Transparent-Object GraspNet–Transformer for Flexible Assembly Tasks

Abstract

Transparent components are becoming increasingly popular in high-precision assembly; however, due to weak texture, unstable refraction and depth measurement, unreliable grasp localization, and potential propagation of pose errors during insertion. This paper develops a transparency-aware GraspNet-Transformer for flexible assembly. The model generates appearance markers based on RGB and polarization cues, generates geometric markers from confidence-completed depth, and fuses them through confidence-gated cross-attention. A grasp decoder proposes collision-aware 6-DOF grasp candidates, and an assembly head selects a grasp quality-compliant and compliant insertion-feasible grasp. A two-stage controller chooses an approach direction and Cartesian impedance according to visual confidence and wrist force. The 12,480 labelled scenes in the system included four transparent-part families and three backgrounds, and they were uniformly illuminated and uncluttered. It achieved an 89.6% grasp success rate and an 84.3% end-to-end assembly completion rate, outperforming the best tested baseline by 6.8% and 8.1%, respectively. The median position translation error was 4.7mm, and the 95th-percentile contact force had decreased from 24.8N to 17.2N. Ablation results show that confidence-gated fusion added 3.9 percentage points of grasp success and constraint coupling increased assembly completion by 3.1 points. The above results indicate that transparent object perception and task constraint compliance should be optimized together, rather than as separate modules, and provide a practical path for a reliable mixed-part production unit.