Executive summary

Humanoid safety depends on context, yet the robot cannot be its only judge. From a conversation with NVIDIA’s Amit Goel, Stéphane Gervais sets out five questions on safe failure, perception, supervision, authority, and deployment evidence

At NVIDIA GTC Berlin, Amit Goel explained why humanoid robots cannot simply be programmed. Stéphane Gervais examines what that changes: when the same movement can be safe or hazardous depending on context, the robot's judgment enters the safety case and can no longer be its only safeguard. He sets out five questions every humanoid safety case must now answer.

That is the uncomfortable distance between what humanoids can demonstrate and what industry needs. At NVIDIA GTC Berlin in October 2026, I asked Amit Goel, NVIDIA's Head of Robotics Ecosystem and Edge AI Product, what is really keeping them from large-scale industrial deployment.

His answer was strikingly pragmatic and began with the machines. "You can have a robot do a really good thing one time, a few times," he said. Deployment at scale, he added, means repeating it across eight-hour shifts, every day of the year, and hardware maturity is the first hurdle.

The second, equally important limit is the brain. "As you can see in a demo, you sometimes don't know how much data was used for it and how the brain generalizes," Goel said. A robot may excel with a familiar object in a familiar setting; change the object, add an obstacle or an unfamiliar human interaction, and its behavior becomes far less predictable.

Industrial automation has long handled uncertainty by controlling the environment. Humanoids target spaces designed for people, where objects move and interactions cannot all be anticipated, so they will have to reason about what is happening around them, Goel explained. "You cannot just program them." Hence the interest in vision-language-action and other robotics foundation models, which combine perception, language, action, and large-scale data into a broader understanding of the world.

Broader intelligence raises another question: when a machine adapts instead of repeating predefined movements, how do we establish that its decisions stay appropriate in situations it has never met? For humanoids, that is a safety question, and it ran through our conversation.

From a reference platform to a specific safety case

"At NVIDIA we build the core technology. We don't build robots," Goel said. His team helps companies building robots and robot brains integrate NVIDIA technologies. Before reaching the hard problems, those developers must assemble hardware, sensors, an operating system, and a credible simulation model, which can take several years, he noted.

Enter Agile ONE SN, launched at GTC Berlin by Germany's Agile Robots with NVIDIA: a reference humanoid built on NVIDIA Isaac GR00T and Jetson Thor, with an integrated body, sensors, onboard computing, and simulation assets. NVIDIA describes Halos integration as planned; the reference platform should not be mistaken for a safety-certified deployment. NVIDIA is "left shifting" that groundwork, Goel said, so developers can go straight to building the robot's brain and testing it in the real world.

The bet is broader. Robotics has usually started from an application and built a machine for it. "Our goal with the reference design is to provide that general platform on which application developers can create specialization on top of it," Goel said: one foundation, many applications through specialized skills.

As I noted in The Platform War Behind Robotics, the computing-platform analogy has limits: a robot cannot download a capability without accounting for forces, movements, environment, and safety, and a specialized machine often does the job faster, more reliably, and more cheaply. Turning a general foundation into dependable applications takes serious engineering and validation, above all in safety, because a safety case is never general: it covers a specific machine, task, environment, and set of people. A general-purpose platform does not remove that specificity; it moves it to the moment each new skill is deployed.

Industrial robotics has long used separation to reduce risk: as Goel summarized it, a cage around the robot, a laser scanner, and a stop when someone comes close. Collaborative applications already require a more detailed assessment of interaction. Humanoids extend that challenge to mobile, adaptive behavior in less constrained environments. His own example captures the shift: approaching a person is acceptable if intended, a hazard if not. Moving toward a technician who asked for help may be part of the robot's task; the same trajectory, unexpected and blind to the person's movements, can create a hazard. Intended interaction still has to respect validated limits on speed, force, separation, and stability. Context changes the risk assessment; it does not make a requested movement inherently safe.

Even the classic safe state changes. For many conventional machines, stopping motion and removing hazardous power are part of reaching a safe state. A robot that must actively balance can become unstable when it loses power, so its safe state needs specific engineering. ISO is writing a dedicated standard for such machines, ISO 25785-1 on dynamically stable industrial mobile robots, bipedal humanoids included, and it is still at committee-draft stage.

Functional safety does not go away, Goel stressed: hardware and software failures must still be handled. But intelligent machines add cases where every component works as designed and the system still decides wrongly. Automotive engineers formalized that risk in ISO 21448 as the safety of the intended functionality: hazards from functional insufficiencies rather than faults, typically in perception or decision. The distinction is useful for humanoids too, although ISO 21448 applies to road vehicles and is used here as an engineering analogy, not as a humanoid certification standard.

The robot cannot be its only safety judge

Goel's answer is to make the robot's brain a source of context within the safety architecture. NVIDIA built Halos, he said, as a framework for programmable, software-defined safety, "so that you can dynamically define what safety means." I examined Halos in IEEE Computer; for robotics, NVIDIA presents it as a full-stack safety system for physical AI.

