Overview
PhORM optimizes a small mobile robot's shape with physics-informed neural networks in the loop: fifteen bounded geometric and mass-layout variables against an objective combining ZMP tipover margins, rolling and drag energy, a shape-conditioned thermal PINN, Helmholtz speaker acoustics, and packaging constraints. The PINN-in-the-loop iteration improved the objective by 15.0% over the stock design (mass -12%, stability margin +24%, energy -11%), and an independent finite-volume certification confirmed the optimized thermal maximum at 54.6°C versus the stock 56.3°C. A multi-seed probe that did not beat the budget seed is reported as an explicit negative, and PINN-to-finite-volume verification errors of -2.0% to -3.1% on T_max are archived in results JSONs.