Interactive Robotic Grasping in Cluttered Scenes via Multimodal Semantic-Spatial Fusion

Abstract

Robotic grasping in cluttered scenes requires both semantic target understanding and spatial safety reasoning. This paper presents SSCM-Grasp, an interactive grasping framework that combines language-guided semantic masks with depth-based affordance estimation. The semantic mask is used as a pixel-level constraint to filter non-target grasp candidates, while a depth-aware occlusion reasoning module estimates whether an obstacle should be removed before grasping the target. A closed-loop state machine further switches between obstacle removal and target grasping according to the spatial overlap and relative depth relationship. Simulation experiments in cluttered scenes show that SSCM-Grasp improves grasp robustness under heavy occlusion and achieves a 62% grasp success rate in 8-object scenes, with an obstacle clearance rate of 73%.