This paper describes methods to develop autonomous flight control systems for an aero-robot using an unmanned helicopter RMAX developed by YAMAHA Co., LTD. Such systems are difficult to develop because helicopter dynamics are nonlinear. We propose an efficient method to design controllers by training neural networks. The proposed method is easy to use with online training or adaptive controllers to compensate for undesirable effects not modeled or sudden changes of the controlled object and environment, making the control system highly reliable. Results of flight experiments demonstrate the effectiveness of our approach.