Coordination and localization patents are the quiet frontier, and the class to watch in 2021 was the G01C 21/38 mapping family. Application US20210063199A1, "Map creation and localization for autonomous driving applications" (published March 4, 2021, with David Nister among the inventors), is a clean example: G01C 21/3811 (map data creation) and G01C 21/3878 (map updating) sit at its core, with G06N 3/02 marking a neural component.
Why flag this class? Because perception — detecting objects in the moment — is necessary but not sufficient for autonomy. The vehicle also has to know precisely where it is and what the road ahead looks like beyond sensor range, and that is the mapping-and-localization problem. A 2021 filing concentrated in G01C 21/38 rather than the perception classes signals investment in the half of the stack that gets less press but is no less load-bearing.
The enabling read is favorable but qualified. The combination of map-creation and localization classes with a neural class describes a learned mapping pipeline — a concrete approach, not a hand-wave. But this is an A1 application, so the claimed scope will narrow in prosecution, and the presence of a neural class does not by itself prove the method works at scale. What it proves is direction: building and using HD maps with learned components.
For CPC Watch purposes, the takeaway is that 2021's autonomy IP was not only about cameras and lidar; the localization classes were heating up in parallel. Filings anchored in G01C 21/38 are the tell that builders understood map-relative positioning as core infrastructure. Read it as a frontier marker, label it correctly as an application, and watch whether the cluster grew in subsequent years — it did.