Features Classification Learning Grasping Selection Evaluation

{ A haptic object recognition methodology suitable for application with a multiingered robot hand is presented. The methodology exploits peculiar features of robot hands such as their distributed sensoriality and parallel kinematics for fast acquisition of contact data, and is based on volumetric representation models for eecient dynamic integration of the perceived information. Experimental results demonstrate the eeectiveness and applicability of the methodology to non-trivial examples. I. Introduction The potential for tactile sensing and haptic perception in general in robotics is well known. Haptics can provide general 3-D perceptual information to strengthen or substitute machine vision (if needed) as well as local, relevant knowledge to be exploited in carrying out complex manipulation tasks. In 1984 Ruzena Bajcsy, in an inspiring paper titled \What can we learn from one nger experiments?" 3], investigated the properties that could be inferred from the exploration of an object carried out using a single nger equipped with several sensors. Research has made signii-cant progress along this \one nger" or \sensor platform" paradigm (e.g. 9]), leading to more accurate and reliable sensing devices, better sensor integration modalities, and sophisticated sensor fusion techniques. Key features in this approach are: a relatively simple probing device (a ngertip or a pinch) equipped with a rich set of state of the art sensors (hence the phrase \sensor platform"); dynamic (\active") exploration strategies, consisting of multiple, sequential probing actions. Availability of robotic multiingered hands has led to investigation of new, diierent tradeoos 1, 2, 14, 20]. On one hand, each nger now can act as a sensor platform by itself, with multiple measures obtained in parallel rather than sequentially. On the other hand, the complexity of promising sensor technologies, such as tactile sensors, connicts with the complexity and miniaturization requirements of the dextrous hands themselves. Psychological studies about the organization of the hap-tic perceptual system in humans 13] have provided useful reference models for the architecture of a robotic counterpart. Yet, substantial diierences in kinematics and available sensing system between human and robot hands do not always permit straightforward emulation of such models. Among the many results from psychological studies 13], one we have capitalized upon to guide the research described in this paper is that enclosure grasps (also called grasps by containment or enveloping grasps) generally provide eeective capabilities for fast gross-shape discrimination. Emulation of a perceptual strategy based on grasp by containement calls for a robotic device provided with …