COMBINATION OF AFFINE DEFORMATION AND DYNAMIC MOVEMENT PRIMITIVE IN LEARNING HUMAN MOTION FOR REDUNDANT MANIPULATOR

LfD(Learning from Demonstration) has the advantage of requiring no expert knowledge about the robot itself, which make large-scale application of robots possible. In this paper, we present a novel approach based on combination of affine deformation and DMP(dynamic movement primitive) to deal with two fundamental problems in LfD: data gathering and policy deriving. In the proposed approach, demonstrated motion data are gathered from optical-based motion capture system and DMP is used to sketch feature and derive policy. Combined with affine deformation, joint trajectory derived from the control policy can be refined so that the manipulator's physical constraints satisfied, the end point accuracy preserved and the execution time optimized. We verify the feasibility of our approach by reproducing a series of different motions with various trajectory profiles on a humanoid robot's arm basing on limited human demonstrations.