This is in part due to an issue in AutoMap:

brandtbucher/automap#5

This is also due to an issue in how NumPy chooses not to handle NaNs in some operations;

>>> np.union1d(np.array((1, np.nan)), np.array((np.nan,)))array([ 1., nan, nan])>>> np.unique((np.array((1, np.nan, np.nan))))array([ 1., nan, nan])

This appears to be a 10-year old issue:
numpy/numpy#2111

As this is an issue common todict as well asAutoMap, as it stands presently this is expected, though confusing behavior for a mapping object. I think the only option, from SF's perspective, is to consider disallowing NaNs inIndex, though that option will cost performance.

>>>dict.fromkeys(np.array((np.nan,np.nan))){nan:None,nan:None}