Reproduce “RAFL: Generalizable Sim-to-Real of Soft Robots with Residual Acceleration Field Learning” on OpenResearch

About this paper

Dong Heon Cho, Boyuan Chen

Researchers at Duke University developed RAFL, a framework that augments differentiable soft robot simulators with a generalizable, element-level residual acceleration field to bridge the sim-to-real gap. The system consistently improves simulation accuracy across varying robot morphologies and mesh discretizations, outperforming traditional system identification and global residual learning methods in zero-shot generalization settings.

Key insights

• Decoupling the learned correction from global mesh structure by focusing on element-level residual acceleration enables transferability across diverse soft robot morphologies and discretizations.
• Employing physically structured input features (e.g., principal stretches, co-rotated strain rates) and enforcing constraints like rotational equivariance and zero residual at rest enhances generalizability and physical plausibility.
• The learned residual forces induced by RAFL are distributed more uniformly throughout the body, reducing sensitivity to specific vertex locations compared to localized forces from global-state residual models.

Summary by alphaXiv.