rollaxis has an optional start parameter that, if specified, moves the axis to the position before start. Or, as the documentation puts it

The axis is rolled until it lies before this position. The default, 0, results in a “complete” roll.

(Just by the way: The last sentence is unclear. What's a "complete roll"? Better be explicit: The default, 0, moves the axis to the beginning.)

E.g.,

a = numpy.ones((1, 2, 3, 4))
b = numpy.rollaxis(a, 0, 3)
print(b.shape)

moves the axis at position 0 to the place before position 3, and hence the above will print (2, 3, 1, 4).

This before logic results in the fact that there are always two start indices which will do exactly the same thing, namely numpy.rollaxis(a, i, i) and numpy.rollaxis(a, i, i+1) – nothing. As a trade-off for this duplication, what you can't seem to do is to move the axis in the last place, since there is no axis that comes after.

I used the word seem, because you can actually do numpy.rollaxis(a, i, len(a.shape)) to insert it axis i before the imaginary axis at position len(a.shape). This is not explicitly documented.

On top of that, why is the parameter called start? I have no idea.

Edit The following suggestion is actually just what moveaxis does.

A more consistent approach would be to have a parameter target (or target_position) that can be used to specify the position of axis i after the roll. E.g.,

a = numpy.ones((1, 2, 3, 4))
numpy.rollaxis(a, 0, target=0).shape  # (1, 2, 3, 4)
numpy.rollaxis(a, 0, target=1).shape  # (2, 1, 3, 4)
numpy.rollaxis(a, 0, target=len(a.shape))  # error
numpy.rollaxis(a, 0, target=-1).shape  # (2, 3, 4, 1)

The last example shows how to move the first axis to the last position without recalling len(a.shape).