Researchers led by Andreagiovanni Reina from the University of Konstanz have shown that a simple interaction rule enables robot swarms to reach clear collective decisions despite faulty or manipulated information. Remarkably, the same basic decision-making logic recurs across very different biological systems.

From searching disaster zones and responding to chemical spills to monitoring fragile ecosystems, future robot swarms may have to act in places where direct human control is difficult or dangerous. To operate autonomously, the robots must be able to decide together which problem to address and where to go next. But collective decision-making creates its own vulnerability: Robots improve their decisions by sharing information, yet faulty machines, inaccurate observations or manipulated messages can mislead the entire swarm.
An international research team led by computer scientist Andreagiovanni Reina from the University of Konstanz has now shown that a simple rule inspired by biological decision-making can help robot swarms reach a clear consensus quickly – even when some of the information circulating in the group cannot be trusted. Their study, published in Nature Communications, suggests that this widespread interaction rule could make robot swarms less vulnerable to the noisy information that characterizes real-world application scenarios.

Copy the message – or pause first?
In more detail, the researchers compared two simple ways in which a robot can react when it receives information that conflicts with its own opinion: Under the first rule, known as direct-switch, the robot immediately abandons its current opinion and adopts the information it has just received. This requires very little memory or computing power, but conflicting or inaccurate messages can leave the swarm repeatedly switching opinions without forming a clear majority.

The second rule, called cross-inhibition, introduces a brief but important intermediate step. “The key is a moment of indecision. Information that conflicts with a robot’s own opinion is not immediately copied. Instead, the robot pauses and temporarily becomes uncommitted before accepting new evidence”, Reina explains. The rule was inspired by honeybee colonies searching for a new nest: Bees supporting one location can send stop signals that inhibit bees advertising competing sites. “This simple behaviour prevents the colony from remaining divided between alternatives, allowing it to reach consensus on a single destination – and the same principle works equally well in robots.”

When a few robots mislead the swarm
The researchers examined several ways in which the information available to a robot swarm may become unreliable. For example, some robots may stubbornly support one option while ignoring all opposing information. Others may occasionally rely on their own limited observations rather than information from the group. In addition, messages exchanged between robots may be altered by technical faults or deliberate external manipulation.
“We are all familiar with situations in which a small but persistent minority strongly influences a collective decision”, Reina says. “Similar effects occur in human societies, animal groups and artificial systems. For robot swarms, the source might be a malfunctioning robot, incomplete sensor data or a cyberattack”. The study revealed that in the presence of such disturbances, direct-switch frequently produced weak majorities or prolonged indecision within the swarm. Cross-inhibition, by contrast, generally enabled the swarm to reach clearer and faster decisions. This advantage even persisted when the robots had to choose among more than two alternatives or as the swarm became larger.
One of the most surprising results was that a moderate amount of disturbance could sometimes improve the accuracy of cross-inhibition. A limited amount of unreliable information prevented the swarm from settling on the worse option, making the better decision more likely. “This means that robust collective systems need not always eliminate every source of noise. Under the right conditions, imperfections can actually help the swarm avoid a poor decision”, Reina explains.

The same solution in cells, brains and bee colonies
Cross-inhibition is not unique to honeybees, but similar patterns of inhibitory interaction appear at very different levels of biological organization – from networks of neurons to the molecular regulatory networks controlling the cell cycle. Despite their differences, these systems share the same “winner-take-all” logic: competing alternatives suppress one another until one clear outcome emerges. “When the same pattern appears at such different scales, it may be pointing us towards a powerful general principle for turning competing signals into one clear decision – and one that we can borrow to build better robot swarms”, says Reina.

The study therefore connects biology and engineering in both directions. Natural systems provide inspiration for designing robust autonomous robots, while mathematical analysis and robotic experiments can help researchers understand why similar decision-making patterns recur throughout nature. For future robot swarms, such bio-inspired mechanisms could support rapid decisions, for example, about which part of a disaster zone to search first, which environmental threat requires the most urgent response or where limited resources should be deployed.

Key facts:

• Original publication: R. Zakir, T. Carletti, M. Dorigo & A. Reina (2026) Bio-inspired decision making in robot swarms under biases. Nature Communications; DOI: 10.1038/s41467-026-76408-4
• Andreagiovanni Reina leads a research team at the University of Konstanz's Centre for the Advanced Study of Collective Behaviour. His interdisciplinary research examines how large groups of autonomous robots can make decisions independently using minimalist algorithms.
• The Centre for the Advanced Study of Collective Behaviour at the University of Konstanz (Germany) is an interdisciplinary research centre that studies the principles behind the collective behaviour of animals and other systems.
• The study was conducted in collaboration with researchers from IRIDIA, the laboratory for interdisciplinary research on AI of the Université libre de Bruxelles (Belgium), and naXys, the Naumur Institute for Complex Systems (Belgium).
• Funding: German Research Foundation (DFG) and the Belgian Fund for Scientific Research (FNRS)