research
A few highlighted research projects.
Data-Enabled Predictive Control (DeePC)
Using raw data directly for predictive control of complex systems, bypassing the need for explicit parametric models.
Data-Enabled Predictive Control (DeePC) is a control framework that uses measured input-output data to directly predict and optimize future system behavior, without first identifying a parametric model. DeePC replaces the traditional model identification step with a data-driven representation of the system’s behavior, enabling real-time optimal control directly from data.
Our work on DeePC spans the full spectrum from theory to practice. On the theoretical side, we have developed regularization techniques that make DeePC robust to noise and model mismatch, drawing connections between regularized DeePC and classical system identification approaches. On the practical side, we have demonstrated DeePC on quadrotors and autonomous excavators, showing that the approach can handle the complexities present in real-world robotic systems.
Robust Data-Driven Control
Designing uncertainty models and robust control methods directly to make data-driven control reliable.
A central challenge in data-driven control is ensuring that controllers designed from noisy, finite data are reliable when deployed on real systems. Our research addresses a key question: how should uncertainty be characterized in data-driven nonparametric models, and how should this uncertainty be used to guarantee robust performance and safety of the controlled systems?
We have developed distributionally robust formulations of data-enabled predictive control that provide probabilistic guarantees on constraint satisfaction. We have also developed methods for behavioral uncertainty quantification which are used to develop reliable control strategies directly from data.
Adaptive Data-Driven Control
Real-time adaptation of data-driven controllers using streaming data.
Real systems change. Dynamics shift, operating conditions vary, and a controller designed from yesterday’s data may not work well today. This project asks: how can a controller continuously learn from new data and adapt in real time?
We develop online subspace tracking algorithms that use streaming data to recursively identify the behavior of an unknown or time-varying dynamical system, even when the system’s complexity (e.g., model order) is itself unknown or changing. These algorithms operate can be integrated directly into data-enabled predictive control frameworks for adaptive, real-time control.