The core message of the provided transcript is that “Physical AI” is not a single recipe, but a set of approaches for getting AI to understand and act in the real world, where data sources, physics, deployment conditions, and feedback loops matter differently by domain. Moderator Philipp Gerbert opened by positioning the panel inside Munich’s deep-tech ecosystem, citing TUM Venture Labs’ roughly 750 incubated ventures, about 150 companies started per year, 12 deep-tech and life-science labs, and €2 billion to €2.5 billion of annual investment. He framed AI’s distinctive power around the number 1,000,000, arguing that electronics, computation, and self-improvement can operate at scales vastly faster than biological human processes. From there, he contrasted large language models’ success in language with the harder ambition of learning from and acting in the real world.

The panel’s first major thread was data. Ralf from Sereact argued that robotics cannot rely on simulation alone because contact forces and contact dynamics are still simulated poorly; simulation is useful for pre-training, but not sufficient. Teleoperation can produce high-quality post-training data, and he described seeing a multi-level Xiaomi teleoperation operation in China, but called that approach “clearly not a sustainable way.” Sereact’s preferred strategy is post-deployment continual learning: 400 robots deployed with customers such as IKEA and Zalando generate both recurring robot-as-a-service revenue and the real-world edge-case data needed to improve the system. Milan from alqem described the opposite data posture for materials prediction: because humans have only made about 50,000 to 100,000 materials, while synthetic datasets number roughly 8 million to 20 million, alqem uses only synthetic data for predictions and experimental data for how to make materials. Lukas from SE3 AI added that defense data is harder to obtain, so his team uses language models connected to 3D environments, giving drones context about where they are, what objects exist, and what tools they can use.

The second major thread was model design and the role of physics. Milan said that adding more heuristics, physics, or chemistry directly into models had not helped alqem; instead, materials expertise goes into creating better synthetic data. Ralf took a more embodied robotics view: models can learn physics from accurate data, but robotic outputs must align with real-world constraints such as gravity and manipulation feasibility. He said physical understanding improves manipulation accuracy, for example when a robot reasons that flipping an unreachable item may create a better future grasp state. Lukas distinguished this from defense autonomy: low-level drone flight control is largely solved, while SE3 AI focuses on higher-level “system two” reasoning for coordinating many autonomous systems across hundreds of kilometers and many hours. The supplied transcript ends at 14:48, so later discussion and closing conclusions are not available in the provided text.