A Novel Dynamic Movement Primitives for Learning from Demonstrations

Dynamic Movement Primitives (DMPs) represent one of the most commonly employed frameworks for Learning from Demonstration (LfD). Nonetheless, a noteworthy drawback of DMPs pertains to their limited ability to learn from a sole demonstration, thus possibly constraining their practical utility in actual robotic scenarios. In this research, a modified DMP approach is proposed whereby the spring and damping term in DMPs are replaced by a pre-set system possessing intrinsic char-acteristics of converging towards the target, while the force term in DMPs is modeled through employment of the kernel trick. Subsequently, the modified DMPs are experimentally validated via simulation experiments and a writing task executed with the Franka Emika robot. The experimental findings affirm that the proposed algorithm not only facilitates learning from multiple demonstrations but also allows for adaptability in traversing via-ooints while preserving the primary motion features.