com/autopia/2012/11/ron-medford-google-nhts...