Isn’t the model needed rather similar to what’s needed for sensor fusion in general? If you can extract features from each sensor that you expect to match to features from a different sensor, then you can collect a bunch of samples of this sort of data and then use it to fit the transformation between one sensor’s world space and another sensor’s world space.