Extended DMPs Framework for Position and Decoupled Quaternion Learning and Generalization

Abstract Dynamic movement primitives (DMPs) as a robust and efficient framework has been studied widely for robot learning from demonstration. Classical DMPs mainly focuses on the in Cartesian or joint space, can't properly represent end-effector orientation. In this paper, we present an extended (EDMPs) both space 2-Dimensional (2D) sphere manifold Quaternion-based orientation generalization. Gaussian mixture model regression (GMM-GMR) are adopted initialization phase of EDMPs to handle multi-demonstrations obtain their mean covariance. Additionally, some evaluation indicators including reachability similarity defined characterize generalization abilities EDMPs. Finally, real-world experiment was conducted with human demonstrations, endpoint poses arm were recorded successfully transferred robot. The experimental results show that absolute errors Riemannian skills less than 3.5 mm 1.0°, respectively. Pearson’s correlation coefficients mostly greater 0.9. developed exhibits superior multi-space skills’ This research proposes fused GMM-GMR which sufficient capability multi-demonstrations.