Robot failure mode prediction with deep learning sequence models

Detecting and preventing an impending robot grasp failure can prevent object damage during robotic manipulation. When a failure is predicted, a human can opt for teleoperation rather than automation of the grasping task. The operator can also intervene to stop the robot from adjusting, such as tightening the grasp or reducing the speed. In this paper, we propose new Machine Learning models for failure prediction based on recurrent neural networks, introducing two innovative approaches: Generative-interpretive grasp prediction (GIGP), which emphasizes both the predictive and interpretive aspects of the model, and adaptive time to failure analysis (ATFA), which also predicts the time to failure. GIGP, can perform an “early” failure prediction, which is not possible with existing models. Both GIGP and ATFA perform competitively or outperform state of the art methods while having the additional desired feature of being interpretable. We evaluate our methods on the problem of early failure prediction in robot grasping. We reach competitive results exceeding 80% in predictive accuracy and agenerative model and full sequence for discriminative modelrea under the ROC curve for the tasks of failure prediction after a limited number of observations of the grasper while being able to make interpretable predictions.