Multifunctional Principal Component Analysis for Human-Like Grasping

In this paper, a method to derive the synergies subspace of an anthropomorphic robotic arm–hand system is proposed. Several human demonstrations of different objects grasping are measured using the Xsens MVN suite and then mapped to a seven Degree-of-Freedom (DoF) robotic arm. Exploiting the anthropomorphism of the kinematic structure of the manipulator, two Closed-Loop Inverse Kinematics (CLIK) algorithms are used to reproduce accurately the master’s movements. Once the database of movements is created, the synergies subspace are derived applying the Multivariate Functional Principal Component Analysis (MFPCA) in the joint space. A mean function, a set of basis functions for each joint and a pre-defined number of scalar coefficients are obtained for each demonstration. In the computed subspace each demonstration can be parametrized by means of a few number of coefficients, preserving the major variance of the entire movement. Moreover, a Multilevel Neural Networks (MNNs) is trained in order to approximate the relationship between the object characteristics and the synergies coefficients, allowing generalization for unknown objects. The tests are conducted on a setup composed by a KUKA LWR 4+ Arm and a SCHUNK 5-Finger Hand, using the Xsens MVN suite to acquire the demonstrations.