COMPUTATIONAL MODELING OF HAND EYE COORDINATION

We are developing a visually guided robot arm that is able to grasp and transport objects across an obstacle-strewn environment. Recent work has shown how analysis of moving image contours can provide estimates of the shape of curved surfaces. This forms the basis of the robot's capabilities which we are seeking to elaborate in three respects. First a computational model - the 'dynamic contour' - is being developed for real-time visual tracking, based on simulated Lagrangian Dynamics. Secondly we are developing a two-and-half-dimensional system for incremental, active exploration of free-space. Thirdly a computational theory of the visual planning of two-fingered grasps has been developed which can be coupled to the 'dynamic contour' model. This paper concentrates on the third of these topics. A qualitative theory for classification of grasps is described and demonstrated. Optimal finger positions are obtained from the mutual intersection of the local symmetry and anti-symmetry sets, and of a third set, the critical set of the grasp map. These three sets themselves form boundaries in a natural partition of the set of all grasps (the configuration space). The partition corresponds to a classification of possible grasps according to their stability properties.