ETH Zurich's FW-NKF Beats KalmanNet by 50x on Drone State Estimation
FW-NKF embeds a learnable frequency filter into the Kalman update step, cutting localization error by up to 10% on drones and wearable IMUs
Takeaways
- FW-NKF adds a learnable IIR frequency filter to the Kalman innovation signal, teaching the filter which frequency bands are noise before updating the state estimate.
- Achieves up to 10% lower localization error and ~50x lower MSE on drone IMU tracking (EuRoC) vs. KalmanNet and other baselines.
- Tested on four benchmarks: Lorenz attractor, nonlinear pendulum, UWB-IMU human pose, and EuRoC MAV drone sequences.
- Key innovation: spectral-domain supervision on the observation model forces the network to reconstruct clean signals, not fit to corrupted ones.
- Code is open source (MIT) with pretrained checkpoints: github.com/eth-siplab/Frequency-weighted-neural-Kalman-filters
- Best suited for systems with structured, frequency-localized noise: drones, wearables, wheeled robots, or any sensor affected by vibration or EMI.
The Kalman filter has been the backbone of robot state estimation for over 60 years. It is elegant, computationally cheap, and mathematically principled. But it was designed for a world where sensor noise is white and Gaussian -- random, flat, and structureless. Real-world sensors are none of those things. A drone's IMU vibrates at motor frequencies. A wearable sensor picks up periodic gait artifacts. An encoder on a wheel accumulates low-frequency drift. In practice, sensor measurements contain structured disturbances such as low-frequency drift, high-frequency jitter, or periodic environmental noise, none of which are captured by the white noise assumption.
Researchers from ETH Zurich's Signal and Interactive Processing Lab have now published FW-NKF: Frequency-Weighted Neural Kalman Filters, a hybrid approach that directly addresses this blind spot. The paper was accepted at ICRA 2026.
The gap that deep Kalman filters left open
The past few years have seen a wave of neural Kalman filter variants -- KalmanNet, Bayesian KalmanNet, Recurrent Kalman Networks -- that replace hand-crafted noise models with learned ones. Deep Kalman Filter (DKF) variants extend the Extended Kalman Filtering (EKF) framework by learning latent transitions, but they lack explicit mechanisms to suppress band-limited noise components that typically corrupt sensor measurements in real-world scenarios.
The core issue is the Kalman innovation -- the gap between what the filter predicted the sensor would read and what it actually read. These spectral artifacts can lead to suboptimal state estimation, as the Kalman gain cannot distinguish between informative signal components and frequency-localized noise. No matter how expressive the neural network wrapping the filter is, if the innovation signal fed into the update step is contaminated with periodic noise, the filter will act on it.