By taking a Stewart platform as an example, this paper presents a novel calibration method by designing a robust joint compensator based on artificial neural networks. In this method, the pose error arising from various time-independent error sources is treated as that produced only by configuration-dependent joint motion errors equivalently, thus allowing the static pose error to be eliminated by directly correcting the nominal joint variables. Hence, the calibration procedure can be implemented in three successive steps: (1) acquisition of necessary joint corrections with point measurement at finite configurations considering near singularity problems, (2) approximation of the function between joint corrections and nominal joint variables using feedforward neural networks with coupled/decoupled architectures, and (3) design of a joint compensator embedded in the numerical control system to realize online real-time error compensation. Experimental results show that the proposed robust compensator based on coupled or decoupled networks can significantly improve the static pose accuracy in comparison with previous methods.