Abstract
While predicting robot grasps with parallel jaw grippers have been wellstudied and widely applied in robot manipulation tasks, the study on naturalhuman grasp generation with a multi-finger hand remains a very challengingproblem. In this paper, we propose to generate human grasps given a 3D objectin the world. Our key observation is that it is crucial to model theconsistency between the hand contact points and object contact regions. Thatis, we encourage the prior hand contact points to be close to the objectsurface and the object common contact regions to be touched by the hand at thesame time. Based on the hand-object contact consistency, we design novelobjectives in training the human grasp generation model and also a newself-supervised task which allows the grasp generation network to be adjustedeven during test time. Our experiments show significant improvement in humangrasp generation over state-of-the-art approaches by a large margin. Moreinterestingly, by optimizing the model during test time with theself-supervised task, it helps achieve larger gain on unseen and out-of-domainobjects. Project page: https://hwjiang1510.github.io/GraspTTA/