Abstract
There is a fundamental limitation in accuracy and robustness of traditional Global Navigation Satellite System (GNSS) positioning techniques that rely on the Weighted Least Squares (WLS) algorithm. Traditional techniques use static heuristic functions for assigning weights to pseudorange measurements in the WLS algorithm. This disclosure addresses these limitations by replacing the static heuristic functions that are used for assigning weights to pseudorange measurements in the WLS algorithm with a trained Neural Network (NN). This NN acts as a dynamic Satellite Weight Prediction Model. Per the techniques described herein, the entire positioning pipeline—from satellite features to final positional error—is treated as a single, differentiable system. This allows the parameters of the NN to be optimized directly to minimize the final positional error (loss function of calculated position vs. ground truth position). This is a key advantage over conventional techniques.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Liu, Yu; Hoang, Duong; Atkinson, Ian; Tawashi, Mohammed; Patkar, Saket; and Sahai, Anant, "End-to-End Differentiable GNSS Positioning with Neural Network-Derived Weights", Technical Disclosure Commons, (December 11, 2025)
https://www.tdcommons.org/dpubs_series/9023