A Self-Supervised Learning-Based 6-DOF Grasp Planning Method for Manipulator

To realize a robust robotic grasping system for unknown objects in an unstructured environment, large amounts of grasp data and 3D model the object are required; sizes these directly affect rate successful grasps. reduce time cost acquisition labeling increase grasps, we developed self-supervised learning mechanism to control tasks performed by manipulators. First, manipulator automatically collects point cloud from multiple perspectives efficiency acquisition. The complete is obtained using hand-eye vision truncated signed distance function algorithm. Then, used generate series six-degrees-of-freedom poses, force-closure decision algorithm add quality label each pose automatic data. Finally, gripper closing area corresponding train grasp-quality classification manipulator. results performing actual experiments demonstrate that proposed method can grasps Note Practitioners—Most existing planning methods based on public datasets or simulation algorithms. Owing limited types objects, amount datasets, lack real sensor noise data, robustness trained insufficient, it difficult apply production environments. solve above problems, propose 6-DOF capture introduce problem scenes. perspectives, performs desktop-level reconstruction, finally uses order scenario. Preliminary show this obtain high-quality be applied operations multi-target cluttered However, has not been tested This paper focuses module framework. In future research, will design more efficient improve