Model-based control uses a "best-guess" plant for use by the controller. If all things were perfect, the best-guess plant matches the actual plant. To test how robust the model-based controller is, it helps to have an actual plant that differs in many ways, e.g., the mass, center of mass, and inertia of bodies, the location of joints, actuator limits, friction constants, sensor models, etc.

Hence, it is helpful to have two plants. One for the controller and one for the actual plant.