Both. In the 2000's, I worked at a startup developing spectrum-sharing radio technology.[1] We ran into a range of problems similar to what I believe will hold up autonomous car development:
1) When you scale from individual autonomous devices to vast networks of them you run into emergent phenomena that you simply could not predict ahead of time.
2) You have to spend enormous engineering effort protecting incumbents. E.g. there has been a ton of work on figuring out how to protect wireless microphones used by churches and stadiums. They are intermittent and have a "dumb receiver" (that never transmits), making them very difficult to handle using a "listen before talk" approach. And that's for fricking church PAs. Imagine the push-back when your technological edge case runs over a kid.
3) You can't deploy commercially until you change the regulatory regime. That can take years to decades if you're asking for seismic shifts in policy.
I would predict another 5-10 years before the technology works correctly at scale, rather than under controlled conditions. Then maybe another 5-10 years after that for limited regulatory approval, and maybe 10-25 years for widespread regulatory approval. I don't expect to be able to catch a self-driving cab in New York or Chicago before 2040.
[1] The end goal was technologies that could obsolete fixed spectrum allocation by letting networks of cognitive radios sort out for themselves what frequencies to use. There are various narrower applications, such as technologies that can take advantage of TV "whitespaces."