Abstraction of Manipulation Tasks to Automatically Generate Robot Motion from Observation

ion of Manipulation Tasks to Automatically Generate Robot Motion from Observation Jun TAKAMATSU and Katsushi IKEUCHI † Institute of Industrial Science, The University of Tokyo †† Graduate School of Interdisciplinary Information Studies, The University of Tokyo Institute of Industrial Science, The University of Tokyo, 3rd Dept. Ikeuchi Laboratory 4-6-1 Komaba, Meguro-ku, Tokyo, 153-8505, JAPAN +81-3-5452-6242 [email protected] Abstract For a framework to automatically generate robot motion to execute various tasks, the Learning from Observation (LFO) paradigm has recently been proposed. In this paradigm, a robot system observes and recognizes a demonstration of a task, builds an abstract representation of the task, and finally generates robot motion which consists of repetitions of movement primitives. Movement primitives are essential operations to realize the situation corresponding to the abstract representation. Here, the key to the successful implementation depends on the representation of the target task and the definition of sufficient movement primitives. The target tasks of this paper are everyday manipulation tasks including an assembly task and manipulation of linkages which are connected by a joint. In assembly tasks, a contact relation between the two objects is employed as the abstract representation. Because of overcoming the difficulty to deal with numerous kinds of contact relations, we define a Motion Degree of Freedom (DOF), which is an index that represents the quality of possible local displacement of an object in each contact relation and define sufficient movement primitives in advance using Motion DOFs. Next, we propose a method to deal with manipulation of linkages which are connected by a revolute joint. Considering the task of rotating a doorknob, which is connected to a door by a revolute joint, it is possible to rotate the knob, if one knows the axis direction and the center of rotation (parameters of the joint). Therefore, we propose a method to estimate the parameters from noise-contaminated observation.