IUCA: A Model-Agnostic In Situ Uncertainty Calibration Framework for High-Dimensional State Estimation in AIoT Robotic Systems

Abstract

Autonomous robots increasingly serve as mobile Artificial Intelligence of Things (AIoT) sensing nodes, where reliable edge perception and state estimation are critical to safe and efficient system operation. In practical deployment, uncertainties arising from perception and decision-making processes (e.g., localization algorithms) often distort the system state and misalign with real-world conditions, leading to overconfident decisions and potential system failures. This article presents the In-situ uncertainty calibration approach (IUCA), a model-agnostic uncertainty calibration framework for AIoT-enabled robotic systems based on high-dimensional conformal prediction (CP). Unlike conventional uncertainty metrics that rely on model assumptions or heuristic confidence measures, the proposed framework provides distribution-free, finite-sample guarantees, ensuring that reported confidence regions reflect actual deployment-time reliability. This framework calibrates joint and correlated state uncertainties without modifying the underlying estimator or requiring retraining, making it suitable for embedded and edge-deployed AIoT systems. As a representative case study, we integrate the framework with adaptive Monte Carlo localization, a widely deployed localization module. The calibrated system constructs conformal confidence ellipses with provable coverage guarantees in the robot pose space, statistically bounding estimation error and mitigating catastrophic failures, such as the Robot Kidnapping Problem. Extensive validation in simulated environments, open-source datasets, and real-world warehouse deployments demonstrates that the proposed framework consistently achieves target coverage while preserving or improving localization accuracy. These results establish IUCA as a foundational reliability layer for trustworthy, robust, and scalable AIoT-enabled robotic sensing systems.