This article presents a methodology for the haptic perception of contour shapes of almost planar objects grasped by a five-fingered robot hand as well as the detection of any object cavity. The originality of our approach resides in (1) finding the reaction force patterns at the fingertips of a five-fingered robot hand that grasps different deformable objects (forward problem) and (2) using these contact force patterns to find the shapes of grasped objects (inverse problem) and (3) to determine material defects such as holes in an object with identified shape. Contact force patterns are generated in the forward problem by the finite element method (FEM) and the shape identification in the inverse problem is realized by a supervised neural network architecture using the backpropagation algorithm. Following shape identification, detection of holes is performed by clustering actual and prototypical contact force patterns using the self-organizing feature maps of neural gas networks as an unsupervised hole-screening method. (C) 1999 John Wiley & Sons, Inc.