Self-driving cars a great example, because in this case there are two competing approaches -- the symbolic one, mostly consisting of handcrafted software encoding human abstractions, and the deep learning one, learned end-to-end. One will get to L4--even L5, the other never will.
It's not that deep learning is intrinsically incapable of driving -- it's that the situation space is extremely high-dimensional (due to edge cases), and that a deep learning system requires to be trained on a *dense sampling* of the same space that the system will operate in.
Because such a representative, dense sampling is impossible to obtain, even when heavily leveraging simulated environments, the symbolic approach will prevail (specifically, an approach that is mostly symbolic but blends human abstractions with learned perceptual primitives)
You can follow @fchollet.