GMGP: Generalized Model for Grasp Planning of Vacuum and Parallel Jaw Grippers

Abstract The capability of robots to adapt to initially unmodeled tasks with minimum human intervention is important to achieve better reconfigurability in manufacturing processes. Using advanced vision systems and existing information like Computer Aided Design (CAD) models, there have been plenty of attempts in manufacturing to build an adaptable robotic system equipped with a certain type of end-effectors for specific manipulation tasks. However, such approaches still require experts to either manually pre-plan how robots should grasp each object or develop customized algorithms for grasp planning, which can be a bottleneck for better usability and scalability of a robotic system in manufacturing. To tackle the problem, this preliminary study presents GMGP, a simple but effective composite model that can be universally used to plan grasps for various vacuum grippers and parallel-jaw grippers. In order to produce diverse grasps for arbitrary combinations of the two most widely used end-effectors and objects efficiently, GMGP initially samples potential contact surfaces on an object and roughly checks their geometrical properties by projecting a deformable circular grid. Depending on the size and number of the surfaces that each end-effector requires, our model performs adaptive grouping of the contact surfaces to remove redundant grasp configurations. Then, for each end-effector configuration, GMGP quickly check if it causes potential collision between an object and a gripper using a Signed Distance Field (SDF). Lastly, the model computes grasp quality metrics for the refined subset of geometrically feasible grasp configurations to provides relative stability of the grasps. For evaluating performance of the model, we test all proposed grasp configurations from GMGP in physics-based simulations while assuming scenarios of offline planning prior to online grasping tasks. Specifically, the diverse grasp set proposed by the framework achieves 92.0% grasp success rate overall even if the combinations of two end-effector and 10 target objects are arbitrary.