Physics-Based Reasoning and Imagination at Inference

What if humanoid robots could imagine possible futures before acting—just like humans do? The talk will include a deep dive into why today’s dominant approaches, from imitation learning to model-free reinforcement learning, struggle to scale toward general-purpose embodied intelligence. It will illustrate a new hierarchical architecture in which a physics-grounded latent world model bridges high-level planning and low-level whole-body control, enabling true model-based reasoning directly on the robot at control frequency. The session will explain how this system learns physics and task rewards from human videos, performs rapid imaginative rollouts in latent space, and selects actions through internal forward simulation—without external simulators or rote trajectory memorization. Attendees will gain a clear understanding of how physics-based reasoning unlocks faster adaptation, stronger generalization, and dramatically reduced real-world data requirements. Join this talk to explore a fundamentally different path toward scalable, embodied AI.