Acies-OS: a twin-assisted systems architecture for edge intelligence

Li, Jinyang

Permalink

https://hdl.handle.net/2142/132532 Copy

Description

Title

Acies-OS: a twin-assisted systems architecture for edge intelligence

Author(s)

Li, Jinyang

Issue Date

2025-11-26

Director of Research (if dissertation) or Advisor (if thesis)

Abdelzaher, Tarek

Doctoral Committee Chair(s)

Abdelzaher, Tarek

Committee Member(s)

• Nahrstedt, Klara
• Caesar, Matthew
• Shenoy, Prashant

Department of Study

Siebel School Comp & Data Sci

Discipline

Computer Science

Degree Granting Institution

University of Illinois Urbana-Champaign

Degree Name

Ph.D.

Degree Level

Dissertation

Keyword(s)

• Edge AI
• Digital Twin
• Cyber-Physical System
• Internet of Things

Abstract

The rapid proliferation of Artificial Intelligence (AI) within the Internet of Things (IoT) and Cyber-Physical Systems (CPS) has created new opportunities for intelligent sensing, perception, and control at the network edge. However, deploying deep learning-based intelligence on embedded platforms introduces fundamental challenges, including limited computational resources, thermal constraints, and unreliable network connectivity. Addressing these challenges requires system-level innovation that spans from performance modeling and resource optimization to workflow orchestration and adaptive learning. This dissertation introduces Acies-OS, a twin-assisted, content-centric middleware framework designed to enhance the efficiency and robustness of distributed edge intelligence systems. Acies-OS unifies data, computation, and control under a structured namespace abstraction, allowing distributed nodes to coordinate seamlessly through declarative data and control interfaces. Its design combines efficient latency and thermal modeling for performance optimization, a digital twin mechanism for cross-layer monitoring and failover recovery, and an extensible control plane for implementing runtime optimization services such as model selection and workflow reconfiguration. Together, these mechanisms enable dynamic, data-driven management of complex sensing-to-decision workflows in resource-constrained environments. The dissertation further introduces an unsupervised collaborative adaptation framework that enables in-situ model refinement at runtime without labeled data. By leveraging spatial and temporal correlations across distributed sensors, this approach improves model robustness against domain shifts encountered in real deployments. The system is implemented and evaluated using a multi-modal, multi-node vehicle classification testbed that has supported numerous research efforts and produced the largest publicly available dataset of its kind. Finally, the dissertation outlines a design for extending Acies-OS toward agentic edge intelligence, integrating large language model (LLM)-based agents with cyber-physical systems through standardized control interfaces. Overall, this work presents an integrated architecture that advances the efficiency, adaptability, and resilience of edge AI systems. By connecting AI optimization, system middleware, and digital twinning within a unified framework, Acies-OS offers a practical step toward enabling intelligent and self-managing IoT and CPS deployments.

Graduation Semester

2025-12

Type of Resource

Thesis

Handle URL

https://hdl.handle.net/2142/132532

Copyright and License Information

Copyright 2025 Jinyang Li

Owning Collections