Hysteresis Compensation in Force/Torque Sensor based on Machine Learning

Abstract

This paper proposes a method to improve the accuracy of the force/torque (F/T) sensor based on machine learning considering time series data. There are several problems with F/T sensors, one of which is hysteresis. Hysteresis is a factor of error dependent on force history. There have been few researches focusing on hysteresis in an F/T sensor. We solved this problem by considering time series data. Time series data was put into machine learning such as linear regression and Support Vector Regression (SVR). We evaluated this method with an existing high dynamic range F/T sensor. We confirmed that the error decreased in both high and low force ranges. Since there is nonlinearity in hysteresis, we predicted that SVR will be more accurate than linear regression. Linear regression considering time series was better than SVR when loading training data at random intervals and loading test data at constant intervals.