Abstract
Non-rigid objects are present everywhere in our daily life as well as industrial contexts. Despite the increasing need for automating their manipulation, there is to date no generally applicable and easily implementable method to control the shape of such objects. Indeed, the high number of Degrees of Freedom (DOF) of soft objects makes it difficult both to track and control the way they deform during manipulation.Uncalibrated visual servoing, a method where the model mapping visual data to robot control is estimated online, has been developped and widely used for years. Yet, the early works implementing visual servoing usually focus only on rigid objects. More recently, researchers have designed shape servoing controllers, relying on model estimation, neural networks or reinforcement learning to compute the command.Although these methods present good results, there are drawbacks - the need for prior knowledge of the object (shape, material parameters), specific models (cables, fabrics), huge datasets and high training time - all limiting the range of application or increasing the difficulty of implementation.This thesis aims to combine visual servoing, data-based and model-based methods with the goal of shaping soft objects with a dual-arm robot. We first present a data-based control framework to shape objects into a desired 3D contour, using cooperative tasks frames to coordinate both arms. This work does not require any knowledge of the manipulated object, but is limited to shaping a contour.With this in mind, we then propose tools to build geometrical and simple mechanical models while making use of the dual-arm robot setup. We implement physics simulations relating the deformation of the obtained models with the real-time displacements of the end-effectors.Finally, we use real-time simulations combined with visual feedback to feed the control framework to shape objects in their full volume. Our framework operates with little knowledge of the manipulated object, allowing a larger range of applications, and the use of simulations remedies the difficulties of getting visual data for control computation. Throughout the developments, we conduct experimental validations with the robot BAZAR composed of two KUKA arms and different soft objects, and a single static RGB-D camera.