Multisensor fusion of contact data

In this thesis, evidence showing the need for robotics is first presented. Then, sensing is shown to be a necessary component of practical robotic systems. Our focus is then narrowed to contact sensing, which is shown to be useful and required in many applications. The need for multiple sensors and for sensor fusion is next presented. Fusion is defined and illustrated by an example. Next, the subject of contact sensing is taken up in some detail. Various methods of detecting contact, force, slip, torque, etc. are described. Criteria for the design of contact sensors are presented. Various strategies which may be used to derive useful information for various intended applications from the use of contact sensors are examined, as well as methods of processing this data and extracting information about object shape, location, etc. Four general classes of data fusion methods are briefly described, and some recent work in sensor fusion is examined. This leads us to consider the fusion process in more detail. It is then seen that fusion methods serve to accomplish either one of two seemingly contradictory purposes: knowledge addition or knowledge verification. (Other desired charcteristics of fusion functions are also given.) Various existing fusion functions are then examined. Then, utilizing the two basic principles, we derive a new analytic method uniting both principles in such a way as to allow any desired "mix" or ratio of them. In order to test these fusion methods, real force data are acquired by means of active robotic contact sensing. Features (edges) are extracted, and the different methods applied to the fusion of the data. The relative amounts of the two principles involved in the various methods are computed using the concepts of ideal volume and actual volume and illustrated in graphs and tables. Comparison of these methods show advantages of the analytic method.