In this article a depth stabilizing system for AUVs is studied. Two feedforward multi layer neural networks are used. The first neural network is utilizied for approximation of AUV dynamics. The second one plays the controller role. Both neural networks are trained in online regime in order to cope with changing AUV dynamics. A cost function is used in control action calculation. The system stability and performance are verifies through simulation carried in Matlab. The simulation results show that the proposed depth control system is able to keep the predifined depth and can be applied in real control systems.