The technology for level four has been around for a few years, just not sufficiently advanced for commercial application, see the google car. Even level 3 has seen a fair amount of actual real world use from Tesla as the article points out. There is still a ways to go before level four cars hit the road, but it is largely refinement existing tech and legal clarity.
R&D has now shifted focus to deal with level five driving. Over at Nvidia they are working on developing learning capabilities that focus on issues arising from removing human control of the car. The example I was given was a kidnapping attempt where human driven cars could abuse the collision avoidance response of an autonomous vehicle to box it in and bring it to a halt or direct it elsewhere. A human would be able to recognize what is happening and respond and they hope to replicate that recognition in the car.