For a decade the question hanging over robots was simple: can the machine actually do the job? Increasingly, it can. A quieter and more awkward question has taken its place: when the machine hurts someone, who is certified, insured, and liable? A dense academic survey and a run of corporate moves this year point to the same conclusion. The binding constraint on physical AI has shifted from capability to paperwork.
The survey, published in Machine Intelligence Research by researchers from the Chinese Academy of Sciences, University College London, and others, reads like a technical audit of how vision-language models can be fooled. Its real argument is structural. In a robot, a mistaken description or a manipulated command does not produce a bad answer on a screen; it produces a physical action, a collision or a dropped load. And the failures cascade: biased training data, a forged sign, a hidden backdoor, or a jailbreak prompt can each break a different layer, so patching one component leaves the others exposed. Safety, the authors conclude, cannot be bolted on late. It has to be defended along the whole path from sensor to motor.
The clearest sign the market has absorbed this comes from where a chip company put its money. In June, NVIDIA launched Halos for Robotics, extending the safety stack it built for self-driving cars to industrial and humanoid machines, and, tellingly, opening an inspection lab to help partners reach third-party certification. Its first taker was Agility Robotics, whose Digit humanoids already work for Amazon, Toyota, and others, customers with mature safety cultures who will not put a machine next to a worker on the strength of a demo video.
Standards have caught up too. ISO 10218, the reference for industrial robots, has been rewritten for the first time since 2011, splitting duties between the firms that build robots and the integrators who install them, folding in cybersecurity, and sorting machines into risk classes. The section on integration, the point where individually safe parts combine into unsafe behaviour, roughly tripled in length. The message is that a clever robot is not the same thing as a safe deployment, and certifiers now want the deployment itself documented.
Insurers are reaching the same conclusion in real time, and it stings. Standard cyber and liability policies increasingly carve out AI altogether, with a new family of exclusions arriving at renewals this year. Lawyers have coined the phrase "silent AI" for the grey zone where nobody has established whether a robot-caused injury is even covered. Into that gap step specialist products such as Munich Re's performance cover, which lets a vendor stand behind its robot without betting the balance sheet on it. Buyers are now told, bluntly, to confirm coverage before a machine reaches the floor.
The strategic implication is that value is migrating away from the robot and toward the permission to run it. Claims histories, regulatory accreditation, and fleet software that has actually operated in production take years to accumulate and cannot be bundled overnight by a hardware startup. That is why a chip firm built an inspection lab and reinsurers are writing AI cover. The hardest environments, a live construction site full of dust, noise, and improvised hand signals, are exactly where laboratory results fail to transfer, and exactly where the certificate, not the capability, will decide who wins the contract.