The proposed robust fixed-structure Gaussian mixture filter, or rfsGMF, recorded the lowest position error in both noise cases of a simulated radar-tracking test. Its ANEES consistency scores were also closest to one among the three filters compared, according to a version 1 arXiv preprint dated 20 August 2026.
A filter built to keep its shape
The paper addresses Bayesian state estimation in discrete-time nonlinear systems when measurements contain outliers. Its predictive Gaussian mixture—a set of probability components used to represent the estimate—is fixed through an offline, axis-aligned decomposition of the transition density, making the structure deterministic and tunable by decomposition fidelity.
For the measurement update, outlier-affected components are modeled with a Student's-t distribution, and variational Bayes is used to approximate each Gaussian-mixture component's update. The nonlinear measurement map is handled with an unscented transform; when Gaussian and Student's-t channels are combined, their updates are applied sequentially to each mixture term.
A test built around radar
The numerical illustration modeled a two-dimensional object tracked with radar range and bearing measurements. Its bearing noise included outliers, and the state also contained an unknown bearing bias; the supplied extraction reports 103 Monte Carlo simulations, each representing a simulated tracking run.
The comparison used the pruned rfsGMF, a UKF with kappa set to 0, and a bootstrap particle filter (BPF) with 105 samples. Student's-t filters were left out because their assumed noise model did not match this example.
Results were reported as time-averaged RMSE and ANEES over steps 1 through 30. RMSE summarizes filtering error, while ANEES is the study's consistency diagnostic.
Across the two cases, rfsGMF had the lowest position RMSE. For x-position, its values were 6.321 in the first case and 8.514 in the second, versus 6.906 and 10.459 for BPF and 8.435 and 9.51 for UKF, respectively.
The ranking for the y-position was the same: rfsGMF was lowest in both cases. For the unknown bias, however, BPF had the lowest RMSE in the first case, while rfsGMF had the lowest in the second.
ANEES also favored rfsGMF by the paper's criterion of being closest to one. Its values were 1.901 and 2.132 across the two cases, compared with 2.584 and 2.308 for UKF and 2.756 and 3.599 for BPF.
A result with a narrow reach
The evidence is confined to one synthetic tracking model, two simulated noise cases, and the three filters in this comparison. It contains no real-world tracking data or human or animal participants.
The reported results are averages without confidence intervals, standard errors, or significance tests, and the paper gives no quantitative runtime, memory, or computational-cost results.
Student's-t filters and a broader range of robust estimators were not part of the comparison. The authors also identify efficient construction of the fixed Gaussian-mixture structure for high-dimensional problems as future work.
The authors describe UKF and rfsGMF as slightly optimistic on the ANEES diagnostic and BPF as too optimistic, then conclude that the robust filter consistently outperformed both comparators in the numerical illustration.
Paper data and sources
Original title: Fixed-structure Gaussian Mixture Filtering with Robust Measurement Updates under Outliers
Authors: Ondřej Straka, Uwe D. Hanebeck
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-20
DOI: Not available
Original paper · Full text