A Parallel Graph Network for Generating 7-Dof Model-Free Grasps in Unstructured Scenes Using Point Cloud

Generating model-free grasps in complex scattered scenes remains a challenging task. Most current methods adopt PointNet++ as the backbone to extract structural features, while the relative associations of geometry are underexplored, leading to non-optimal grasp prediction results. In this work, a parallelized graph-based pipeline is developed to solve the 7-DoF grasp pose generation problem with point cloud as input. Using the non-textured information of the grasping scene, the proposed pipeline simultaneously performs feature embedding and grasping location focusing in two branches, avoiding the mutual influence of the two learning processes. In the feature learning branch, the geometric features of the whole scene will be fully learned. In the location focusing branch, the high-value grasping locations on the surface of objects will be strategically selected. Using the learned graph features at these locations, the pipeline will eventually output refined grasping directions and widths in conjunction with local spatial features. To strengthen the positional features in the grasping problem, a graph convolution operator based on the positional attention mechanism is designed, and a graph residual network based on this operator is applied in two branches. The above pipeline abstracts the grasping location selection task from the main process of grasp generation, which lowers the learning difficulty while avoiding the performance degradation problem of deep graph networks. The established pipeline is evaluated on the GraspNet-1Billion dataset, demonstrating much better performance and stronger generalization capabilities than the benchmark approach. In robotic bin-picking experiments, the proposed method can effectively understand scattered grasping scenarios and grasp multiple types of unknown objects with a high success rate.