The issue here was that the calibration using the quadratic regression worked okay for yaw=0. However, as soon as there is a small tilt of the head base frame (i.e., the robot root frame and the head root frame where not parallel as in ARMAR-6), this calculation failed and resulted in a sinuidal movement of the head when the platform was rotated. To fix this, the gaze targets are now transformed to a gaze root frame which in turn is relative to the head frame. Additionally, the target yaw is compensated before calculating the pitch.
/cc @reister FYI