Pseudo-instancing applies well to geometry with a small number of per-instance attributes.
The technique does not scale well to complex mesh rendering techniques like skinning; there
are not enough vertex attributes to store all of the bone transforms for each instance.

Direct3D instancing reduces the number of DrawIndexedPrimitive() calls from many to one.
This DrawIndexedPrimitive() call reduction has a large performance benefit in Direct3D.
In OpenGL, the application still calls glDrawElements() (or the like) for every instance.
This isn’t too much of a performance hit because glDrawElements() is very efficient in OpenGL.

This points out that instancing solves different problems in GL and DX - in DX, each draw call is very slow, so instancing is used to group them together. In GL, draw calls are much faster, so instancing doesn't help as much here.
Furthermore pseudo instancing doesn't actually reduce draw calls at all - you need the real deal (in GL3.1) to achieve that. Pseudo instancing is only used to slightly speed up the calculation of the world-view-projection matrix.

As you can see, in OGL instancing has only an advantage for models with a small number of polygons. Besides that it will not work with characters that require all shader registers for the joint matrices. Real or pseudo instancing could be a nice option for a vegetation system extension (anyone willing to write one? ) but I think we do not require it in the engine core at the moment.

It should be pretty easy to add to the engine - the only modification is the way in which the world matrix is sent to the shader. Shader resources would require a flag of some sort to indicate to the engine which method to use.

LE: I came across this extension, ARB_draw_instanced This is the real deal. Even used with opengl 2.1. An extension build on this could be plausible.