Probabilistic Modeling of Reaction Force/Torque through Data Transformation and Entropy Analysis

Abstract

In this study, we propose a method for improving the recognition performance of a probabilistic model through entropy analysis after transforming the time-varying reaction force/torque (F/T) signals. To conduct a task, it is important for a robot to recognize the reaction forces/torques from physical interactions with objects or the environment. The reaction force/torque signals measured using an F/T sensor contain a large number of noise components owing to the sensitivity of the sensor. Therefore, the recognition performance depends on how the noise components included in the training and test datasets are dealt with. For this purpose, the reaction force/torque signals are transformed from time-domain signals to noise-reduced and/or noise-robust features through transformation techniques. Herein, we apply three different transformation techniques: fast Fourier transform, discrete wavelet transform, and moment transform. Next, taskrelevant features are selected from all these transformed features based on entropy analysis, after which the features are used to learn a hidden Markov model. To evaluate our proposed method, several robot manipulation tasks (approaching, transferring, and positioning) are conducted using an open dataset with the reaction force/torque signals.