Models need to generalize over what they've been shown though. Most people won't learn to drive near an airplane or a wheelchair either, but would have no problem avoiding those.
Humans generate general 3d representation of the world. You don't know what this object is, yet you see it's borders. Tesla seems to generate understanding only of the objects that it has already seen.
I can imagine a PM meeting at Tesla triaging the issues - "severe bug, yet very rare, waiving it for this release."
One can also imagine their modelers populating the synthetic training world with all of the planes right now. Like a "Mria" coming right in front of you on 101, as well as F-35 passing you on the left, and of course Millennium Falcon dropping right on you, and all that happening on a flight deck of huge aircraft carrier.
"Training corpus did not include airplanes, people in wheel chairs,..."