The 'rmsold' and 'rmsnew' fields were mostly put into the output of 3dvolreg for debugging purposes, and then were never removed. I know that some people, such as Michael Beauchamp, like to use these as extra baseline regressors in 3dDeconvolve. Personally, I do NOT like that idea.
It is hard to say that any particular value of rmsnew should be used to censor the data. Typically, you have a large time series (100 or 100s of points), so a few (< 10) mildly bad points will not cause huge problems. The thing to worry about is very large spikes in the data time series. You can search for those by doing something like
3dToutcout -automask -save Tout -polort 2 dset+orig
and then look at the time series Tout+orig in one AFNI window along with the input dset+orig in another. If you have large spikes that occur at the same time point, then censoring is an option. If you have large spikes that occur sporadically in space and time, then 3dDespike might help. But I do not recommend automatically censoring or filtering your data -- LOOK AT IT! You spent some effort to acquire your subjects and then the imaging data. Spend some effort to look at the data and decide on its quality.