Co-design of communication, computing and control for real-time interactions in industrial cyber-physical systems

Chen, Kan (2026) Co-design of communication, computing and control for real-time interactions in industrial cyber-physical systems. PhD thesis, University of Glasgow.

Abstract

As intelligent systems evolve toward tighter integration of communication, computation, and control, Cyber-Physical Systems (CPS) and the Industrial Metaverse have emerged as the next frontier for real-time human–machine collaboration. These systems promise immersive, taskaware, and adaptive interaction between humans, robots, and digital twins. Yet, existing frameworks remain fragmented: communication networks, control algorithms, and human feedback mechanisms are often optimized in isolation, making it difficult to guarantee task-level performance and human trust under dynamic conditions. This thesis develops a unified, task-oriented, and human-centered co-design framework that bridges these domains to enable real-time, adaptive, and trustworthy CPS operation.

The research begins by establishing a task-oriented cross-system design framework that integrates communication scheduling, computation offloading, and control policy optimization into a single learning-based architecture. A Human-in-the-Loop Reinforcement Learning (HITLRL) mechanism is developed to enable adaptive policy refinement through interactive human feedback. Experimental validation on a teleoperation platform demonstrates that the proposed HITL-RL achieves an average position RMSE of 0.023 m, outperforming the data-based RL baseline by 47.3%, and significantly enhancing control smoothness and trajectory stability under stochastic delay conditions. These results confirm that incorporating human corrective inputs effectively improves policy robustness and task execution accuracy in dynamic environments.

Building on this foundation, a Human-in-the-Loop Meta-Learning (HITL-MAML) framework is introduced to enhance adaptability across operators and task contexts. By leveraging Model-Agnostic Meta-Learning, the system learns a transferable initialization that can be efficiently adapted with limited human feedback. To ensure that the framework is generalizable and robust, we propose the HITL-MAML algorithm, dynamically adjusting prediction horizons. To verify the proposed framework and algorithm, we build a prototype including a real-world robotic arm and its digital model in the CPS. The results demonstrate that our approach reduces the weighted sum of the Root Mean Squared Error (RMSE) from 0.0712 m to 0.0101 m, significantly outperforming various baseline methods. This substantial improvement enhances both the responsiveness and reliability of real-time CPS interactions.

To further align system behavior with human experience, a preference-driven reinforcement learning approach is developed, incorporating Reinforcement Learning from Human Feedback (RLHF) into the co-design loop. By modeling implicit human responses as latent reward signals, the framework learns control strategies that optimize both task efficiency and perceived comfort. We validate our framework using a UR3e robotic arm for reactor tile inspection in a nuclear decommissioning scenario. Compared to baseline methods, our approach enhances scene representation while optimizing trajectory efficiency. The RLHF-based policy consistently outperforms baseline selection, prioritizing task-critical details. By unifying explicit 3D scene representations with implicit human-in-the-loop optimization, this work establishes a foundation for adaptive, safety-critical robotic perception systems, paving the way for enhanced automation for remote maintenance and other high-risk environments.

Comprehensive simulations and hardware-in-the-loop experiments confirm that the proposed frameworks achieve consistent improvements in latency, reliability, adaptability, and human satisfaction. Collectively, these contributions establish a theoretical and experimental foundation for scalable, task-oriented, and human-adaptive CPS. Looking ahead, this work envisions the evolution of co-design methodologies toward multi-human collaboration—where multiple operators interact with distributed agents through shared intent inference—and the incorporation of foundation model-driven optimization, leveraging large pre-trained models for semantic reasoning and zero-shot adaptation. Together, these directions point toward the realization of humancentric CPS and Industrial Metaverse ecosystems.

• Item Type: Thesis (PhD)
• Qualification Level: Doctoral
• Subjects: Q Science > QA Mathematics > QA75 Electronic computers. Computer science
• Colleges/Schools: College of Science and Engineering > School of Computing Science
• Supervisor's Name: Li, Dr. Emma Liying, Zhao, Professor Dezong and Lan, Dr. Jianglin
• Date of Award: 2026
• Depositing User: Theses Team
• Unique ID: glathesis:2026-85847
• Copyright: Copyright of this thesis is held by the author.
• Date Deposited: 31 Mar 2026 09:59
• Last Modified: 31 Mar 2026 10:00
• Thesis DOI: 10.5525/gla.thesis.85847
• URI: https://theses.gla.ac.uk/id/eprint/85847

Related URLs

• Conference proceeding
• Article DOI