Non-zero mean noise, or estimating measurement offsets

[st fn]

Hi,

I've a similar problem as Pose2SLAMExample: A wheeled robot with some GPS-like and some IMU/Odometry-like sensors.

My problem is, that the odometry sensor seems to have systematic errors, like a additive offset (odom = odom + a) or a scale offset (odom = odom * b), or both, or something more complex (odom + a *sin(odom)^b + c) .. but lets just assume scale and additive offset for now.

This problem renders itself by observing the results: all trajectories estimated are a tiny bit too short (given you trust odometry a lot, and gps not so much), so odom underestimates speed in some systematic way.

How would I model this in my factor graph, coming from the Pose2SLAMExample? It seems this ploblem is similar as estimating the camera intrinsics (as single factor shared over the graph)? Resp. similar to parameter estimation in general. I could also think of the additive part as gaussian noise with non-zero mean, but with unknown mean that has to be estimated.

Could s.o. point me into the right direction?

thanks :)