Event

PartialBiGrasp: Dual-arm Grasp Generation from Partial Point Clouds

1 reports · 1 sources

Researchers including Ayush Kaura propose PartialBiGrasp, a dual-arm grasp generation framework that directly takes partial point cloud observations as input and no longer requires complete point clouds. The framework uses a convolutional occupancy network to implicitly learn geometric features, inferring graspability, collision-free contact regions, and object thickness, based on which it generates grasp pairs satisfying force closure, and uses sampling-based optimization to correct ambiguities caused by incomplete geometry. The authors state that previous dual-arm grasping methods assume complete point clouds can be obtained, which may not be available in real scenarios; for large objects, graspable regions are often sparse and depend on local geometry such as thickness, edge structure, and gripper clearance. Evaluation uses analytical force-closure metrics, large-scale simulation experiments, and tests on a real robot with noisy partial point clouds of unfamiliar objects.