Real- Time Failure/Anomaly Prediction for Robot Motion Learning Based on Model Uncertainty Prediction

Abstract

End-to-end robot motion generation methods using deep learning have achieved various tasks. However, due to insufficient training or the occurrence of abnormal conditions, the model sometimes fails tasks unexpectedly. If failures/anomalies can be predicted before occurring, irreversible task failures can be prevented. In this paper, we propose a method of predicting model uncertainty to predict failures/anomalies in real-time. For a naive method, we used a model that predicts the robot's actions stochastically and also tried a method that predicts failure/anomaly on the basis of the variance. However, it was experimentally shown that the variance due to the variation of the training data and the uncertainty of the model cannot be distinguished. Therefore, by predicting the likelihood of the model, which corresponds to the degree of discrepancy between the model and observations, in real-time and treating it as the uncertainty of the model, we applied it to the prediction of failure/anomaly. The method's effectiveness was demonstrated by achieving a high judgment accuracy rate of 85% (17/20 cases) in an object-picking task.