A Variational Bayes approach for reliable underwater navigation

This paper presents a filtering algorithm for non-linear systems in the case of sensor degradation. The algorithm adapts the relative importance of the sensor measurements, compared to the model predictions, in real time; yielding a filter that is robust to noisy observations and sensor blackouts. The filter is constructed using a Variational Bayes Approximation of the conditional probability distribution of the system's state; i.e., the probability distribution of the state, given the measurements from the sensors. The algorithm is evaluated both in simulation and experimentally on a robotic platform. In the experiments, the sensor measurements from an Autonomous Underwater Vehicle (AUV) are altered artificially. The sensor output is either corrupted with outliers or manually stuck to a constant value; simulating in this fashion a sensor defect. In both cases, the filter reconstructs the robot's state accurately, thus enabling the vehicle to resume with mission execution.