I have serious, serious doubts about the whole approach of "identify known objects in a camera feed so you can try not to hit them". I don't know what the current approaches are but there needs to be some kind of subsumption setup where if you don't recognize an object, at the very least you assume it's a stationary solid object and you don't hit it. It doesn't matter if your classifier says it's a brick, or a grandma, or a hibiscus, or (unknown). Unless you've positively identified it as something that's safe to hit (say, a plastic bag wafting along in the breeze) then don't hit it.
If you do have an identification then you can layer behaviours on top of this (eg. Is it a person? They usually walk forward or in the direction they're looking, so anticipate this.) But the default behaviour cannot be "dunno what that is so I'll ignore it".
"Don't hit the thing" is the most basic, fundamental behaviour for a self driving vehicle.
And this is why I like to call them "Self-Crashing Cars".
I'm fully capable of ploughing in to moving and stationary objects at high speed, under acceleration, myself. Thank you very much.
And OTA updates that can change the behaviour of the vehicle between uses? Just no. I have enough trouble switching between my Japanese and European cars (one of each) where the indicator stalks are on opposite sides of the steering column. I'm forever indicating my intention to turn with the wipers!