Physics-Informed Data-Driven Control of Nonlinear Polynomial Systems with Noisy Data

Abstract

This work addresses the critical challenge of guaranteeing safety for complex dynamical systems where precise mathematical models are uncertain and data measurements are corrupted by noise. We develop a physics-informed, direct data-driven framework for synthesizing robust safety controllers (R-SCs) for both discrete- and continuous-time nonlinear polynomial systems that are subject to unknown-but-bounded disturbances. To do so, we introduce a notion of safety through robust control barrier certificates (R-CBCs), which ensure avoidance of (potentially multiple) unsafe regions, offering a less conservative alternative to existing methods based on robust invariant sets. Our core innovation lies in integrating the fundamental physical principles with observed noisy data which drastically reduces data requirements, enabling robust safety analysis with significantly shorter trajectories, compared to purely data-driven methods. To achieve this, the proposed synthesis procedure is formulated as a sum-of-squares (SOS) optimization program that systematically designs the R-CBC and its associated R-SC by leveraging both collected data and underlying physical laws. The efficacy of our framework is demonstrated on four benchmark systems, three discrete-time and one continuous-time nonlinear polynomial systems, confirming its ability to offer robust safety guarantees with reduced data demands.