Controllers for swarms of robots are hard to design as swarm behaviour emerges from their interaction, and so controllers are often evolved. However, these evolved controllers are often difficult to understand, limiting our ability to predict swarm behaviour. We suggest behaviour trees are a good control architecture for swarm robotics, as they are comprehensible and promote modular reuse. We design a foraging task for kilobots and evolve a behaviour tree capable of performing that task, both in simulation and reality, and show the controller is compact and understandable.

Designing the individual robot rules that give rise to desired emergent swarm behaviors is difficult. The common method of running evolutionary algorithms off‐line to automatically discover controllers in simulation suffers from two disadvantages: the generation of controllers is not situated in the swarm and so cannot be performed in the wild, and the evolved controllers are often opaque and hard to understand. A swarm of robots with considerable on‐board processing power is used to move the evolutionary process into the swarm, providing a potential route to continuously generating swarm behaviors adapted to the environments and tasks at hand. By making the evolved controllers human‐understandable using behavior trees, the controllers can be queried, explained, and even improved by a human user. A swarm system capable of evolving and executing fit controllers entirely onboard physical robots in less than 15 min is demonstrated. One of the evolved controllers is then analyzed to explain its functionality. With the insights gained, a significant performance improvement in the evolved controller is engineered.

Area coverage and collective exploration are key challenges for swarm robotics. Previous research in this field has drawn inspiration from ant colonies, with real, or more commonly virtual, pheromones deposited into a shared environment to coordinate behaviour through stigmergy. Repellent pheromones can facilitate rapid dispersal of robotic agents, yet this has been demonstrated only for relatively small swarm sizes (N < 30). Here, we report findings from swarms of real robots (Kilobots) an order of magnitude larger (N > 300) and from realistic simulation experiments up to N = 400. We identify limitations to stigmergy in a spatially constrained, high-density environment—a free but bounded two-dimensional workspace—using repellent binary pheromone. At larger N and higher densities, a simple stigmergic avoidance algorithm becomes first no better, then inferior to, the area coverage of non-interacting random walkers. Thus, the assumption of robustness and scalability for such approaches may need to be re-examined when they are working at a high density caused by ever-increasing swarm sizes. Instead, subcellular biology, and diffusive processes, may prove a better source of inspiration at large N in high agent density environments.