SGPD: Stable Grasp Pose Detection in Complex Environment

In an unstructured environment, there are problems such as mutual occlusion between objects, and the inability to accurately obtain object category, quality, and shape information, which bring huge challenges to stable grasping. To this end, this paper proposes the stable grasp pose detection (SGPD) method based on Gaussian clustering. This method can establish the sampling distribution function of grasping points based on the geometric characteristics of the object, thereby ensuring that more sampling points are obtained at the center of the object, and solving the problem of poor grasping stability caused by unreasonable sampling in existing grasp pose detection methods. Besides, this method uses a convolutional network model to score candidate poses, and obtains the pose with the highest grabbing success rate and executes it. In addition, SGPD method uses an iterative closest point registration method to fuse multi-frame point clouds to solve the problem of poor quality candidate poses caused by incomplete observation input. In conclusion, the SGPD method is capable of handling challenges such as incomplete observations, unknown object labels and shapes, and has excellent generalization performance to new objects. Finally, this paper conducts random grasping experiments for unknown objects in complex scenes, the probability that the SGPD method retrieves the optimal grasp pose is 89.74% and the grasping success rate is 79.48%, which is better than the random sampling methods proposed in existing research.