Exercising Affordances of Objects: A Part-Based Approach

Abstract

This study shows how learning relations between affordances facilitates performing robotic tasks. Tasks usually involve multiple affordances. For example, for pounding a nail with a hammer, grasp-ability and pound-ability of the hammer are important for performing the pounding task successfully. Furthermore, these affordances are associated with parts of the hammer. In the pounding task, the head of the hammer affords pounding and the handle of the hammer affords grasping. We propose an Red Green Blue-Depth (RGB-D) part-based approach for performing tasks. In this paper, affordances are linked to object parts. We learn affordances associated with manipulation and execution of the tasks, i.e., grasping for manipulation and pounding for execution in the task of pounding a nail. Since affordances are associated with parts, tasks can be executed directly on the objects. Our approach is evaluated in six different robotic tasks on a real robot. We obtained an average of 65% task detection rate superior to the baseline methods and an average of 77% task success rate.