Deployment intelligence for physical AI

Abstract

Robotics is entering a foundation-model era, in which robot policies can generalize across tasks, objects and embodiments. Yet benchmark success does not show that a system is ready to act in factories, hospitals, warehouses or public spaces. Deployment requires a different kind of evidence: whether embodied intelligence remains valid, safe and accountable after its assumptions, sites and models change. This Perspective proposes deployment intelligence as a measurement science for that evidence. Rather than replacing task success, deployment intelligence extends evaluation to the model-in-system: the coupled policy, body, site, human workflow, safety infrastructure and governance process. We distinguish capability evidence from deployment evidence, define a readiness vector for physical AI, extend operational design domains into dynamic site envelopes and propose deployment-evidence cards as a reporting norm. The goal is to make deployability measurable before robots are treated as real-world infrastructure.