I don't believe it's a fad, but it will take time and companies/institutes who pursue it need to exploit the best technologies possible, i.e. active range-sensing/LIDAR rather than just passive (stereo) methods, while also furthering the computer vision techniques needed. It's tempting for companies like Tesla to decide that LIDAR is too expensive and that they can get by with software and a pair of cameras, but the laser impinging on the scene obviates difficult/slow correspondence problems that are onerous for computers to solve.
Meanwhile, deep learning has to emphasize object recognition based on 3D point clouds (PointNet-style) much more as opposed to just 2D image processing, to which the bulk of research on DL applies currently.
Finally, I also think that these companies and institutes need to accelerate testing of their autonomous navigation algorithms based on traffic simulation (SUMO-style), which makes it possible to validate autonomous nav algos faster than real time as well as on a massively parallel scale.
I like it is very much. I think this will become a sound reality in the near future. Self-driving cars mean no or almost no level of "road rage", an ugly reality of cars with some drivers of kind.
29 months ago