Robustness really depends on the computer these days, I think. Flash drives already route around bad sectors, and most computer chips are capable of gracefully degrading performance as sections fail (one of the ways chip yields go up: over build the chip so failures can be routed and sold as a cheaper variant).
To your first line, the micromachine-for-micromachines idea often reminds of Feynman's talk "There's Plenty of Room at the Bottom" [0], which is great at least in how long ago it was made. Turns out it's a really hard problem. There have been some functional versions of the micro arm-onna-arm designs, but they're just stupendously hard to control. I'd love to see some progress on that front, though, because it really is cool when it works.