Virtual-hand input realizes the natural and dexterous functionality in direct humancomputer interaction. The recognition of virtual-hand models is difficult due to two reasons: the complexity of hand structure and the lack of accurate measure, which may be caused by either mechanical noise or human factors. This paper presents a novel fuzzy-logic recognition system that can effectively deal with imprecise hand data. The system consists of three components: the classifier, identifier, and posture database. Fuzzy-logic processing is applied in both the classifier (to build class-indexing structure in the posture database) and in the identifier (to find the most likely match from the database in real-time VR applications). The posture database provides a GUI interface for the user to browse the posture images and interactively update the database by adding and deleting the sample postures, as well as adjusting the certainty threshold in recognition. Our experiments show that the fuzzy-logic method can keep nearly constant performance in recognition using tens of microseconds, even as the size of posture database increases. The recognition rate declines only slightly when the standard derivation of the noise distribution (Gaussian distribution) in the input parameters is below 15 deg. The bounded value in uniform distribution to achieve accurate recognition is below 20. The results show that the fuzzy-logic processing can improve the tolerance of noise or imprecise data in an efficient way. A free-hand modeler based on fuzzy-logic processing has been developed for creating virtual objects and Web-based 3-D VRML worlds.