Skip to main content

Applied Analysis Seminar: Adam Thorpe (UT Austin)

Thursday, October 22
1:00 PM - 2:00 PM
301 TMCB

Title:  Runtime Intelligence and Adaptation for Embodied Systems

Abstract: Embodied systems face an unavoidable gap between the conditions anticipated during training and those encountered during deployment. Environmental conditions, system dynamics, and physical properties may only become apparent through interaction, requiring autonomous systems to adapt their models, policies, and decisions during operation. Yet adaptation from limited observations remains computationally demanding, particularly when decisions must be made under strict time and resource constraints. My research develops mathematical and computational methods that make adaptation tractable by reformulating how new information is represented and incorporated at runtime. In this talk, I discuss these ideas through examples in learned function representations, robotic autonomy, and physical world modeling. Using a technique known as function encoders, I learn neural bases for spaces of dynamics and control policies, reducing adaptation to low-dimensional estimation and optimization. In robotics applications, these representations support online dynamics and policy identification, terrain adaptation, and high-speed off-road autonomy, where rapid adaptation is essential. In robot manipulation, physical interaction and weak-form momentum balance can be used to identify deformable material properties through linear least squares, which enables more accurate prediction and manipulation planning. Together, these results demonstrate how mathematical structure can make runtime adaptation tractable under real-world constraints.

Categories