A hierarchical world model with sensor- and task-specific features

Abstract

The design of mobile robots which can cope with unexpected disturbances like obstacles or misplaced objects is an active field of research. Such an autonomous robot assesses the situation by comparing data from one or more sensors with an internal representation of its environment. In this paper we present a hierarchically structured world model that combines a general geometric object representation with sensor- and task-specific features and therefore can be used for various sensors and perception tasks. By predicting only those features that, first, can be detected by the sensor and, secondly, are relevant for the current perception task, sensor data interpretation gets faster and more robust. This is illustrated at an exemplary perception task concerning the video-based determination of the joint states of articulated objects.