Modified Local Gaussian Process Regression for Inverse Dynamics Learning

Robots have been playing an important roles on our life. In various field such as entertainment, military, space and medical fields, the precise control is required according to the increase of the importance on robot. In robot manipulator position control problem, modeling robot inverse dynamics is important because it can allow accurate robot control using computed torque control and PD control with computed feedback. However, modeling rigid-body inverse dynamics is not simple and not accurate in some case, because of unmodeled nonlinearities such as hydraulic cable dynamics, complex friction or actuator dynamics. Instead of rigid-body dynamics, nonparametric regression such as Locally Weighted Projection Regression (LWPR), Gaussian Process Regression (GPR) is proposed as alternative. Locally Weighted Projection Regression is fast, but it is difficult to tune because of many user-parameters. Gaussian Process Regression has high accuracy but low computation speed. In other word, high complexity of computation is drawback of Gaussian process regression. In Gaussian Process Regression, the inverse of Gram matrix is a significant problem and it dominates the computation time. To improve the low computation speed, there are many methods such as approximation method and Local Gaussian Process Regression (LGPR). In approximation method, the approximation of inverse of Gram matrix is proposed and in local Gaussian Process Regression, the training data is divided into local training data using Gaussian kernel. It generates M local models. After partitioned the training data, the local model is trained. When test data is given, each local model predicts the local prediction. The total prediction value is weighted average of M local prediction values. The weight is a similarity measure and it can be calculated by Gaussian kernel. Local Gaussian Process Regression In this paper, Modified Local Gaussian Process Regression (MLGPR) is suggested for improving accuracy and computation time of Local Gaussian Process Regression. Modified Local Gaussian Process Regression is used adaptive method for partitioning the training data. Modified Local Gaussian Process Regression uses multiple model generation threshold w(gen) values depending on the local target variance. Proposed method is demonstrated by 2-dimension regression example and learning inverse dynamics of SARCOS arm. The result of simulations will be compared with other method such as Gaussian Process Regression, Local Gaussian Process Regression. As a result, the accuracy of Modified Local Gaussian Process Regression is improved and computation cost is reduced. The result represent Modified Local Gaussian Process Regression has low computation cost as compared with Local Gaussian Process Regression.