An estimator based exploration engine for autonomous learning systems

The problem of an intelligent system which acquires knowledge through active sensing is examined. Such a system can exert some control over the sensing apparatus in response to past sensor data and intelligent decision-making. A system consisting of a robotic arm, a fingertip sensor, and an intelligent controller is considered. When presented with an unknown object to be identified, the system intelligently explores the object so that convergence to an accurate shape representation is as fast as possible. The system thus `seeks´ out the features which it finds most useful in its sensing task. The exploration technique estimates system unknowns with a Kalman filter. The technique is formulated in terms as general as possible, so that applications may be found for many forms of active sensing, and for learning problems not necessarily restricted to unknown shape representation, such as efficient problem solving or assembly