Connectivity-Aware Model Predictive Control for Swarms of Dynamically Constrained Vehicles Engaged in Multi-Target Persistent Observation
This paper introduces and experimentally evaluates five control strategies for swarms of dynamically constrained, forward-only vehicles, demonstrating that centralized Model Predictive Control achieves the lowest mission cost by optimizing curvature and acceleration while maintaining connectivity, whereas distributed MPC maximizes network connectivity at a higher cost and heuristics prioritize specific targets at the expense of overall coverage.
Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Imagine a team of robots sent into a city to watch over several important locations at once. They must stay close enough to talk to each other, forming a chain of communication back to a central base, but they also need to spread out to keep an eye on different targets. This is a difficult balancing act. If they cluster too tightly to keep the conversation going, they miss the targets. If they spread out too far to see everything, they lose contact with each other. The problem becomes even harder when the robots cannot stop, turn around, or spin in place. They are like cars that can only move forward and turn in wide arcs, or like airplanes that cannot hover. They must keep moving, and their paths are curved, not straight lines. This makes planning their movements a complex puzzle: how do you guide a group of forward-only machines to watch many things at once without breaking their connection or crashing into one another?
Researchers at the École Polytechnique Fédérale de Lausanne tackled this exact challenge. They wanted to see if they could teach a swarm of such robots to coordinate themselves effectively. To do this, they built a system that could predict the future movements of the robots. Instead of just reacting to where the robots are right now, the system looks ahead, imagining where the robots will be in the next few seconds. It then calculates the best path for the whole group, ensuring they stay connected while moving toward their goals. The team tested five different ways of making these decisions. Some methods relied on a central computer to tell every robot what to do, while others let each robot make its own local decisions based on what its neighbors were planning. They also tested methods that were purely reactive, which simply reacted to the current situation without looking ahead, and methods that used simple math to find a good, but not perfect, solution.
The researchers put these five strategies to the test with six small, wheeled robots in a controlled indoor arena. The robots were programmed to move only forward and turn in circles with a fixed minimum radius, mimicking the limitations of real-world vehicles like fixed-wing drones or marine craft. The mission involved watching six stationary points of interest. Every twenty seconds, the importance of one of these points would change, forcing the robots to quickly rearrange themselves. One target would become the most important, requiring the robots to focus their attention there, while the others remained secondary. The robots had to constantly shift their formation, sometimes acting as observers and other times as relays to keep the communication chain intact. The team ran fifty separate trials, ten for each strategy, to see which approach worked best in the real world.
The results revealed a clear trade-off between efficiency and safety. The strategy that performed best at minimizing the overall cost of the mission—meaning the robots stayed close to their targets and used their energy efficiently—was the centralized model predictive control. In this approach, a single powerful computer calculated the best path for the entire group at once, taking into account the curved paths and speed limits of every robot. This method kept the mission cost about six percent lower than a simpler, static mathematical approach and more than fifty percent lower than a hybrid method that mixed central planning with local rules. However, this efficiency came at a price. The robots using this method did not maintain the strongest possible connection to each other.
In contrast, the distributed approach, where each robot solved its own small piece of the puzzle and shared its plans with neighbors, produced the strongest communication links. The robots using this method stayed much more tightly connected to one another, maintaining a robust network even as they moved. But this extra safety and connection came with a significant cost: the mission was much less efficient, with the robots spending more time further away from their targets. The researchers found that the robots using the distributed method were willing to sacrifice some observation quality to ensure the group stayed together. This confirmed that there is no single perfect solution; the best choice depends on whether the priority is watching the targets closely or keeping the team connected at all costs.
Another critical finding concerned how the robots avoided crashing. The researchers implemented a safety check that looked at the final destination points the robots were aiming for. If two robots were heading toward points that were too close, the system would nudge them apart before sending the command. This worked well enough to prevent the robots from being told to go to the same spot, but the researchers discovered it was not a perfect guarantee. Because the robots move in curves and cannot stop instantly, they could still cross paths dangerously even if their final destinations were safe. The study showed that checking the destination points is not the same as checking the entire path the robot will travel. This distinction is vital for real-world safety, as a command that looks safe on paper can still lead to a collision if the robot's curved trajectory brings it too close to another machine during the turn.
The experiments also highlighted the gap between computer simulations and real-world performance. In previous simulations, the distributed method had shown promise, balancing cost and connection well. But on the physical robots, the distributed method became much more conservative. The robots stayed closer together than necessary, which increased their connection strength but made the mission significantly more expensive in terms of distance traveled and time spent. This happened because the real robots had to deal with the physical reality of turning in arcs and the slight delays in communication, which made the local decision-making process more cautious. The researchers noted that while the distributed system worked, it was not as efficient as the central computer version when faced with the messy reality of physical movement.
Ultimately, the study demonstrated that it is possible to coordinate a swarm of forward-only, curved-path robots to watch multiple targets while staying connected. The researchers proved that a central computer can calculate the most efficient paths, but a distributed system where robots talk to each other can provide a stronger, more robust network. They also showed that while simple safety checks can prevent obvious conflicts, they do not replace the need for more complex path planning to ensure safety during the turn. The work provides a practical blueprint for how to manage teams of robots that cannot stop or turn on a dime, offering clear evidence of the compromises required to keep them both connected and effective in the real world.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.