Comparative Analysis of Post Hoc Explainable Methods for Robotic Grasp Failure Prediction

In human–robot collaborative environments, predicting and explaining robotic grasp failures is crucial for effective operation. While machine learning models can predict failures accurately, they often lack transparency, limiting their utility in critical applications. This paper presents a comparative analysis of three post hoc explanation methods—Tree-SHAP, LIME, and TreeInterpreter—for explaining grasp failure predictions from white-box and black-box models. Using a simulated robotic grasping dataset, we evaluate these methods based on their agreement in identifying important features, similarity in feature importance rankings, dependency on model type, and computational efficiency. Our findings reveal that Tree-SHAP and TreeInterpreter demonstrate stronger consistency with each other than with LIME, particularly for correctly predicted failures. The choice of ML model significantly affects explanation consistency, with simpler models yielding more agreement across methods. TreeInterpreter offers a substantial computational advantage, operating approximately 24 times faster than Tree-SHAP and over 2000 times faster than LIME for complex models. All methods consistently identify effort in joint 1 across fingers 1 and 3 as critical factors in grasp failures, aligning with mechanical design principles. These insights contribute to developing more transparent and reliable robotic grasping systems, enabling better human–robot collaboration through improved failure understanding and prevention.