In this paper we propose a framework for task allocation in human-robot collaborative assembly planning. Our framework distinguishes between two main layers of abstraction and allocation. In the higher layer we use an abstract world model, incorporating a multi-agent human-robot team approach in order to describe the collaborative assembly planning problem. From this, nominal coordinated skill sequences for every agent are produced. In order to be able to treat humans and robots as agents of the same form, we move relevant differences/peculiarities into distinct cost functions. The layer beneath handles the concrete skill execution. On atomic level, skills are composed of complex hierarchical and concurrent hybrid state machines, which in turn coordinate the real-time behavior of the robot. Their careful design allows to cope with unpredictable events also on decisional level without having to explicitly plan for them, instead one may rely also on manually designed skills. Such events are likely to happen in dynamic and potentially partially known environments, which is especially true in case of human presence. © 2017 IEEE