Toward a learning robot

The problem of robot manipulation-of planning how to successfully grasp, push, and pull arbitrary objects to reconfigure them as desired-is an unsolved problem in robotics. We are presently developing a hand-eye system which is intended to begin with simple knowledge about how to manipulate its world, and to itself acquire increasingly refined knowledge about manipulation in specialized contexts. For example, the robot may begin with knowledge such as "to move an object somewhere to the right, push from some point on the left", and may refine the knowledge through experience into more specific, more precise assertions such as "to move an object which is in contact with a wall to the right in a straight line, push from a point on the left which is 2/3 of the distance from the wall" The learning method is a type of explanation-based generalization driven by an incomplete theory of the domain-in particular by a theory for qualitative differential analysis. This paper presents an approach to constructing a planning/learning system of this kind, and is primarily organized around a hand-generated example of the type of behavior our system is intended to exhibit. This research has been supported by NASA Ames under grant NCC 2-463, and in part by NSF under grant IRI-8740522. 1 Table of