I find their approach of end-to-end machine learning absolutely fascinating. Since the video is very long, I'll explain the gist: instead of the traditional approach of labelling everything in the scene using neural nets and then trying to divine a driving plan based on this information, comma is trying to behave more like a human and going straight from perception input to neural net to car controls output (expressed as a path).
This to me is a fascinating approach and raises interesting questions about the observability of the system and how much we can trust systems like this, where validation is based on testing of the behaviour of the whole system rather than validation of its subsystems.