Research On Migration Framework Of Manipulators Based On Neural Network Dynamics Modeling

With the rapid development of artificial intelligence,robot technology has gradually entered human life,and its related scientific research has also been rapid development.Although the types and models of robots are increasingly rich and their mechanical structures are becoming more and more complex,most of the traditional control methods are only suitable for specific and structured application scenarios.Moreover,they often require a complex modeling process,which makes the methods have great limitations in control accuracy,execution efficiency,system robustness and individual generalization,so that they can not meet human diverse control requirements for robots.Because the end-to-end training of neural network technology has high efficiency and accuracy,and can avoid individual differences in the trained system.Therefore,the deep neural network technology is used to realize the dynamics modeling of the manipulator,the physical constraints are introduced and the model-based robust controller is designed in this thesis.On this basis,a multi-platform migration control framework based on the constraint model is constructed to further improve the control generalization of the method for different robot platforms.First of all,the thesis briefly describes the research basis of the self-made 7-DOF humanoid robot platform,including the forward and inverse kinematics modeling analysis of its arm structure.Then,the Lagrangian dynamics equation is derived from the perspective of energy,and a parametric neural network is constructed based on it as the prior knowledge.Dynamics modeling is carried out in the joint space of generalized series manipulators,which lays a solid theoretical foundation for subsequent research.Secondly,aiming at the interference of physical factors in the control process of manipulators,physical constraints including internal friction constraint and external contact force constraint are introduced on the basis of inverse dynamics to improve the control accuracy and anti-interference ability of our method.Then,the joint data are collected from the physical manipulator to train the inverse dynamics neural network with constraints.The essence of this method is to solve the current positively correlated with the joint torques into the actual joint torques,so that the data acquisition process does not need the torque sensor to save the technical cost.Next,aiming at the system parameter disturbance in the training process of the dynamic network,a model-based robust controller is designed in this thesis to eliminate the uncertainty error generated in the training process of dynamic parameters,and the Lyapunov stability theorem proves that the controlled system can achieve a variety of stable states under its control.Then,the constraint on the upper bound of the parameter uncertainty is loosened in the model-based robust control law,so that the robust control effect of the underlying joint can be displayed on different manipulator platforms.After that,based on the characteristics of end-to-end training,the acceleration-level joint configuration adjustment technology is introduced to propose an efficient multi-platform dynamic modeling method,which provides a model basis for the general effect of the robust controller.Finally,the thesis proposes a multi-platform migration control framework based on the constrained dynamics model by using the characteristics of end-to-end training,which includes joint physical constraints and general robust control effects.At the same time,the general potential energy function is introduced,which includes four kinds of high generalization tasks based on parameter properties,realizing the technical transition from the bottom joint control to complex task decomposition planning.Then,the additional joint technology is introduced in the simulation environment to realize the migration of the control framework to the redundant manipulator platform.Experiments are designed to further verify the joint control accuracy of the method,as well as adaptability and generalization for different platforms and different tasks.