RMSD is “root mean squared deviation” (https://en.wikipedia.org/wiki/Root-mean-square_deviation). It’s a measure of structural similarity. It doesn’t per se correlate with affinity. Instead it tells you how close a given structure is to a reference structure.
Often times we use the native, bound structure as the reference structure, in which case models which are closer (lower RMSD) to the native are more likely to be “better models”, but that’s sort of coming at it backwards — we already know what the “best” structure is, and are judging how good a structure purely by how close it is to the reference structure. If you don’t have a gold-standard reference structure, then RMSD isn’t necessarily going to tell you how good the structure is, just that it’s close to the arbitrary reference structure you picked.
I believe for the ROSIE docking runs, it will normally use your starting structure as the reference structure. If your starting structure is arbitrary, there’s not necessarily and correlation between the RMSD and the qualtity of the model. RMSD is simply provided in those cases so you can judge how much the structure has moved from the starting structure, if there’s a particular reason you think the end result should be close-ish to the starting structure.
.
