Neural Aided Adaptive Innovation-Based Invariant Kalman Filter

Researchers Barak Diker and Itzik Klein have introduced a novel navigation algorithm, the Neural Aided Adaptive Innovation-Based Invariant Kalman Filter, which merges classical control theory with modern machine learning. The core innovation is a theoretical extension of innovation-based process noise adaptation formulated directly within the Lie-group framework, a principled geometric space that offers favorable error dynamics. This is paired with a lightweight neural network designed to estimate process noise covariance parameters directly from raw inertial sensor data.

Trained entirely using a sim2real (simulation-to-reality) framework enhanced with domain adaptation, the neural network learns to capture complex, motion-dependent and sensor-dependent noise characteristics without needing labeled real-world data. This addresses a critical gap, as adaptive noise estimation methods had remained largely unexplored in the tangent Lie space. The team validated their approach on the challenging real-world scenario of autonomous underwater navigation.

Experimental results demonstrated that the proposed filter achieves superior performance compared to existing methods, specifically in reducing position root mean square error (RMSE). These findings validate the effectiveness of the sim2real training pipeline and confirm that geometric invariance significantly enhances learning-based adaptation. The work establishes adaptive noise estimation in the tangent Lie space as a powerful mechanism for improving navigation accuracy in complex, nonlinear autonomous systems like drones, robots, and underwater vehicles.