Online human in-hand manipulation skill recognition and learning

This work intends to contribute to transfer human in-hand manipulation skills to a dexterous prosthetic hand. We proposed a probabilistic framework for both human skill representation and high efficient recognition. Gaussian Mixture Model (GMM) as a probabilistic model, is highly applicable in clustering, data fitting and classification. The human in-hand motions were perceived by a wearable data glove, CyberGlove, the motion trajectory data proposed and represented by GMMs. Firstly, only a certain amount of motion data were used for batch learning the parameters of GMMs. Then, the newly coming data of human motions will help to update the parameters of the GMMs without observation of the historical training data, through our proposed incremental parameter estimation framework. Recognition in the research takes full advantages of the probabilistic model, when the GMMs were trained, the log-likelihood of a candidate trajectory can be used as a measurement to achieve human in-hand manipulation skill recognition. The recognition results of the online trained GMMs show a steady increase in accuracy, which proved that the incremental learning process improved the performance of human in-hand manipulation skill recognition.