Myoelectric classification has been widely studied for controlling prosthetic devices and human computer interface (HCI). However, it is still not robust due to external conditions: limb position changes, electrode shifts, and skin condition changes. These issues compromise the reliability of pattern recognition techniques in myoelectric systems. In order to increase the reliability in the limb position effect when a limb position is changed from the position in which the system is trained, this paper proposes a myoelectric system using dynamic motions. Dynamic time warping (DTW) technique was used for the alignment of two different time-length motions, and correlation coefficients were then calculated as a similarity metric to classify dynamic motions. On the other hand, Fisher's linear discriminant analysis was applied on static motions for the purpose of dimensionality reduction and Nave Bayesian classifier for classifying the motions. To estimate the robustness to the limb position effect, static and dynamic motions were collected at four different limb positions from eight human subjects. The statistical analysis, t-test (p<0.05), showed that, for all subjects, dynamic motions were more robust to the limb position effect than static motions when training and validation sets were extracted from different limb positions with the best classification accuracy of 97.59% and 3.54% standard deviation (SD) for dynamic motions compared with 71.85% with 12.62% SD for static motions.