Research On Data Driven Robot System Modeling Method

The current modeling methods are mainly mechanism modeling,knowledge modeling,and data-driven modeling methods.The accuracy of the model of mechanism modeling is high,but the modeling period is particularly long and difficult,and the parameters in the model are difficult to identify.The model form of the knowledge model modeling method is simple and easy to implement,but the accuracy is low and it is difficult to obtain knowledge.In addition,these two modeling methods are difficult to model when there are many degrees of freedom,the dynamic system itself or the environment changes,and the existing motion methods cannot meet the motion.For a dynamic system,a model built using input and output data that does not require process structure information,and can better describe a dynamic data-driven modeling method has higher research value.The main research work is as follows:Because the deep neural network has the ability to approximate arbitrary precision functions,two modeling methods based on the deep neural network direct modeling method and indirect numerical integration method are proposed.The former directly maps the current state/input to the new state.The latter approximates the functional relationship between the current state/input mapping and the state derivative,and then uses the numerical integration method to obtain the new state.For dynamic systems,the function form required in the latter is more concise.Deep neural network relies heavily on training samples.The quality of the sample data directly determines the accuracy of the model.The subject studies three sampling methods:random sampling,gridded sampling,and weighted gridding sampling.In order to facilitate research,a simulation environment for single pendulum and primate bionic robots was established based on the Gym platform.Experimental work was performed on the proposed method on the simulation platform to evaluate the accuracy and generalization ability of the two models.The results show that although such methods can establish effective data-driven dynamics models,they require a large number of samples and are difficult to apply in physical systems.Aiming at the shortcomings of deep neural network modeling methods such as long time,prone to overfitting,and network parameters that need to be repeatedly debugged,a nonparametric modeling method based on Gaussian process regression is proposed.This method requires almost no prior process knowledge and can obtain A measure of model uncertainty.However,the computational complexity of Gaussian process regression will increase as the data set increases.For some high-dimensional and massive data sets,the increase in calculation amount will seriously affect the performance of the algorithm and affect the real-time performance of the control system.Aiming at this problem,algorithms such as KD-tree,kernel interpolation scalable structured Gaussian process and Lanczos variance estimation are studied to improve the efficiency of Gaussian process modeling.Based on a primate bionic robot and a pendulum simulation experimental platform,the Gaussian process regression modeling method is tested,and the results show that it has strong accuracy and generalization ability.Adopting an improved Gaussian process regression modeling method can greatly improve the calculation efficiency while hardly affecting the accuracy.The purpose of modeling is for real-time control.The subject uses model predictive control algorithms for control research.First,for the simple pendulum system,the predictive control strategy based on the Gaussian process regression model is studied to make it stand upside down.For primate bionic robots,it takes a lot of time to pump energy into the system for the vertical position of the pendulum.A fast energy algorithm based on window search is studied,but this algorithm can only increase the system's energy quickly,and cannot reach the target position periodically.Cantilever grab was performed,so energy and angle were added to the objective function of the model prediction algorithm,and the results were verified on the simulation platform.A semi-physical experimental platform for primate bionic robots was built,and Gaussian process regression modeling was performed on real sample data.Based on this model,a physical experiment of cantilever motion of primate bionic robots was performed.The experimental results show the effectiveness of the algorithm.Finally,in order to solve the problem that the traditional modeling method is difficult to perform modeling control in the case of sudden changes in the environment of the robot,a data-driven Gaussian process regression modeling combined with a model predictive control algorithm was studied,and on the V-REP simulation platform,The emergency control is performed in the case of a wheel failure of the omnidirectional trolley,and the results show the effectiveness of the algorithm.