Two men who build robots for a living disagree about what stops a humanoid from walking onto a factory floor, and the disagreement is instructive.
Knut Sandven, CEO of Sonair, says the bottleneck "isn't intelligence or speed. It's safe perception; knowing reliably under any condition, that a human is nearby."
German Aguirre, systems manager for robotics at Texas Instruments, says something that sounds compatible but isn't: humanoids "do not have a centralized safety controller" and "continuously adapt their motion, making failure modes harder to define."
Sandven is describing a problem you can certify a sensor against. Aguirre is describing a problem you can't. That gap is the whole story.
Sandven is describing a problem you can certify a sensor against. Aguirre is describing a problem you can't. That gap is the whole story.
This week the same industry moved up a layer, into the paperwork that decides whether any of it is allowed to operate near people. And a coordinated policy campaign from Agility Robotics turned that paperwork into a live political question.
Start with the July wave, because it reads like an implicit chorus agreeing with Sandven. Everyone is shipping certified sensing, not smarter brains.
Sonair unveiled the ADAR One, which it calls the world's first safety-certified 3D ultrasonic sensor. It delivers 180°×180° spatial awareness using sound instead of light, catching the obstacles above and below the single plane that 2D laser scanners miss.
The certification is the product. ADAR One was assessed under IEC 61496, the standard for electrosensitive protective equipment, plus IEC 61508 and ISO 13849. It carries an EC type-examination certificate from exida, a notified body under the Machinery Directive.
The result is a rating of SIL 2 and PL d, with a probability of dangerous failure reportedly below 1.5 x 10⁻⁷ per hour. Sandven's framing is precise: ADAR One is "a certified 3D perceptual foundation that sits beneath any camera, AI stack, or motion system."
Beneath. That word is doing enormous work, and we'll come back to it.
ExRobotics shipped the same week with the ExR-2.5, the only UL-certified autonomous inspection robot for potentially explosive atmospheres in North America. UL 6260 certifies that the robot won't become an ignition source in an oil-and-gas facility.
Aptiv and Robust.AI announced a radar-vision fusion path for the Gen 3 Carter warehouse robot, explicitly targeting PL(d) certification under ISO 13849-1 — which, per Aptiv, is the "high-reliability safety classification used for hazardous robotics applications."
Three announcements, three certificates, one pattern. Nobody claimed a smarter robot. Everybody claimed a certified sensing primitive.
So the question the press releases don't ask: certified against what?
Read the standards, not the headlines. IEC 61496 governs protective devices — the light curtains and scanners that trip a machine when a person crosses a boundary. UL 6260 governs ignition risk in a fixed hazardous zone. ISO 13849 governs the safety-related parts of control systems.
Every one of these standards was written for machines with a perimeter. A fenced cell. A defined zone. A boundary that, when crossed, triggers a deterministic stop.
They certify components against fixed geometry. The sensor detects a body in a plane, or a volume, and something stops. That is a well-defined failure mode, and it is exactly the kind of thing a notified body knows how to test.
Now put Aguirre's sentence back in. Humanoids have no centralized safety controller. Safety lives at the joint, the sensor, the system level, distributed. And the motion continuously adapts, so the failure modes are "harder to define."
You cannot certify against a failure mode you cannot define. Certification is a test, and a test needs a fixed thing to measure.
This is the structural crack under the July optimism. The industry is assembling a stack of narrow, single-modality certificates — a SIL 2 ultrasonic backstop here, a UL 6260 chassis there, a PL(d) radar-vision fusion elsewhere — and each one genuinely attests to something real.
But every one certifies a component behaving like a fixed-perimeter machine. The ultrasonic sensor certifies that if a human enters the volume, the reading is reliable. It does not certify what the whole body does next.
And the whole body doing something next is the entire point of a humanoid. The behavior that makes a humanoid a humanoid — continuously adapting, whole-body motion in close proximity to people — is precisely the behavior no standard exists to certify.
Sandven's "beneath" is honest about this. ADAR One sits underneath the AI stack and the motion system. It is a backstop, not a governor. It tells you a person is there. It does not certify that the robot's emergent response is safe.
That is not a criticism of Sonair. It is the correct scope for an electrosensitive protective device. The point is that the certified layer stops exactly where the hard problem begins.
The academic literature says the same thing from the other direction. A recent Physical AI synthesis reportedly found humanoid limits are physical, not cognitive — data scarcity, sim-to-real gaps, and the difficulty of safely coordinating whole-body interaction. It notes plainly that learned behaviors are "difficult to formally verify," which makes certification slow.
Difficult to formally verify is a polite way of saying there is no test to pass. And no notified body will stamp a certificate on a behavior it cannot bound.
Which brings us to Agility, and why its posture reads differently once you see the gap.
Agility is, per its own social posts, advocating for "voluntary, industry-driven safety standards," positioning itself as the launch partner for NVIDIA Halos and publishing a white paper titled "Safeguarding America's Humanoid Future."
CEO Peggy Johnson frames the philosophy this way: "You can't build your robot and then make it safe. You have to have all of the safety certified — the electrical system, all of the parts, and the software to support all of that."
Notice what her list contains. Electrical system. Parts. Software. Components — exactly the things the existing certificate stack already knows how to certify. The list conspicuously does not contain the emergent whole-body behavior, because there is no certificate to name.
Read "voluntary, industry-driven standards" not as confidence but as a race. There is no standard for the defining humanoid behavior. Someone will write one. Agility would rather it be the industry, with a voluntary test the incumbents can already pass, than a regulator with a mandatory one written after an incident.
The logic is sound. The problem is timing.
Self-regulation works when the writers understand the failure modes better than anyone else and can freeze a testable definition. But the whole reason there is no standard is that the behavior keeps changing — Aguirre's point exactly. Each capability advance redraws the failure surface.
You cannot write a durable voluntary standard against a moving target, and capability is the thing moving. That is why this is a race self-regulation is structurally unlikely to win: the test has to hold still, and the behavior won't.
So here is the deployment reality beneath the July press releases. You can buy a pre-certified 3D ultrasonic backstop today. You can buy a PL(d) perception module and an Ex-rated chassis.
You cannot buy — because no one can sell — certification of the one thing that makes the machine a humanoid. The stack certifies everything around the hard part and stops at its edge.
Sandven is right that safe perception is the bottleneck for coexistence. He is describing the layer that can be certified. Aguirre is describing the layer that can't. Until someone writes a standard for adaptive whole-body motion near people, the certificates will keep accumulating around a hole in the middle — and the fight over who gets to define that hole has already begun.