Robot Learning for Deformable Object Manipulation Tasks

Abstract: Deformable Object Manipulation (DOM) is a challenging problem in robotics. Until recently, there has been limited research on the subject, with most robotic manipulation methods being developed with rigid objects in mind. Part of the challenge in DOM is that non-rigid objects require algorithms capable of generalizing to changes in shape as well as different mechanical properties. Machine Learning (ML) has been shown successful in fields, such as computer vision and natural language processing, where generalization is important thus encouraging the application of ML to robotic manipulation. This thesis tackles DOM problems using ML techniques for tasks with Deformable Linear Objects (DLOs), e.g. ropes and cables, found in a variety of industrial applications. DLOs encapsulate a lot of the general challenges in DOM, making them good case studies on the effectiveness of ML for other types of deformable objects. Typically, ML algorithms require large amounts of data that are better satisfied in simulation. Therefore, the ReForm simulation sandbox is introduced, which includes six DLO manipulation tasks. ReForm aims to facilitate comparison and reproducibility of robot learning research on tasks where the goal is to control the shape of a DLO. Such shape control tasks are categorized as: explicit , if a precise shape is to be achieved; or implicit , if its deformation is dictated by a more abstract goal. Two representative DLO manipulation tasks are addressed: (i) shape-servoing (explicit) and (ii) cable-routing (implicit). For shape-servoing, special emphasis is given to Reinforcement Learning (RL) methods. Initial work tackles shape-servoing of an elastoplastic DLO towards a unique goal, using online RL with ReForm. Subsequent work moves towards a multi-goal task in a real-world experimental setup, using offline RL methods to learn directly from real data. In the cable-routing works, the aim is to lay the groundwork for solving this type of task through motion primitives, with limited use of ML. First, a vision-based approach is presented, which is able to route a cable through randomly placed fixtures. Then, a force-based approach is introduced for a similar problem, in which the state and stiffness of a DLO can be estimated through contact with fixtures.