RGB-d image-based real-time 6-dof grasp pose estimation network for robots

Abstract A lightweight 6-DoF grasp-pose estimation framework guided by RGB-D images is presented for robotic manipulation in unstructured settings. The pipeline first employs a Key-Region Grasp Guidance Model (KRGGM) to generate heat-maps that highlight candidate grasp regions and provide grid-based 2-D pose priors. These regions are refined by a Local Point-Cloud Grasp Network (LPCGN) that predicts precise 6-DoF poses, while a Local Geometric Attention module achieves efficient feature fusion from 2D image features to 3D point clouds. Moreover, extensive experiments on the dataset indicate that our approach achieves 3-4 times faster inference on industrial edge devices compared to existing methods while reducing training time by more than 50%, and maintains competitive accuracy. Real-world experiments on a robotic platform further validate our method’s effectiveness, achieving a 94.7% average grasp success rate in single-object scenarios and a 92% scene completion rate in cluttered environments.