The ability of an ML system to learn to mimic what the manual "virtual unrolling" process is doing, from a small number of examples-to-follow, is growing.
Each bit of success, once confirmed by other experts or correlation with other texts, improves the training data.
Eventually a fully-software pushbutton pipeline of "raw imaging to likely texts" should be possible.
And if, say, some of the scrolls are sufficiently 'read' nondestructively to embolden teams to risk destructive techniques – such as incremental ablation while reading the exact chemicals at every coordinate – even higher-resolution data could become available.
When you are discovering to do a new thing that nobody knows how to do, you first try a bunch of things manually to learn what works and what doesn't.
Only after you have a procedure that reliably works, you automate it.