Closed-loop robotic interactions reveal dynamic social filtering in schooling fish

Collective motion emerges from local interactions among individuals, yet whether interaction rules inferred from trajectory data correspond to the mechanisms actually used by animals remains unresolved. Here, we address this question using an autonomous closed-loop robotic fish implementing a data-driven model of social interactions reconstructed from the schooling fish Hemigrammus rhodostomus. The robot continuously updated its behavior from real-time tracking of freely swimming fish while reproducing spontaneous locomotion, wall avoidance, and anisotropic attraction and alignment. We compared entirely biological groups, biohybrid groups containing one robotic fish, and numerical simulations using identical behavioral descriptors across isolated individuals, pairs, and groups of five fish. The robotic fish successfully integrated into natural schools and reproduced the principal signatures of collective coordination, providing the first direct causal validation of interaction rules reconstructed from behavioral trajectories. Biohybrid experiments showed that a robot responding only to its single most influential neighbor was sufficient to sustain natural collective coordination. By contrast, numerical simulations reproduced the behavior of biological groups most accurately when each fish interacted with its two most influential neighbors. This discrepancy identifies the contribution of hydrodynamic interactions, which remain available to living fish but are absent from the robotic controller, demonstrating that physical and behavioral interactions jointly shape collective organization. These findings establish closed-loop biohybrid robotics as a powerful framework for experimentally testing the mechanisms underlying collective animal behavior.