Well, I’ve seen it very confidently and correctly navigate rural unmarked roads, and drive around rocks and logs across the road and things like this. I find it hard to believe it would drive through a huge stationary object like a house, which I imagine is definitely in the training data.
I guess I’m not saying it’s impossible, but I think at this stage it’s improbable.
The computer does not see things like you see things ;)
The fact that you are saying "but houses are in the training data!" demonstrates your naivete.
Do you think there's some magic going on here?