Learning Physics-Guided Residual Dynamics for Deformable Object Simulation

arXiv cs.AI · 18 hours ago

Researchers developed Physics-Guided Residual Dynamics, a hybrid simulation system that pairs a physics-based spring-mass model with a neural network trained to predict corrections to its predictions for deformable object dynamics. The system uses a sliding-window transformer architecture and velocity-based formulation to improve accuracy over purely physics-based or learning-based methods. The framework enables applications including manipulation planning with Model Predictive Control and interactive simulation through action-conditioned video prediction.