Where safety sits is changing too. "Safety used to mean you build something and then you used to put safety on top," Goel said.

"But now the functionality and safety have to coexist together."
The robot's brain, he explained, can now run on the same computer as its safety system, the two communicating. Co-location can reduce integration overhead, but it raises the burden of proof. An assessor needs evidence of freedom from interference, including protection against memory corruption, timing interference, and shared-resource failures. Running two functions on one computer does not, by itself, establish their independence.

NVIDIA also pursues the complementary principle of independence. Its open-source Halos Outside-In Safety Blueprint, in early access, uses facility cameras, AI perception, and safety logic to see beyond a robot's onboard sensors: "a brain that's outside the robot body that's monitoring the robot and then making a decision whether it's doing the right thing or not," as Goel put it. Its warehouse workflow illustrates how external perception could detect a worker emerging from behind shelving and request a slowdown or stop. The repository explicitly describes the blueprint as early-access material for prototyping and evaluation. Its safety notice makes no claim of functional-safety compliance or suitability for production safety-related systems; a deployment needs its own appropriately engineered safety layer.

The value goes beyond visibility. As autonomy grows, the robot's intelligence and its supervision need demonstrable separation of responsibilities and protection against common-cause failures: a robot may interpret its surroundings and choose an action, but its own interpretation should not be the final authority on whether that action is safe. The monitor must itself be designed and validated to the relevant safety requirements, and it can add a second point of view. Whether its blind spots are sufficiently different must be demonstrated. NVIDIA is pursuing on-robot integration and external supervision; the safety case must show how those layers remain effective when sensors, models, communications, or power supplies fail.

Independence also redistributes responsibility. If cameras installed in a building become part of a safety function, the safety case covers the installation, not only the robot, and robot makers, integrators, and site operators must agree on who owns each layer. ISO 25785-1 already reserves a second part for the integration of applications.

Safety engineering has a reference for this boundary. ISO/IEC TR 5469, the 2024 technical report on functional safety and AI, distinguishes AI used inside a safety-related function from non-AI safety-related functions used to keep AI-controlled equipment safe. It is guidance, not a standalone product-certification standard. European law now attaches consequences to that line: from 20 January 2027, the Machinery Regulation requires a notified-body conformity-assessment route for the self-evolving machine-learning safety components listed in Annex I, Part A, point 5, and the embedded systems covered by point 6. For point 6, that requirement concerns those systems, not automatically the whole machine. It does not apply to every use of AI in robotics. Where the robot's brain ends and its supervision begins is therefore not only an engineering choice. It shapes the certification route.

What evidence survives the real world?

I asked Goel how developers move from demonstrating capability to proving safety.

"The evidence has to be shown in the real world,"
he said, yet nobody wants to put people or equipment in harm's way to prove it. Simulation resolves part of that tension by recreating, repeating, and varying hazardous scenarios. Goel pointed to Halos SAFE, the Safety Assurance Framework for Evaluation that NVIDIA presented for autonomous vehicles at GTC Berlin, as a model for robotics, and expects simulation to play "a very big role."

Simulation is only as good as its fidelity: a robot can behave perfectly in a virtual world and fail when real friction, lighting, sensor noise, or human movements differ. Goel pairs it with design discipline: assume the system can be unsafe, build in ways to detect and respond when something goes wrong, and add redundancy. "Redundancy is a critical piece of it."

For industrial buyers, a claim that a robot is intelligent, adaptable, or safe is not evidence that it consistently operates within defined limits. That evidence must cover the whole system, from tasks and environment to failure modes and human interaction, which is why safety, cybersecurity, sensor reliability, AI validation, and operational monitoring can no longer be treated separately.

Recognizing a person does not grant authority

Goel placed AI agents in the same safety conversation: in agentic robotics, agents can write their own scripts and create their own connections. NVIDIA's answer is OpenShell, a secure runtime that treats the model with zero trust and enforces programmable, software-defined policies outside it to monitor agents and control their access. Its documented controls constrain file, network, and process access. They do not establish who is entitled to issue every instruction or certify the safety of the robot's motion. I examined it, and where its guarantees stop, in a separate analysis. The question here is narrower: who is entitled to ask a robot to act?

I had asked him a concrete version: how does an autonomous humanoid decide what it is authorized to do depending on whom it recognizes, an elderly person, a child, or a professional? His answer centered on reasoning models, which bring internet-scale knowledge into robotics and let a robot distinguish a person to interact with from one to avoid.

"The semantics matter, the context matters."
That capability matters: a hospital robot may need to adapt to patients, clinicians, and children; a factory robot to operators, maintenance staff, and visitors. The safety and authorization requirements depend on the task; appearance is not a reliable basis for assigning a person's role.

