FW-NKF: Frequency-Weighted Neural Kalman Filters
2026-06-01 • Robotics
RoboticsArtificial Intelligence
AI summaryⓘ
The authors tackled a common problem in robotic state estimation where traditional methods struggle with certain types of noise like vibrations and periodic disturbances. They created a new method called Frequency-Weighted Neural Kalman Filter (FW-NKF) that combines neural networks with a technique to filter out specific noise frequencies. By doing this, their approach better cleans up sensor data and improves accuracy in estimating positions and orientations, especially in complex systems like chaotic models and full-body motion tracking. Their tests showed up to a 10% decrease in errors compared to previous methods.
Kalman FilterExtended Kalman Filter (EKF)Deep Kalman Filter (DKF)frequency-dependent noisespectral shapinglatent state representationsensor vibrationsinertial pose estimationchaotic systemslocalization error
Authors
Adnan Harun Dogan, Berken Utku Demirel, Christian Holz
Abstract
Robust state estimation is central to robotic autonomy, yet classical Kalman filters struggle with frequency-dependent disturbances and model mismatch such as sensor vibrations, electromagnetic interference, and periodic noise. Although Deep Kalman Filter (DKF) variants extend the Extended Kalman Filtering (EKF) framework by learning latent transitions, they lack explicit mechanisms to suppress band-limited noise components that typically corrupt sensor measurements in real-world scenarios. We introduce the Frequency-Weighted Neural Kalman Filter (FW-NKF), a unified hybrid approach that embeds a causal spectral-shaping operator into the Kalman measurement residual and jointly learns observation, and transition networks. By adapting both the filter spectrum and the latent state representation, FW-NKF attenuates the noise-dominated frequency bands while capturing complex residual structures. We conduct extensive experiments on four heterogeneous benchmarks, including chaotic systems such as multi-dimensional Lorenz systems and full-body inertial pose estimation, and find a reduction in localization error of up to 10% as well as marked improvements in orientation accuracy. Our ablation studies confirm that frequency weighting and deep latent-state modeling contribute to overall performance.