When talking about self-driving cars and autonomous construction machinery, many people refuse to understand that HD maps and point clouds from previous mappings are sooo 80s!. Well, back then and basically until the early 2010s it was clear that computational power and algorithms were not quite suitable for driving without some reference to reduce computational requirements. First, we have to have a look at what HD maps represent.
HD maps contain many features that are considered as useful, e.g traffic light locations, lane lines, static obstacles and often a local set of point clouds. The resolution of these features is on a cm level. A free course on the Baidu apollo platform on Udacity covers mapping and the apollo map open service in great detail. To understand why most self-driving car platforms use and really depend on HD maps we have to go back in time a bit. You may not believe it but there was time without deep learning and with less computational power than on our phones nowadays. If we build robots in extremely defined environments, then reference data is really handy but even in those times it worked without as well. Further, HD maps and other hand-engineered stuff is really useful to obtain seemingly working results fast to get either research funds or venture capital. However, after such an initial demonstration we’re ending up going nowhere, fast.
The most important reason why I think (educated guess - not a religious believe!) that HD maps (and point cloud maps) will disappear quite rapidly:
- We have to deal with transient elements anyhow! We have to predict what other entities are going to as well. We have to deal with new construction sites that might be only temporary (e.g. fixing broken water pipes).
- Acceleration chips (e.g. the Tesla driving computer) will increase computational
- Everyone will end up with enough suitable training data to train neural networks that are capable of doing things which currently require a lot of hand-engineering.