Prediction of Robot Execution Failures Using Neural Networks

Abstract

In recent years, the industrial robotic systems are designed with abilities to adapt and to learn in a structured or unstructured environment. They are able to predict and to react to the undesirable and uncontrollable disturbances which frequently interfere in mission accomplishment. In order to prevent system failure and/or unwanted robot behaviour, various techniques have been addressed. In this study, a novel approach based on the neural networks (NNs) is employed for prediction of robot execution failures. The training and testing dataset used in the experiment consists of forces and torques memorized immediately after the real robot failed in assignment execution. Two types of networks are utilized in order to find best prediction method - recurrent NNs and feedforward NNs. Moreover, we investigated 24 neural architectures implemented in Matlab software package. The experimental results confirm that this approach can be successfully applied to the failures prediction problem, and that the NNs outperform other artificial intelligence techniques in this domain. To further validate a novel method, real world experiments are conducted on a Khepera II mobile robot in an indoor structured environment. The obtained results for trajectory tracking problem proved usefulness and the applicability of the proposed solution.