“…In contrast, we formulate this contact clustering as a datadriven machine learning process trained to match the experimental data given the dynamic information (i.e., pose, twist, and wrench) of all the contact points, thereby, significantly improving the contact simulation accuracy. We also adopt IN architecture [15] for this to accommodate vastly-varying contact point set for the two tasks due to their complex/non-convex geometries, that cannot be addressed with other typical learning architectures (e.g., MLP (Multi-Layer Perceptron [16])); • Constraint/energy-based robust contact solving: Given the clustered points (and normals), the process of con-tact solving computes the contact force for each point using relevant physics principles. For this, two main methodologies exist: penalty-based method (e.g., [17], [18]) and constraint-based method (e.g., [19], [20]), with the latter preferred when the accuracy is concerned, since the former, from its relying on virtual springs, typically cannot precisely enforce physical principles of contact (e.g., Signorini condition for no penetration [21],…”