Autonomous acquisition of pushing actions to support object grasping with a humanoid robot

Abstract — There are many situations in which an object that needs to be grasped is not graspable, but could be grasped if it was situated at a different location. By applying nonprehensile manipulation actions such as poking, the object can be moved to a new location without first being grasped. We consider these issues in the context of an artificial cognitive system. The goal of the paper is twofold; firstly, we study how the robot can acquire nonprehensile manipulation knowledge by observing the outcomes of exploratory movements on objects. We propose a learning process that enables the robot to acquire a general pushing rule describing the relationship between the direction of poke and the observed object motion for a class of objects. In this way the robot acquires new action knowledge without having any specialized prior model about the action. Secondly, we investigate how the acquired action knowledge can be used to realize grasping in complex situations where the robot could not grasp the object without moving it to a new location. Here the learned poking behavior serves as a support action for robot grasping. The proposed approach has been implemented and tested on a humanoid robot Armar III. I