The advancement of robotics as a field towards new applications and domains has always been accompanied with and enabled by advances in robot software architecture. This is true for the first plan-based systems, introducing flexibility in action sequences, through behavior-based architectures, introducing speed, agility, and emergent behavior, to hybrid layered architectures with their combination of aspects of both. Most recently, a focus on the software engineering aspects and component-based architectures aims to increase engineering efficiency and maintainability, while also studying the architectural proposals in more empirical depth.

This thesis contributes to this development by advancing architectural principles needed for better Human-Robot-Interaction (HRI) systems, in particular, situated, social HRI. These systems are characterized by a tight integration of interaction and action, and by the need to rapidly iterate designs, particularly the interaction, during research and development.

Specifically, the well-known principle of describing action execution with finite-state machines has been developed into a general, abstract coordination interface, the Task-State Pattern. Firstly, the historical emergence of a commonality in otherwise different coordination systems has been identified and formalized as architectural

pattern. Secondly, a means of implementation has been proposed that enhances abstract observability, allowing diverse tasks to be observed in a general fashion. Thirdly, the benefits of implementing the pattern in a service-level toolkit have been demonstrated.

Moreover, the efficient design and engineering of components based on the data-flow principle has been studied. In particular, a decomposition method to increase

the re-usability of the constituent nodes in such components has been proposed, based on ideas from event-based software integration. The suitability of the method for the robotics domain has been studied, to contribute to the design knowledge, and to suggest possible remedies for situations where data-flow alone is not adequate, such as maintaining global state.

All of these contributions have been integrated, and empirically evaluated, in a comparative fashion on multiple successive iterations of an HRI scenario, the "Curious Robot". Such comparative evaluation is, to the best of the authors knowledge, a first in this area. The identified pattern is applied here to provide tight integration between coordination components in the interaction and manipulation sub-system without large changes to each. As a result, the scenario realizes detailed mixed-initiative, multimodal interaction in an object-learning setting.