Vision-based three-finger grasp synthesis constrained by hand geometry

Learning of 2D grasping strategies from box-based 3D object approximations

“…For grasp hypotheses from 2D contours, we will use an algorithm that is closely related to the work of Morales [4]. This algorithm involves a four step procedure for finding a number of grasp hypotheses, followed by a fifth to disqualify unfeasible grasps and selecting the best of the hypotheses.…”

“…We use the notion of grasp region, as defined in [4] and assume that a good region for grasping is a region that is as straight as possible. The fact that studies have shown that slightly concave curvature may be better suited [15] is left as a possible extension to the work.…”

“…In the case of nonempty intersections, the centroid of the intersection area is calculated and projected back to the regions. These points will be used for finger positions, as discussed in [4].…”

“…[2,3], most techniques rely on 3D data. Due to the complexity of the task, much work has been done for simplifications of 3D shape, such as planar [4] or 3D-contour-based [5] representations. Other approaches involve modelling an object perfectly, i.e.…”

“…Secondly, our work makes use of an algorithm for finding and predicting the success of a grasp, but for planar objects, as proposed by Morales et al . [4,13]. The approach uses 2D image analysis to find contact point configurations that are valid given specific kinematic hand constraints.…”