This paper deals with the problem of dynamic hovering of a bio-inspired robot with undulatory fins. First, the mechanism and dynamic model of the underwater robot named RobCutt-II is briefly introduced. Next, an neural network visual servo control is presented to achieve hovering control. In particular, a dual-loop visual hovering control is designed. At first, we analyze the relationship between the image feature error and the direction of the controllable velocity. In the outer loop, a PID controller is used to calculate the current controllable velocity based on image feature error in real time. Meanwhile, the inverse model of undulatory fins is established using the neural network. The velocity error is hereby mapped to the control parameters of the undulatory fins of the RobCutt-II. The simulation results of dynamic hovering are provided to illustrate the validity of the proposed method.