Development and implementation of a framework for Model Predictive Control of discrete-event manufacturing systems

In the vast, humming world of modern manufacturing, factories are no longer just lines of machines moving parts from one station to another. They are complex, living networks where decisions must be made in split seconds to keep production flowing. Imagine a factory floor where a robot arm must decide whether to pick up a part now or wait, or where a conveyor belt needs to reroute a package because a machine ahead has stalled. For decades, engineers have relied on a powerful method called Model Predictive Control to manage these decisions. This approach works like a chess player looking several moves ahead: it uses a mathematical model of the factory to predict what will happen in the future, calculates the best sequence of actions to reach a goal, and then executes the very first step before starting the process all over again. While this method has revolutionized industries like chemical processing, where reactions follow smooth, continuous laws, it has struggled to take hold in discrete manufacturing. Here, the "moves" are distinct events—a part arriving, a machine finishing a cycle, a robot grabbing an object—rather than smooth flows. The sheer number of possible combinations and logical rules in these systems makes building the necessary mathematical models incredibly difficult, often requiring teams of experts to write custom equations for every single machine and robot.

A team of researchers from Italy has developed a new way to solve this problem, creating a flexible, open-source framework that allows factories to use this advanced control method without the heavy burden of manual modeling. Instead of asking engineers to write complex equations for every piece of equipment, their system looks at the physical layout of the factory—the "topology"—and automatically builds the control logic based on how the machines and robots are connected. The researchers, working with the National Research Council and the Polytechnic University of Milan, designed a software tool that treats the factory like a map of interconnected blocks. These blocks represent everything from the machines themselves to the buffers where parts wait, and even the robots that move them. By simply defining how these blocks are linked together, the software automatically generates the rules needed to control the flow of production. It then uses a powerful solver to calculate the best possible sequence of moves to keep the factory running efficiently, minimizing delays and energy use.

The core of this innovation is a language the team created to describe the factory's structure, which they call the Plant Topology Modeling Language. In this system, a factory is not a list of numbers but a graph of relationships. There are items, which are the physical objects being made, and containers, which are the places where those items sit, such as a buffer zone or a machine's loading port. There are also resources, like robotic arms, that move items from one container to another. The software understands that a robot cannot pick up a part from an empty container, or that a machine cannot start a new cycle until it has finished the previous one. These are not hard-coded rules for a specific factory; they are general principles that the software applies automatically once it sees the connections between the blocks. When the system runs, it constantly checks the current state of the factory—where every part is, which machines are busy, and which are free. It then looks ahead over a specific period, known as a prediction horizon, to see what will happen if it takes different actions. It calculates the best path forward, sends the first command to the machines, and then immediately starts the calculation again with the new information, adapting instantly to changes like a machine breaking down or a new order arriving.

To make this work in a real-world setting, the researchers integrated their software into a layered automation system that mirrors how modern factories are actually organized. At the top is a planning layer that decides what needs to be made and in what order. In the middle, a supervisor layer acts as the bridge, translating those high-level plans into specific instructions for the machines. At the bottom, the machines and robots execute the commands. The beauty of this framework is that the middle layer can talk to the machines whether they are real hardware or simulated versions on a computer. This allows engineers to test the control system thoroughly in a virtual environment before ever touching a physical factory floor. If a machine goes offline or a robot fails, the system detects the change, updates its internal map of the factory, and recalculates the best course of action without needing a human to reprogram anything. This adaptability is crucial because manufacturing environments are rarely static; they change constantly, and a control system that cannot adjust to those changes is of little use.

The researchers tested their framework on a simulated factory containing twenty-seven different control units, forty-four containers for parts, and ninety-six possible actions the system could take. The results showed that the system could successfully navigate the complex web of decisions required to keep the factory running. However, the tests also revealed a fundamental limit: as the number of parts in the factory increases, or as the time the system looks into the future gets longer, the time required to calculate the best solution grows very quickly. This is a known challenge in complex optimization problems, where the number of possible scenarios explodes exponentially. The researchers found that to keep the system fast enough for real-time use, it is necessary to limit the number of parts the system tries to optimize at any one moment. They demonstrated that by hiding parts that are not immediately needed from the main control loop, the system could remain efficient without losing sight of the overall production goals. This approach allows the factory to handle a large number of items by only focusing on the ones that need immediate attention, effectively managing the complexity.

One of the most significant findings was how the system handles potential gridlocks. In a factory, a deadlock occurs when two or more machines are waiting for each other to release a part, creating a traffic jam that stops production entirely. The researchers noted that while their system cannot guarantee that a deadlock will never happen in every possible scenario, it can be tuned to avoid them. By adjusting the priorities in the cost function—the mathematical measure of what the system is trying to achieve—the engineers can penalize paths that are likely to lead to a jam. For example, the system can be instructed to avoid moving a part into a buffer that is already full, or to prioritize machines that are less likely to get stuck. The simulations showed that with careful tuning, the system could navigate around these potential bottlenecks, keeping the flow of production smooth. The framework also includes safeguards at the machine level; if a command is sent that a machine knows it cannot execute, perhaps because a sensor has failed or a part is missing, the machine can reject the command and ask for a new plan, preventing the system from crashing.

The work represents a significant step toward making advanced control systems accessible to a wider range of manufacturers. By removing the need for custom mathematical modeling, the framework lowers the barrier to entry for factories that want to optimize their operations but lack the resources to hire specialized teams of mathematicians. The software is open-source, meaning it is freely available for anyone to use, study, or improve. The researchers have already begun testing the framework in a real-world setting at a contact lens manufacturing plant, moving beyond the safety of computer simulations to see how it performs with actual hardware. While the system is not a magic solution that solves every problem instantly, it offers a scalable and cost-effective way to bring intelligent, adaptive control to the chaotic and dynamic world of discrete manufacturing. The future of this technology lies in refining the way the optimization problems are formulated to make them even faster and more capable, ensuring that as factories become more complex, their control systems can evolve right alongside them.