In this paper a behaviour coordination manager for a mobile manipulator is proposed. The manager activates and coordinates purposive perception-action units, so-called behaviours, of a mobile manipulator based on a task provided by a deliberative planning system. As a decision making mechanism for the behaviour coordination Bayesian belief networks are applied. The Bayesian belief networks are trained to learn the influence of the behaviours on the manipulators movement in each situation. After development, training, and testing in a virtual environment, the implemented module is transferred to a real mobile manipulator and evaluated there. Finally, experimental results are presented.