PhysCoRe: New AI model predicts deformable object physics with 2x accuracy

Predicting how deformable objects behave under robotic manipulation remains a core challenge in robotics — soft materials like dough, cloth, or sponges don't follow rigid-body physics. Existing methods either require slow per-object optimization to estimate material parameters (and can't generalize) or use end-to-end learned models that extrapolate poorly and often violate physics. PhysCoRe, introduced by Yin, Tao, Chen, and Gan in a July 2026 arXiv paper, solves this by coupling a differentiable Material Point Method (MPM) simulator with two feed-forward neural networks: one infers material properties from visual observations (Material from Motion, MfM), and the other learns and corrects systematic biases in the simulator (Residual from Dynamics, RfD).

This hybrid approach allows PhysCoRe to adapt online to novel objects: MfM infers per-particle elasticity from limited visual interactions, and its predictive uncertainty actively guides further exploration toward regions where the estimate is least confident. In experiments on real deformable-object manipulation sequences, PhysCoRe significantly outperformed state-of-the-art baselines in prediction accuracy. The model also outputs a reliable per-region confidence distribution across the object's geometry, providing a natural signal for future active learning or safe manipulation planning.