$$ % Define your custom commands here \newcommand{\bmat}[1]{\begin{bmatrix}#1\end{bmatrix}} \newcommand{\E}{\mathbb{E}} \newcommand{\P}{\mathbb{P}} \newcommand{\S}{\mathbb{S}} \newcommand{\R}{\mathbb{R}} \newcommand{\S}{\mathbb{S}} \newcommand{\norm}[2]{\|{#1}\|_{{}_{#2}}} \newcommand{\pd}[2]{\frac{\partial #1}{\partial #2}} \newcommand{\pdd}[2]{\frac{\partial^2 #1}{\partial #2^2}} \newcommand{\vectornorm}[1]{\left|\left|#1\right|\right|} \newcommand{\abs}[1]{\left|{#1}\right|} \newcommand{\mbf}[1]{\mathbf{#1}} \newcommand{\mc}[1]{\mathcal{#1}} \newcommand{\bm}[1]{\boldsymbol{#1}} \newcommand{\nicefrac}[2]{{}^{#1}\!/_{\!#2}} \newcommand{\argmin}{\operatorname*{arg\,min}} \newcommand{\argmax}{\operatorname*{arg\,max}} \newcommand{\dd}{\operatorname{d}\!} $$

Robust Machine Learning

Neural controller design, energy shaping, and Bayesian learning for uncertain dynamical systems.

How can learning help us design controllers that account for uncertainty?

A learned controller acts on a dynamical system: its decisions influence the states it will encounter next. Our research connects neural approximation and Bayesian learning with nonlinear control, focusing on controller design, stability, and sensitivity to uncertain models and measurements.

Learning to design controllers

Passivity-based control provides a connection between a system’s energy and its stability. Applying this approach can require solving nonlinear partial differential equations, which may be difficult to solve in closed form.

We investigate neural networks as approximations within this design process. In our work on interconnection and damping assignment, the required equations become part of a data-driven optimization problem. The physical structure of the control problem guides what is learned.

Representative work

Data-Driven Passivity-Based Control of Underactuated Mechanical Systems via Interconnection and Damping Assignment

Wankun Sirichotiyakul and Aykut C. Satici · International Journal of Control

Publication and Scholar search

Bayesian learning and uncertainty

A controller is often designed from imperfect information about the system and its state. Bayesian learning represents uncertainty through distributions, providing a way to examine controller design beyond a single parameter estimate.

Our preprint Robustness of Control Design via Bayesian Learning compares deterministic and Bayesian controller design for a one-dimensional, open-loop unstable stochastic system. The Bayesian formulation accounts for uncertainty in both the system parameter and state. This deliberately simple setting allows the role of uncertainty to be studied directly.

Read the preprint · Publication and Scholar search

Connecting learning and nonlinear control

A related line of work brings neural approximators and Bayesian inference into passivity-based control of underactuated systems. Together, these studies explore how flexible approximation methods can be used within a control-theoretic framework.

Representative work

Robust Passivity-Based Control of Underactuated Systems via Neural Approximators and Bayesian Inference

Nardos Ayele Ashenafi, Wankun Sirichotiyakul, and Aykut C. Satici · IEEE Control Systems Letters, 2022

Combining Energy-Shaping Control of Dynamical Systems with Data-Driven Approaches

Wankun Sirichotiyakul and Aykut C. Satici · CCTA, 2021

Questions that guide this direction

• How can physical structure make a learning-based control problem easier to solve?
• How should uncertainty in models and measurements enter controller design?
• How can stability analysis inform the design and evaluation of learned controllers?

These questions connect this research direction to the lab’s work in robotics and control.

Research overview → · All publications →