Title: FW-NKF: Frequency-Weighted Neural Kalman Filters

Abstract

While robust state estimation is a cornerstone of robotic autonomy, traditional Kalman filters often falter when confronting model inaccuracies and frequency-specific disturbances, including electromagnetic interference, sensor vibrations, and periodic noise. Although Deep Kalman Filter (DKF) variants build upon the Extended Kalman Filtering (EKF) framework by learning latent transitions, they do not offer explicit methods for filtering out band-limited noise that frequently degrades sensor data in practical applications. To address this, we present the Frequency-Weighted Neural Kalman Filter (FW-NKF), a cohesive hybrid methodology that integrates a causal spectral-shaping operator directly into the Kalman measurement residual while simultaneously training both observation and transition networks. By optimizing both the latent state representation and the filter’s spectrum, FW-NKF effectively dampens noise-heavy frequency bands and preserves intricate residual patterns. Extensive testing across four diverse benchmarks—ranging from full-body inertial pose estimation to chaotic multi-dimensional Lorenz systems—demonstrates that the proposed method reduces localization error by as much as 10% and significantly enhances orientation accuracy. Furthermore, ablation studies validate that the combination of deep latent-state modeling and frequency weighting is essential for achieving these performance gains.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC