I have noticed that the "reason" step in ROBOT takes much longer than it does in Protege, for the same version of Hermit and the same ontology (and same hardware).

The difference isn't in the reasoning step per se (i.e., computation of the class hierarchy, and subsequent sub-steps) - rather it is the fact that ROBOT does separate steps for Checking for unsatisfiable classes and Checking for unsatisfiable object properties (in my hands, it's the 2nd one that can be very expensive - as much as 60 minutes in a CI environment).

This was mentioned a few years ago in #368, but wasn't really the focus of that ticket (though it triggered some discussion).

@cmungall noted at the time that omitting those steps altogether would be unwise - but on the other hand, as far as I can tell, Protege accomplishes the same result, at least for unsatisfiable classes, more or less as a side-effect of computing the class hierarchy. So for classes at least, it seems that ROBOT's logic could place the class-unsats check after the class-hierarchy computation, and then implement it simply as a check for classes equivalent to owl:Nothing.

For object-properties, I'm less certain about the opportunities for a speedup - in a way, the question comes down to "does Protege effectively compute unsatisfiable object properties as an artifact of running a reasoner, the way it does for unsatisfiable classes?". If so, it'd be nice to have ROBOT skip "Checking for unsatisfiable object properties" and instead take advantage of reasoning results to infer it, saving lots of time when doing robot reason.