Contact-centric deformation learning

We propose a novel method to machine-learn highly detailed, nonlinear contact deformations for real-time dynamic simulation. depart from previous deformation-learning strategies, and model in contact-centric manner. This strategy shows excellent generalization with respect the object's configuration space, it allows simple accurate learning. complement learning two additional key ingredients: continuous vector field of deformations, instead discrete approximation; sparsifying mapping between deformations. These ingredients further contribute accuracy, efficiency, method. integrate our learning-based deformation subspace dynamics, showing simulations fine detail.