Perception as stochastic grammar-based sampling on dynamic grahp spaces

Abstract

This thesis develops and studies novel techniques that allow robots to properly model their environments autonomously. For this purpose it is possible and feasible to use all the available information that robots can use. Generally this information results in if-then-else constructs that are hard to understand then the environments of the robots are considerably complex. It is proposed to use �Active Grammar-based Modeling� (AGM), a new technique developed within this thesis. It uses very high-level descriptions that allow developers to achieve higher flexibility and scalability, as well as reducing the development time and the amount of programming errors. The solution consists on describing the grammar of the environment using a domain-specific language that is compiled into PDDL, allowing AGM-based systems to use classic AI planners to decide what robots should do to achieve their goales and incrementally verify that the model generated is valid according to the grammar described. Moreover, AGM can coordinate different particle filters so they can work simultaneously, allowing to choose the most appropriate filters depending on the context. This enhances the accuracy and effectivenes of the perceptual systems of the robots Along AGM, this thesis also presents the different algorithms used by AGM, as well as different experiment related to active indoor modeling using RGBD cameras.