Intelligent grasping using neural modules

Abstract

This paper presents a three-fingered robot gripper which is able to capture objects of arbitrary shape. To handle such objects, we propose a system formed by two stages: image processing and object contact points definition. A vision system captures a top image of the object and uses the nearest-neighbor method to define a set of points representing the object outline. In the second stage, two neural network architectures work together to select three contact points in the outline for the gripper. The first neural network (competitive Hopfield neural network) executes a polygonal approximation over the set of points, reducing the number of points to be processed. The second neural network (radial basis function-global ridge regression) determines three contact points from the approximated polygon. This system yields stable contact points for objects with arbitrary shapes and performs within time intervals compatible with online applications.