Pose Estimation of a Meso-Scale Robot using Magnetic Fields

Abstract

Accurate pose estimation is a fundamental need for autonomous robotic systems, particularly for small scale robotic platforms operating in confined, cluttered, or sensing deprived environments. Conventional localization approaches based on Global Navigation Satellite Systems, vision, or LiDAR often become unreliable or impractical in such small scale settings due to line-of-sight constraints, environmental conditions, or computational and payload limitations. This dissertation addresses a series of pose estimation problems motivated by these challenges, with a particular focus on actively controlled sensing strategies that improve observability and estimation robustness under severe system constraints. The first part of the dissertation investigates two-dimensional position and orientation estimation of a robotic platform using an actively controlled external electromagnetic field. By dynamically controlling the excitation source, the proposed approach enhances the information content of magnetic field measurements and enables reliable pose estimation without requiring a global magnetic field map. This framework is subsequently extended through the integration of vision sensing, resulting in a hybrid system that combines an actively controlled magnetic field with a camera to achieve robust two-dimensional position estimation. The dissertation then explores learning based magnetic field modeling for pose estimation. A deep learning based magnetic field representation is developed and embedded as the measurement function within a recursive state estimation framework, enabling accurate two-dimensional position and orientation estimation under active magnetic field excitation. Building on this foundation, the methodology is further extended to three-dimensional localization. A three-dimensional deep learning magnetic field model, combined with active magnetic field control and Unscented Kalman Filter based state estimation, enables robust estimation of robot three-dimensional position in environments where conventional localization methods are ineffective. Finally, the dissertation investigates two-dimensional position and orientation estimation using an actively controlled camera and a laser sensor for a larger ground robot, demonstrating the broader applicability of active sensing and control principles beyond magnetic localization. Across all scenarios, active control of the sensing modality plays a central role in improving observability, reducing estimation ambiguity, and enhancing localization accuracy. The proposed methods are validated through rigorous experimental studies using controlled testbeds. The results demonstrate that actively controlled sensing, when combined with model based and learning based estimation techniques, provides a powerful and general framework for reliable robot pose estimation in challenging environments.