Good point, but if you consider the Tesla dataset... it's formidable. Every day they could collect data enough for ~55 years of driving a lot every day. Even if you never encountered this case, if it happens at all it's likely to be seen many times (probably 100+ in a few months) in that dataset. After driving cars have gone mainstream, this may start to be seen as a design problem by traffic agencies: they might standardize ways to deal with traffic a little more.
Ultimately as long as the cars driving autonomously is small enough and procedures change slowly enough you should be able to continously update the driving system.
But let me reinforce that a pure learning approach even with very large datasets may not be efficient as one would like -- the curation of signs is a good idea, and manually reviewing accidents and near misses (a highly human-intensive task) and perhaps flagging bad driving behavior (probably after some outlier screening, which can be good or bad) will be important to get it really good with the training-intensive approach (and not the top down optimal path planning and control approach).
EDIT: Mobileye CEO discusses some interesting design issues and manual validation (and shows they have lots of data, good sign) https://www.youtube.com/watch?v=kp3ik5f3-2c&feature=youtu.be...