CRAGE and GNN-based AI architecture for autonomous transitions in modular robots

Abstract

The Moonshot R&D project, “Self-evolving AI robot system for lunar exploration and human outpost construction,” aims to realize AI modular robots that integrate advanced physical capabilities with self-evolving AI learning. Modular robots face significant challenges in achieving optimal and rapid shape transitions due to the vast number of possible configurations. To tackle this challenge, we first developed a Graph Neural Network (GNN) designed to learn abstract structural features inherent to modular robots—such as functional equivalence and geometric symmetry. To enable effective training of such a GNN, we developed a novel isomorphism determination method called Canonical Robot Adaptive Graph Encode (CRAGE), which converts structural graphs into canonical strings for efficient and consistent comparison. This graph-to-string encoding makes it possible to identify structurally equivalent configurations with high accuracy and speed. Moreover, CRAGE allows for the rapid and scalable generation of high-quality training data for robot structures, isomorphism classification, and deformation path planning—providing essential input for the GNN. Experiments show that CRAGE achieves 100% accuracy in small- to medium-scale structures while reducing processing time by over 90% compared to conventional methods. GNNs trained on CRAGE-generated data accurately predicted Lv2 structural equivalence from Lv1 inputs, achieving up to 100% classification accuracy and demonstrating strong generalization across structural variations. Together, CRAGE and the GNN form a unified framework for fast and scalable transition planning, contributing to the realization of autonomous self-evolving modular robots.