Model Predictive Supervisory Control for Hierarchical and Distributed UAS Traffic Management

The sky above our cities is becoming increasingly crowded, not with birds, but with small, autonomous aircraft designed to deliver packages, inspect infrastructure, and provide emergency services. As these unmanned vehicles move from military and academic experiments into everyday logistics, they face a fundamental problem: the low-altitude airspace is a shared, physical bottleneck. Unlike a highway where cars can simply change lanes, the air requires precise coordination to prevent collisions, ensure that no two drones occupy the same space at the same time, and manage limited resources like charging stations. This challenge belongs to the field of traffic management for unmanned systems, a discipline that blends the logic of computer science with the physical constraints of flight. At its core, this work relies on two distinct but complementary ideas. The first is a method of "supervisory control," which acts like a strict traffic warden, defining the absolute rules of the road to guarantee safety and prevent deadlocks where vehicles get stuck waiting for each other forever. The second is "predictive control," a strategy where each vehicle looks ahead a few steps to choose the most efficient path, weighing factors like battery life and delivery time. The question researchers have long grappled with is how to combine these two approaches for a large fleet of drones without creating a computer system so complex it cannot run in real time.

In a recent study, researchers from the Federal University of Minas Gerais in Brazil proposed a solution that bridges this gap, creating a framework they call Model Predictive Supervisory Control. They designed a system that allows a large group of drones to operate safely in a busy urban environment by separating the job of "keeping everyone safe" from the job of "finding the best route." Imagine a city where a central authority draws the boundaries of every street and intersection, declaring which paths are open and which are closed to prevent accidents. This authority does not tell every driver exactly which turn to take; instead, it simply forbids dangerous moves. Each driver, knowing the rules, then looks at the map and chooses the fastest, most fuel-efficient route to their destination that stays within those safe boundaries. This is exactly how the new system works. A central controller, acting as the supervisor, monitors the entire fleet and the shared airspace. It uses a mathematical model to identify which movements would cause a collision or violate a rule, such as entering a restricted zone or trying to land at a charging station that is already full. It then broadcasts a list of "forbidden moves" to every drone in the fleet.

Each drone receives this list and uses its own onboard computer to solve a short-term optimization problem. It looks at its current location, its battery level, and its assigned mission, then calculates the sequence of moves that will get it to its goal with the least amount of time and energy, strictly avoiding the moves on the forbidden list. Once the drone executes the very first move in that sequence, the process repeats. The drone sends its new position back to the central supervisor, the supervisor updates the list of forbidden moves based on the new reality, and the drone plans its next step. This cycle happens continuously, allowing the fleet to adapt instantly to changes, such as a new delivery request or a sudden battery warning. The researchers tested this system in a simulated urban environment using a standard robotics software platform. They created two scenarios: a smaller test with two drones navigating a simple map with nine nodes, and a larger, more complex test with four drones navigating a map with nineteen nodes across two different altitude layers.

The results of these simulations demonstrated that this hierarchical approach solves the problem of complexity that has plagued previous attempts. When the researchers tried to build a single, all-knowing computer model to manage the entire fleet at once—a method known as monolithic synthesis—the system became overwhelmed. For the smaller two-drone scenario, the single model required nearly 445,000 internal states to describe every possible combination of positions and actions, a number that would grow explosively if more drones were added. In contrast, the new hierarchical system reduced the complexity of the supervisor to just 570 states, a reduction of roughly 780 times. This massive simplification meant the system could run efficiently without needing supercomputers. In the larger four-drone scenario, the system successfully completed five out of six assigned missions, with an average mission time of 44.42 seconds and a throughput of 2.46 missions per minute. The central supervisor had to forbid an average of 19.28 moves per time step to maintain safety, proving that the system was actively managing conflicts rather than just letting the drones fly freely.

The study confirms that separating the global safety rules from local decision-making allows for a scalable solution. The central supervisor does not need to know the specific details of every drone's battery or cargo; it only needs to track the occupancy of the airspace and the availability of shared resources. The drones, in turn, are free to optimize their own paths as long as they respect the boundaries set by the supervisor. This architecture ensures that the system remains safe and non-blocking, meaning no drone gets stuck in a situation where it cannot move forward or backward. The researchers found that even as the number of drones and the complexity of the airspace increased, the system maintained its ability to coordinate the fleet without collisions or delays. The work suggests that this method is a viable path forward for the future of urban air mobility, where thousands of vehicles might one day share the same low-altitude corridors. By using a framework that combines strict safety enforcement with flexible, local optimization, the researchers have shown that it is possible to manage a complex, multi-agent system in a way that is both computationally efficient and operationally robust. The study concludes that while the system is currently validated through simulation, the principles established offer a concrete foundation for developing real-world traffic management systems that can handle the dynamic demands of a busy city sky.