Enabling Physical AI: The Role of Simulation Technology in Bridging the Physical and D

In this invited talk introduced by Dennis Hoppe of HLRS, Christoph Heinrich of Siemens AG discusses how simulation technology enables physical AI, the class of AI that lets machines perceive and interact with physical surroundings. He contrasts impressive humanoid robot demonstrations with industrial requirements for safety, precision, and repeatability, and distinguishes physical AI from physics AI, models trained on physics principles such as fluid dynamics or structural mechanics. The central challenge is the sim-to-real gap: robot policies trained on scaled synthetic simulation data often degrade when deployed in reality. Heinrich presents two remedies. Simulation enhancement uses high-fidelity, physics-based simulation, long established in product development, to generate more accurate training data and to validate AI systems on rare scenarios outside the training distribution; a geometric deep learning surrogate trained on classical simulations of a robotic gripper can predict mechanical stresses for new CAD designs in seconds, and a pipeline built with NVIDIA Isaac Lab scales imitation learning including deformable objects. Real-world data integration infuses physics into world models: an executable digital twin of a milling process running on edge hardware compensates industrial robot flexibility, improving milling accuracy from about one millimeter to 0.1 millimeter. In the Q&A he notes GPU-accelerated solvers such as Simcenter STAR-CCM+ achieving roughly 30x speedups, and highlights Siemens participation in European initiatives including EuroHPC, Catena-X, and Manufacturing-X.