Hi Ross,

IMHO you should have a look at the parameters lstep, astep and 
iterations. By choosing low values for these three, you can limit the 
impact of scan matching on the estimated robot pose.

GMapping's gridslamprocessor class contains a parameter minimumScore, 
which could be interesting for you as well: If scan matching returns a 
score less than this parameter, the pose refined by scan matching is 
discarded altogether and GMapping relies on odometry only.

Unfortunately, the gmapping wrapper for ROS does not support setting 
this parameter at the moment. Adding it should not be a big problem, though.

Happy hacking,
Sebastian