Robotic Manipulation of Deformable Linear Objects : A Model-Free Control Using Reinforcement Learning Algorithms

This thesis focuses on the robotic manipulation of deformable linear objects (DLOs) in applications such as assembling deformable wire harnesses and cables in manufacturing. Addressing the limitations in existing studies, this research presents a comprehensive investigation into the perception of DLOs using both single-arm and dual-arm robots. DLOs pose challenges in both perception and manipulation for automated robotic systems due to their lack of distinctive features and intrinsic deformability. The research explores the modeling of DLOs and employs reinforcement learning techniques to tackle tasks such as unknotting, untangling, and shape control. The goal is to contribute to the understanding and application of reinforcement learning in solving challenges related to DLO manipulation. In this work we investigate a new observation method of DLOs and compare its efficiency in relative to the one that is popular in the litrature (images). Then also we see the significance of using different reinforcement learning methods with these observation methods. The use of an RL model like SAC to control a the DLO through a better action space that allows the agent to adabt easily with the environment. Also, the use of an observation space that is more reliable than the images while the robot is interacting with that environment. Finally we present a reliable reward function that can be used for training the agent on the tasks such as achieving a target configuration and a target orientation. The use of this reward function not only has a positive effect on the new observation space investigated here, but also on the one that is popular in the literature. iii