Improving Bayesian inference efficiency for sensory anomaly detection and recovery in mobile robots

Abstract

For mobile robots to operate in real environments, it is essential that basic tasks such as localization, mapping and navigation are performed properly. These tasks strongly rely on an adequate perception of the environment, which may be challenging in some cases due to the nature of the scene itself, the limited operation of some sensors, or even both. A mobile robot should be able to intelligently identify and overcome abnormal situations efficiently in order to avoid sensory malfunctioning. We propose in this work a novel methodology based on Bayesian networks, which enables to naturally represent complex relationships among sensors, to integrate heterogeneous sources of knowledge, to deduct the presence of sensory anomalies, and finally to recover from them by using the available information. The high computational cost of inference is addressed by a new algorithm that takes advantage of our model structure. Our proposal has been assessed in several simulations and has also been tested in a real environment with a mobile robot. The obtained results show that it achieves better performance and accuracy compared to other existing methods, while enhancing the robustness of the whole sensory system.