Jaguar Land Rover создал навигационную систему на базе графов факторов для работы без GPS

Jaguar Land Rover (JLR) has demonstrated a proof-of-concept navigation system based on factor graph optimization (FGO), targeting vehicle localization in environments where GNSS signals are degraded, obstructed or unavailable. FGO represents navigation as a graph linking vehicle states with sensor measurements and constraints, which are optimized jointly over a time window rather than processed sequentially as in conventional recursive filtering. The work was conducted under the European Space Agency NAVISP project ‘Proof-of-concept of advanced navigation algorithms based on factor graph optimization’. It comprised the development of a complete FGO-Nav unit implementing a factor-graph-based navigation pipeline. The modular architecture supports GNSS and inertial measurement units (IMUs), with provision for integration of additional sensors including LiDAR, cameras and radar. A sensor abstraction layer (SAL) provides hardware-agnostic sensor integration and supports reproducible offline replay, while an integrated KPI suite and graphical user interface enable configuration, performance analysis and comparison between estimation methods. Testing factor-graph navigation In the JLR FGO implementation, GNSS, IMU and motion-model factors are incorporated into a graph containing position and velocity states. A nonlinear least-squares optimizer then estimates the state by minimizing the residuals associated with the factors, using either Gauss-Newton or Levenberg-Marquardt optimization. Researchers collected more than 24 hours of synchronized GNSS/INS data across suburban, light-urban and deep-urban environments, including tunnels, non-line-of-sight zones, high-rise corridors and tree-shadowed sections. The datasets included real-world recordings, synthetic data and the public UrbanNav dataset. FGO implementations were benchmarked against weighted least squares (WLS), extended Kalman filtering (EKF) and a commercial GNSS/INS RTK receiver. Results varied with the test environment. In a Birmingham light-urban dataset, tightly coupled FGO (TC-FGO) produced a 2D root mean square error (RMSE) of 2.02 meters, compared with 2.03 meters for WLS and 2.11 meters for the EKF-based solution. In the deep-urban datasets, the advantage became more pronounced: one UrbanNav dataset produced a 2D RMSE of 9.82 meters for loosely coupled FGO, compared with 11.07 meters for the corresponding EKF solution. The project also examined the effect of the FGO optimization window. Testing indicated that a 10-epoch window provided a good balance between accuracy, robustness and computational load, with only small differences in 2D RMSE between two-, 10- and 30-epoch configurations. The final system was delivered as a modular proof-of-concept incorporating the data logger, SAL, FGO engine, KPI framework and GUI. JLR identified improved vertical accuracy, adaptive window management, expanded datasets and integration of vehicle odometry, LiDAR, camera and radar as areas for further development.