Consistent Integration and Propagation of Disparate Sensor Observations

Abstract

To operate efficiently in an unknown or uncertain environ ment, robots must take observations from many different sensors to provide information with which to build a robust world model. We describe a methodfor integrating partial, uncertain, geometric sensor observations into a robust, con sistent estimate of the state of the environment. The integra tion process uses a Bayes procedure for comparing disparate observations of geometric features, rejecting spurious mea surements, and providing partial updates of object locations to a world model. This integration mechanism can combine any number of observations from sensors that provide mea surements of different geometric features. The invariant topology of relations between uncertain geometric features is used to develop a method for propagating observations through the world model. This propagation mechanism forces a consistent interpretation of the environment to be main tained and makes maximum use of sensor information. The method described has been applied in a distributed sensing system, consisting of an active stereo camera platform and a robot-mounted force-tactile gripper. The results demonstrate the usefulness of the proposed techniques.