A generalizable neural operator for full-field deformation prediction in robotic design (March 2026)

This paper will included in the Advanced Engineering Informatics Journal, Volume 73, accessible via the following links https://www.sciencedirect.com/science/article/pii/S1474034626002922 and https://doi.org/10.1016/j.aei.2026.104600

Abstract

Robotic design and optimization often require extensive simulation-based iterations, particularly for predicting static deformations under varying postures, external loads, and gravitational conditions. Conventional finite element simulations can achieve high accuracy but are computationally expensive, limiting their use in large-scale design optimization. Data-driven approaches have emerged to accelerate prediction; however, most existing methods are restricted to simplified scenarios, such as fixed geometries with varying poses or fixed poses with varying loads, and typically predict only partial deformation at points rather than the full deformation field. To enable fast and generalizable deformation prediction, a new neural operator architecture named Robot-NO is developed. A geometric–load graph based on the finite element mesh connectivity, where each node encodes spatial coordinates with local load conditions is constructed. A graph neural network is then employed to extract a unified global representation that captures both geometric structure and external force distribution. A query–decoder module combines global features with arbitrary query-coordinates to predict the full deformation
field. This query-based formulation enables accurate continuous deformation prediction across varying robot geometries, postures, and load conditions. Experimental results demonstrate that the proposed model achieves a root mean square error of 0.91 μm in deformation prediction and is 6000 times faster than conventional finite element simulations. This significant improvement enables rapid design, analysis and optimization for robotic mechanisms.