For a joint performance of a dialogue between a human partner and a robot, the robot analyzes phrase and action of the partner to detect a recognized behavior of the partner and analyze a state of audience listening to utterances from the partner and the robot to detect a recognized state of the audience. A scenario describing the dialogue is stored in entries of a memory. The memory is successively referenced entry by entry and a check is made for a match between an utterance by the partner or the robot to a reaction from the audience. Responsive to a currently detected audience state, a corresponding robot behavior is determined. Preferably, possible partner's behaviors and expected audience states are mapped in a database to specified robot behaviors. The database is searched for a specified robot behavior corresponding to a currently sensed partner behavior or a currently sensed audience state.