In a world where the complexity and performance requirements of the tasks requested from micro aerial vehicles are continuously increasing, smooth design and deployment of multi-robot systems are gaining more significance. This paper tackles such challenging requirements by firmly adopting an architecture based on Nonlinear Model Predictive Control (NMPC). In order to efficiently design such architecture, we propose an approach emphasizing a closure of the reality gap between algorithmic design and physical experiments. More specifically, we use canonical system identification methods combined with additional calibration effort to enhance the faithfulness of our model in a high-fidelity simulation environment. By employing the accurate model obtained, we prototype our decentralized NMPC algorithm in a real-time iteration scheme. To improve further the performance, multi-modal, multi-rate, decentralized extended Kalman filters are integrated to the architecture. While experiments involving up to three quadrotors in high-fidelity simulation and reality outlined the approach’s validity, they also pointed out its limitations when subtle effects generated by aerodynamic interactions among quadrotors are not taken into account in the control design.