The overall objective of this project is to increase the capability of dextrous robot hands to flexibly manipulate unknown objects in unstructured environments. Tactile sensing will be used to determine global properties of an unknown object using class models, rather than specific object models. Specific goals of the project include: 1) extracting sufficient knowledge from local surface information to grasp an manipulate unknown objects, and 2) determining as much about the object as is possible from sparse sensory information, particularly from only an initial grasp. Cylindrical tactile fingertips have already been developed, and algorithms for extracting local contact properties have been demonstrated, namely position, orientation, total force, and radii of curvature for a convex second-order surface. Global properties including size, location, and orientation will be determined from the local contact shape at multiple simultaneous finger contacts. An object class of cones will be assumed, and geometric constraints will be derived from the measurements. In experiments, an unknown cone will be touched with 3 tactile sensors. Additional touch probes may be required to find the cone's global properties.
