Soft Robotic Sim2Real via Conditional Flow Matching

Modeling soft robots remains a significant challenge due to high computational costs and frequent mismatches with real-world behavior, a phenomenon known as the Sim2Real gap. This paper addresses the Sim2Real gap through conditional flow matching (CFM), which learns a mapping between the simulation domain and the real-world experimental domain. A neural network learns a conditional probability path that transforms simulated states into real-world observations, conditioned on control inputs, thereby minimizing simulation inaccuracies. The method is demonstrated through benchmark Sim2Sim and Sim2Real tensile tests, and additionally demonstrated in the domain of soft gripping using fin-ray grippers. A novel encoder architecture is introduced that learns a representation of the contact state, enabling the model to generalize to previously unseen interactions. The model provides a highly accurate prediction of force and deformation, successfully capturing complex elastic behaviors, including hysteresis and force fluctuations. Experimental results validate that CFM can bridge the Sim2Real gap for various soft robot morphologies, without requiring large datasets, and with strong generalization capabilities. Notably, the findings indicate substantial generalization capabilities in Sim2Real scenarios.