Robotics Seminar @ Illinois: Structure-Preserving Learning for Autonomy in Complex Physical Environments

Title: Structure-Preserving Learning for Autonomy in Complex Physical Environments

Abstract: Autonomous robots are increasingly expected to operate in environments where the surrounding physics cannot be ignored, whether ground vehicles traversing unstructured terrain or aerial and marine vehicles navigating geophysical flows. In these settings, the robot and its environment form a tightly coupled dynamical system, and the performance of the robot depends on how well this coupling is understood. These interactions are often highly nonlinear, time-varying, and difficult to model from first principles alone, which makes reliable autonomy fundamentally challenging. Advances in data-driven methods have opened new pathways to modeling these complex environments and interactions. However, these approaches have limited interpretability and generalizability, require large amounts of training data that is often costly to obtain, and fail to preserve the very mathematical properties that make real-time planning and control tractable. This raises fundamental questions about how learning should be done for autonomy: what should be learned, what should be known a priori, and what structure must be preserved. In this talk, I will show that embedding physical structure and preserving mathematical structure are key to answering these questions. Using tight formation flight of quadrotors as an example, I will present our knowledge-based neural ordinary differential equations (KNODE) framework for learning residual dynamics and show how it can be combined with adaptive control to yield sample-efficient and robust model predictive controllers. I will then show how learning residuals that preserve differential flatness leads to controllers that are both computationally lightweight and high-performing. Finally, I will discuss how these ideas extend to modeling the environment itself, from learned finite element models that exactly preserve conservation laws for multi-robot source localization in complex flows to learned Lagrangian models that enable stable long-horizon forecasting. I will conclude by showing how preserving structure can lead to new planning strategies that combine the guarantees of model-based approaches with the generalizability of learned ones.

Bio: M. Ani Hsieh is an Associate Professor in the Department of Mechanical Engineering and Applied Mechanics at the University of Pennsylvania. She is also the Deputy Director of the General Robotics, Automation, Sensing, and Perception (GRASP) Laboratory and Program Chair for the Robotics MSE Program. Her research interests lie at the intersection of robotics, multi-agent systems, and dynamical systems theory. She received her B.S. in Engineering and B.A. in Economics from Swarthmore College and her PhD in Mechanical Engineering from the University of Pennsylvania. Prior to Penn, she was an Associate Professor in the Department of Mechanical Engineering and Mechanics at Drexel University. Hsieh is the recipient of a 2012 Office of Naval Research (ONR) Young Investigator Award and a 2013 National Science Foundation (NSF) CAREER Award.

Location: CSL Studio 1232 at 2:00 pm— please enter through the center doors in the middle of the parking garage's south face if you do not have card swipe access to CSL Studio.