Learning cooperative dynamic manipulation skills from human demonstration videos

This article proposes a method for learning and robotic replication of dynamic collaborative tasks from offline videos. The objective is to extend the concept demonstration (LfD) scenarios, benefiting widely available or easily producible To achieve this goal, we decode important information, such as Configuration Dependent Stiffness (CDS), which reveals contribution arm pose endpoint stiffness, three-dimensional human skeleton model. Next, through encoding CDS via Gaussian Mixture Model (GMM) decoding Regression (GMR), robot’s Cartesian impedance profile estimated replicated. We demonstrate proposed in sawing task with leader–follower structure, considering environmental constraints uncertainties. experimental setup includes two Panda robots, replicate roles profiles extracted two-persons video.