Based On Reinforcement Learning Dexterous Hand Control Algorithm And Application

Dexterous hand operation,one of the most challenging robot control tasks,remains largely unsolved.In this paper,a complete control system is implemented based on the real Baxter robot,and the ability of practical is experimental demonstrated on the actual robot platform.It can effectively control the robot to reach the specified position and follow the given joint position trajectory.The main contribution of our work in this paper as following:Tracking Algorithm with Structured Constraints is designed for the auxiliary module of the control system.The visual tracking algorithm improves the tracking performance in robot graspping scenes.Tracking algorithm,as the auxiliary module of the control system,the location and shape of target provided for the control system at runtime.In our work,the cross-skeleton model and the soft segmentation model are proposed to develop the structural information and appearance information.Our method is tested in the visual tracking database compare with many excellent trackers,and got an excellent performance.The local controllers are constructed for the manipulation task.The task is divided into several simple states.The local controller is trained by the optimal control or reinforcement learning algorithm in this simple case.For example,LQR for follow joint trajectory will be train using the differential dynamic programming.After the local controllers has been trained,a global policy will be trained through the supervision learning technology.The global strategy ensures finish the execution of the entire manipulation task,whatever the initially station.In this paper,the third research is that the control system has implemented on the real robot in the real world.Control algorithm is often easier finished in simulation enviroment which can ignore a variety of interference factors,execution accuracy,unknown dynamics,cost of sampling.Finally,experiments were performed on the Baxter robot at the end of this paper.The local controller and global strategy will be tested respectively.The cost and error plot in the paper.