The research team, led by Anne M. Mayo, Craig Carrigee, Audrey B. Harrison, Michael L. Mayo, and Kevin R. Pilkiewicz, set out to test a hypothesis about distributed computing in biology. Insects and other cognitively simple organisms routinely solve what is, from a computational standpoint, a fiendishly difficult problem: coordinating the movement of many individuals through a constrained environment without grinding to a halt. The researchers reasoned that such feats must be achieved by breaking the complexity down into a vast number of tiny local decisions, each demanding only a small amount of information about the immediate surroundings. To probe this idea, they turned to an unlikely pair of experimental subjects: American cockroaches and vibrating brush-head robots known as bristlebots.

The experimental arena was a plexiglass enclosure roughly four feet long, with a six-inch-wide path running down its center. Glued to the floor of this corridor was a diamond-shaped arrangement of four cylindrical obstacles, each three inches in diameter, forcing any traveler to weave around them. At one end sat a holding chamber where the agents began each trial; at the other, a bait chamber containing food and shelter for the roaches, or simply the downhill end of a gentle incline for the robots. A camera mounted on a tripod above the enclosure recorded every trial at thirty frames per second, and the open-source tracking software TRex converted the footage into precise trajectories for every individual.

The two agent types were driven through the course by very different means. The cockroaches, coaxed forward by a gentle fanning with a piece of cardboard, relied on their antennae: previous work has shown that these sensory organs detect walls and obstacles, and the roach steers away by rotating its body whenever an antenna makes contact. The bristlebots, commercially available Hexbug Nano devices, propel themselves through the asymmetric angling of rapidly oscillating bristles, and their elongated bodies cause collisions to exert a torque that rotates them away from whatever they strike. Despite these mechanical differences, the researchers found that the two agents navigated in strikingly similar ways. Heat maps built from thirty-four solo trials of each type showed nearly overlapping patterns of space use, and the average time spent crossing the enclosure did not differ significantly between bugs and bots.

Individually, then, a roach and a bristlebot are roughly equally competent navigators. The dramatic divergence appeared only when the agents were tested in groups. In trials with ten agents, the two species performed comparably; a permutation test yielded a p-value of 0.988, indicating no statistically significant difference. But when the group size was raised to twenty and the local density in the obstacle corridor climbed, the roaches pulled decisively ahead. The normalized effective crossing times of the two agent types diverged with a p-value of 0.029, a statistically significant gap attributable to one behavioral difference: the roaches, unlike the robots, were willing to wait.

To understand and generalize this difference, the team built an agent-based model that reduces navigation to its informational bare minimum. Each simulated agent is a circle with two protruding antennae, and at every time step the agent checks whether each antenna is touching something. In the asocial version of the model, each antenna registers a simple binary state, contact or no contact, meaning the entire movement decision rests on just two bits of information, one per antenna, regardless of how many other agents crowd the arena or how complicated the environment is. If neither antenna touches anything, the agent moves straight ahead. If one antenna is in contact, the agent rotates away from it until both are free. If both are in contact, the agent considers the smallest rotations in either direction that would free them, with a bias toward continuing forward.

The social version of the model adds a single refinement with profound consequences. The agents now distinguish between antennal contact with a wall or obstacle and contact with another agent’s body, and they further distinguish mutual contact, where each agent’s antenna touches the other’s body, from non-mutual contact, where only one agent touches the other. Non-mutual contact typically means one agent is approaching another from behind, and in that situation the follower has a tunable probability of simply stopping and waiting for the leader to move. Mutual, face-to-face contact triggers no such patience, since two agents blocking each other head-on gain nothing from waiting; instead they maneuver around one another. This distinction raises the informational cost from two bits to four bits per time step, but crucially the cost remains fixed no matter how dense the crowd or how complex the terrain.

The payoff of this stop-and-wait rule was quantified with a new metric the authors call the effective mean crossing time. The ordinary average crossing time is misleading, because it ignores agents that give up, turn around, and exit the way they came. The new metric divides the average crossing time by the fraction of agents that successfully traverse the enclosure, so a group in which half the members fail effectively pays double the time. In simulations of one hundred agents, increasing the waiting probability monotonically reduced this metric, and the improvement grew more dramatic as the emission rate, and hence the local agent density, increased. The benefit persisted, and even intensified, when a quarter of the agents were assigned slower speeds, mimicking the timid individuals in real roach groups that act as moving roadblocks. The authors frame the result in game-theoretic terms: although pausing in the open carries a temporary cost in vulnerability, the strategy ultimately improves each individual’s odds of a swift, successful crossing, echoing the well-documented slower-is-faster effect seen in human crowds and vehicle traffic.

The researchers also draw an evocative analogy to fluid mechanics. At high agent densities, the waiting probability behaves like viscosity: large values produce smooth, laminar flow around obstacles, while small values allow reflections and backflows that collide with oncoming traffic, generating something akin to turbulence. Compared with classic flocking frameworks such as the Vicsek model, where each agent must estimate the orientations of all neighbors within its sensory radius, an amount of information that scales linearly with local density, the four-bit stop-and-wait scheme is extraordinarily frugal. The authors even sketch a hardware realization: a cockroach-shaped robot with two wire antennae wired directly to the circuits driving its legs, where pressure on one antenna halts the opposing legs and turns the robot, and pressure on both antennae stops it entirely, embodying the patience rule without any advanced sensors. If such a design proves viable, swarms of cheap, sensor-poor robots could one day navigate disaster zones and cluttered environments by socializing with anything that moves, provided they are programmed with the one virtue the parable of the long chopsticks has always recommended: a little regard for others.

Subject of Research: Collective navigation and jam avoidance in cockroach groups and bristlebots through minimal social interaction

Article Title: Patience is a virtue: How nature achieves group navigation with minimal cooperativity

Article References: Mayo, A. M., Carrigee, C., Harrison, A. B., Mayo, M. L., & Pilkiewicz, K. R. (2026). Patience is a virtue: How nature achieves group navigation with minimal cooperativity. PLOS Complex Systems, 3(10), e0000136. https://doi.org/10.1371/journal.pcsy.0000136

Image Credits: AI Generated

DOI: 10.1371/journal.pcsy.0000136

Keywords: collective navigation, cockroaches, bristlebots, swarm robotics, agent-based model, jamming, stop-and-wait interaction, distributed computing, slower-is-faster effect, crowd dynamics, minimal cooperativity, effective mean crossing time