Towards a grasping-oriented categorization of rigid industrial objects to support automatic grasping of 3D CAD models in virtual ergonomics

The use of automatic analysis of 3D CAD models to identify a variety of plausible grasp locations on unknown objects with complex shapes and no affordances is a current concern in virtual ergonomics. In robotics, researchers approached this problem by grouping objects into categories related to geometric characteristics. Category-level grasping approaches have been little explored in virtual ergonomics. This paper introduces grasping-oriented categorization for rigid industrial objects to support the identification of a variety of plausible grasp locations in virtual ergonomics. First, we performed a rough analysis of objects' convexity and define two high-level categories: global dominant shapes and local dominant shapes. Second, we propose a more detailed categorization of objects to sort them into four specific categories: 1- threadlike objects, 2- thin objects, 3- small objects, and 4- large and thick objects. Then, we explode those categories based on the presence or lack of holes or protrusions that are useful for grasping. This grasping-oriented categorization of objects is used to manually categorize a subset of 242 real rigid industrial objects. We obtained 27.3% threadlike objects, 43.8% thin objects, 2% small objects, 24.3% large and thick objects and 2.6% uncategorized objects. Our future work will pertain to develop approaches for automatically categorize unknown objects and approaches for automatically identifying variety of plausible grasp locations on objects of given categories in virtual ergonomics.