Pose-independent recognition of convex objects from sparse tactile data

The paper describes a technique for pose-independent recognition of convex, free-form objects from tactile sensing. The technique is based on internal and external volumetric approximations of the unknown object built from sensed contact information. Center of mass and principal inertia axes of the volumetric approximations are then exploited to estimate object pose. Recognition is carried out based on features extracted from the approximations considered in their currently recovered pose. The paper illustrates attained pose-independent recognition performance, with the help of automated exploration strategies, on a problem involving a set of 20 objects. Registration performance is largely dependent upon the accuracy of the developed volumetric approximations and hence upon the extent of the exploration. The main advantages of the overall recognition technique are its ability to deal with sparse sensory data and its limited computational requirements.