EvoPhys-World, the first 5D world model for Physical AGI

Talk overview

This talk proposes EvoPhys-World, a five-dimensional world model that adds causal worldlines to three-dimensional space and time. Inspired by human motor control and multisensory feedback, it uses branching and collapsing worldlines, a dual world-engine and world-policy design, Phy-JEPA representation decoupling, and self-supervised 4D memory to learn physical intelligence from large-scale human-interaction video.