Morphological-Symmetry-Equivariant Heterogeneous Graph Neural Network for Robotic Dynamics Learning

Fengze Xie, Sizhe Wei, Yue Song, Yisong Yue, Lu Gan

Keywords: Neural-symbolic learning, Morphological symmetry, Graph neural network

TL;DR: This work propose MS-HGNN as a general and efficient deep learning model for robotic dynamics learning, that integrates both robotic kinematic structures and morphological symmetries.

Abstract: We present a morphological-symmetry-equivariant heterogeneous graph neural network (MS-HGNN) for robotic dynamics learning. MS-HGNN unifies robotic kinematic structures and morphological symmetries within a single graph-based neural-symbolic architecture. By embedding these structural priors as symbolic constraints in the network design, MS-HGNN achieves strong generalization, high sample efficiency, and compact model complexity. This neural-symbolic integration enables the model to reason over the physical structure of multi-body dynamic systems while retaining the flexibility of data-driven learning. We formally prove the morphological-symmetry-equivariant property of MS-HGNN and empirically validate its effectiveness on a range of quadruped robot learning tasks using both real-world and simulated datasets. Code is publicly available at https://github.com/lunarlab-gatech/MorphSym-HGNN/.

Submission Number: 34