AURORA: Autonomous Updating of ROM and Controller via Recursive Adaptation
Jiachen Li, Shihao Li, Dongmei Chen
TL;DR
AURORA introduces a five-agent, LLM-enabled framework for autonomous design and online updating of ROM-based controllers for high-dimensional nonlinear systems, integrating a shared Code Agent for iterative generation and validation. The approach addresses the ROM-control co-design problem by enabling autonomous method selection, online diagnostics, and adaptive updates, validated across eight benchmarks with five LLMs. GPT-5 demonstrates the strongest performance, achieving high success and autonomy, while the framework identifies ROM construction as a key bottleneck and underlines the need for robust evaluation and real-world integration. Overall, AURORA represents a practical path toward fully autonomous, online ROM-based control design with formal stability considerations and modular, traceable workflows.
Abstract
Real-time model-based control of high-dimensional nonlinear systems faces computational intractability, while traditional reduced-order model (ROM) control requires manual expert tuning without online adaptation. We propose AURORA (\textbf{A}utonomous \textbf{U}pdating of \textbf{RO}M and Controller via \textbf{R}ecursive \textbf{A}daptation), a multi-agent LLM framework automating ROM-based controller design with online adaptation. AURORA employs five specialized agents collaborating through iterative generation-judge-revision cycles, with an Evaluation Agent diagnosing degradation sources and routing corrections appropriately. Validated on eight benchmark systems spanning mechanical assemblies, thermal PDEs, and robots. Comparative evaluation across five state-of-the-art LLMs demonstrates high autonomy with minimal intervention, establishing practical viability for autonomous control design.