assert_array_equal does not compare NaNs as numbers when dtype is object

OS: macOS Sierra (10.12.5)
Python version: 2.7.12
numpy version: 1.13.0

I notice that numpy.testing.assert_array_equal should compare NaNs as numbers unlike elsewhere in numpy, which makes sense to me. I verify this with the following test:

>>> import numpy.testing as npt
>>> import numpy as np
>>> a = np.array([1, 2, np.nan])
>>> npt.assert_array_equal(a, a)

However, this does not work with arrays that have dtype object:

>>> b = np.array([1, 2, 'a', np.nan], dtype=object)
>>> npt.assert_array_equal(b, b)

AssertionError: 
Arrays are not equal

(mismatch 25.0%)
 x: array([1, 2, 'a', nan], dtype=object)
 y: array([1, 2, 'a', nan], dtype=object)

Is this the desired behaviour? The only workaround in my case is to cast the array as dtype str.

Start at numpy.testing.assert_array_equal and reproduce the two examples from the issue, comparing regular and object-dtype arrays containing NaN. Trace the object-array comparison path and add a regression test showing the intended NaN comparison behavior, then run the relevant NumPy testing tests.