Region-Aware Grasping for Stacked Workpieces: A 6D-Wise Label Self-Generation Method and Robust Evaluation Strategy

The high-quality datasets and generalized network model combined with robust evaluation strategies serve as a significant benchmark for developing new policies for industrial bin-picking. In this paper, we propose the concept of region-aware grasping, a sim2real cutting-edge system to generate and evaluate 6D poses for robots to pick up novel workpieces from stacked environments. It consists of Region-Aware-Dataset, a large-scale synthetic point cloud grasp dataset; and Semantic-Graspnet, a 6D-wise affordance policy that predicts full 6D grasp pose for stacked workpieces. The introduced Semantic-Graspnet transforms the 6D pose prediction problem into semantic categorization via point cloud encoding and decoding. Meanwhile, we propose a robust evaluation strategy based on pose evaluation and mechanical grasping evaluation, which enhances the robot’s grasping success rate and sorting efficiency. In real industrial tasks, the robot achieves a grasp completion rate of 91.3% in cluttered scenes and 89.2% in densely stacked scenes, demonstrating state-of-the-art results in robotic picking-and-placing applications.