Fall 2026 GRASP SFI: Benjamin Shaffer, University of Pennsylvania, “Structured learning for Robotic Navigation, Modeling, and Sensing”

Modern learning methods offer powerful flexibility for robotics challenges involving complex physical and geometric structure, but often struggle to generalize and provide few guarantees on the resulting behavior. This motivates approaches that embed learning within structured computational models, using methods from PDEs and discrete geometry to encode geometry, physics, directly into the solution process and naturally enforce constraints. Building on ideas from implicit neural networks, we incorporate these structures into learned models to enforce properties such as conservation, stability, and safety while improving performance over black-box approaches. We illustrate this perspective through several recent developments, including Neural Navigation Functions, which embed learning inside a PDE-based planner to achieve zero-shot geometric transfer with collision avoidance and global convergence guaranteed by construction. We further demonstrate how the same principles extend to physics simulation and robotic sensing.