One-shot Learning Closed-loop Manipulation of Soft Slender Objects Based on a Planar Polynomial Curvature Model

Many are the challenges that make robotic manipulation of deformable objects such a complex task. For example, to properly plan and execute a control action, a robot needs to understand how external forces will modify the deformation states of the object. Creating such an internal representation is even more complex in the typical situation where the robot is interacting for the first time with the object. In this paper, we look at this challenge when controlling the deformation states of a planar and slender object. Leveraging soft robots' modelling and control, we show that the only non-geometrical information needed to perform this task is the stiffness distribution. We thus propose a strategy to learn this function from a single interaction with the object, testing it experimentally. We then propose a closed-loop controller that exploits this learned information to perform the manipulation task and test it with simulations.