There are only a handful of experts who dare to say that we do not need HD maps to cope with autonomous driving. I count myself among this small group, and believe that high definition maps contradict the idea of fully automated driving.
HD maps contain many useful features modelling the real world, much better and more accurate than navigation maps do today. The road geometry, carriageways, road lanes, dividers, bridges, tunnels, traffic lights and signs, crosswalks, sidewalks and many other features are stored in a geospatial database on cm-level. However, the time an HD map database is shipped or streamed, it’s already outdated. I agree that HD maps are useful to look around the corner, ahead of the visual horizon, like an additional virtual sensor. Since there will be no error-free HD map database there will be no 100% confidence in the contained map features and attributed
One of the most significant reasons why I think that HD maps will disappear sooner or later is that autonomous driving must learn to deal with a constantly and frequently changing roadside elements, road users and traffic situations, as we human driver need to. Providing a real-time, high-accurate, high-detail, complete and error-free map that reflects the real-world is utopia. Instead of relying on data extracted from a geospatial database driving algorithms need to carefully observe and predict the vehicle and its environment, especially other road users.
Of course, this requires exceptional onboard computing power embedded into silicon. Neural networks are data hungry and require billions of kilometers of training data in order to reach error-free driving in all conceivable situations.
Yes, there was a time where there was no usable artificial intelligence, no deep learning, and not enough computational power and storage to train computers to drive autonomously.