[Paper] Robust Unscented Kalman Filtering via Recurrent Meta-Adaptation of Sigma-Point Weights

Source: arXiv - 2603.04360v1

Overview

The paper presents Meta‑Adaptive Unscented Kalman Filtering (MA‑UKF), a new way to make the classic Unscented Kalman Filter (UKF) far more resilient to changing dynamics and non‑Gaussian noise. By treating the sigma‑point weights as learnable hyper‑parameters and updating them with a recurrent meta‑learning module, the authors achieve a filter that can “learn how to trust” its predictions versus measurements on the fly.

Key Contributions

• Reformulation of sigma‑point weight selection as a hyper‑parameter optimization problem rather than a fixed, hand‑tuned rule.
• Memory‑augmented meta‑learning architecture (a Recurrent Context Encoder + policy network) that ingests the entire history of measurement innovations to produce context‑aware weight vectors at each time step.
• End‑to‑end differentiable UKF loop, enabling gradient‑based training of the weight‑generation policy through the filter’s recursive equations.
• Robustness to heavy‑tailed and glint noise, demonstrated on maneuvering‑target tracking benchmarks where MA‑UKF outperforms standard UKF, Adaptive UKF, and particle‑filter baselines.
• Generalization to out‑of‑distribution dynamics, showing that the learned weight policy can handle motion patterns not seen during training.

Methodology

1. Problem Setup – The classic UKF propagates a set of sigma points through nonlinear dynamics and recombines them using static weights derived from scaling parameters (α, β, κ). These weights implicitly assume Gaussian noise and stationary dynamics.

2. Meta‑Adaptation Layer –

   • Recurrent Context Encoder: A gated recurrent unit (GRU) processes the sequence of innovation vectors (measurement – prediction) and compresses the past into a low‑dimensional latent state (h_t).
   • Policy Network: A small multilayer perceptron takes (h_t) and outputs two vectors: the mean‑weight vector (w^{(m)}_t) and the covariance‑weight vector (w^{(c)}_t). These replace the fixed UT weights at time t.

3. Differentiable UKF Loop – The UKF equations (sigma‑point generation, propagation, measurement update) are written in a way that gradients can flow through the weight vectors. During training, the loss is the mean‑squared tracking error over a simulated trajectory.

4. Training Regime – The system is trained offline on a diverse set of simulated motion models (constant‑velocity, coordinated turns, abrupt accelerations) and noise conditions (Gaussian, heavy‑tailed, glint). The recurrent encoder learns to capture “how surprising” recent measurements have been, and the policy learns to down‑weight unreliable sigma points accordingly.

5. Inference – At runtime, the MA‑UKF runs exactly like a standard UKF, except that before each update it queries the encoder‑policy pair for the current weight set. No extra computational overhead beyond a forward pass through a tiny neural net.

Results & Findings

• Robustness: MA‑UKF maintains low error even when measurement noise deviates dramatically from the Gaussian assumption.
• Adaptivity: The learned weight policy quickly reduces the influence of outlier measurements, effectively performing an online “robustification”.
• Consistency: Covariance estimates remain well‑calibrated, avoiding the over‑confidence that plagues many adaptive filters.
• Computation: The added neural‑net forward pass adds < 0.5 ms per filter step on a modern CPU, preserving real‑time capability.

Practical Implications

• Autonomous Vehicles & Drones – State estimation for navigation (e.g., GPS/IMU fusion) often suffers from multipath or sensor drop‑outs. MA‑UKF can automatically down‑weight corrupted measurements without hand‑tuned gating thresholds.
• Radar / Sonar Tracking – In cluttered environments, glint and spurious returns are common. The meta‑adaptive weights give a principled way to stay locked on the true target.
• Industrial IoT & Robotics – Low‑cost sensors produce non‑Gaussian noise; deploying MA‑UKF can extend the usable lifespan of legacy hardware.
• Edge Deployment – Because the policy network is tiny, the approach fits on micro‑controllers or embedded GPUs, enabling robust filtering on the edge.
• Reduced Engineering Overhead – Engineers no longer need to manually tune α, β, κ for each new platform; the filter learns the optimal weighting strategy from a short offline simulation phase.

Limitations & Future Work

• Training Dependency – The meta‑learner must be pre‑trained on representative dynamics; extreme OOD scenarios (e.g., completely new sensor modalities) may still degrade performance.
• Memory Footprint – The recurrent encoder stores a hidden state; while modest, it adds a small memory requirement compared to a vanilla UKF.
• Theoretical Guarantees – The paper provides empirical consistency but lacks a formal proof of filter stability under the learned weights.
• Future Directions – Extending the framework to handle multi‑sensor fusion with heterogeneous update rates, exploring transformer‑based encoders for longer histories, and deriving analytical bounds on estimation error with meta‑adaptive weights.

Authors

• Kenan Majewski
• Michał Modzelewski
• Marcin Żugaj
• Piotr Lichota

Paper Information

• arXiv ID: 2603.04360v1
• Categories: cs.LG, eess.SP
• Published: March 4, 2026
• PDF: Download PDF