A car will rarely trust a map over its sensors. (The map might say, "This situation is confusing to your sensors, if your sensors show X, we assure you it's Y.) It's key that the map tells the car what the sensors will see if the road is as mapped. That's why you need details in the map, and Elon doesn't seem to get that. If the map is out of date, of course trust sensors. If the map is valid, then trust it because it was made with more data than you have, by computers smarter than yours, who had all the CPU time and memory of the cloud to figure out this situation, and saw every part of the road up close from multiple directions, and you only see it from a distance. And in many cases the map was drawn with better sensors than you have as well as different angles and distances, including aerial views and stuff you could never know, because you are just one car.
There are many contradictions in this. First, the car sensors win over the map. But, then trust the map because it was made with more data and more CPU. But, then maybe the map is out of date. But then the map was made with better sensors and angles (from non-frequently visiting sources).
And then the comment that somehow the map knows to "say" something to the car about the car's sensors being confused? A-what? Please show me an HD map that has "confusing to vehicle sensor" as metadata for a road segment. Yeah, not going to happen.
It's all so contradictory that what is described can not be programmed. And to top it off, you completely ignored the neural net confidence score that is the actual important aspect, as I started to describe. This score is really the only opportunity for the car to trust the map more than its own sensors. Ever.
If we break were to break down FSD failures (assuming we could agree on a proper definition of "failure"), we might see failures in lane assignment determination, but then also potential failures for traffic light state recognition, construction sign recognition, dynamic actor recognition (other cars, pedestrians, motorcyclists, bicycles, animals, road workers, etc.(), and there are surely even more aspects of total scene recognition to handle. But, of all of these, only lane assignment determination could be supplied by a map (others would be lane availability and speed limit, for instance). BTW, all this is without considering driving policy failures, which are separate and subject to their own list of potential bugs/failures.
I agree that there are situations where advance knowledge of available lanes and lane assignment are useful. Your route was no doubt intended to hit on those pretty hard - the shifting from left lane to right lane (and back again?) of which lane goes straight, for instance. The need to make a quick left hand turn after making a right hand turn. Knowing what to expect from the road ahead can only be helpful, but at the end of the day, the car has to trust what it's seeing over what the map describes. Map outdatedness is the easy kill-factor here, as no map will describe a fallen tree or power pole or the utility person holding a stop sign or a second ago putting up a "Lane Closed" sign or a construction lowered speed limit sign.
But, maps can be wrong in other ways as well. Mobileye's REM system, for instance, does not use "computers smarter than yours" for obtaining map data. As Shashua describes, no images nor video are uploaded to ME's cloud, only what vehicles have themselves computed on-board and summarized in 10kB/km chunks. Maybe ME is clever and assigns different vehicles to report different aspects of the same road segment, but even so the result is at best the well-known multiple blindfolded people feeling different parts of the same elephant.
And even if such a system could get it right, that might be fine in Munich or Paris or New York, but what about Peoria, Ill or even some of the roads in the Santa Cruz Mountains above Silicon Valley? Those roads get a lot less traffic and the prior vehicle might have been more than several minutes earlier. Or, for Hwy 1 in Big Sur, where there are miles and miles of roads with no cellular connectivity?
At best, the map is a declaration of what was true at some point in time. The vehicle can use that to help with anticipation that makes future maneuvers smoother - but that must be subject to what the vehicle itself is seeing. Given
sufficient neural net confidence scores, the vehicle should trust what it is seeing over what the map is telling it was there. And there are many dynamic factors for which no map can ever provide anything use, as listed above, and so the vehicle sensors must almost always be trusted.
It would be an interesting exercise to come up with situations where the vehicle should trust the map over what it's seeing. Unless the confidence score is very low, I can't see that ever being the safer choice.
Maps are not the be-all and end-all for autonomous driving. At best I think they are helpful as predictive elements to avoid "Sunday Driver" behavior, which is useful.