For concurrent consideration of reliability and safety of manufacturing systems on the one hand, and flexibility and performance on the other, a specialized control architecture was proposed recently. It consists of a strategic controller for online computing (sub-)optimal control strategies under consideration of dynamic models, safety specifications, and learned knowledge. The second building block is an operational controller for providing certified safety when dealing with behaviors that are not completely predictable, for instance in case of faults. This paper extends the architecture by a specific technique for the strategic controller: For the case of an intelligent manufacturing system which reacts adaptively to a human operator, control trajectories for a robot arm are computed online such that collision with the operator is excluded. The computation is based on solving a mixed-integer programming problem that takes a dynamic safety area around the operator into account as constraint.