Active Tactile Feature Extraction
Jonathan D. Layes and Roger A. Browse
Abstract
This research proposes that practical tactile object recognition will depend not on high resolution
information contained in individual tactile images, but on generalizing tactile images by basic
features. These features can then be used to build an approximate representation of the object.
The problem of classifying tactile images by a novel set of compound primitives is examined.
We present an implementation of a robotic system capable of exploring the workspace, acquiring
tactile images, and using a backpropagation neural network to extract information from the
collected images. The system is comprised of distinct components, consisting of a robotic
manipulator interface, a tactile acquisition interface, a path planning and control overseer, and a
network server to allow message passing between the other clients. Using this networked client-server approach, the system's components were distributed across several workstations to
provide extensibility and to allow computationally intensive processes to be shared among
several processors.