Navigation system learns when to trust each sensor

Researchers at Beijing Institute of Technology have published a navigation method that estimates how much to trust GNSS, LiDAR and vision in real time, then uses that belief state to reweight sensor fusion. In tests on a complex urban route, the system beat several common baselines and kept working even after the GNSS branch was removed.

Why it matters: - Unmanned vehicles often lose reliable positioning when GNSS drops out, cameras struggle in low light or blur, or LiDAR becomes less useful in sparse or repetitive environments. - A system that adjusts sensor trust in real time could improve navigation in tunnels, urban canyons, forests and other difficult settings. - The method may also help drones and robots keep position when one sensing mode degrades.

What happened: - Researchers from the School of Automation at Beijing Institute of Technology published a study online Aug. 3, 2026 in Satellite Navigation, Volume 7, Article 21. - The study introduces a quantitative belief-guided method for adaptive multi-sensor fusion navigation. - The system builds a belief state and uses a learning-based decision agent to choose how to allocate weight across sensors in real time.

The details: - The belief state has two parts. - Exogenous availability belief measures the quality of environmental geometric constraints. - Endogenous credibility belief measures the reliability of internal estimation information. - Exogenous availability belief propagates landmark uncertainty through an information matrix, normalizes the constraint distribution and uses Shannon entropy to estimate a degree of availability for each sensor. - Endogenous credibility belief uses a sliding-window observability matrix to compare reconstructed state errors with theoretical noise propagation and produce a degree of credibility. - The decision agent uses a double deep Q-network with long short-term memory to encode temporal belief patterns. - The action space includes 66 prototype strategies for distributing weight among GNSS, LiDAR and vision sub-filters. - The model was trained on about 80% CARLA simulation data and about 20% real-world platform data, with quasi-hardware-in-the-loop noise augmentation to narrow the simulation-to-reality gap. - The system ran at about 33.6 Hz on the test platform. - In real-world tests on an urban route that included parking lots, gardens, main roads, breezeways and sidewalks, the method achieved 1.330 meters RMSE. - That beat the adaptive fusion baseline AFN at 1.518 meters, LIO-SAM at 2.061 meters and VINS-Fusion at 5.934 meters. - When the GNSS branch was removed, the system still reached 1.876 meters RMSE, better than LIO-SAM. - The study is available at the full paper. - The work was supported by the Basic Science Center Program of the National Natural Science Foundation of China and the National Natural Science Foundation of China.

Between the lines: - The main shift is from fixed fusion rules to a learned policy that adapts to changing sensor quality. - The temporal memory component matters because navigation conditions do not change in a single step; they evolve across a route. - The GNSS-off result suggests the decision layer adds value beyond any one sensor stack.

What's next: - The team plans to explore lighter versions for resource-constrained miniaturized platforms. - The team also plans to extend the framework to distributed multi-robot navigation, where multiple agents must coordinate sensing and positioning under shared uncertainty. - If the approach scales, it could become a practical tool for autonomous vehicles, drones and other systems that move through unpredictable environments.