The newly emerged and quickly growing science of nanotechnology has been recognized as one of ``the twenty-first century's great leaps forward in scientific knowledge''. Self-assembly provides a powerful enabling technique for nanotechnology by providing a bottom-up solution as an alternative to the conventional top-down approach in nano-fabrication. Employing self-assembly in nanotechnology seems in fact inevitable. As we try to build ever smaller structures as big as only a few atoms, utilizing tools for putting the molecular building blocks together proves more and more inefficient and impractical. Alternatively, we may let the building blocks put themselves together, let the molecules do what they do best, self-assembling themselves into useful structures. The big question today is thus, can we learn to build things the way nature does? A core element of our work is the experimental robotic system. With the goal of realizing a distributed robotic system in which the resource-constrained robotic modules build pre-defined target structures through programmable stochastic self-assembly, our developments are centered around the 3-cm-sized water-floating Lily robotic module. Furthermore, we implement a controllable setup around the Lily robotic modules where several environmental features such as the fluidic flow in the environment as well as the ambient luminosity perceived by the modules can be controlled in order to influence the self-assembly process towards the target structure. The experiments reported in this dissertation has been carried out with up to 15 Lily robotic modules. Developing models that accurately describe the assembly process dynamics is a key component in studying programmable stochastic self-assembling systems. Such models help in: (1) accurately predicting the performances (assembly rate and yield) of the distributed system, and (2) evaluating and optimizing control strategies, whether distributed (e.g., ruleset controllers programmed on the modules) or centralized (e.g., modulating environmental features such as mixing forces deriving random interactions among modules), based on model predictions. We develop models at three abstraction levels, namely submicroscopic, microscopic, and macroscopic. Programmable self-assembly defines a subclass of self-assembly processes where the building blocks carry information about the final desired target structure. It is through modifying this information that the outcome of the self-assembly process can be programmed. The problem of distributed control for programmable self-assembly is thus one of designing a global-to-local behavioral compiler. The problem of ruleset synthesis for programmable self-assembly of bodiless modules has been studied in the literature by employing graph grammar formalism. We extend the graph grammar formalism and take into account the morphology of the robotic modules. This allows for formulating automatic rule synthesis methods for self-assembly of robotic modules, where the synthesized rules can be directly deployed on the robotic modules, with no further tuning. Moreover, we propose a new rule synthesis algorithm for synthesizing assembly rules which further promote parallelization in the self-assembly process without losing guarantees on the completeness of the achieved target.