Physical AI systems that operate at the edge—in factories, warehouses, or hospitals—face a fundamental tension: they need to understand a dynamic scene, anticipate what happens next, and decide how to act, all while running on memory-constrained hardware far from a data center.

The real problem isn’t just perception or planning in isolation; it’s that edge devices lack the compute headroom to run separate large models for vision, reasoning, and control.

The article argues that what’s missing is a unified world model compact enough to run on-device, yet rich enough to connect visual observation directly to action generation.