Automatic preshaping for a dextrous hand from a simple description of objects

Human grasping has been studied intensively and has shown that during the reaching phase of grasping, the hand preshapes in order to prepare the 'shape matching' with the object. A preshape represents a class of configurations, among which one is executed by adjusting the fingers against the object. Preshaping is guided by visual data providing with the object description. The expected object modelling for preshaping is very far from classical CAD models: the shape must be described as being rather round, flat, thin, long, among others; size may be, for example, rather small or large. A major issue addressed in this work is the automatic extraction of such a model from low-level visual data. A system for preshaping a planar bi-fingered hand with four joints grasping planar objects is described: the input is a bitmap of the object from which size and shape descriptions are extracted and used for preshape selection; a preshape is executed using a potential field approach.<>