Robotics & Embodied AI
A new adaptive Kalman filter estimates both measurement and process noise covariances and remains accurate in simulations, a 2025 study reports
In a 2025 paper in IEEE Access, researchers at Universität Ulm and Friedrich-Alexander-Universität Erlangen-Nürnberg present an adaptive Kalman filter that updates two kinds of noise terms at once. A Kalman filter is a widely used method for estimating the hidden state of a changing system from noisy measurements. It normally requires fixed noise covariance values, which describe how uncertain the measurements and the system's own changes are. Those values are often unknown or may shift while a system runs. The new method adds a Kalman smoothing step and uses estimators derived from that smoothed information to adjust both noise terms. In the simulation tests described in the abstract, the approach showed accuracy and robustness.
How does this method differ from earlier adaptive filters?
Earlier adaptive Kalman filters often adapt only the measurement noise covariance or only the process noise covariance. They can drift or diverge when both are unknown. Most also provide no convergence result or only a local one, so a poor starting point can hurt the state estimate. The method in this paper uses an additional Kalman smoothing pass and analytically derived covariance estimators to adapt both noise terms.
What did the simulations show?
The abstract reports a detailed evaluation across various simulation scenarios. In those simulations, the method demonstrated accuracy and robustness. The abstract does not give specific error figures or scenario details.
FAQ
What does a Kalman filter need noise covariance for?
A Kalman filter combines a model of how a system changes with noisy measurements. The noise covariances tell it how much weight to give the model and how much to give the measurements. If those values are wrong, the estimate of the system's state can suffer.
Does this algorithm adapt both kinds of noise?
Yes. The abstract says the method estimates the measurement noise covariance and the process noise covariance at the same time, using Kalman smoothing and derived estimators.
Has it been tested on real systems?
The abstract describes simulation testing only. It does not say whether the method has been tested on physical systems or field data.
Limits and what's next
The evaluation described in the abstract is based on simulations. The abstract does not give the number of simulation scenarios or the specific systems used. This is one study, and the abstract reports no independent replication. Convergence is shown only for time-invariant systems with one-dimensional measurement and process noise. For higher-dimensional problems, the abstract says convergence can be tested in simulation for the specific dynamical system rather than shown generally.
Source: Theresa Kruse et al., “Adaptive Kalman Filtering: Measurement and Process Noise Covariance Estimation Using Kalman Smoothing,” IEEE Access, 2025.