From Acoustic Object Recognition to Object Categorization by a Humanoid Robot

Abstract

Abstract — Human beings have the remarkable ability to categorize everyday objects based on their physical and functional properties. Studies in developmental psychology have shown that infants can form such object categories by actively interacting and playing with objects in their surroundings. It is infeasible to pre-program a robot with knowledge about every single object that might appear in a home or an office. If robots are to succeed in human inhabited environments, they would also need the ability to form object categories and relate them to one another. In this work, we present an approach to interactive object categorization in which the robot uses the natural sounds produced by objects to form object categories. The method is evaluated on an upper-torso humanoid robot which performs five different manipulation behaviors (grasp, shake, drop, push, and tap) on 36 common household objects (e.g., cups, balls, boxes, pop cans, etc.). Using unsupervised hierarchical clustering, the robot is able to form a hierarchical taxonomy of the objects that it interacts with. The results show that the formed categories capture certain physical properties of the objects and allow the robot to quickly recognize the correct category for a novel object after a single interaction with it. I.