Learning Rope Manipulation Policies Using Dense Object Descriptors Trained on Synthetic Depth Data

摘要: Robotic manipulation of deformable 1D objects such as ropes, cables, and hoses is challenging due to the lack high-fidelity analytic models large configuration spaces. Furthermore, learning end-to-end policies directly from images physical interaction requires significant time on a robot can fail generalize across tasks. We address these challenges using interpretable deep visual representations for rope, extending recent work dense object descriptors manipulation. This facilitates design transferable geometric built top learned representations, decoupling reasoning control. present an approach that learns point-pair correspondences between initial goal rope configurations, which implicitly encodes structure, entirely in simulation synthetic depth images. demonstrate representation -- (DDODs) be used manipulate real into variety different arrangements either by demonstrations or policies. In 50 trials knot-tying task with ABB YuMi Robot, system achieves 66% success rate previously unseen configurations. See this https URL supplementary material videos.