A control method that trains robot to learn motion skills by natural language is explored in this paper, where fuzzy neural networks are used as a general structure of controller for different motion primitives. For any specified motion primitives, the controller acquires knowledge and is trained by instructions of natural language from a supervisor. This might be a more natural way to construct a motion controller. Based on the trained motion primitives, more complex motion tasks can be implemented. Experiments on natural language training for motion primitives have been done on a wheeled robot.