Support-Aware Geometric Reasoning for Ordered Robotic Grasping in Stacked-Tabletop Scenes

Ordered grasping in stacked-tabletop scenes requires a robot to infer both local grasp poses and object-support order. This paper presents an annotation-guided geometric support-reasoning framework based on multi-view red–green–blue–depth (RGB-D) point clouds. Twelve preset views are registered using point-pair-feature matching followed by iterative closest-point refinement. In the reported experiments, manually supplied 2D labels and bounding boxes are back-projected into the fused cloud and refined by normal-based region growing. Each object is represented by an oriented bounding box (OBB). Candidate contacts are filtered by OBB collision tests, and candidate pairwise support directions are inferred by comparing downward-normal scores on the detected contact faces; this score rule is a geometric heuristic and is not a force-equilibrium, friction-cone, or center-of-mass stability proof. A directed graph of these candidate relations then determines the robot removal order. The reported segmentation macro-average is 86.91% mean intersection over union (MIoU) under manually supplied instance boxes. The historical robot-trial summaries contain 53 trials recorded as successful out of 90 for template matching and 75 out of 90 for the proposed pipeline, a descriptive difference of 24.4 percentage points. The historical recorded-success summaries favor the proposed pipeline configuration in the tested scenes; however, they provide only a descriptive contrast between complete pipeline configurations and do not establish accurate directed support reasoning, isolate the support rule’s contribution, or establish end-to-end autonomous perception.