Research On The Intelligent Grasping For Dexterous Hand

Intelligent grasping is an important technology in natural human-computer interac-tion,which simulates human behavior to complete interactions in the virtual world and increase user's immersion and participation.How to generate natural,precise,and real-time interaction processes has always been an important research area in computer graph-ics,human-computer interaction,and robotics.Grasping interaction based on dexterous hand is the research trend of intelligent grasping.In the process of grasp planning and interaction based on dexterous hand,there are problems such as the high complexity of hand model,difficult simulation of interaction details,time-consuming computation,and the strong dependence of the input device,which brings significant challenges for natural human-computer interaction.Considering about the above difficulties,this thesis divides the intelligent grasping according to the human grasp process into four parts: Firstly,the human can estimate the pose of the hand before grasping the object.Then,the human plans a set of feasible grasp poses and executes the grasping process.After grasping the object,the human can manip-ulate the object on the palm to change the posture of the object.Finally,the human can transfer grasping experience to similar objects.The main problems involved in the above process,including penetration depth computation,grasp planning and interaction,in-hand manipulation planning,and grasp transfer,are studied separately.The specific ideas are:1.Computing approximate contact space for the articulated model in configuration space,which can be used to compute penetration depth and reduce the computation space of grasp configurations? 2.Integrating the approximate contact space,optimization method,and the physical rules of grasping to compute grasp space and manipulation planning? 3.Con-sidering the surface mapping between similar models to implement grasp space transfer.The main contributions of this thesis include:(1)We propose an algorithm for computing the global penetration depth for the artic- ulated model.To solve the problem that the collision space cannot be accurately computed in the high-DOF configuration space,we adopt the support vector ma-chine method to compute the approximate collision space in the configuration space.It performs real-time computing by transforming the optimization problem into a query problem.This method not only can handle the articulated model with hy-brid joints but also guarantee that the configuration used to compute the penetration depth is collision-free.This method also can handle articulated models with a high degree of freedom,such as applied to dexterous hand grasp planning and simulation.(2)We present a grasp planning and interaction algorithm based on dexterous hand using learned grasp space method.For grasp planning,we use the support vector machine to compute the approximate contact space and reduce the search area in the configuration space.Then we adopt the particle swarm optimization algorithm and combine with the constraints of grasp stability and collision-free to compute grasp spaces.For object models with different complexity,we can generate a lot of grasp configurations.At runtime,we rely on the pre-computed stable grasp con-figurations,and non-penetration constraints along with motion planning techniques to compute plausible looking grasps.We have integrated our grasping algorithm with virtual reality and evaluated its performance for different tasks corresponding to grasping virtual objects and placing them at arbitrary locations.(3)We propose an in-hand manipulation planning using dexterous hand based on ma-nipulation graph construction and exhaustive precomputation method.Our main technique is to exhaustively compute a primary object motion graph and secondary manipulation graph to actively explore physically feasible states of the hand and the object.For a given two objects poses,this method uses an adaptive operational planning algorithm,combined with kinematics and dynamic constraints,to achieve operational path planning.To reduce the motion discontinuity caused by the com-bination of different branch motions during the graph expansion,we reduce the cost between the motion transitions.Based on the manipulation graph,we can imple-ment in-hand manipulation planning and animation generation.(4)We present a grasp space transfer algorithm for dexterous hand based on bijective contact mapping and local replanning.For grasping planning,there are problems such as high computational cost,and a similar object needs to compute grasp poses repeatedly.To solve the above problems,we project contact points into similar ob-jects by constructing bijective contact mapping and then compute grasp poses using the local replanning method.We implement grasp transfer between similar objects.Through the integration of object parts,we can even obtain grasp configurations for the new objects,which can be used to increase the diversity of the data.These methods solve the key problems in interaction from the penetration depth com-putation for articulated model,grasp planning and interaction using dexterous hand,in-hand manipulation planning and grasp space transfer.This thesis integrates these methods into a virtual reality interactive system and a good interaction effect is achieved.