Pre-Training of a Large Robotic Design Model

Abstract

Robotic design is a complex, multiparametric, and nonlinear process characterized by the intricate mapping between design requirements and solutions. Traditional methods often face limitations due to sequential workflows and human-induced biases, while conventional artificial intelligence models struggle to generalize across diverse design tasks. To address these challenges, we propose a novel cross-modal pretraining framework: robotic language-graph pretraining (R-CLGP). This framework bridges unstructured natural language requirements with structured robot designs, leveraging large-scale datasets for pretraining and flexible adaptation to various design requirements. The R-CLGP model utilizes a graph-based representation method that captures both non-Euclidean and Euclidean features and contrastive learning to enhance the mapping of textual requirements to robot topologies, significantly improving design efficiency and enabling intuitive design interaction. Through use cases such as requirement–topology retrieval, topology–topology retrieval, and performance prediction, the framework demonstrates its ability to streamline robotic design by minimizing manual intervention and improving scalability. This work not only advances methodologies in robotic design but also offers a transformative and adaptable framework for broader applications in automation driven by artificial intelligence.