Resumen de Inference over bayesian networks for the diagnosis and sensory enhancement of mobile robots
Mobile robots are nowadays present in countless real-world applications, aiding or substituting human beings in a wide variety of tasks related to scopes as diverse as industrial, military, medical, educational and many others. The use of mobile platforms in all these contexts is revolutionizing their respective fields, overcoming previous limitations and offering new possibilities. However, for a mobile robot to work properly, it is essential that its sensory apparatus provides correct and reliable information, which is often challenging due to the complexity of the physical world and its uncertain nature. To address that, this thesis explores the possibilities of the application of Bayesian networks (BNs) to the problem of sensory diagnosis and enhancement in the context of mobile robotics. Arised from the realm of artificial intelligence, Bayesian networks constitute a rigorous mathematical framework that enables both the integration of heterogeneous sources of information and the reasoning about them while taking their uncertainty into account. The thesis first analyzes different sensory anomalies in mobile robots and the impact of such abnormal behavior on the performance of these platforms. Given the wide variety of existing sensory devices, the analysis is focused on range sensors, since they are essential to many robotic tasks also grounded on probabilistic frameworks such as Bayesian estimators. Specifically, the thesis contributes with a rigorous statistical study of the influence of abnormal range observations on the performance of Bayesian filters, addressing the problem from a generic perspective thanks to the use of BNs. The conclusions obtained serve to illustrate the importance of sensory abnormalities beyond the pervasively studied issue of noisy observations. The treatment of sensory anomalies in mobile robots with Bayesian networks is then addressed. The thesis contributes with a novel modeling paradigm, a so-called Bayesian sensory architecture, that enables the representation of any robotic sensory system, allowing the identification of anomalies and the recovery from them. The main drawback of this proposal is the potentially high computational cost of inference with Bayesian networks, which is addressed with a novel, approximate algorithm that leverages the structure of the proposed model. Both the sensory architecture and the corresponding inference algorithm are implemented for different robotic tasks, and are validated through different sets of both simulated and real experiments. One of the implementations is aimed at analyzing the performance of the proposed algorithm in terms of error and computation time. The results obtained from the experiments show that the cost of inference is significantly reduced, and that the approximate queries produced still serve to perform sensory diagnosis and recovery adequately. Another implementation is proposed for the problem of robotic navigation in human environments. In this case, the experimental results prove that the use of the architecture manages to increase the safety and efficiency of navigation. Lastly, a new inference approach based on the use of feedforward neural networks is implemented and tested for this problem, showing that it is possible to reduce, even more, the cost of inference with Bayesian networks, enabling real time operation.
