Robot swarms that move too purposefully jam up. A new Harvard study finds the fix is not more AI: it is the right amount of randomness. The math also predicts that adding capacity sometimes makes things worse.
image from Gemini Imagen 4
Researchers at Harvard and Eindhoven University found that optimal robot swarm coordination emerges from a 'Goldilocks zone' of environmental noise rather than individual robot intelligence or centralized control. The counterintuitive finding shows that sufficient randomness enables statistical averaging across the swarm, making collective dynamics mathematically tractable in ways impossible for deterministic systems. Physical experiments with wheeled robots confirmed simulations, demonstrating that simple local navigation rules can achieve coordinated task completion at scale without complex path-planning.
When Lucy Liu runs her robot swarms in the lab at Eindhoven University of Technology, she watches for the moment of gridlock. Set the noise too low and the robots march in straight lines toward their goals, slamming into each other in dense traffic jams. Set the noise too high and they zigzag uselessly forever. Somewhere between those two failures there is a Goldilocks zone: just enough randomness so the robots bump into each other, form temporary jams, and slip past. Flow emerges. Task completion climbs.
It is not a result about robots. It is a result about coordination.
Liu, an applied mathematics PhD student at Harvard's School of Engineering and Applied Sciences, led a study published in the Proceedings of the National Academy of Sciences that asks a deceptively simple question: given a fixed area, how many robots should you deploy on a time-sensitive task, and how should they move? The answer turns out to be less about the intelligence of the individual robot and more about the collective dynamics of the group. Specifically, it is about noise.
The counterintuitive part: randomness makes the math tractable. When Liu and her collaborators modeled thousands of agents moving toward random goal locations, they found that a high enough noise level lets you take averages across distances, times, and behaviors. Those averages unlock predictions that would be impossible in a fully deterministic system, where every interaction between every robot has to be tracked separately. "How could randomness make things easier to work with?" Liu said in a Harvard press release. "But in this case, when you have a lot of randomness, it becomes possible to take averages."
The researchers tested the simulations against real hardware in the lab of physicist Federico Toschi at Eindhoven University of Technology: small wheeled robots carrying QR codes, tracked by an overhead camera and reassigned to new goal positions the moment they arrived. The physical robots turned and moved more slowly and imperfectly than the simulations predicted. The core finding held anyway. Short-lived jams appeared at the optimal noise level, and flow kept moving.
The result that will matter most to robotics engineers: you do not need a central computer orchestrating the fleet, and you do not need ultra-intelligent robots making complex path-planning decisions. A simple local set of navigational rules, at least up to certain densities, may be all you need. As the researchers put it: a powerful central computer or ultra-intelligent robots are not necessary to achieve coordinated tasks.
This maps onto a known puzzle from infrastructure engineering called the Braess paradox: adding a new road to a traffic network can make congestion worse, not better. Adding optimization capacity to a robot swarm can produce the same counterintuitive result. More robots, or smarter pathfinding, does not automatically mean better throughput. Liu's study offers a mathematical framework for finding the actual optimal configuration, and it is not the one that intuitive engineering would produce.
L. Mahadevan, Liu's advisor and the Lola England de Valpine Professor of Applied Mathematics, Organismic and Evolutionary Biology, and Physics at Harvard, framed the broader relevance. "Understanding how active matter, whether it is a swarm of ants, a herd of animals, or a group of robots, become functional and execute tasks in crowded environments using the principles of self-organization, is relevant to many questions in behavioral ecology," he said.
A separate line of research at Georgia Tech takes the same problem in a different direction. Bolei Deng and Xinyi Yang built robotic particle swarms that coordinate without any electronics, sensors, or processors at all: the robots are geometric structures that bend, latch onto each other, and release using only mechanical interaction. No computation, no communication, no central controller. Coordination emerges from shape and contact. Deng described the approach as a form of mechanical intelligence, where dumb units produce emergent collective behavior purely through physical interaction. That work, published in Advanced Intelligent Systems, represents a different answer to the same underlying question: how much intelligence does coordination actually require?
The Harvard and Georgia Tech results are not the same mechanism, but they point in the same direction. Liu's study shows that simple local rules can replace central planning at the right noise level. The Georgia Tech group shows that geometry alone can replace computation entirely. Both are arguments against the assumption that smarter individual robots are the only path to better collective outcomes.
The Harvard study has limits. The researchers explicitly note that simple local rules work up to certain densities. Beyond that threshold, the math changes. And the framework was validated in a controlled lab with wheeled robots on a flat surface: warehouse floors, yes, but not the unstructured environments where most deployment difficulty lives. Real deployment means dealing with sensor noise, mechanical failure, and surfaces the simulation never accounted for.
What the study does provide is a design principle. For engineers building warehouse fulfillment fleets, drone swarms for search and rescue, or coordination systems for autonomous vehicles in urban environments, the lesson is structural: before investing in smarter individual agents or more sophisticated central planning, check whether the noise level is right. The bottleneck may not be intelligence. It may be that the robots are moving too purposefully.
The research was funded by the National Science Foundation Graduate Research Fellowship Program under Grant No. DGE 2140743, along with grants from the Simons Foundation and the Henri Seydoux Fund.
Story entered the newsroom
Assigned to reporter
Research completed — 4 sources registered. No-electronics robotic swarm: particles coordinate via mechanical design (bend/latch/release), not code or sensors. Intelligence is built into geometr
Draft (723 words)
Reporter revised draft based on fact-check feedback
Reporter revised draft based on editorial feedback (723 words)
Approved for publication
Published (897 words)
Get the best frontier systems analysis delivered weekly. No spam, no fluff.
Robotics · 13h 20m ago · 4 min read
Robotics · 14h 26m ago · 3 min read
