Manipulation of Objects using an AIBO

Present techniques for manipulation with an AIBO involve specialized routines written for specific object shapes and domains, such as ball manipulation in RoboSoccer, where fixed sets of key-frames or fixed motion sequences have often been used, which the AIBO performs without feedback, similar to acting out a pre-defined script. This approach is difficult to generalize, and thus the AIBO is unable to adapt to new situations. This research focuses on general methods that allow an AIBO to manipulate classes of objects, such as boxes, spheres, and cylinders. The manipulation techniques are as general as possible, so that the AIBO is able to act on a variety of objects. Visual feedback is used to guide the AIBO as it performs the desired action. This allows the AIBO to adapt to changing circumstances and subtle differences in the objects being manipulated. In this paper, we first describe a technique we use to create a coherent model of a box object by composing various camera images, so as to allow the AIBO to have some understanding of the object. Next, we describe an approach technique that allows the AIBO to approach an object from any direction, while maintaining visual contact with the object and keeping a safe distance from it. Finally, we describe a general manipulation technique that uses the AIBO’s body to push the object, such that the AIBO can effectively manipulate the object’s position and orientation. SCHOOL OF COMPUTER SCIENCE, CARNEGIE MELLON UNIVERSITY ii