Structural bias enables reliable sim-to-real transfer in reinforcement learning for robot swarms

Reinforcement learning offers a powerful framework for automatically generating control policies for robot swarms, yet its application to physical multirobot systems has been limited by the simulation-to-reality gap. Despite substantial advances in multiagent reinforcement learning, systematic demonstrations of gradient-based sim-to-real transfer to physical robot swarms remain scarce. Most work remains confined to simulation, and the few real-robot demonstrations lack evaluation across multiple runs and tasks. We show that this barrier can be overcome by embedding structural constraints into policy learning. We introduce attention-based counterfactual policy gradients for behavior selection in robot swarms (SwarmACB), a multiagent actor-critic method in which robots arbitrate among predefined behavior modules-reactive sensorimotor routines such as phototaxis or aggregation-rather than learning low-level motor commands. This architectural choice reduces the variance of learned solutions, which we argue mitigates overfitting to simulation artifacts. Through systematic evaluation on five collective missions using 20 physical e-puck robots (250 real-robot runs), we demonstrate that SwarmACB transfers successfully from simulation to reality, substantially outperforming continuous-action baselines. By progressively constraining policy architecture, perception, and model capacity, we show that smaller, structured policies transfer more reliably than large, unconstrained ones. These findings challenge the assumption that scale is necessary for successful transfer, suggesting instead that principled structural design enables reinforcement learning to produce control software for robot swarms that is both learned automatically and directly deployable in the physical world.