As the field of ethology advances, especially over the past two decades, the role of animal-robot interaction tools has increasingly become essential. This importance arises from the need for controlled, repetitive, repeatable, and long-duration experiments, which not only relieve human researchers from tedious tasks but also offer novel insights into the rules that govern collective behavior. Such devices can infiltrate groups of animals (in this case, fish) and engage in closed-loop interactions, eliciting responses that range from biomimetic to modulated behavior. However, constructing harmonious biohybrid groups of animals and robots is an intricate task. Despite significant progress in the domain, many questions remain unanswered, necessitating further research and development in both robotics and collective behavior modelling. This thesis delves into the intersection of collective behavior phenomena and robotics. It capitalizes on the advancements in electronics manufacturing, cutting-edge algorithms, and increased accessibility to computational power that have reshaped the field, resulting in a mixed society methodology that achieved unprecedented levels of biomimicry. Initially, a comprehensive exploration was conducted to understand how fish groups interact with artificial agents. We showed that models displaying active, bidirectional interactions lead to a higher probability of integrating the artificial agent into the fish group. This realization underlined the necessity of transitioning to more detailed and accurate models of interaction that can withstand comparisons to spontaneous fish interactions. Addressing these limitations, we carried out an extensive study on the key design factors that enhance the performance of social interaction models, also revealing the need for a rigorous spatio-temporal benchmarking metric set. This set ensures these models successfully generate realistic short- and long-term social dynamics. However, these models uncovered a secondary engineering problem. Transferring high-fidelity models back to reality demanded highly agile and responsive robotic equipment, a requirement unmet by the current state-of-the-art. Accordingly, within this thesis, we designed a novel framework, inclusive of an experimental setup, a mobile robot, and ancillary software (e.g., for robot control, artificial intelligence and analytical behavioral models...) to overcome these limitations. This approach subsequently enabled us to make substantial strides towards bridging the "biomimicry gap" by transferring high-fidelity models from simulation to reality. The culmination of this thesis outlines our success in progressively bridging this gap, demonstrating unprecedented similarity between simulations, biohybrid, and spontaneous fish-only interaction experiments. By open-sourcing the entirety of the developed software and hardware tools, we aim to lay a solid foundation for future research in the realm of robot-animal interaction. In conclusion, this thesis significantly contributes to our understanding of the underlying rules that govern collective behavior. We hope it will pave the way for the creation of truly biohybrid groups and set the stage for future explorations into models of social interactions and their exploitation as robot controllers.