Grasp Quality Evaluation Network for Surface-to-Surface Contacts in Point Clouds

The problem of grasping is essential in robotics, in which robotic hands grasp target objects by deploying contact force. The contact between robotic hand and target object is always reduced to point-to-point while generally the actual type should be surface-to-surface. In this paper, we propose a novel surface contact model to parameterize the contact area. It computes the grasp quality of the contacts between the robotic hand and target object, then determines whether or not the surface-contact-based end effectors can resist an external wrench on target object. The key idea of the proposed approach is that the contact modeling is based on surface-to-surface contact rather than equivalent contact points which is more in line with the actual situations. Then we propose a grasp quality evaluation network based on surface-to-surface contact to evaluate grasp motion, which can capture the geometric feature of the actual contact area and classify the quality level of input grasp. Experimental results and comparisons with state-of-the-art methods (e.g. GPD and PointNetGPD) demonstrate that our approach could achieve superior performance.