According to this github issue, the
max_correspondance_distance corresponds to : the maximum distance in which the search tried to find a correspondence for each point.
So I think the value really depends on the scale and the expected distance of your volumes. In case you’re not sure, I think having a higher threshold would help the ICP converge. For example, I’m registering two large point clouds that are placed at a large distance one from another, so I decided to set the threshold to the distance between the mean points of my point clouds, assuring that the ICP will converge. So far, it seems to work okay.
I hope it helps