Learning Pixel-Wise Suction Grasp Representation for Cluttered Environments

Robotic vacuum grippers provide distinct advantages for handling objects with complex geometries or compliant materials. Existing data annotation methods for suction grasp planning typically sample object-level grasp labels and transfer them to scene-level annotation on sensor data (e.g., depth images or point clouds). However, this process suffers from sparse annotations due to random sampling and label domain shifts, degrading downstream model performance. To overcome these limitations, we propose a suction grasp evaluation framework that directly assesses grasp feasibility at every sensor pixel. Our approach introduces an orthographic ray projection module to sample pixel-aligned grasp poses, followed by a physics-informed metric to evaluate suction grasp quality. This pixel-aligned annotation pipeline ensures a direct bijective mapping between sensor pixels and grasp labels. Additionally, we present a modular robotic system that utilizes depth data to perform object-agnostic seal map prediction and RGB data for grasp pose refinement. Real-world robot experiments demonstrate that our pixel-wise annotations align well with practical scenarios, and the learned grasp planning model outperforms existing baselines.