Collective motion in animal groups is typically inferred from trajectory data, but whether those inferred rules reflect the mechanisms animals actually use has remained uncertain. This study tested interaction rules reconstructed from behavioral trajectories by embedding an autonomous closed-loop robotic fish in groups of the schooling fish Hemigrammus rhodostomus. The robot implemented a data-driven model of social interactions and updated its behavior in real time from continuous tracking of freely swimming fish, allowing a causal test of whether trajectory-derived rules can reproduce natural collective coordination.
The robotic fish reproduced multiple behavioral components observed in the species under study. It implemented spontaneous locomotion patterns, strategies for wall avoidance, and anisotropic rules of attraction and alignment derived from the trajectory analyses. Critically, the robot operated in a closed-loop configuration: it continuously received tracking information from the live animals and updated its motor output online so that its movements reflected the same descriptors used to model biological individuals.
The authors compared three experimental conditions using identical behavioral descriptors whenever possible: entirely biological groups of fish, biohybrid groups containing a single robotic fish among live conspecifics, and numerical simulations in which digital agents followed the same reconstructed rules. These comparisons were performed across different social contexts, including isolated individuals, pairs, and groups of five fish. Using the same descriptors allowed direct assessment of how well the reconstructed rules and their implementations reproduced signatures of collective behavior.
The robotic fish successfully integrated into natural schools of Hemigrammus rhodostomus and reproduced the principal signatures of collective coordination observed in biological groups. This result provides a direct causal validation that the interaction rules reconstructed from trajectory data can generate natural collective organization when implemented in an embodied agent. Because the robot responded in real time to living fish, the biohybrid experiments demonstrate that trajectory-derived rules can be sufficient to drive coordinated group dynamics in a naturalistic setting.
A notable difference emerged between biohybrid experiments and numerical simulations. In biohybrid trials, the robot needed to respond only to its single most influential neighbor to sustain natural collective coordination among live fish. In contrast, numerical simulations reproduced the behavior of biological groups most accurately when each simulated agent interacted with its two most influential neighbors. This disparity highlights a context-dependent aspect of social filtering in schooling fish: the number of neighbors whose influence must be accounted for differs between physical robotic implementation and purely simulated agents.
The authors interpret the discrepancy between robot and simulation as evidence for the contribution of hydrodynamic interactions. Living fish experience fluid-mediated forces and cues that are not captured by the robotic controller or by the behavioral descriptors alone. Because the robot lacks these physical hydrodynamic couplings, it could match natural coordination while attending to only one neighbor, whereas simulations that omit hydrodynamics required interactions with two neighbors to reproduce group-level patterns. Thus, the work indicates that both physical (hydrodynamic) and behavioral interactions jointly shape collective organization in schooling fish.
By demonstrating that an embodied closed-loop robot can causally reproduce key features of collective motion, the study establishes closed-loop biohybrid robotics as a powerful experimental framework for testing mechanisms underlying collective animal behavior. This approach complements observational and simulation-based inference by allowing controlled perturbations and direct manipulation of sensory-motor rules in an embodied agent interacting with live animals. The findings underscore the importance of considering both behavioral rule reconstruction and organismal physical interactions when interpreting models of collective systems.
The preprint declares funding from Campus France and the Swiss National Science Foundation. The authors declared no competing interest. The manuscript is a preprint and has not been peer reviewed; all methodological and result summaries above reflect information reported in the source preprint.