Robot perception and manipulation of deformable linear objects

Robot manipulation of Deformable Linear Objects (DLOs) spans applications in industry, healthcare, and daily life. Long-term DLO manipulation is challenging due to complex object dynamics and logical reasoning requirements. A task common in both everyday life and manufacturing is Shoe Lacing (SL). This thesis focuses on the perception and manipulation of DLO s under the task of SL. First, this thesis approaches SL is by intuitively decomposing it into primitives and planning with symbolic planners. The task is structured across three levels: high-level optimisation of SL patterns, mid-level task planning, and low-level execution. Experiments are conducted using different shoes and conditions to evaluate system performance. This thesis proposes a simulation environment and benchmarking framework for bimanual SL systems. The environment supports controlled training and testing, while the framework offers evaluation metrics and task protocols to ensure performance comparison and repeatability. Baseline results demonstrate the effectiveness of the framework and the environment. Existing tracking algorithms overlook interactions with scene objects. This thesis introduces a method that tracks shoelaces while preserving topological relations with the shoe. Two evaluation metrics are proposed: sequential H-Signature, representing topological state, and Dynamic Time Warping (DTW) distance, quantifying tracking error. The algorithm consistently yields correct signatures and achieves errors comparable to state-of-the-art methods. Learning from Demonstration (LfD) offers an efficient way to learn robot skills. While existing work often focuses on learning policies, few consider the operator’s role. This thesis presents a bimanual teleoperation system and analyses how first- and third-person perspectives affect completion time, trajectory diversity, and policy quality. Overall, this thesis advances Deformable Object (DO) manipulation by addressing the full pipeline of robot SL, including task planning, perception, simulation, and primitive learning. These developments offer foundational tools and insights for robust, generalisable DLO manipulation in real-world settings.