Human Task Reproduction with Gaussian Mixture Models

This paper proposes a new motion-copying system which uses statistical approaches for recording and reproducing of human tasks. In conventional motion-copying systems, haptic data of human motions is recorded directly to the database at every sampling. As a result, the amount of haptic data for the database is large in general. In addition to that, it is hard to segment and reorganize the recorded human motions. Therefore, the motion-copying system proposed in this paper uses Gaussian mixture model (GMM) to model human motions for the recording. The modeled GMM are recorded in the database instead of raw haptic data. Therefore, the recorded data size is reduced compared with conventional methods. Furthermore, the automatic segmentation and reorganization of recorded human motions are possible. Proposed method uses Gaussian mixture regression (GMR) to retrieve haptic information from GMM for the reproducing. The validity of the proposed method was confirmed through 1DOF motion-copying experiment.