A Study of Effective Prediction Methods of the State-Action Pair for Robot Control Using Online SVR

Prediction of future state and action In order to work effectively, a robot should be able to adapt to different environments by deciding its correct course of action according to the situation, using determinants other than pre-registered commands. For this purpose, the ability to predict the future state of a robot would be effective. On the other hand, the future state of a robot varies infinitely if it depends on its current action. Therefore, it is difficult to predict only the future state. Thus, it is important to simultaneously predict the state and the action that the robot will adopt. The purpose of this study was to investigate the prediction of the advanced future state and action of a robot. In this paper, the results of the study are reported and methods that allow a robot to decide its appropriate behavior quickly, according to the predicted future state are discussed. As an application example for evaluating the proposed method, the inverted pendulum model is used and the prediction results are compared with the robot’s actual responses. Then, two methods will be discussed for predicting the robot’s state and action. To perform state and action prediction, two methods are used, firstly the Online SVR (Support Vector Regression) and secondly Online SVR and the LQR (Linear Quadratic Regulator).