A robotic in-hand manipulation dictionary based on human data

We propose a method to extract the bases of human in-hand manipulation actions that can be the core of a path planner algorithm for multi-fingered robotic hands. We consider in-hand manipulation actions as a combination of motion primitives. This work concerns the extraction of the motion primitives from human in-hand manipulation data as a single dictionary. The motion primitives dictionary will allow to re-create in-hand manipulation paths and to generate new paths by combining motion primitives. To extract motion primitives, we use Non-negative Matrix Factorization (NMF) as a tool. The human in-hand manipulation data is tuned in order to meet the NMF method properties and the in-hand manipulation case. The implementation for the NMF algorithm to obtain one dictionary representing the whole hand motion is explained and demonstrated. Motion primitives are able to re-create paths from testing data with a high efficiency and a low standard deviation of the error between the fingertips positions and their estimations based on the learnt dictionary.