We propose a neural network controller using only joint position measurements for rigid robot manipulators. The joint velocity needed for the control law is estimated using an observer based on sliding mode. A decomposed structure neural network approximates the unknown model of the system. Each neural network (MLP) approximates a separate element of the dynamical model. These approximations are used to conduct an adaptive stable control law. The TaylorYoung series was used to solve the nonlinearity problem of the MLP and to lead to the parameters adaptation algorithm. The corresponding parameters are the weights of the neural net. They are updated via the adaptation algorithm derived from stability study of the system in closed loop using the Lyapunov approach and intrinsic properties of robot manipulators. Simulations were conducted to show the conductance of the proposed controller.