Obviously they've already tried that, and it doesn't work. The map is not the territory.
In theory, there is no difference between theory and practice. In practice, there is.
What I mean is that. Do not "teach" the thing in real time. Instead collect the sensor data from the cars a human is driving (and also collect the human input also), and train the thing on it, safely inside the lab.
You say, they have done it already. But I am asking if they have done it enough. And if that is so, how come the accidents such as these are possible, when the situation is pretty out of a text book in basic driving?