Motion optimization based on hierarchical iterative parameter learning for complicated trajectory

Abstract

PurposeThe purpose of this paper is to propose an accurate and practical imitation learning for robotics. The modified dynamic movement primitives (DMPs), global fitting DMPs (GLDMPs), is presented. Framework design, theoretical derivation and stability proof of GLDMPs are discussed in the paper.Design/methodology/approachBased on the DMPs, the hierarchical iterative parameter adaptive framework is developed as the hierarchical iteration stage of the GLDMPs to tune the designed parameters adaptively to extract richer features. Inspired by spatial transformations, the coupling analytical module which can be regarded as a reversible transformation is proposed to analyze the high-dimensional coupling information and transfer it to trajectory.FindingsWith the proposed framework and module, DMPs derive majority features of the demonstration and cope with three-dimensional rotations. Moreover, GLDMPs achieve favorable performance without specialized knowledge. The modified method has been demonstrated to be stable and convergent through inference.Originality/valueGLDMPs have an advantage in accuracy, adaptability and practicality for it is capable of adaptively computing parameters to extract richer features and handling variations in coupling information. With demonstration and simple parameter settings, GLDMPs can exhibit excellent and stable performance, accomplish learning and generalize in other regions. The proposed framework and module in the paper are useful for imitation learning in robotics and could be intuitive for similar imitation learning methods.