Speakers
Reinforcement learning-based adaptive control algorithm for PNT
A robust navigation engine is critical to ensure a reliable PNT solution, especially in environments where sensor quality and availability vary over time. One of the primary challenges is determining the appropriate level of trust to assign to each sensor. Effectively managing this trade-off is necessary to maintain the overall robustness of the PNT solution. Fixed-configuration systems often struggle in time-varying scenarios. These include deep urban canyons, dense forests, and highly dynamic platforms such as launch vehicles. Most efforts have been focused on the dynamic adjustment of measurement and process covariance matrices. In many cases, measurement covariance is adjusted based on the carrier-to-noise density ratio (C/N₀) of incoming pseudorange signals.
This presentation introduces a reinforcement learning–based adaptive control algorithm that dynamically adjusts the covariance matrices of a navigation unit to achieve robust PNT solutions. The approach builds upon recent work that developed an RL-based adaptive control solution to optimise the response time of tracking loops in a GNSS receiver and now extends this technique to navigation engines. In this framework, the environment represents the navigation unit of the GNSS receiver, while the agent embodies the adaptive control mechanism. The reinforcement learning algorithm learns the optimal parameters for the adaptive control system by leveraging selected features extracted from the environment.
To optimise the policy parameters (i.e., weights of the control algorithm), the REINFORCE method is used. Overall, this work highlights the significant potential of reinforcement learning to enhance adaptive control strategies in PNT systems.
