Optimizing Local Explainability in Robotic Grasp Failure Prediction

This paper presents a local explainability mechanism for robotic grasp failure prediction that enhances machine learning transparency at the instance level. Building upon pre hoc explainability concepts, we develop a neighborhood-based optimization approach that leverages the Jensen–Shannon divergence to ensure fidelity between predictor and explainer models at a local level. Unlike traditional post hoc methods such as LIME, our local in-training explainability framework directly optimizes the predictor model during training, then fine-tunes the pre-trained explainer for each test instance within its local neighborhood. Experiments with Shadow’s Smart Grasping System demonstrate that our approach maintains black-box-level prediction accuracy while providing faithful local explanations with significantly improved point fidelity, neighborhood fidelity, and stability compared to LIME. In addition, our approach addresses the critical need for transparent and reliable grasp failure prediction systems by providing explanations consistent with the model’s local behavior, thereby enhancing trust in autonomous robotic grasping systems. Our analysis also shows that the proposed framework generates explanations more efficiently, requiring substantially less computational time than post hoc methods. Through a detailed examination of neighborhood size effects and explanation quality, we further demonstrate how users can select appropriate local neighborhoods to balance explanation quality and computational cost.