14 October 2026 in SSRN Electronic Journal
Reliable semantic segmentation of cable-driven manipulators is essential for safe operation in complex tabletop environments; however, their slender structures, weak visual textures, and cluttered backgrounds often cause missed regions, discontinuous boundaries, and background confusion in RGB-only perception, while existing RGB-D fusion methods frequently introduce substantial model overhead without fully exploiting depth geometry. This article presents Geo Fusion, a lightweight geometry-guided module that assigns complementary roles to the two modalities and performs bidirectional spatial attention without expanding the feature-channel dimension: RGB semantics first suppress task-irrelevant depth responses, and the refined depth geometry then enhances boundary and spatial-structure representations in the RGB features. Experiments on NYU Depth V2, SUN RGB-D, and a real tabletop-manipulator dataset show that Geo Fusion achieves 39.14% and 40.92% mIoU in the lightweight framework and 45.86% mIoU in ESANet-R34 with only approximately 200 fusion parameters, while improving F1-score, Arm IoU, overall mIoU, and Boundary F1 on real scenes to 93.1%, 86.8%, 72.9%, and 84.0%, respectively. These results demonstrate that Geo Fusion provides an accurate and computationally efficient RGB-D perception solution for boundary-sensitive segmentation on resource-constrained robotic platforms.