[LudoA]

Yes, I'm also confused about this.

Existing maps should be enough, and LIDAR, etc. ensure that self-driving cars can respond to road conditions in an ad-hoc manner, instead of relying on a predefined detailed map.

The article doesn't make a good point of why this is needed. OTOH, given the money involved in the technology & companies mentioned in the article (like Here), I presume there must be a need for this.

[gbin]

You can see it as a simple optimization. The map can be considered mostly invariant.

For example, why would you need to waste resources to check over and over if this tree is a pedestrian, a car, a dog or a cyclist ?

[a_t48]

How can a combination of existing maps/car technology and LIDAR pinpoint at high resolution where a car is? Imagine a scenario where a car needs to merge across two lanes of traffic and turn into an alley. The car has to be able to plan the merges before it gets to the alley, and know that it is at the proper place to turn. Before you say "GPS" think of how often your phone's GPS messes up.

[badestrand]

Google Maps tells me a few hundred meters in advance so the car AI can know, too. The location does not need to be precise down to a centimeter, it should be enough to know "the alley is somewhat near" so it can start to get into the rightmost lane. Just as we humans do it. There are enough artifical obstacles on the road anyway (wrongly parked cars etc).

[maxerickson]

There is a rough line between the task of understanding the immediate environment and the task of driving through it.

I get the impression that many of the efforts have used detailed mapping to assist the first task so that they can get started on the second. I don't think this means they are stuck with it.