The slamtoolbox documentation states that the scan matcher of Karto SLAM is extremely good; however, it seems the original version of Karto SLAM (which supposedly uses ICP for scan matching) is not as good for large-scale environments, so it has been modified for use in slamtoolbox. Is the scan matching algorithm used in the slam_toolbox implementation of Karto SLAM a variant of ICP or some other method? (It looks like Ceres is used for pose graph optimization when loop closure is detected but not for scan matching itself)