It is well known that computed torque robot control is subjected to performance degradation due to uncertainties in robot model, and application of neural network (NN) compensation techniques are promising. In this paper we examine the effectiveness of neural network (NN) as a compensator for the complex problem of Cartesian space control. In particular we examine the differences in system performance of accurate position control when the same NN compensator is applied at different locations in the controller structure. It is found that using NN to modify the reference trajectory to compensate for model uncertainties is much more effective than the traditional approach of modifying control input or joint torque/force. To facilitate the analysis, a new NN training signal is introduced and used for all cases. The study is also extended to non-model based Cartesian control problems. Simulation results with three-link rotary robot are presented and performances of different compensating locations are compared.