ACTIVE SENSING STRATEGIES WITH NON-CONTACT COMPLIANT MOTIONS FOR CONSTRAINT BASED RECOGNITION

Data driven active sensing strategies for the adquisition of geometric features are presented. In a object recognition and localization process, these techniques allow to exploit the discriminant power of the features although they are partially occluded to some fixed sensors. Geometric features are represented taking into account the location uncertainty due to the measurement errors. To drive the sensors to observe the features, a control mechanism is required. We use non-contact compliant motions with proximity sensors to sense features and reduce uncertainty when their location is partially known. The control algorithms have been implemented in a experimental multisensorial system, using a PUMA 560 robot.