A Robust Pose Estimation Method for Robot Grasping in Bin-Picking Scenarios Using Point Cloud

Industrial robot grasping in bin-picking scenarios is challenging. This is mainly due to the need for robots to extract individual parts and find suitable grasping poses accurately and efficiently. This paper addresses this challenge by focusing on the complex morphology of injection-molded corner pieces and proposing a noise-robust pose detection model (NRP-Net) for suction-based grasping. We introduce a directional encoding module to enhance the perception of local structures. We also present an instance segmentation method based on differential features, which we integrate with pose space and visibility attention mechanisms to improve the accuracy of pose estimation. To ensure the correctness of the suction area, we design a sealing detection algorithm suitable for cluttered scenes. Validation in practical scenarios shows an 87.4% success rate in grasping. This demonstrates the effectiveness of our method in bin-picking scenarios and offers a viable solution for industrial robot grasping tasks.