NIH barely applies here at all. If we take into account the constraints they were under and the absolute state of desktop Linux when Android started, it makes total sense.

Starting a lot of the components as greenfield projects will fortunately help a ton with shedding years and years of legacy cludge, hacks and desktop first design. Without mentioning the colossal amount of friction they would've faced if they tried to upstream any of the code they would've produced by the end. "We don't care about that issue. Closing."

Because of the above, they still would've needed to have their own fork regardless and still would've eventually diverged further and further -- so why bother? With a greenfield project, you can just optimize to your specific constraints. If the maintenance becomes too burdensome, just pick up the best of your design and reimplement them onto the non-greenfield project, everyone wins :D