In order to grasp parts we approximate the object with its Bouding box.

Bounding box: The smallest box that can contain the object

So when we grasp an object like the one above it is possible to find postures like that:

This problem only happens with big objects that are not well represented by there bounding box.

In this example it is also happening because when grasping an object with one hand, we will always try to grasp it on the top first.

A solution would be to grasp this object with two hands, because when using two hands, we first try to grasp it on the sides.

P.S: This is just a work around and we are currently working on a solution to this software limitation.