IEEE.org | IEEE Xplore Digital Library | IEEE Standards | IEEE Spectrum | More Sites
Convex optimization has long been a cornerstone of classical optimal and robust control, enabling strong guarantees and reliable computational tools. At the same time, modern control design increasingly relies on nonconvex optimization, especially direct policy search and learning-based methods, which have shown striking empirical success and are now supported by a growing body of theory. This talk highlights a unifying perspective that connects these two worlds. We begin by reviewing benign nonconvex landscapes that arise in several benchmark control problems, explaining why simple first-order methods can succeed despite nonconvexity. We then introduce an Extended Convex Lifting (ECL) framework that exposes hidden convexity in classical control from a modern optimization viewpoint. This lifting provides a principled bridge between nonconvex policy optimization and convex reformulations. When an ECL exists, it can (i) transform the original design problem into an equivalent convex program and (ii) certify global optimality for a class of stationary points. We conclude by discussing algorithmic implications and recent scalable methods that make these ideas practical at larger scales. Looking ahead, these connections suggest a promising path to combining the global guarantees of convex optimization with the flexibility of policy optimization for robust and learning-enabled control of modern dynamical systems.
Uncertainty propagation and mitigation is at the core of all robotic and control systems. The standard approach so far has followed the spirit of control of a system “with uncertainties,” as opposed to the direct control “of uncertainties” affecting the system itself. Covariance control, developed by Bob Skelton and his colleagues in the early 80’s was introduced as a principled approach to handle uncertainty with guarantees in the asymptotic case. The finite-time case has only recently been addressed under the name covariance steering. Borrowing ideas from the classical Optimal Mass Transport (OMT) and the Schrödinger Bridge problems, covariance steering provides a new tool to control stochastic systems with strict performance guarantees that go beyond classical controllability results, which hold only for deterministic systems. In this talk, we will review some recent results on covariance and distribution control for stochastic systems subject to probabilistic (chance) constraints, robust and data-driven distributional control, and will demonstrate the application of the theory to solve problems on control, robotics, and generative AI applications.
Data-driven and learning-based methods have attracted considerable attention in recent years both for the analysis of dynamical systems and for control design. While there are many interesting and exciting results in this direction, our understanding of fundamental limitations of learning for control is lagging. This talk will focus on the question of when learning can be hard or impossible in the context of dynamical systems and control. In the first part of the talk, I will discuss a new observation on immersions and how it reveals some potential limitations in learning Koopman embeddings. In the second part of the talk, I will show what makes it hard to learn to stabilize linear systems from a sample-complexity perspective. While these results might seem negative, I will conclude the talk with some thoughts on how they can inspire interesting future directions.
The main focus of this talk is to present mathematical fundamentals, state-of-the-art, technical challenges and open problems in control of flight for atmospheric vehicles, such as aircraft and other aerial platforms. Reduced order modeling and flight simulation key features for control applications will be discussed. The emphasis is on the theoretical and engineering aspects of creating and transitioning to practice guidance and flight control systems with guarantees of closed-loop stability, robustness and performance.
This talk is motivated by applications that require a new perspective on elementary control/filter design and algorithms. In particular, I will motivate why in certain scenarios, it becomes important to relax the knowledge of exact parameterization of the system, noise statistics, and/or disturbances, in order to develop feedback control and filters. As a guiding example, I will consider clock synchronization in satellite constellations, which is essential for positioning, navigation, and timing. This application highlights the need for filtering techniques that do not rely on precise knowledge of noise and measurement covariances. Building on this motivation, I will then step back and introduce a calculus for control and filter design based on direct policy optimization using first-order methods, while revisiting several outstanding problems in control. The goal of this plenary is to highlight key analytic and geometric features of this perspective on system design and provide pointers to some of the ongoing efforts in this rapidly growing area.
More information provided here.
