TBH, since it's a training based system it's "just" a matter of making sure the training set is large enough, including the situations you mentioned (assuming the training method is robust, generalizes well, etc). I would love someone knowledgeable to give an estimate, but I would guess you need at least a handful (10+?1000+?) of examples of each edge case (involving bicycles, pedestrians, weird road designs, street signs, and so on) -- and there are many of them I suspect (at least 100s?). Estimating you'd take about 1 hour between experiencing a tricky scenario while driving around, this should put the number of hours at something like 100,000+ -- not easy to come up with by himself (that's about 50+ years of driving 6 hours a day).
Mobileye is doing something interesting by curating the reliable parts of the dataset (e.g. they have curated databases of traffic signs for each region) -- again not something you could do own your own, and seemingly archaic (hence GeoHot's criticism), but if you can afford it can speed up the training significantly.
Tesla is a massive resource here because they already have a huge fleet of internet connected cars proving enough data to fill the aforementioned training set in a matter of days or months: let's estimate their fleet at 40,000 cars -- then they could fill that minimum dataset in less than a day, and in a month they might have a 100x safety margin. Of course, there's a big technical problem of relaying all that video (maybe they just relay prediction failures), but the data is there.
Another fundamental problem with exclusively hands-off training (and little optimal control theory, etc) is picking up bad habits from drivers -- even the best algorithms will have a hard time and be only about as good as a good driver in each scenario, in the best case -- since the training data is acting as a ground truth.