Efstratios Gavves presents Cyberphysical World Models & Agents: Toward Embodied General Intelligence
On 2026-06-10 - 2026-06-10 14:00:00 at G205, Karlovo náměstí 13, Praha 2
Current AI systems excel at pattern recognition yet remain fundamentally
disconnected from the physical world. Bridging the chasm between digital
intelligence and the dynamics of nature is the defining challenge on the path
toward general intelligence. In this talk, I present my research program on
Cyberphysical AI: algorithms that understand cause-and-effect and physical
dynamics, enabling embodied agents that can reason, imagine, and act safely in
the real world.
Moving from world models to robot world models that understand and interact
with
the physical world, I introduce a framework built on three pillars: world,
physical, and interaction priors, integrated toward Physical AI System
Orchestration. I will present my recent research on these pillars, including
learnable digital twins and object replicas that allow robots to reimagine the
physical world (DreMa, 2025; RecGen, 2026); physics-informed and mechanistic
learning and evaluation (MechNN, 2025; Morpheus, 2026); and end-to-end robot
learning pipelines that close the real-to-sim-to-real loop (Demo2Reward, 2026;
DreMa2, 2026).
I argue that compositionality and cross-modal grounding — not scale alone —
are the keys to escaping purely correlational learning. Integrating explicit
physical and causal mechanisms into world models yields AI that is not only
accurate, but controllable, auditable, and safe. I conclude with my vision for
the road ahead toward Physical and Embodied General Intelligence.
disconnected from the physical world. Bridging the chasm between digital
intelligence and the dynamics of nature is the defining challenge on the path
toward general intelligence. In this talk, I present my research program on
Cyberphysical AI: algorithms that understand cause-and-effect and physical
dynamics, enabling embodied agents that can reason, imagine, and act safely in
the real world.
Moving from world models to robot world models that understand and interact
with
the physical world, I introduce a framework built on three pillars: world,
physical, and interaction priors, integrated toward Physical AI System
Orchestration. I will present my recent research on these pillars, including
learnable digital twins and object replicas that allow robots to reimagine the
physical world (DreMa, 2025; RecGen, 2026); physics-informed and mechanistic
learning and evaluation (MechNN, 2025; Morpheus, 2026); and end-to-end robot
learning pipelines that close the real-to-sim-to-real loop (Demo2Reward, 2026;
DreMa2, 2026).
I argue that compositionality and cross-modal grounding — not scale alone —
are the keys to escaping purely correlational learning. Integrating explicit
physical and causal mechanisms into world models yields AI that is not only
accurate, but controllable, auditable, and safe. I conclude with my vision for
the road ahead toward Physical and Embodied General Intelligence.