Physical AI in Practice: Scaling Robots from Lab to Factory Floor

This panel discussion brings together three experts to examine where physical AI stands today, what is blocking industrial-scale deployment, and what the next five years may look like for robotics and automation. The session features Dr. Kal Mos, Head of Research and Predevelopment at Siemens; Prof. Dr. Alin Albu-Schäffer, DLR and Director for Institute of Robotics and Mechatronics at German Aerospace Center and TU Munich and Christoph Berlin, VP Cloud+AI Engineering at Microsoft. The conversation opens with concrete examples of physical AI already in operation. These include a robot arm guided by a vision language action model — a neural network enabling a robot to see, reason, and act — being tested in Siemens facilities, as well as humanoid robots developed at DLR being used to support astronauts and assist people with physical impairments in daily life. The panel then addresses the primary obstacles standing between current demonstrations and reliable deployment at scale. Three challenges are highlighted: the high cost and complexity of generating meaningful training data for physical systems, the absence of scalable processes for validating and evaluating AI models across different robots and use cases, and the difficulty of integrating individual robotic systems into broader factory automation pipelines. Hardware complexity receives particular attention. The discussion notes that a capable robot hand contains more actuators than an electric car, and that the tight integration of hardware and software is frequently underestimated, especially by software-focused organizations. On safety, the panel describes a layered approach that separates a probabilistic cognition layer from a deterministic safety layer, allowing innovation to proceed within defined guardrails. Digital twins are identified as a key tool for accelerating testing in virtual environments before physical deployment. The concept of compliance AI, which refers to using supervising agents to ensure models remain within defined operating procedures, is also introduced. Looking ahead, the speakers discuss the gap between trade fair demonstrations and certified, production-ready systems, referencing lengthy certification timelines in autonomous driving, surgical robotics, and collaborative robots. The panel points to industrial settings as the near-term priority, with a future vision of manufacturers using digital twins to select, commission, verify, and continuously improve robotic systems before physical deployment. This panel was recorded during Hannover Messe 2026.