Teaching robots real physics by learning what simulators miss

Simulators assume materials are perfectly uniform, but real objects have hidden irregularities that break this assumption. MoSA starts with a standard physics simulator and then learns additional stress patterns that capture these residual effects—mild warping, uneven material properties, localized stiffness variations. By treating these corrections as learnable layers guided by motion constraints, the method achieves better accuracy and generalization than pure neural networks while staying interpretable. Robot manipulation experiments show the improved simulator trains controllers that transfer to real hardware more reliably.