Yup, I have seen the ardupilot. A friend of mine used one on his UAV with good results, I almost went with that rather than building my own, but part of this is just for me to learn, and I did want to tweak a few things in their design anyway. I'm thinking about using some of their code, haven't looked yet to see how much I can share. I'm making my own to save a little cost by removing the barometer, and have a faster onboard mcu, and have direct USB connection. When the 9 axis (accel+gyro+magnetometer) invensense chip comes out I am going to use that as well. I'm looking to do full inertial navigation w/ lots of filtering (it seems the MPU6000 has pretty terrible zero rate drift and temperature effects), so I figured it would be better just to over design a little than have an underpowered chip (and its still only 5$! things have advanced so much : ) ). I'm planning on running the accelerometers as fast as they can go, would like the sensor to be the limiting factor (might switch to more accurate gyros later, if I can find relatively cheap fiber ring gyros... was looking at KVH's website and they all say "ask for price", not a good sign).
The Kalman filter and quaternions do different things. Quaternions represent a rotation about a line, similar to a euler angle / rotation matrix (and good point, they are better because there is no singularity, but they take more processor time as the matrix is larger), and the Kalman filter is a way to combine data from different sensors, sensor error models, and vehicle dynamics model to perform data fusion and more robust inertial nav solutions with accuracy information (the all important standard deviation in addition to the "you are here").