Learning motion from images

摘要

Describes a method of determining a robot end-effector's motionrequired to achieve a standard position and orientation relative to anobject through learning. By using a back-propagation network, theauthors establish the direct mapping from `what is seen' to `what shouldbe done'. The method does not need camera calibration, nor hand-eyecalibration, nor explicit object model. Some general rules for correctlearning are presented. A recursive scheme of movement control isdesigned with convergence proof. The method is simulated on anapplication object and shows promising application potential