Learning how to grasp under supervision

Abstract

The problem of grasping a generic sphere is addressed. A supervised learning approach using a multilayer neural network for learning the position in 3D space and the radius of the sphere is introduced. Learning is based on laser range finder measurements of the surface of spheres of known radii at known positions. The problem is first formulated. An analytical solution for a set of four laser range finders and a solution based on supervised learning are then given and compared. Experimental results showing the feasibility and novelty of the approach are reported.<>