Intelligent Navigation in Urban Environments Based on an H-infinity Filter and Reinforcement Learning Algorithms

In urban areas, robustness of a positioning solution suffers from relatively unpredictable reception of attenuated, non-line-of-sight and multipath-contaminated signals. To reflect a GNSS signal propagation environment, parameters of a state estimation filter need to be adjusted on-the-fly. A mixed H2/ H∞ filter has been considered here to address the vulnerability of a minimum error variance estimator to measurement outliers. An emphasis between the H2filter and the H∞ filter (minimizing the worst-case error) is continuously adjusted by a reinforcement learning (RL) model. Specifically, a continuous action actor-critic RL model with eligibility traces is implemented. The Cramér-Rao lower bound is considered for the filter performance evaluation allowing for the RL reward computation. The algorithm has been tested on a real-world dataset collected with mass-market hardware applying tightly-coupled IMU/GPS sensor integration. A positive RL model learning trend has been identified in two segments of the trajectory with the highest obstruction environment, suggesting the applicability potential of the technique.