The study, published in Frontiers in Robotics and AI, tackles a problem that has haunted swarm robotics for years, known as the reality gap. Simulations are fundamental for planning experiments with real robots, but the dynamic models they use cannot fully capture the complexity of the environment a robot actually operates in. What makes this work unusual is that the gap did not merely change absolute performance numbers. It changed the relative ranking of two competing bio-inspired behaviours, meaning that a team relying on simulation alone would have drawn a qualitatively different conclusion about which strategy to deploy on physical hardware.

The biological inspiration comes from ant colonies, where foragers share information about profitable food sites. In tandem running, a leader ant that knows the location of a food source physically guides a naive recruit to the site, a form of teaching documented in species such as Temnothorax ants. The strategy allows important information to be shared among colony members, but it carries costs: the leader must first find and memorize the route, the recruit may wait inside the nest when no leaders are available, and the tandem couple travels more slowly than a single individual. The alternative, lone scout foraging, is the more typical strategy, where a single naive ant explores the surroundings using a random search without any prior knowledge of the food location.

To compare these strategies, the team used kilobots, each with a diameter of 3.3 centimetres, communicating over short range with infrared. In the tandem running experiments, stationary kilobot beacons were arranged in a line to simulate the visual landmarks that a leader ant memorizes along the route. Each beacon transmitted a unique ID, with the food source carrying the lowest ID, and the leader constantly searched for a lower ID than the one it was receiving. The pair orbited around each beacon in turn until the next one was detected, with the follower keeping behind the leader and maintaining a close distance. The lone scout, by contrast, was programmed with an isotropic random walk, turning in a random direction for 2.5 seconds and then moving straight for 5 seconds before turning again, a simple stochastic search chosen for its equal probability of turning left or right.

The experiments were conducted on a tabletop arena bounded by PVC planks, with the arena length increasing from 25.5 to 35.5 centimetres as the number of beacons grew from one to three, corresponding to longer memory paths. Every moving kilobot had to be individually calibrated using the kilogui software, because strong inter-individual variations in motor output mean each robot needs its own parameters to walk straight and turn reliably. Trials were recorded with an overhead camera, and the kilobots’ LEDs indicated their state for accurate timing. Tandem pair runs were repeated five times, while the lone scout sample was increased to 15 runs given the high variance observed in initial trials.

The simulations in CoppeliaSim, using the K-Team kilobot model with the Bullet 2.78 physics engine, reproduced the behaviours as closely as possible, with only minor adjustments to orbiting distances and turning times. The results revealed the reality gap in full. In real-life experiments, median search times for tandem pairs were 41.60, 72.50 and 120.30 seconds in the small, medium and large arenas, compared with 42.10, 77.00 and 156.60 seconds for lone scouts, so tandem pairs grew progressively faster as the arena expanded. In simulation, the pattern reversed: tandem pairs took 158.21, 268.01 and 317.05 seconds, while lone scouts needed only 111.70, 165.61 and 259.01 seconds. Notably, the average time for tandem pairs was roughly three times higher in simulation than with real kilobots.

The team also quantified the gap directly by subtracting real-life times from simulation times. For lone scouts, the gap increased regularly with arena size, from 69.60 to 102.41 seconds, while for tandem pairs it started higher at 116.61 seconds and jumped to around 196 seconds in the larger arenas. Interestingly, the gap varied more in the less stochastic tandem running algorithm than in the more stochastic lone scout search, which led the researchers to deduce that stochastic decision-making is not the main source of the reality gap. Instead, the discrepancy is more likely caused by epistemic uncertainty, such as the accuracy of the dynamic models used in simulators, including friction and contact dynamics that the virtual world treats as ideal.

A combination analysis strengthened the findings. Because a tandem leader must first discover the food source before recruiting a follower, the researchers combined lone scout runs with tandem pair runs, producing 3003 unique combinations per arena size using MATLAB’s nchoosek function. This approach included the leader’s initial exploratory trip and mitigated the extremely high variance of individual random searches. The combination tests confirmed the pattern: in real-life kilobot combinations, the tandem-plus-scout strategy became progressively faster than two lone scouts as arena size increased, with all comparisons statistically significant, while in simulation the lone scout strategy became faster in the medium and large arenas.

The authors identify several technical reasons for the discrepancy. Real surfaces have small imperfections in which kilobot legs can get stuck, legs can be bent affecting speed and turning, real kilobots need charging so their behaviour degrades over time, and real communication relies on the surface being reflective enough for infrared signals, whereas simulation bases it on proximity. One of the most striking examples was the orbiting functionality, which was not performed by real kilobots as it was in simulation. The study also notes that three out of eight tandem pair tests with real kilobots failed, a 37.5 percent failure rate, possibly caused by malfunctioning infrared proximity sensors, compared with a 23.2 percent breakup rate observed in nature.

The implications extend well beyond the robotics laboratory. The researchers suggest immediate applications in selecting robot navigation strategies when the target location is unknown, and in swarm intelligence, collective decision making and navigation techniques. More distant fields such as logistics, agriculture and search and rescue could also benefit. But the central message is a cautionary one: even in the oversimplified foraging arena of a robotic lab, simulation alone was not enough to predict which bio-inspired strategy would win, and experiments with real kilobots will also be needed. For a discipline that increasingly relies on virtual testing before physical deployment, this study is a vivid demonstration that the reality gap can do more than distort numbers, it can rewrite the conclusions themselves.

Subject of Research: Swarm Robotics

Article Title: Comparing simulation and physical kilobots performance in two ant-inspired foraging behaviours

Article References: Ellis, N., Kelly, J., Caravaggio, A., Young, I., Manfredini, F., & Giannaccini, M. E. (2026). Comparing simulation and physical kilobots performance in two ant-inspired foraging behaviours. Frontiers in Robotics and AI, 13, Article 1828495. https://doi.org/10.3389/frobt.2026.1828495

Image Credits: AI Generated

DOI: 10.3389/frobt.2026.1828495

Keywords: swarm robotics, kilobots, reality gap, ant-inspired foraging, tandem running, CoppeliaSim, bio-inspired robotics, Comparing, simulation, physical, performance, ant-inspired