Also, not only do they rely on “just vision”, crucially they rely on real-time processing without any memory or persistent mapping.

This, more than anything else is what bewilders me most.

They could map an area, and when observing a construction hazard save that data and share it with other vehicles so they know when route setting or anticipate the object. Not they don’t. If it drives past a hazard and goes around the block it has to figure out how to navigate the hazard again with no familiarity. That’s so foolish.