Robotic Perception and Manipulation of Deformable Linear Objects

At LAR, we are actively researching approaches to effectively perceive DLOs using deep learning methods, while also developing smart strategies to enable the automatic generation of training datasets. For the robotic manipulation, our focus is on combining both model-based and learning-based approaches, with the aim of addressing some of their individual limitations.

01

Survey and Tutorials

3 papers
Two robot arms holding a cable, with the survey's three areas highlighted: perception, modeling, and estimation, planning and control
Journal Survey · The International Journal of Robotics Research · 2026

Robotic perception and manipulation of deformable linear objects: A survey

Alessio Caporali, Ignacio Cuiral-Zueco, Gonzalo López-Nicolás, Gianluca Palli

Read the survey
Figure 1 of the tutorial: the shape-servoing loop, and two robot arms deforming a cable toward a target shape
Preprint Tutorial · two parts · 2026

A Tutorial on Shape Servoing of Deformable Objects

Ignacio Cuiral-Zueco, Alessio Caporali, Miguel Aranda, Gianluca Palli, Gonzalo López-Nicolás, Youcef Mezouar

Part I: Fundamentals Part II: Advanced Concepts
02

Vision-based Dataset Generation

4 papers
Semi-automatic labelling of cables: a VR tracker traces the cable, producing pixel-wise masks and labelled camera images

Deep-learning perception of deformable linear objects needs large labelled datasets, but annotating thin, deformable and often overlapping cables by hand is slow and error-prone. We develop strategies that generate training data with little manual effort: chroma-key acquisition with background replacement to obtain domain-independent images, semi-automatic labelling with a VR tracker that blends real and synthetic samples, dense annotations obtained from a few sparse inputs, and copy-and-paste augmentation for wiring-harness components.

Key papers

More papers

03

2D Perception

6 papers
Photos of cables on different backgrounds, including a switchgear cabinet, next to their instance segmentations

Cables and wires lack a distinctive shape, colour or texture, can take countless configurations and often cross each other, which makes them hard to segment for general-purpose detectors. We developed a family of instance-segmentation methods that combine deep networks for semantic segmentation with graph-based reasoning, separating each cable and describing it with a spline: Ariadne+ covers the binary mask with paths on a graph, FASTDLO solves the intersections with skeletonization and a similarity network at more than 20 FPS, and RT-DLO reasons on a graph of centre-line nodes in real time. DLO Perceiver segments the specific cable described by a text prompt, and we also estimate the orientation of the connector at the cable end from a single camera.

Key papers

More papers

04

3D Perception

3 papers
Real cables in a switchgear and on a desk, next to their 3D reconstructions

Grasping and manipulating cables requires knowing their shape in 3D. We estimate and track the 3D shape of deformable linear objects from multiple 2D views acquired by a compact camera mounted on the robot: key points extracted from each image are modelled with B-spline curves and matched across views with a multi-view stereo approach, reaching sub-millimetre accuracy, and multi-view model-based visual tracking closes the loop for the robotic manipulation of the cables.

Key papers

More papers

05

Manipulation

7 papers
A red cable on a green surface, shaped into four different forms in manipulation experiments

Manipulating cables is hard because of their unpredictable configuration, high-dimensional state and complex nonlinear dynamics. We combine model-based and learning-based approaches: neural-network models of the cable whose physical parameters are estimated online while the robot controls its shape, spline-based dynamic models, learning-based pipelines that reshape a cable in cluttered scenes, and shared-autonomy frameworks in which the robot assists a human operator.

Key papers

More papers

06

Robotic Assembly and Industrial Applications

15 papers
Two robot arms with grippers above a wiring harness

From perception and planning to complete robotic cells: wire and pin insertion in connectors with vision and tactile sensing, dual-arm systems for wiring-harness manufacturing and for handling several cables at once through graph-neural-network representations of their topology, switchgear cabling, and the manipulation and quality check of medical hoses.

Key papers

More papers

07

Cloth and Other Deformable Objects

3 papers

Several of our perception and manipulation methods extend beyond linear objects, to cloth and garments.

Papers