The DARPA grand challenges were an important step for autonomy, by demonstrating basic capabilities and bringing attention and industry funding to the field, but overhyping the expertise gained there may have held the field back over the years. The hardest part of building a self-driving car has proven to be robust perception (processing sensor data to create a very accurate representation of the dynamic environment), which is an aspect that was not heavily emphasized in the challenges: for instance, the winning entry in 2005 had $500K worth of Lidar on it, and the cars were given a GPS map, which allows very accurate localization with yet another expensive sensor (differential GPS). By all accounts I'm aware of, the perception software involved in the challenges was rudimentary, and a lot of the effort was spent on mechanical engineering and path planning. Most importantly, robustness cannot be demonstrated in a one-off competition. Building robust perception is a hard AI problem, and the expertise required to tackle this aspect is orthogonal to the robotics aspects demonstrated in the challenges (which are also obviously very important).