>simply failures of scene construction

Isn't "scene construction" part of AI?

From a technical angle: computer vision and object recognition are AI.

From a philosophical angle: part of human conscientiousness, and intelligence, is to exist in a spatiotemporal frame. This is one of Kant's points[0]. Our intelligence constructs scenes from our sensory input, but we have things like object permanence which have a deeper understanding than simply what our senses show.

[0] https://plato.stanford.edu/entries/kant-spacetime/

I Kant argue with that. But you would at least agree that basic 3d modeling of your world is at a lower level of difficulty than say understanding the intention of drivers or someone giving directions?

So the “simple” failures are due to the very hard problem of reconstructing a “mental” model from multiple 2D images, and actually attributing different properties to objects recognized? Like you don’t have to stop at a plastic bag?

People learn these things for a long time and we actually have a proper model of the world for exactly that reason.

Yes but the notable incidents are not "car stopped for plastic bag"

The incidents that I have described would actually be resolved by 3d scanning technology (not saying that thats the correct approach either)

No they wouldn’t be solved. And there was actually an incident like that with a plastic bag.

Link or it didnt happen ;)