Absolutely stop and go in this situation.
I think it's easy to misinterpret meaning in a forum context like this. I think we're saying essentially the same thing. You proceed, possibly extremely slowly, so as to aid visibility. You can't just sit there before the crosswalk on the street you're coming from hoping that you'll be able to see around the massive metal box.
So, it may involve coming to a stop. But since the light is green, there's no reason to stop, unless you determine there is someone in the crosswalk you must yield to. And it would be legal to proceed into the intersection (since you must do so in order to be able to see) as you turn, and then stop, to allow the pedestrian to cross.
I certainly would not stop in this situation. But I'd probably be going 3-5 mph, or slower, possibly going down to 0 at some point during the turn, not 14mph. And I'd be going as fast as I can, up until the point when I should be going much slower. The speed would be whatever is appropriate, between 15mph and 0mph, and would vary continuously between those limits based on circumstances.
Nobody should follow you so closely as to come and hit from behind.
No one
should, but defensive driving requires that you drive in such a way so as to minimize the chance of them doing so, no matter how closely they are following. That behavior is all I was describing. There's not any formula for it. You just make sure (as much as possible) they don't hit you.
They should expect you to stop for pedestrians at cross walk anyway.
They
should, but often you can identify in advance the driver who is not going to expect you to stop.
Are any autonomous vehicles able to make those kind of predictions?
I sure hope so. Is any AV company attempting to develop "understanding" human behavior? Perhaps not. But probably there are AVs looking at behavior, orientation wrt crosswalk, current speed, and hand signals from pedestrians and making decisions accordingly. It's essentially required from an AV to be able to understand and interpret human signals. I don't know what is the state-of-the-art currently, but all of this is certainly going to be required - there is no way to avoid it.
. It seems to forget about objects that it's detected as soon as they are blocked from view. Object persistence doesn't appear to be part of its current strategy.
That garbage truck video somewhere still showed the garbage cans on the opposite side of the truck after it passed it. Not sure if that is because it was able to see under the truck (I doubt it...it was a garbage truck!). Object persistence is really important, of course. The idea is to build as good a model of the entire environment as possible. For that you have to have object persistence.
For example, imagine FSD identifying a child running towards the street, but as the child enters the street, she passes in front of a parked car, which blocks your view (and FSD's view), and the parked car is low enough that it is not possible to see the child's feet. In that case, you have to have object persistence. The child is not gone simply because she is no longer in view. The car (just like a human) must react accordingly. This sort of situation happens all the time, many times a day. It's a super easy problem for a human, and if the car is high enough, the child is not lost from view and can be tracked via the feet. I'm not sure if FSD has this capability yet - I am surprised no one with the alpha/beta has set up situations specifically to test for it, but it will need it before wide release. (I apologize if I've missed the video of this!)
In the end, as was mentioned, driving behavior is often determined by what you can't see currently but have seen in the past. So persistence is very important.
we need a release of FSD to the public soon
Release to the public soon? No, I don't think I'm making that point for you. Quite the opposite!