Learning by doing-an approach to robotic skill acquisition

Abstract

An approach to skill acquisition and knowledge representation for the control of vision-based calibration-free robots is introduced. It allows a robot to collect experiences and knowledge automatically during its normal operation, and to adapt them to changing conditions. This, in turn, makes the robot improve its skills and operation speed over time without any operator intervention and gives it self-learning characteristics in a form of learning by doing. The concept has been successfully realized and tested in real-word experiments involving the grasping of a variety of differently shaped objects by a visually guided calibration-free manipulator.