AssistDLO: Assistive Teleoperation for Deformable Linear Object Manipulation

Figure 2: Overview of the proposed DLO trace extraction pipeline from an RGB view captured from a single wrist camera. (a) Original image I_{t} . (b) Refined rope segmentation M_{t} shown as a green overlay, with sparse contour samples C_{t} (blue dots). (c) Voronoi diagram computed from the contour samples (cells between edges rendered with low transparency with different colors) to reveal candidate medial-axis structures within the mask. (d) Resulting vertices (yellow dots) in the pruned graph G_{t} after removing vertices outside the rope mask, yielding a clean set of rope trace points for

Figure 9: Task success rates categorized by user expertise (Left: Naive, Right: Expert). Bar heights represent the mean success rate across trials, and error bars indicate the 95% confidence intervals.

Figure 10: Task Completion Times categorized by user expertise (Left: Naive, Right: Expert), computed only over successful trials. Bar heights represent the mean TCT across trials, and error bars indicate \pm 1 standard deviation.

Figure 4: Shared Autonomy via Linear Blending (SA-LB). (a) Top view and (b) Front view of the grasping sequence. The U-shaped grippers represent the task space poses of the end-effector, where the solid black gripper denotes the human input ( \mathbf{u}_{h}^{j} ) at time t=0 , and the red gripper indicates the autonomous grasping target ( \mathbf{u}_{a}^{j} ), which is kept constant during the approach for visualization. The gray border outlines the table workspace, and the green dots represent the perceived DLO point set \mathcal{R}_{t} . To clearly illustrate the system’s behavior, we fixed

Figure 6: A sequential four-frame progression of an overhand knot untangling task. The manipulation is executed via our bimanual teleoperation setup and captured from the perspective of the fixed central workspace camera ( C_{m} ). Multiple demonstrations of the task are provided in the supplementary material.

Figure 1: System overview of AssistDLO . Dual wrist-mounted RGB-D cameras ( C_{l} and C_{r} ) continuously estimate the DLO state \mathcal{R}_{t} in real time. By combining human input with \mathcal{R}_{t} to estimate the operator’s intention, AssistDLO provides two forms of support: Visual Assistance (VA) to augment perception of the human operator through visual cues \mathcal{R}_{t} , and Action Assistance via our novel control barrier function (CBF)-based shared autonomy controller ( SA-CBF ) for geometry-aware DLO grasping.

Figure 3: Representative samples from the labeled dataset, featuring four distinct rope textures and properties and 41 images annotated with positive point prompts (indicated with green plus markers).

Figure 5: Shared Autonomy via Control Barrier Functions (SA-CBF). (a) Top view and (b) Front view of the grasping sequence. Visual notations for the human input ( \mathbf{u}_{h}^{j} ), autonomous target ( \mathbf{u}_{a}^{j} ), and environment remain consistent with the Figure 4 . Subjected to the exact same fixed human input trajectory (a direct path from \mathbf{u}_{h}^{j} to \mathbf{u}_{a}^{j} ), the wireframe mesh visualizes the dynamic height-field safety barrier z_{b} , constructed using spatial funnels centered on the rope’s perception points \mathcal{R}_{t} . The resulting trajectory de

Figure 12: Task Completion Time (TCT) by Rope Type. The longest DLO (Blue) required significantly more manipulation time than the shorter baseline (Green, p=0.0001 ), clearly demonstrating the severe impact of object length on teleoperation complexity. (Left: Naive, Right: Expert) Bar heights represent the mean TCT across trials, and error bars indicate \pm 1 standard deviation.

Figure 8: To allow for a direct comparison of physical properties (length and bending behavior), (a) displays all four ropes stacked one on top of the other in a relaxed position (stretched tightly at both ends and laid on a flat surface) and (b) explains our measurements for DLO bending behavior using a large-deflection cantilever test. While the figure displays the measured horizontal projection distances ( b_{\text{proj}} ) for each rope, these values are mapped to the Heavy Elastica boundary value problem to estimate the effective flexural rigidities.

Figure 11: Subjective rope metrics, as defined in Table I . Left (R-EXP): Compared with the short, compliant Green rope, both the high-rigidity Orange rope and the much longer Blue rope were perceived as behaving less as expected during the trial. (Higher is better) Right (R-PROP): Participants reported that the physical properties of the Blue and Orange ropes made the task harder than those of the Green and Red ropes. (Lower is better)