Multi-Fidelity Robotic Co-Design via the Refinement Limit

Gordei Verbii, Kirill Bunin, Ilya StarikovL2P @ RSS 2026 · OpenReview

Abstract

Translating robot co-design from the lab to production demands representations that support both rapid design iteration under tight timelines \emph{and} trustworthy high-fidelity simulation for contact-rich tasks such as packaging, cable routing, and flexible assembly. Pipelines today face a dilemma: continuum simulators are accurate but too slow for design loops, while cheap discrete graph models (mass--spring, truss) may not \emph{converge} to the physics they approximate -- and a model that mis-predicts stiffness fails to transfer from simulation to the line. We supply the missing bridge. We formalize \emph{graph-like} robots (1-skeletons of nodes and edges) as discretizations of free-form/continuum bodies at scale $h$, equip the hierarchy with prolongation operators, and establish: (i) geometry and the Laplace--Beltrami spectrum converge at measured rates; (ii) the discrete elastic energy $\Gamma$-converges to a continuum beam/plate limit \emph{if and only if} refinement is all-dimensional and renormalized -- otherwise the effective bending stiffness $EI_h$ diverges or over-stiffens without bound. Where the limit exists, coarse optima prolong to near-optimal fine designs and warm-started multigrid topology optimization converges; scale-invariant controllers transfer zero-shot across resolution. These guarantees turn a naive ``factor-of-two'' refinement -- a source of sim2real error -- into a \emph{certifiable} multi-fidelity pipeline: cheap discrete co-design with a static fidelity certificate before committing to continuum-accurate validation.
BibTeX
@inproceedings{verbii2026multi,
  author = {Gordei Verbii and Kirill Bunin and Ilya Starikov},
  title = {{Multi-Fidelity Robotic Co-Design via the Refinement Limit}},
  booktitle = {L2P @ RSS},
  year = {2026},
  url = {https://openreview.net/forum?id=7MXofCEzNF},
}

Loading related content...

Figure