ARC: Adaptive Robust Joint State and Covariance Estimation
Proposes ARC: a unified block-coordinate descent framework combining adaptive robust loss, IRLS, and MWCD for joint state and covariance estimation, robust against outliers and non-Gaussian noise, with automatic parameter tuning.
Key Findings
Methodology
The ARC framework employs a block coordinate descent (BCD) strategy, integrating a norm-aware adaptive robust loss function, iteratively reweighted least squares (IRLS) for state updates, and a minimum weighted covariance determinant (MWCD) estimator for covariance re-estimation. Initially, the loss shape parameter α is adaptively tuned by minimizing the negative log-likelihood of residuals, allowing the model to automatically adjust robustness levels. During each iteration, residuals are weighted using the norm-aware function, which accounts for the residual distribution, and these weights are used to refine the state estimate via IRLS. Subsequently, the covariance is re-estimated by applying MWCD on residuals after outlier removal, leveraging the same weights to ensure consistency. The process iterates until convergence criteria—based on changes in state, covariance, and loss parameters—are met. This integrated approach enables the estimator to adaptively handle outliers and unknown noise statistics simultaneously.
Key Results
- In Monte Carlo simulations with up to 50% outliers, ARC maintains a position RMSE below 0.1 meters, outperforming traditional L2 and fixed-loss robust methods, with a significant reduction in covariance estimation error as measured by Wasserstein-2 distance, which remains close to zero. The adaptive estimator accurately recovers the true inlier measurement covariance across varying contamination levels, with W2 distances less than 0.02, indicating high fidelity in noise modeling.
- In real-world ultra-wideband (UWB) indoor localization experiments, ARC effectively identifies and suppresses heavy-tailed NLOS errors, achieving a standard deviation estimate close to the true inlier noise (~0.16m). It outperforms fixed-loss schemes such as Charbonnier and Cauchy, which tend to over- or under-estimate noise levels, leading to miscalibrated uncertainty. The method demonstrates robust performance across multiple trajectories, both in uncluttered and cluttered environments, with average RMSEs around 0.097 meters, significantly better than baseline methods.
- Ablation studies confirm that the continuous norm-aware weighting scheme in MWCD enhances covariance recovery and state accuracy, especially under high outlier contamination. The combination of adaptive loss and covariance estimation provides a resilient framework that maintains high estimation accuracy and reliable uncertainty quantification, even in challenging environments.
Significance
This work advances the field of robust state and noise covariance estimation by introducing a self-tuning, unified framework capable of handling complex, real-world noise conditions without manual parameter tuning. It addresses a long-standing challenge in robotics and sensor fusion: how to reliably estimate states when measurements are contaminated by outliers and the noise statistics are unknown or time-varying. The proposed ARC method not only improves estimation accuracy but also enhances the reliability of uncertainty quantification, crucial for safe autonomous operation. Its ability to adaptively recover measurement noise covariance in the presence of non-Gaussian, heavy-tailed noise significantly broadens the applicability of robust estimation techniques, making it highly relevant for indoor navigation, autonomous vehicles, and sensor fusion systems. The experimental validation on both simulated and real-world datasets demonstrates its practical viability and potential for industry deployment.
Technical Contribution
The core technical innovation lies in the integration of a norm-aware adaptive robust loss function, IRLS-based state refinement, and MWCD covariance estimation within a single block coordinate descent (BCD) framework. Unlike prior methods that treat robustness and covariance estimation separately or require manual tuning, this approach jointly optimizes all components, ensuring consistency and adaptability. The adaptive loss function dynamically adjusts its shape parameter α based on residual distribution, enabling the estimator to handle diverse noise profiles. The IRLS procedure leverages the norm-aware weights, which are designed to account for the residual distribution's mode and tail behavior, effectively downweighting outliers. The MWCD estimator combines residual-based outlier rejection with continuous residual weighting, providing a robust covariance estimate that is less sensitive to outliers and clusters. The theoretical convergence is guaranteed under standard BCD assumptions, and the framework's flexibility allows extension to various sensor modalities and noise models.
Novelty
This research is the first to unify adaptive robust loss tuning, residual-based covariance estimation, and outlier rejection into a single, self-tuning framework for joint state and covariance estimation. While previous works addressed robustness or covariance estimation independently, none integrated these components with automatic parameter adaptation in a unified optimization scheme. The introduction of the norm-aware residual weighting, combined with the MWCD covariance estimator within a BCD approach, represents a significant leap forward in robustness and adaptability. This innovation enables the estimator to automatically respond to changing noise environments, outperforming fixed-parameter methods and static robust estimators in both simulated and real-world scenarios.
Limitations
- Despite its robustness, the computational complexity of the iterative framework may hinder real-time deployment in large-scale or high-frequency applications, necessitating further optimization.
