This Dagstuhl Seminar focuses on a shift in autonomous cyber-physical systems (CPS): from self-contained, edge-centric intelligence toward distributed Physical AI systems whose sensing, reasoning, learning, and control span devices, edge infrastructure, networks, and the cloud. Traditionally, cyber-physical systems have assumed that critical perception, decision-making, and control must reside close to the physical process because of latency, reliability, privacy, and safety constraints; cloud resources play a complementary role.

The rapid emergence of large AI models for multimodal perception, reasoning, and world modeling is challenging this design point. These models enable perception, reasoning, planning, adaptation, and coordination that are difficult to realize on resource-constrained platforms. Meanwhile, advances in accelerators, infrastructure, and low-latency networking increasingly allow physical systems to draw on remote intelligence during latency-sensitive operation. Intelligence may therefore be dynamically distributed across the computing continuum rather than confined to an individual device.

Why now, and why is this different from traditional edge–cloud computing? Classical offloading largely treats remote execution as an optimization, moving computation to improve latency, energy, or resource utilization. For emerging Physical AI systems, access to remote models and capabilities may become integral to what the system can perceive, reason about, and accomplish. The question shifts from whether and where to offload computation to how to architect a physical system whose intelligence is inherently distributed.

This shift challenges existing abstractions. AI inference may be stateful, iterative, multimodal, and input-dependent; models differ in capability, cost, and latency; compute and network availability may change; and AI outputs may influence closed-loop physical actions. Systems must jointly address capability, placement, communication, state, adaptation, uncertainty, and safety. Key questions include partitioning intelligence across heterogeneous resources; how systems should degrade gracefully when connectivity or remote capabilities disappear; how state and context should move across models; and how security, privacy, predictability, and safety can be maintained when physical control depends on distributed AI.

The seminar aim to convene researchers from robotics and CPS, machine learning and Physical AI, cloud and distributed systems, networking, computer architecture, and security. These communities approach the problem with different abstractions and assumptions. We aim to identify where existing approaches fall short, develop new systems abstractions and architectural principles, and articulate key research questions at the intersection of AI and networked physical systems. The seminar seeks a shared research agenda for the next generation of distributed, adaptive, and dependable Physical AI systems.

Creative Commons BY 4.0

Erik Elmroth, Lili Qiu, Nader Sehatbakhsh, Prashant Shenoy, and Mani B. Srivastava