Genetic algorithms and unsupervised machine learning for predicting robotic manipulation failures for force-sensitive tasks

Abstract

Recent advances in the state-of-the-art force-torque sensors have improved the close-loop control of robotic manipulators. However, it is still challenging to perform a force-sensitive pick-and-place task, unless a considerable number of sensors monitor the process. The predictive capability of failures in conventional robotic object-sorting systems are limited. Using fifteen force-torque samples from the University of California-Irvine (UCI) database, we demonstrate the viability of failure prediction using an unsupervised Machine Learning (ML)-based method, whose learning parameters were optimised via Genetic Algorithms (GAs). This hybrid algorithm was deployed for discriminating between manipulation failure and successful object placement. GA was used to avoid overfitting or overtraining. The proposed model could detect robotic manipulation failures with 91.95% classification accuracy, thus improving on the performance of previous classification methods. This study validates the use of GAs and unsupervised ML to predict the extent of success for force-sensitive object placement using information on forces and torques alone.