I have rebuilt my balancing robot. The reason was that I wanted to fit more sensors and functions, and there where no space for that in the previous design. I have also rewritten all the code and replaced the Kalman-filter I previously used for angle estimation with a complimentary filter instead. This in combination with a higher center of mass have made the robot a lot more stable and tolerant against pushes and other disturbances.
A Kalman-filter should work better than a complimentary filter, but the Kalman-filter is a lot more complicated. Since I do not understand exactly how the Kalman-filter worked, it was to hard for my to tune it properly. The estimated angle of the robot was reacting to slow. My current solution with a complimentary filter is much more responsive. That allowed me to increase the parameters of my PID regulators to make the robot more stable.