Material-Aware First-Grasp Target Preselection for Mixed Rigid-Deformable Technical-Waste Mock-Ups

Robotic sorting of mixed rigid–deformable technical waste requires the selection of a graspable first target while minimizing disturbance to surrounding objects. This paper proposes a lightweight material-conditioned target-preselection method for cluttered RGB-D scenes. Object instances are detected using YOLO26n and segmented using SAM2-B, while their material categories are predicted using DPB-CNN. Depth-consistency filtering is subsequently applied to refine the candidate masks. Each candidate is evaluated according to geometric accessibility and local overlap, while visually inferred rigid-over-deformable relations are used to penalize or exclude deformable objects constrained by rigid-like objects. AnyGrasp generates a 6-DoF grasp pose using dense target points together with a downsampled workspace cloud retained for collision checking. Experiments were conducted on a UR5 platform using 40 predefined layouts covering four representative rigid–deformable interaction patterns, with three trials per method for each layout. The proposed method achieved a target selection accuracy (TSA) of 86.7%, a first-grasp success rate (FSR) of 80.8%, and a severe-disturbance rates (SDR) of 8.3%. After Holm correction, TSA was significantly higher than for both baselines, and SDR was significantly lower than for Native-AnyGrasp. The observed FSR improvements did not reach the corrected significance threshold. These results support improved first-grasp decision-making and reduced disturbance relative to Native-AnyGrasp in the evaluated technical-waste mock-up scenarios.