Bridging the Gap in Grasp Quality Evaluation and Grasp Planning

摘要

Robot grasp planning has been extensively studied in the last decades often consisting of two different stages determining where to grasp an object and measuring the quality of a tentative grasp. Additionally, because these two processes are computationally demanding, form closure grasps are more widely used in practice than force closure grasps, even though the latter is, in many cases, preferable. In this dissertation, we introduce our framework to improve grasp quality evaluation by increasing the speed of evaluating a grasp and developing more informative metrics. Specifically, we accelerate the computation of the grasp wrench space, used to measure the grasp quality, by exploiting some geometric insights in the computation of a convex hull through identifying a cutoff sequence to terminate the convex hull calculation with guaranteed convergence to the quality measure. Furthermore, we go into detail about the metric improvement for the grasp quality. Specifically, we study how noise at each joint of the manipulator affects grasp quality and how different arm configurations will generate different noise distributions at the end-effector, which has a huge impact in the robustness of grasping. Moreover, we illustrate our method that evaluates arm configurations based on the probability of achieving a force closure grasp. Then we introduce our work taking into account the hand structure and the local geometry of the object to be grasped as the second aspect for improving grasp quality metrics. In particular, for concave objects, we exploit the fact that grasping the concave region can make the grasp more robust. These insights are explored through theory and then validated on an experimental platform. Finally, we present three grasp planners we developed. We constructed two planners taking advantage of the negative curvature feature. The first the planner uses the geometry model of the object and constructs a database for online use. The second planner does not require the model but instead, detects negative curvature features on the fly and calculates candidate grasps in real time. Lastly, our third grasp planner searches through the objects' surface, represented as a triangular mesh, and tries to find the global optimal grasp.