MoSA: Motion-constrained Stress Adaptation for Mitigating Real-to-Sim Gap in Continuum Dynamics

Continuum robots—such as soft manipulators, endoscopes, and growing robots—suffer from a severe real-to-sim gap because standard physical simulators assume homogeneous, isotropic materials. MoSA (Motion-constrained Stress Adaptation) closes this gap by learning a neural residual stress field that captures anisotropic material behavior that physics engines inherently miss. The method uses a neural network trained on visual observations of a deforming soft body. By comparing real deformation sequences with simulator predictions, the network learns to predict a residual stress field that, when added as a correction force in the simulator, brings its dynamics into close agreement with reality. Crucially, this approach does not require explicit calibration of material parameters such as Young's modulus or Poisson's ratio, which are difficult to measure for soft tissues and custom elastomers. Experiments are conducted on a cable-driven continuum manipulator performing trajectory tracking and contact-rich tasks. MoSA reduces trajectory tracking error by over 60 percent compared to uncorrected simulators and enables successful model-predictive control (MPC) on hardware that fails when using the baseline physics model. The learned residuals generalize across different cable tensions and external loads, suggesting that the network captures material-level structure rather than merely memorizing specific trajectories.