Pushing using learned manipulation maps

Abstract

Robot haptics ultimately consists of a set of models which interpret and predict a robot’s physical interaction with the world. In this paper, we describe one approach to modeling support friction within a two-dimensional environment consisting of a single robot finger pushing objects on a table. Instead of explicitly modeling the friction distribution between the object and the table, we learn the mapping between pushes and the motion of the object using an online, memory-based model using local regression. The resulting manipulation map implicitly describes the support friction without a complex model. We also describe methods of acquiring object shape and localizing the object using a proximity sensor. Results are presented for objects with different friction distributions.