Learning from demonstration using improved dynamic movement primitives

Abstract

It is important to endow the robot with the ability of learning the complex motion sequences and thus adopt such motions when facing to changeable environment. This paper proposes an improved Dynamic Movement Primitives (DMP) method. In order to solve the problem of invalidation of forcing term in conventional DMP, the improved DMP approach, i.e., the DMP together with Deep Neural Network (DNN), is proposed. Specially, DNN is introduced to fit the target nonlinear function with the demonstrated trajectory information, instead of using a specific formula to describe the forcing term in DMP. Thus, improved DMP method can avoid the drawback of conventional DMP. Simulation work is conducted and the results show that the invalidation performance of forcing term is improved compared with conventional DMP. In addition, the generalization property of improved DMP is also beneficial to work environmental adaptability.