Inverse Dynamics Modeling Of Robot Based On Sparse Spectrum Gaussian Process Regression

Inverse dynamics models are widely used in robotics and related fields.In order to establish accurate inverse dynamics model of robot,various modeling methods are emerging.In addition to the traditional modeling method based on dynamic principle,the method based on data driven modeling has gradually entered people’s field of vision.Gaussian Process Regression(GPR)is a data-driven modeling method with high accuracy for small data sets.However,as the computational complexity of GPR model increases with the increase of training samples,it is not suitable for large-scale data sets.In order to reduce the computational complexity,this dissertation adopts the Sparse Spectrum Gaussian Process Regression(SSGPR)method to establish the offline model and online model of the inverse dynamics of the robot,and the SSGPR algorithm is optimized.The model training and prediction speed are improved comprehensively,and the model updating function of online learning is realized.The main contents of this dissertation are as follows:(1)Aiming at the problem of excessive computational complexity of Gaussian Process Regression in processing large-scale training data,this dissertation proposes to establish inverse dynamics model of robot by Sparse Spectral Gaussian Process Regression method.In this method,the spectrum of Gaussian process is sparsed by the feature mapping of finite dimension so as to improve the speed of model training and prediction.Genetic algorithm was used to optimize the initial hyperparameters and the prediction error of the model was reduced effectively.(2)In order to improve the prediction accuracy of sparse spectral Gaussian process regression method,this dissertation proposes a local SSGPR modeling method based on Self-Organizing Map(SOM)neural network: Through SOM network,the training data were mapped to the local training subsets corresponding to the winning neurons in the output layer,and the corresponding local SSGPR model was established and trained.Then the local subset of the prediction sample is calculated,and the prediction results are obtained according to the corresponding local SSGPR model.Experimental results show that the proposed method has high prediction efficiency and accuracy.(3)Based on the dynamic characteristics of the robot will change with the change of the time and environment problems,put forward the modeling method based on local SSGPR online: online SOM network is adopted to establish the local model online,according to the new sample,the corresponding weights of each local subset of neurons to update online,in order to more accurately predict the robot joint torque.(4)Aiming at the problem that gaussian process regression retrains the model according to the newly acquired samples,which leads to too much computational complexity,the local model uses the method of recursive update: Through Residual Sum of Squares(RSS)criterion to select the new sample and delete the old sample,can not only control the size of the local model within a certain range,also can ensure the updated model can adapt to changes in the robot dynamics;Then,the updated local SSGPR model is obtained by recursive calculation based on the results of the original model,which reduces the computational burden of training the new sample set.Experimental results show that the model prediction performance of the proposed method is better than that of other modeling methods.In this dissertation,the inverse dynamics model is evaluated on two datasets:Sarcos and Sarcos＿inv.Experimental results show that the proposed inverse dynamic modeling method can achieve high prediction accuracy and efficiency.