Stretch_navigation localization difficulty

I’ve been working with the stretch_navigation recently, but run into some issues with the localization with amcl. I thought I could solve this by using the tool to calibrate the wheel separation. However, my workspace has multiple different floors, and each different floor calibrates the wheel separation value slightly differently. When the robot is on the same floor type that the wheels were calibrated with, amcl can localize pretty well. However, on a different floor type, amcl has a great deal of difficulty localizing.

Does anyone have any experience/insight how to overcome this kind of issue?

As an update, to troubleshoot the issue, I was following along this ROS Navigation tuning guide (section 2.2). Moving forward or backward does not cause an issue, but turns/rotations result in quite a bit of rotational drift.
I tried to use the stretch_scan_matcher and even the stretch_ekf without success to improve the odometry or localization.

Given the poor localization, the autonomous navigation is not working (which is ultimately what I’ve been trying to set up). Any help or suggestions would be much appreciated!