Apple researchers train AI drivers to merge lanes in a simulated environment

I'm curious about how "relaxed adherence to rules" will be implemented.

If you're merging from an onramp into the rightmost lane of traffic, you'll get in either in front of someone or behind someone. Meanwhile, the merge lane that you're in is narrowing.

Say that after the merge lane disappears, there's a nice, fat shoulder, divided from the highway by a white stripe. You're not supposed to cross that stripe.

If things go wrong, you might sometimes elect to continue on into the shoulder, and then merge. Or you might just stop in the merge lane, which has its own problems.

So, if traffic is thick and stubborn, will an autonomous vehicle stop in the merge lane, or continue into the shoulder?