"Until these problems are solved, human beings will remain an integral part of the driving task, monitoring the AI system as it performs anywhere from just over 0% to just under 100% of the driving. The governing objectives of the MIT Autonomous Vehicle Technology (MIT-AVT) study are to (1) undertake large-scale real-world driving data collection, and (2) gain a holistic understanding of how human beings interact with vehicle automation technology."
(1) So much for driverless cars, I guess. (2) I guess a study is needed when aerospace and rail has, oh, 50+ years or so of cumulated data about the problem.
MIT or not, these people seem pretty confused about the rules governing safety in automation, and the abstract is just inaccurate. The driving task can be formalized. It has been done before.  Maybe it's a boring driver, but it's possible. Part of this formalization requires the safety control functions to be separated from the driving and navigation functions. AFAIK this principle was formulated first by Georges Westinghouse when he patented his air brake system in 1873 .
For example, RFIDs in the road or on the road side that tell the cars (or actually, allow them to look it up in a database) the exact layout of the road around them.
For instance what SpaceX is trying to do with retropulsive landings. If they can master this in a variety of conditions then the technology can be adapted outwards to no end. On the other hand, if you focus exclusively on landing back on Earth then landing on another planet is suddenly a huge ordeal. So in this domain, if we master the car driving without too many cues, then a self driving vehicle ought be able to drive in Timbuktu as well as LA, which I think is much more desirable - even if it's vastly more challenging.
But really I just want a segway city :(