cs.ROSep 24, 2026
SaveTRACE: Interactive Bi-Directional Tracing of Monochrome Cables Amid Clutter
Organizations: AUTOLab at the University of California, Berkeley
Abstract
Accurate state estimation (tracing) of Deformable Linear Objects (DLOs) such as cables is a critical challenge for data centers, manufacturing, construction, homes, and surgery, where precise cable management directly impacts operational safety and efficiency. However, resolving the state of multiple monochrome cables amid foreground and background clutter poses challenges due to occlusions, overlap, and ambiguous crossings. We present Two-way Routing And Cable Estimation (TRACE), which combines bi-directional cable tracing with interactive perception primitives-Divergence Push and Cluster Dilation-to actively resolve ambiguities. Evaluation with 110 physical experiments suggests that TRACE can increase the percentage of cable length correctly traced in complex scenarios (with up to 4 cables and 40 crossings) from ~60% with the strongest prior method, HANDLOOM 2.0, to ~90%, outperforming RT-DLO, Nano Banana Pro, and ChatGPT 5.2 as well. For a trial run on a workstation with an NVIDIA GeForce RTX 4090 GPU, the average computation time is 0.4 seconds per cable. Project website: https://trace-paper.github.io/.
Figures & tables
Fig. 2 : Overview of the TRACE architecture using the MANIP [ 23 ] framework. The system first processes the overhead RGB image to detect occlusions by comparing the TRACE-predicted cable state with object masks. If an occlusion is detected, a foreground object decluttering primitive is executed. Then, divergence points are identified via bi-directional tracing and classified as tangential crossings or cable clusters, activating the corresponding interactive perception primitive.
Fig. 3 : (Left): the Cable Distance Transform (CDT) , similar to a Voronoi diagram, maps each pixel value to the Euclidean distance between the pixel and the nearest cable (white). Ridgelines (red) are from the CDT using a Hessian-based filtering algorithm and highlight potential paths for robot trajectory planning while minimizing unintended cable interaction. Divergence points (circled) indicate ambiguous intersections, resolved through targeted push primitives. (Right): a Divergence Push primitive starts along a ridgeline at one end and moves through the divergence point to attempt separation at tangential crossings.
Fig. 4 : TRACE’s bi-directional tracing that identifies divergence points on 2 cables. (Left): cable traces initialized starting from 2 connectors of the same cable. (Right): overlapping cable traces are shown in orange; a divergence point in blue arises when the traces diverge into two non-overlapping paths.
Fig. 5 : Four examples; each column represents a different tier. The first row shows the initial scene with cluttered cables and external objects, the second row displays the initial predicted traces with occlusions, the third row shows the completed traces after object decluttering and interactive perception primitives.
| Tier 1 | Tier 2 | Tier 3 | Tier 4 | |
|---|---|---|---|---|
| # Cables | 2 | 2 | 3 | 4 |
| Avg # Crossings | 12 | 15 | 30 | 40 |
| # Tangential Crossings | 2 | 3 | 3-4 | 4-5 |
| # Foreground Objects | 3-4 | 3-4 | 3-4 | 3-4 |
TABLE I : Tiers of complexity for evaluation
| Primitive | Tier 1 | Tier 2 | Tier 3 | Tier 4 |
|---|---|---|---|---|
| Bimanual Decl. | 1.19 (36.5) | 1.53 (34.8) | 1.33 (19.2) | 1.27 (16.1) |
| Cluster Dilation | 0.19 (5.8) | 0.53 (12.1) | 1.00 (14.4) | 2.80 (35.6) |
| Divergence Push | 1.88 (57.6) | 2.33 (53.1) | 4.60 (66.4) | 3.80 (48.3) |
| Total | 3.25 (100) | 4.40 (100) | 6.93 (100) | 7.87 (100) |
TABLE II : Average number of interactive perception primitives used per trial (over 60 trials), reported as count (% of total). Cluster dilation is used more as scene context increases.
| Tier 1 | Tier 2 | Tier 3 | Tier 4 | |
|---|---|---|---|---|
| TRACE (No Clutter) | 99.1% | 91.2% | 94.2% | 89.4% |
| TRACE (Background Clutter) | 78.7% | 77.3% | 66.2% | 65.3% |
TABLE III : Average percent of cable length correctly traced (20 trials).
Fig. 6 : (Left) Without foreground clutter—Average percent of cable length correctly traced comparing TRACE and HANDLOOM 2.0 for 50 physical experiments. Tracing results for HANDLOOM 2.0 were sourced from MANIP [ 23 ] . (Right) With foreground clutter—Average percent of cable length correctly traced using TRACE with bimanual decluttering for 60 physical experiments. Vertical bars represent one standard deviation of the mean.
Fig. 7 : Colored cables are relatively easy to trace in contrast to monochrome cables: TRACE execution with background clutter, with colored cables (Top) and white cables (Bottom). (Top Left) Initial scene with colored cables and background clutter. (Top Right) Color-thresholded segmentation masks correctly trace 100% of cables. (Bottom Left) The initial configuration of white cables placed on a visually complex tabletop with distractor textures and objects. (Bottom Right) The reconstructed cable traces after bi-directional tracing and the application of interactive perception primitives, demonstrating robustness to spurious edges and background variability.
Fig. 8 : Baseline method RT-DLO performs poorly even on non-cluttered backgrounds, where cables are visually unambiguous. (a) On a cropped region, RT-DLO correctly traces only 48.2% of the total cable length while TRACE traces 100.0%; the red arrow marks the divergence point at which RT-DLO’s trace fails. (b) On a full cable scene, RT-DLO recovers only a fraction of the cable length: (left) original image, (center) RT-DLO trace overlaid on the original, and (right) the isolated RT-DLO result.
| Method | Computation Time | Initial | Final | Improvement |
|---|---|---|---|---|
| RT-DLO | 0.05s | 57.1% | 76.0% | 18.9% |
| TRACE | 0.40s | 68.3% | 97.5% | 29.2% |
| Increase | 19.6% | 28.3% |
TABLE IV : Average percentage of cable length correctly traced in small image crops, comparing TRACE and RT-DLO over 10 trials without background clutter.
Fig. 9 : Nano Banana Pro outputs on complex cable scenes, showing fragmented and topologically inconsistent traces compared to TRACE. (Left) Initial scene. (Right) Nano Banana Pro output. Note the hallucinated green cable appearing from the bottom of the right USB hub.
| Method | Initial | Final |
|---|---|---|
| Nano Banana Pro | 37.5% | 34.8% |
| ChatGPT 5.2 | 19.5% | 26.8% |
| TRACE | 40.2% | 89.3% |
TABLE V : Average percentage of cable length correctly trace in full-scene images with foreground clutter, comparing TRACE to Nano Banana Pro and ChatGPT 5.2. Initial and Final correspond to images taken before and after the interactive perception primitives.
Fig. 10 : TRACE on 2 scenes with increased cable count (6 cables, left; 8 cables, right). Cables densely cover nearly the entire workspace, leaving minimal visible background and creating highly occluded scenes.