Vision-only versus mapped is a cost bet, not a faith, and US10620317B1 is the mapped side's bet written down. Granted in April 2020, "Lidar-based high definition map generation" describes turning lidar sweeps into the centimeter-accurate prior map a robotaxi localizes against.
The edge case the patent quietly admits is the one mapped autonomy lives or dies on: a car that localizes against a stored HD map is only as good as the map's freshness. Classified under G01S 17/89 (lidar systems) and G05D 1/024 (position control using stored maps), the filing is about generating those maps efficiently — a tell that generating and regenerating them is expensive enough to need patenting.
Here is the trade, fairly stated. A mapped vehicle gets an enormous prior: it knows where every curb, lane line and stop bar should be before its sensors confirm them. That prior makes perception easier and the driving policy calmer. The vision-only camp rejects exactly this dependency, betting that a camera stack reading the world fresh every frame scales anywhere without a map-maintenance bill.
The risk factor is where the honesty lives, and the honesty here is in the word 'generation.' If maps were a one-time cost, you would patent the format, not the generation pipeline. Patenting the pipeline concedes that maps are a recurring operational expense — every road that changes is a re-mapping job, and a city's roads change constantly.
This 2020 grant sits at the headwaters of a debate that defined the decade: Waymo, Cruise and their peers paid the mapping bill and got reliable localization; Tesla refused it and bet on vision generalizing. Neither camp was practicing a religion. Both were pricing a trade-off, and this patent is the mapped camp's invoice.
For readers tracking the robotaxi race, the question to carry forward is not 'whose sensors are better.' It is 'whose cost structure survives scale' — and the answer starts with whether you have to keep a high-definition map of every street you drive.