Online adaptation of direct data-driven LQR methods

Develop a direct data-driven methodology that adaptively updates the linear quadratic regulator (LQR) state-feedback gain from online closed-loop data, rather than relying on offline or episodic data batches, so as to enable real-time controller adaptation.

Background

The paper surveys indirect and direct approaches to data-driven LQR design. While indirect methods identify a model and then design a controller, direct methods bypass identification and typically use offline or episodic data. The authors note that existing direct data-driven LQR methods rely on data collected in episodes and lack mechanisms for real-time updates.

Within this landscape, the authors explicitly emphasize that enabling online adaptation—updating the controller directly from streaming closed-loop data—has been acknowledged as an open problem. They motivate their proposed DeePO approach as a step toward addressing this gap.