Jointly Learning Position and Orientation in Cartesian Space with Dynamic Movement Primitive

Dynamic movement primitive (DMP) is an effective method widely used for learning human manipulation. Position trajectory in human demonstration is relatively easy to learn, but orientation is difficult to model due to its inherent complexity. Also, the inconsistency in time scales between translation and rotation changes can also lead to a decrease in task performance. In this paper, a method jointly learning position and orientation in Cartesian space with dynamic movement primitives is proposed to learn human manipulation. A DMP based on screw displacement is established to model both translation and rotation as a whole, with dual quaternions applied for a concise representation. Furthermore, a newly designed delayed goal is introduced to make the demonstration trajectory more aligned with human operational habits, achieving better learning performance. And validations are put forward to illustrate its humanlike characteristics, consistency and generalization ability.