The debate is almost as old as (fMRI) time, with some people suggesting that you never regress out motion. But the AFNI way is to do both censor and regress out motion. Note that censoring means that those TRs don't contribute to the final stats, so the motion related to those TRs is also not considered.
I've played with motion censoring and regressors a decent amount, and while it's an imperfect solution, it works better than not including them in the model. There are suggestions of using more parameters to model motion, and in some AFNI processing there are derivatives and per-run things modeled.
You might start with
this paper, and then farm the references. There's also some handy references mentioned in the
ABIDE project info.
-Peter