For folks building balacing robots like Dave Anderson's marvelous nBot or Larry Barello's gyrobot, I've written an extensively commented Kalman Filter that uses a dual axis accelerometer and a single axis angular rate gyro to accurately determine tilt angle, angular rate and gyro bias. The filter is able to be tuned for different amounts of noise in the sensors and to place varying trust in the different sensors. It also automatically tracks the gyro bias by comparing the covariance of the state estimate with the state measurement.

The full source code is in tilt.c and tilt.h. It is available under the GPL for your use and education.