Gmapping failure culprit -- parameters, odometry, sensors?

Hello, I am running ROS Hydro on a Turtlebot 2 (Kobuki + Kinect) I have been trying to use teleop + gmapping to get a decent map of my cubicled office building, but after slowly perimeter-driving and looping back through several hallways, inevitably, the map will eventually do something absurd like this:  I started with the first two hallways on the right, constantly weaving in and out of cubicles and repeating areas. The picture of those hallways really isn't bad, but when I started exploring the third vertical hallway after 30+ minutes, gmapping suddenly decided to create that horizonal hallway beginning to form on the bottom-left, which is extremely off. Anything after that point in time got progressively worse. This happens fairly frequently when things seem to be going otherwise well. The following are my gmapping parameters: I have changed very little from the default turtlebot configuration, besides lowering the linear/angular update, slightly increasing the number of particles, and decreasing the map update interval. Since I know that a failure in scan matching causes gmapping to rely on odometry, I wonder if my odometry is decent enough. I ran the test suggested in the navigation tuning guide (http://wiki.ros.org/navigation/Tutorials/Navigation%20Tuning%20Guide): - I booted up `minimal.launch`, `keyboard_teleop.launch` , `gmapping_demo.launch`, and `view_navigation_app.launch` - In rviz, I set the global frame to odom, turned off all costmaps, and put the decay time on /scan to 600s. I did all the following tests at the lowest teleop speed. **This is the result of performing three rotations in my standard-sized cubicle:**  **Entrance to cubicle, staring at wall ~5ft away:**  **Drove straight towards wall:**  **Circled in cube, drove outside, turned around, drove back inside:**  So my question is, how do I get better gmapping results? Is this odometry decent enough? Do I need to tweak gmapping SLAM parameters? "srr", "srt" error parameters? Is it time to get a sensor better than the Kinect or find an algorithm better than gmapping? Am I doomed to having to map only a few halls at a time and then manually stitch the PGMs together? Additionally, if anyone has some helpful tips on how to drive to get the best map (speed, looping, etc), I would love to hear them. Thanks in advance!