Deep Dexterous Grasping of Novel Objects From a Single View

Dexterous grasping of a novel object given single view is an open problem. This paper makes several contributions to its solution. First, we present simulator for generating and testing dexterous grasps. Second data set, generated by this simulator, 2.4 million simulated grasps variations 294 base objects drawn from 20 categories. Third, basic architecture generation evaluation that may be trained in supervised manner. Fourth, three different evaluative architectures, employing ResNet-50 or VGG16 as their visual backbone. Fifth, train, evaluate seventeen variants generative-evaluative architectures on showing improvement 69.53% grasp success rate 90.49%. Finally, real robot implementation the four most promising variants, executing 196 total. We show our best architectural variant achieves 87.8% seen view, improving baseline 57.1%.