Learning From Noisy Data: A Direct Data‐Driven Disturbance‐Observer‐Based Controller

Disturbance observers (DOBs) are widely employed in control engineering for real‐time estimation and compensation of external disturbances and model uncertainties. However, most existing DOB designs rely on an explicit plant model, which may be unavailable or inaccurate in practice. To overcome this limitation, we propose a novel end‐to‐end data‐driven framework that synthesizes a controller directly from open‐loop data, uniquely eliminating the need for system identification. A primary contribution lies the derivation of a sufficient upper bound on the noise norm to guarantee closed‐loop input‐to‐state stability (ISS), thereby providing a quantitative criterion for data preprocessing. Moreover, the impact of noisy data on observer performance is characterized, and computable convergence bounds for the closed‐loop trajectories are established under realistic data conditions. The effectiveness of the proposed approach is validated through numerical simulations and experiments on a robotic manipulator platform, demonstrating superior disturbance rejection performance compared to classical data‐driven control methods.