Application of machine learning for isomorphic analysis of robotic wrist mechanisms

Abstract. Robotic wrist mechanisms are essential for achieving dexterity, precision, and orientation control in manipulators. Traditional isomorphic analysis, based on graph-theoretic tools such as connectivity matrices, articulation points, and Wiener numbers, is effective but limited by scalability, efficiency, and the need for manual interpretation. This paper proposes a machine learning (ML) framework that represents wrist topologies as graphs and applies Graph Neural Networks (GNNs) to identify structural equivalence. A benchmark dataset was created from Tsai’s atlas of spherical and non-spherical wrist mechanisms, augmented with synthetically generated topologies. The preliminary simulation results demonstrate that the GNN-based model achieves over 95% accuracy in classifying isomorphic and non-isomorphic designs, while reducing classification time by approximately 60% compared to the Wiener number method. This preliminary study relies heavily on a data set, which may misclassify symmetrical structures. With this, the future work aims to expand the data sets, conduct detailed computational comparisons, and utilise other learning approaches.