- The current model assumes zero-mean noise; in scenarios with bias or systematic errors, the estimator's performance could degrade, requiring extensions to handle non-zero mean noise.
- The convergence guarantees depend on proper initialization and parameter tuning of thresholds; in highly non-stationary environments, adaptive thresholding strategies might be necessary.
Future Work
Future research will focus on reducing computational overhead through approximation techniques and parallelization, enabling real-time applications. Extending the framework to handle biased or non-zero mean noise will broaden its applicability. Additionally, integrating deep learning-based residual modeling could further enhance robustness against complex noise patterns. Exploring multi-sensor fusion scenarios and extending the approach to nonlinear dynamic systems will also be key directions, aiming to develop a comprehensive, scalable, and fully autonomous robust estimation system.
AI Executive Summary
In the rapidly evolving fields of robotics, autonomous vehicles, and indoor positioning, the ability to accurately estimate a system's state amidst noisy and contaminated measurements remains a fundamental challenge. Traditional estimation methods, such as least squares, rely heavily on assumptions of Gaussian noise with known covariance, which rarely hold true in real-world environments. When sensors encounter obstacles, multipath effects, or environmental interference, measurement errors deviate from Gaussianity, often exhibiting heavy tails and outliers that can severely bias estimates and undermine system reliability.
Addressing this critical issue, the paper introduces ARC (Adaptive Robust Joint Estimation), a novel framework that seamlessly integrates adaptive robust loss functions, residual-based covariance estimation, and iterative optimization strategies within a unified block coordinate descent (BCD) scheme. Unlike conventional methods that treat robustness and noise modeling separately or depend on manual parameter tuning, ARC dynamically adjusts its robustness level by tuning the loss shape parameter α based on residual distribution. Simultaneously, it estimates measurement noise covariance in a data-driven manner, leveraging the Minimum Weighted Covariance Determinant (MWCD) algorithm, which robustly identifies inlier measurements and accurately recovers the underlying noise statistics.
The core of ARC’s innovation lies in its iterative process: in each cycle, the framework refines the system state via iteratively reweighted least squares (IRLS) using a norm-aware weighting scheme that accounts for residual distribution characteristics. Concurrently, it updates the measurement covariance by applying MWCD on residuals after outlier removal, ensuring that the covariance estimate remains robust against heavy-tailed noise and outliers. The loss shape parameter α is also adaptively tuned by maximizing the residual likelihood, enabling the estimator to respond to changing noise environments without manual intervention.
Experimental validation spans both Monte Carlo simulations and real-world ultra-wideband (UWB) indoor localization experiments. In simulations, ARC maintains position RMSE below 0.1 meters even with 50% outliers, outperforming fixed-loss and traditional robust estimators. It also accurately recovers the true measurement covariance, with Wasserstein-2 distances close to zero, demonstrating its high fidelity in noise modeling. In real indoor experiments, ARC effectively identifies and suppresses non-line-of-sight (NLOS) heavy-tailed errors, achieving a standard deviation estimate nearly identical to the true inlier noise (~0.16m), significantly better than baseline methods. Ablation studies confirm that the continuous norm-aware weighting scheme in MWCD enhances covariance recovery and overall estimation robustness.
This work marks a significant step forward in robust sensor fusion and state estimation, providing a fully autonomous, self-tuning solution capable of adapting to complex, dynamic noise conditions. Its ability to accurately estimate both state and measurement noise covariance without manual parameter tuning makes it highly suitable for deployment in autonomous navigation, indoor localization, and multi-sensor fusion systems. Moving forward, efforts will focus on optimizing computational efficiency, extending the framework to handle biased noise, and integrating deep learning techniques to further enhance robustness and adaptability, paving the way for more reliable autonomous systems in challenging environments.
Deep Dive
Abstract
Sensor measurements are frequently corrupted by outliers and non-Gaussian noise. These imperfections in the sensor data can cause classical state estimators to generate biased and unreliable state and uncertainty estimates. Robust estimators reject or downweight outliers but do not perform measurement covariance estimation, whereas joint state and covariance estimators assume Gaussian residuals and fixed loss shape parameters. Integrating these two capabilities into a single framework is an opportunity to simultaneously estimate both state and covariance in the presence of outliers. This paper proposes a unified Block-Coordinate Descent framework that combines a norm-aware adaptive robust loss, an Iteratively Reweighted Least-Squares state update, and a Minimum Weighted Covariance Determinant covariance estimator, yielding a self-tuning joint state and covariance estimator. The framework is evaluated in a Monte-Carlo simulation and on real-world ultra-wideband localization experiments in cluttered non-line-of-sight environments. Results show that the proposed estimator consistently recovers the true inlier measurement covariance and matches or exceeds the state estimation accuracy of all baselines, without requiring any manual parameter tuning.