> It looks like the estimated pose is displaced by about half
> height of your map.  Where are you getting the "true"
pose for this  
> plot.? Keep in mind that amcl uses the *center* of the map as
> not a corner.  Also, amcl uses a right-handed coordinate
system, in  
> which Y increases up (as opposed to a graphics coordinate system,
> which Y increases down).

It's not a problem with the map coordinates, because if the position covariance at the beginning is very low, the real position is found very fast and both plots are the same then.
I also think that the displacement has nothing to do with the size of the map, because the displacement differs. In value, and direction (sometimes only in X or Y direction or in both directions).
It seems like the algorithm is only adding the odemetry data to the hypothese if the weight of the hypothese is 1.0 . Otherwise the algorithm should recognize his wrong estimated pose (at the latest if it's going through walls).