Imitation Learning Study for Robotic Peg-in-hole Assembly

Imitation Learning is a class of algorithms with which a robot acquires movement skills by learning demonstration trajectories. In response to the difficulty of modeling the uncertainty of multiple demonstration results by Dynamical Movement Primitives (DMP), this paper introduces Kernelized Movement Primitives (KMP) to fit multiple demonstration trajectories. The main steps are as follows: Firstly, the optical motion capture system is used to acquire multiple demonstration trajectories. Secondly, the KMP to fit the numerous demonstration trajectories obtained at each demonstration point into one demonstration trajectory. Thirdly, the DMP is used to extract the feature parameters of the multiple demonstration trajectories. Finally, the extracted feature parameters are used to reconstruct the feature motion traj ectories under the new task parameters. The experiments of peg-in-hole assembly on the KUKA-IIWA robot show that the method can quickly and effectively generalize the new motion trajectory to complete the assembly task under the new task scenario.