Annoyingly, object arrays occur frequently because there is some text in the array when loading it, and in that context something like nan_to_num is often needed.

>>> arr = ndtest((2, 3)).astype(object)
>>> arr['a0', 'b1'] = la.nan
>>> arr
a\b  b0   b1  b2
 a0   0  nan   2
 a1   3    4   5
>>> nan_to_num(arr)
a\b  b0   b1  b2
 a0   0  nan   2
 a1   3    4   5

The workaround is to convert to float:

>>> nan_to_num(arr.astype(float))
a\b   b0   b1   b2
 a0  0.0  0.0  2.0
 a1  3.0  4.0  5.0

I think I remember there was another issue regarding nan and object arrays, but I cannot find it anymore. I guess I fixed that problem without creating an issue about it, because general_isnan and obj_isnan in larray code were created exactly to alleviate this issue.