Learning Control Policies for Variable Objectives from Offline Data