Enhancing Robotic Grasp Failure Prediction Using A Pre-hoc Explainability Framework^*

摘要

Enhancing the explainability of machine learning (ML) models is crucial for bridging their use gap for practical applications of robotic autonomy. The common standard for adding explainability has been to use a post-hoc explanation approach. However, post-hoc approaches have recently attracted criticism for their lack of transparency, specifically because the post-hoc methods learn a surrogate model after the predictive model has already been learned, which means that the explanations are not authentically explaining the model’s behavior. This study aims to add explainability to the task of robot grasping failure prediction using Factorization Machines as the predictive model and using a novel pre-hoc explainability framework to learn an explainable Factorization Machine model. Unlike post-hoc methods, pre-hoc explainability starts with learning the explainable model before training the black-box model and then provides guidance while learning the latter to make predictions that are faithful to the explanations through a regularization mechanism. Through a detailed case study, we explore the trade-off between prediction accuracy and explanation fidelity and show that our framework is able to make predictions that are more accurate than an explainable white-box model while simultaneously learning a model whose pre-hoc explanations achieve a high level of fidelity relative to the predictions. Results show that our framework can predict the robustness of the grasp with 83% accuracy while explaining that increased effort exerted in Joint 2 of Finger 3 contributes tremendously to producing grasp failure, which is in contrast to increased efforts exerted at Joint 2 of Fingers 1 and 2 and at joint 1 of Finger 3, that all lead to reducing grasp failure.