All (but perhaps Bill P and Paul B most of all :-),
Based on earlier threads, I had the following idea for the orientation estimation of a quadrotor using only 3 gyroscopes and 3 accelerometers. I'm not sure if it would work.
I propose to track the "gravity vector". I'll update the vector as per the usual DCM algorithm (iterate the vector each timeframe via the cross product of omega and the vector, where omega is the gyro output compensated via a PI feedback of the cross product of itself and the accelerometer outputs). This effectively gives two of the three orientation components, and has already been described elsewhere as "DCM lite". However, I would only update the PI loop when the magnitude of the gravity vector is sufficiently close to 1 g.
The third component comes from integrating the dot product of the compensated gyro signals with the gravity vector - i.e. the component of angular velocity that is around the gravity vector. I'm pretty confident that this gives me a uniquely defined orientation, without singularities (assuming I restrict the angle to +/- 180 deg or equivalent). This scheme seems analogous to the "axis-angle" formulation or even the quaternion formulation in that 4 parameters are used to specify the 3 degrees of freedom (3 components of a vector and an angle).
My experience is that the MEMS gyro drift is pretty small and slowly varying, so that as long as the quad is maneuvering (tilting), all 3 gyros will have some compensation from the PI feedback of the accelerometer error, even without magnetometer feedback (it won't be perfect, but it might be good enough for the 15-30 minutes that the battery will last).
Does this sound like it might work? If so, how would I transform the 4 parameters into a representation of orientation in the inertial frame (DCM, Euler angles, etc) for use in groundstation displays? My idea was to take the cross product of the gravity vector with the body z axis and reverse it to indicate how the body moved, then simply rotate the inertial body around the gravity vector by the integrated angle. However, this did not appear to work in my simulation. Perhaps my code had an error, or perhaps there is something wrong with these ideas.
As for control, my idea is that the pitch & roll r/c joystick inputs would serve as virtual angular velocity commands to move a "desired" gravity vector (for a rate command system). Then the cross product between the desired and estimated gravity vector can be used to form the error signal that drives the PID loops. I could have the yaw joystick command the rate about the gravity vector, but I think it would be more intuitive for it to command the body axis yaw rate (and easier to implement)