Learning Grasps Using Back Propagation in a Simulated Neural NET: A Cautionary Note.

We performed an experiment to test learning in neural nets. Our objective was to form a judgement as to whether the technology had advanced sufficiently to be applicable to teaching tasks to a robot hand. Specifically, we sought to determine whether a simulated neural net could learn, under ideal conditions, the attributes of stable grasp from a set of training data. The training data consisted of the perimeters of two-dimensional star-like objects, and the orientation that a gripper composed of three frictionless fingers would take to grasp them. The hand was always centered over the center of mass of the object, and the circumstances of the experiment were such there was always exactly one stable grasp position.