COMMITTEE CHAIR: Dr. Lijun Qian

TITLE: ENERGY-EFFICIENT TASK OFFLOADING FRAMEWORKS FOR MIXED REALITY & EXTENDED REALITY IN EDGE AI ENVIRONMENTS

ABSTRACT: Mixed Reality (MR) and Extended Reality (XR) technologies are increasingly integrated into real-time applications such as remote maintenance, spatial navigation, and immersive learning. However, the computational intensity of deep learning workloads, especially object detection models like YOLO, poses significant challenges for wearable XR devices with limited processing power and battery life. To address these challenges, this thesis presents a comprehensive study of adaptive offloading strategies that optimize task execution between XR devices and edge servers. We first propose the Binary Spatial Allocation Framework (BSAF), which introduces a real-time binary offloading decision mechanism for XR systems. The BSAF framework evaluates scene complexity, network conditions, and system constraints to dynamically select either local execution on HoloLens 2 or edge execution using high-precision YOLOv11 models. A closed-form utility model is developed to balance accuracy, latency, and energy consumption, and Lagrangian relaxation is employed to maintain constraint feasibility. Experimental results across 100 XR scenes show that BSAF reduces latency by up to 37% and energy usage by 44% compared to full offloading, while preserving over 90% detection accuracy. To further enhance adaptability, we propose SPARL-CBF, a Safe Partial Offloading framework that integrates Model-Based Reinforcement Learning (MBRL) with Control Barrier Functions (CBFs). Unlike binary strategies, SPARL-CBF learns a continuous offloading policy that dynamically adjusts task partitioning based on runtime system states such as battery level, bandwidth, and CPU load. CBFs ensure real-time constraint satisfaction by projecting unsafe actions back into feasible regions. Implemented using Proximal Policy Optimization (PPO) and deployed on a live edge-XR testbed, SPARL-CBF achieves over 94% accuracy, reduces latency by 22%, and prolongs device battery life by intelligently modulating the offloading ratio. Together, BSAF and SPARL-CBF offer a unified, constraint-aware framework for intelligent task offloading in XR systems. This thesis provides mathematical formulations, system implementations, and empirical evaluations that demonstrate how adaptive offloading can significantly improve energy efficiency and responsiveness in edge-assisted XR environments. The proposed approaches lay the groundwork for scalable, real-time XR applications in emerging domains such as healthcare, smart cities, and the metaverse.

Room Location: Electrical Engineering Conference Room 315D