As robotic systems become more and more pervasive in our daily lives, automating the design of their controllers is a highly desirable and important goal. There are, however, significant computational challenges due to the costly evaluation of candidate control solutions through high-fidelity simulations or physical hardware. Another critical aspect is the need for controllers, independent of their design method, to be understood, verified and analyzed. Within the realm of Multi-Robot Systems (MRSs), this is a challenging endeavor, due to the complex inter-robot interactions involved. The use of multi-level modeling is a promising direction for both challenges mentioned, as it provides a well-defined transition from physical robotic systems to more abstract, easily interpretable and computationally efficient models. In a first phase, this thesis investigates the use of modeling to mitigate the computational cost of automatic control synthesis. Specifically, a novel, controller-agnostic multi-level modeling technique is introduced and compared to a well-established, but controller-dependent method. Leveraging bespoke modeling technique, we present an arbitrator-based control synthesis framework, which, using a library of hand-coded behaviors, is capable of automatically defining both structure and parameters of the arbitrator, as well as optimizing the parameters of the behaviors. The framework employs a noise-resistant particle swarm optimizer and incorporates two levels of abstraction to synthesize competitive controllers while minimizing computational costs. After validation in a single-robot case study, we successfully extend it to a MRS scenario. Subsequently, we dive into the automatic design of the underlying behavior library. We investigate algorithms based on reinforcement learning to automatically generate libraries of behaviors that can be used by the synthesis framework. After successfully demonstrating their use for an arbitrator-based controller, we attempt to integrate automatic behavior generation into the multi-level model-based control synthesis framework to demonstrate the feasibility of such a fully automatic design approach. In a second phase, this thesis investigates the generalization of our combined multi-level modeling and data-driven approach. To keep this generalization effort within bounds, we delve into its applicability to spatially coordinated collective behaviors. Such behaviors considered in the first axis are concerned with the maintenance of specific spatial relationships among team members, which is of interest for numerous MRS applications. Collective behaviors often involve behavior-based architectures allowing for concurrent behavioral activity and are by nature not well-mixed systems, both aspects severely challenging the modeling and synthesis framework mentioned above. We start by applying our modeling and control synthesis framework to weakly spatially coordinated collective behavior (flocking), before gradually increasing the spatial coordination through our hybrid algorithms, which allow for a smooth transition to formation-based movement. We find that our approach proves advantageous and effective in scenarios with sufficiently high stochasticity and state discretization in the robot controllers.