I ran a group-wise registration and compared it with a pair-wise one. In certain cases, the former can go beyond points where the latter gets stuck. So that's good news...
Also, I ran experiments with mdl_matcher. I tried ssd, and I also tried res_info as the OF. Something is happening, but not necessarily an improvement.
Compare_pair shows something we might expect - an overlap that is about as good as that which is obtained by standard registration.