Gravity shouldn't be a filter state variable

[badicsalex]

I have found that gravity estimation interferes with finding accelerometer bias, and makes the filter stability way worse. In my opionion it should instead be a builder parameter and supplied by the user based on geographic information. (Or simply left as a constant, as the error of that is not that much compared to other error sources)

I do have a patchset that removes gravity from the state, which reduces the size of the covariance matrix (and thus makes the calculations faster), and also made convergence much faster for my use-cases. I've ran tests and compared carla plots, and this doesn't seem to affect performance.

Do you have any opinion about this?

[badicsalex]

BTW, I've re-read the paper, and this is actually mentioned at the bottom of section 5.3.1

We do so to improve linearity: indeed, equation (235b) is now linear in g, which carries all the uncertainty, and the initial orientation q0 is known without uncertainty, so that q starts with no uncertainty.

For what it's worth, in my experience it's not better at all.