Multi-sensor information fusion for state estimation is a well studied problem in robotics, with many applications and well known benefits. While classical methods for information fusion, based on information theoretic frameworks, provide good performance in optimal operation context, they show a deficiency in their ability to evaluate and take into account sensor measurements validity. As a consequence, these approaches may use erroneous information from a sensor and ruin the benefits of sensor redundancy.
This work begins to address this problem by learning context-dependant knowledge about sensor reliability. This knowledge is later used as a decision rule in the fusion task in order to dynamically select the most appropriate subset of sensors. For this purpose we consider the use of the Mixture of Experts framework (ME), traditionally applied to regression and classification problems. In our context each expert is a Kalman filter fed by a subset of sensors, and a gating network serves as a mediator between individual filters, basing its decision on sensor inputs and additional information needed for reasoning about the operation context. The performance of this model is evaluated in the context of UAV take-off/landing task for altitude estimation.
