Modeling and Control of Robot-Enabled Flexible Manufacturing Systems: System Properties and Knowledge-Guided Machine Learning Approaches

Fueled by rapid advances in machine learning and robotics, manufacturing automation is entering a new era of intelligent, flexible production. In this landscape, mobile collaborative robots offer unprecedented adaptability, underscoring the need for structured frameworks to guide their integration. This dissertation meets that need by proposing a mathematical and control framework for mobile robot-operated Flexible Manufacturing Systems (FMSs).

The dissertation begins by formulating a mathematical model that captures the core dynamics of an FMS operated by mobile, multi-skilled robots. From this foundation, performance metrics are derived to solve a real-time robot assignment problem using Reinforcement Learning (RL). The model is extended to incorporate maintenance scheduling and tool change dynamics, creating a holistic representation of the manufacturing process. To address partial observability in real-world systems, the dissertation proposes a heterogeneous graph-based control framework. Using heterogeneous graph neural networks (HGNNs) paired with multi-agent RL, the framework fuses diverse information sources, enabling agents to infer hidden states and generate coordinated, data-driven decisions in real time. To support arbitrary robot skillsets and scalable configurations, the framework introduces indicator matrices and monotonically decreasing functions that describe how workstation cycle time varies with robot deployment. The resulting control problem—defined as ideal clean configuration mimicking—is solved using a predictive, moving window approach paired with graph search algorithms. Finally, to reduce retraining burdens of RL in dynamic environments, a transferable learning strategy is developed. This approach allows models trained on a characteristic three-workstation system to generalize across systems of varying sizes, cycle times, buffer capacities, and reliability parameters.

In summary, this dissertation lays the groundwork for intelligent, scalable, and adaptive robot-operated manufacturing systems. It contributes robust modeling tools, performance-driven control strategies, and practical solutions for real-world deployment—bringing the vision of autonomous, agile production closer to reality.