Date of Award

Fall 11-15-2025

Document Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

Department

Computational Analysis and Modeling

First Advisor

Pradeep Chowriappa

Abstract

This dissertation focuses on designing a robust and uncertainty-aware framework for autonomous systems operating in GPS-denied environments, such as indoor infrastructures, underground tunnels, and lunar surfaces. The proposed framework addresses the challenges posed by multi-modal uncertainties, including sensor noise, distributional shifts under adverse conditions, and conflicting decision-making preferences. These challenges compromise the reliability and adaptability of autonomous platforms. To overcome these challenges, the proposed framework adopts a layered architecture that integrates advanced methodologies across the sensing, perception, and decision-making layers. At the sensing layer, an Edge-Kalman Filter combined with a density ratio-based update mechanism is employed to reduce aleatoric uncertainty caused by noisy and inconsistent measurements from BLE, Wi-Fi, ZigBee, and IMU sensors. In the perception layer, a multiresolution analysis using a filter bank—comprising low, medium, and high-pass filters—captures the spatial-frequency characteristics of LiDAR point clouds to detect out-of-distribution inputs, thus reducing epistemic uncertainty. For the decision-making layer, a Prioritized User Preference-based Multi-Objective Reinforcement Learning (PUPMORL) approach is proposed, which selects policies based on KL-divergence to align system behavior with predefined user preferences, thereby addressing preference uncertainty. Together, these methodologies create a resilient system capable of operating reliably in complex and uncertain environments without GPS. The effectiveness of the proposed framework is evaluated through a series of experiments conducted in both simulated and real-world GPS-denied environments. To assess aleatoric uncertainty, indoor localization experiments are performed using BLE, Wi-Fi, ZigBee, and IMU data collected in controlled testbeds, demonstrating significant improvements in distance estimation accuracy when using the Edge-Kalman Filter with density ratio adaptation. Epistemic uncertainty is examined by applying synthetic fog conditions to LiDAR point clouds using the LISA simulator, where the multi-resolution filter bank combined with KL divergence-based confidence estimation successfully detects out-of-distribution patterns and enhances robustness in adverse conditions. For preference uncertainty, the PUPMORL approach is validated using benchmark multi-objective decision-making environments, showing improved policy alignment with user-defined preferences and lower variance in policy selection stability. These results confirm that the proposed methods work effectively across sensing, perception, and decision layers to ensure reliable autonomous performance under uncertainty. This framework is designed to support a wide range of autonomous systems operating in environments where GPS signals are unavailable or unreliable, such as indoor navigation, planetary exploration, underground inspection, and disaster response. Its modular architecture and sensor-agnostic design make it adaptable to diverse platforms and operating conditions. Future work will focus on extending the proposed framework with real-time adaptive preference learning, integrating vision-based sensing to complement LiDAR under degraded conditions, and deploying the system on edge devices to enable low-latency, uncertainty-aware decision-making in mission-critical scenarios.

Download

Share

COinS