A self-organizing neural-net-controller system (SONCS) is proposed as an adaptive system to handle unknown dynamics of a controlled object and unpredictable environmental conditions. SONCS consists of a controller, forward models, and evaluation and adaptation mechanism. The forward models are neural networks which express the forward dynamics of the controlled object. The controller is a neural network which provides control inputs to the object and the forward models. The evaluation and adaptation mechanism is a modification tool which calculates the difference between actual and desired motion of the object and modifies the system to get a specific function. The basic idea of this system is to adapt the controller network with this evaluation and adaptation mechanism observing the object's motion with the forward models. SONCS is also equipped with an initiation tool called, premature controller, which can be easily constructed with just qualitative information on the object's dynamics. Efficiency of the developed SONCS is demonstrated in the application to 'PTEROA60', a test-bed for underwater vehicles. A suitable neural-net-controller, which can let the vehicle swim stably in desired state, is organized in this system through free-swimming in a tank. It is concluded that an adaptive system, which can generate automatically a desired controller to keep the controlled object in target state, can be constructed using neural network technology. The proposed SONCS has a great possibility to realize highly sophisticated control system which can deal with complicated control problems.