Abstract
It is important for robots that act in human-centered environment to manipulate objects using visual information, where adaptability to unknown factors such as properties of robot body and objects is required. In this paper, we propose a learning method to acquire visual representation of robot body and object that is suitable for motion learning in a bottom-up manner. An advantage of the proposed framework is that it does not require specific hand-coding depending on the visual properties of objects or the robot, such as colors, shapes and sizes. Objects are extracted by a subtraction technique and the state space is constructed by SOM based on the images of extracted objects. Motion of the robot is planned based on reachable set that expresses a region where the object can reach. The task to move an object to a target position is divided into two phases, one to reach a position that is suitable for starting pushing and pulling motion and the other to push and pull the object to the target. The proposed method is verified by experiment of pushing and pulling manipulation of an object with a robot arm.
