Upper-limb myoelectrical prosthesis provides an intuitive interface for prosthesis users to restore an acceptable level of grasping function. However, the existing EMG classifi-cation algorithms are still incapable of decoding various human grasping gestures that are important to control dexterous multi-finger hands fully. Here, we present a novel method to interpret sophisticated grasping motions, using a combination of unsupervised clustering with non-mandatory leaf-node prediction (NMLNP) hierarchical classification. Results show that the optimized subset of gestures specific to the individual selected by clustering leads to higher classification accuracy. In addition, hierarchical classification can achieve better performance than flat classifiers and provide an adaptive interface that automatically outputs a larger set of gestures for subjects with better EMG quality.