Deformable linear object routing with mounted brackets: fast-deployable DLO perception and occlusion-robust bracket pose estimation

Purpose The purpose of this study is to address deployment-oriented perception for single-branch deformable linear object (DLO) routing in cabinet-like industrial cells with pre-mounted arrow-shaped brackets, focusing on two practical bottlenecks: obtaining real-time, background-robust DLO masks after one-time initialization and estimating a graspable bracket pose under self-occlusion when the dominant residual ambiguity is the roll angle about the local DLO axis. Design/methodology/approach A deployment-oriented closed-loop pipeline is proposed: Grounded Segment Anything Model (GSAM, a zero-shot foundation model) is used only once to initialize the target DLO mask from a single RGB image of the scene; afterwards, a Gaussian mixture model (GMM)-based color-domain filter with lightweight post-processing (GMM-POST) performs real-time online segmentation with fixed parameters. The segmented mask is skeletonized to localize the bracket center, and a calibrated mapping from left-right visible area asymmetry is used to estimate the roll angle about the local DLO axis under top-view self-occlusion. The estimated grasp pose is integrated into sequential grasp-adjust-insert execution, forming a perception-to-action routing loop. Findings Mask IoU of GMM-POST exceeds 0.73 on simple textures DLO and 0.69 on complex textures DLO – outperforming RT-DLO and Grounded-EfficientSAM in accuracy while achieving real-time performance (34 FPS), a large speed improvement over GSAM’s 0.09 FPS. Maximal roll angle-estimation error is 3.2°, even when most of bracket is occluded. In the original cabinet-like setup, the end-to-end three-bracket routing task succeeds in 15 of 20 trials. Additionally, the authors conducted 30 end-to-end grasp-adjust-insert runs in an independent testbed with different robot/camera/DLO, achieving 23/30 successes, which further supports repeatability across setups and extends the evaluation to a softer-to-stiffer cable regime. Practical implications The proposed framework is readily transferable to other DLO, brackets and industrial scenarios without additional large amount of data-collection or new hardware. Originality/value This paper presents a deployment-oriented perception-and-routing pipeline for DLO with pre-mounted brackets. Its value lies in combining one-time single-image initialization with lightweight online perception, an interpretable roll-aware grasp-pose estimation module for self-occluded brackets and end-to-end validation across two physical setups with the same bracket type.