| Robot grasping is an important and long-standing research problem in robotics.Datadriven grasping methods have received more and more attention in recent years due to their stronger generalization performance and the ability to grasp unseen objects.In term of current data-driven grasping algorithms,the state representation for grasping is either object-centric representation or the images output by vision sensors.These representations focus on describing the features of obj ects,but under-characterize the gripper features and the interaction between the gripper and the object.Therefore,these methods are difficult to perform in a dexterous hand with a high degree of freedom,and will also fail to produce rich grasps for the same objects.We innovatively use interaction geometry as grasp representation to solve the above problems.Interaction geometry can selectively capture object information and gripper information,as well as their local spatial interaction information.Specifically,we introduce the Interactive Bisector Surface(IBS),which is the voronoi diagram between two close-by 3D geometric objects,as a novel grasping representation.We propose a deep reinforcement learning method based on the above representation to learn a joint planning policy which can resolve the problem of high-DOF reaching-and-grasping.Using the method proposed in this paper,the dexterous hand can not only grasp a variety of objects,but also grasp the same object in a variety of ways.The main innovations of this paper are as follows:1.A Sampling-Based Interactive bisector Surface Computation Method:In order to make the computing speed of interactive bisector surface meet the requirements of real-time applications,we propose a novel sampling-based computing method.The experimental results show that this new computing method can reduce the time cost while ensuring the quality of surface.2.A real-time grasp planning method for dexterous hand based on reinforcement learning and interactive bisector surfaces:In order to overcome the problem of insufficient characterization of the gripper and the interaction information between the gripper and the object in the previous methods,We use an interactive bisector surface(IBS)as the grasp representation.We have also improved existing reinforcement learning algorithms by using vectorized reward functions to balance different goals and using imperfect examples to speed up training.Experiments demonstrate that our method can generate high-quality grasps and smooth motion trajectories for complex shapes. |