But his answer addressed the recognition half of my question. A robot may infer from clothing, behavior, and location that someone is a maintenance engineer; that proves neither identity nor permission. A high-visibility vest is a signal, not a credential, and understanding a spoken instruction does not establish that the speaker may give it. Cybersecurity has long kept recognition, authentication, and authorization apart; physical AI risks collapsing them, because the model that recognizes a person may also decide whether to obey them.

An operator may ask a robot to move materials but not to override a safety restriction; a maintenance engineer may inspect equipment while some operations stay prohibited until the machine is in a safe state. Such conditions cannot rest on what a model finds plausible. They require explicit policies, identity verification, and enforceable boundaries linking request, person, and context. As autonomy grows, preserving that chain of human authority becomes part of engineering trustworthy systems.

Goel pointed to one more frontier: physical contact. Cameras show a robot its surroundings and encoders report what its body is doing, but seeing an object is not feeling it. Asked whether humanoids still lack key sensing, he answered: "One of the modalities that we believe is going to be very important is the capability of touch or tactile feedback." With touch in the models, he expects robots to have the sensing needed for a lot of interaction and physical work.

A person lifting a glass constantly adjusts grip pressure; someone inserting a component feels resistance before seeing a problem. Vision estimates how an object should behave, while contact shows what is actually happening, letting a robot detect slipping, adjust force, and respond to resistance. Reference platforms are starting to reflect that priority: NVIDIA's description of Agile ONE SN specifies five-finger hands with tactile sensing and 1 kHz force control. Those are announced technical features, not evidence that the sensing or control loop is safety-rated.

Touch matters for safety as much as dexterity: unexpected resistance is often the first sign that something, or someone, is where it should not be. Collaborative robotics already treats contact as a safety variable: ISO 10218-2:2025 incorporates collaborative-application safety requirements previously addressed in ISO/TS 15066, including the assessment of contact forces and pressures. It addresses industrial applications; it is not a complete safety standard for every humanoid use case. A poor decision may stem from reasoning or from missing sensing, and a larger model cannot compensate for information the system never measured. Progress will depend on integrating perception, touch, motion, reasoning, and control into one system, because the physical world, unlike text or images, imposes constraints that no model, however capable, can ignore.

Five questions for a humanoid safety case

Taken together, the conversation points to what a humanoid's safety case will have to demonstrate. A safety case is the structured argument, backed by evidence, that a system is acceptably safe for a given use in a given environment. When safety depends on context, I see five questions it must answer.

  1. Does it fail safely? Hardware and software faults remain the foundation: diagnostics, redundancy, and safe states, in the tradition of IEC 61508 and the 2025 revision of ISO 10218. For a robot that must actively balance, even the safe state has to be engineered rather than assumed.
  2. Does it read the situation correctly? Hazards can arise without any fault, from what the robot perceives or infers. This is the kind of insufficiency addressed by SOTIF in road vehicles and, by analogy, an important challenge for humanoid safety engineering. Context-dependent safety architectures such as those envisaged for Halos need evidence that their use of AI remains within validated limits. It requires defining the conditions the robot has been validated for, and detecting when it leaves them.
  3. Who checks its judgment? The robot's interpretation cannot be the final authority on its own safety. Supervision needs justified independence and an analysis of shared blind spots, whether it runs beside the AI on the same computer, with demonstrated freedom from interference, or outside the robot. An outside-in viewpoint alone is not proof of that independence.
  4. Who is entitled to ask? Recognizing a person is neither verifying their identity nor confirming their authority. What a robot may do at someone's request, and what its agents may do on their own, must be bounded by enforceable policies outside the model.
  5. What evidence proves it, for this deployment? Validated simulation, closed by real-world evidence. And because each new skill or site changes the task, the environment, and the people involved, the case must be revisited at every deployment, not only at every robot generation.

The first question is where industrial safety has always started. The next three gain renewed urgency with intelligent, embodied machines. The fifth binds them together: without evidence for the specific deployment, the other answers remain claims.

Asked what might surprise us most by 2030, Goel avoided the familiar promise of a humanoid in every home: "So are we going to have robots in the home? Maybe not, but I think in a lot of environments like factories and manufacturing and logistics, we will see a lot more adoption of robots that are very intelligent and have the reasoning capability to actually go across different environments and different tasks."

The economics explain why. Factories and warehouses let companies define tasks, measure productivity, organize maintenance, and phase deployment; homes, where objects, people, and activities vary endlessly, are far harder to characterize. That matches my reading of NVIDIA's GTC Berlin briefing in Edge AI Is Starting to Move Beyond Pilots: scale emerges first where tasks are bounded and reliability can be demonstrated.

For Europe, with its automation, machinery, and engineering base, the opportunity lies beyond humanoid hardware. If safety is becoming contextual, integrating AI, robotics, functional safety, and industrial processes, and proving that the result stays within defined limits, could become a real differentiator, without building foundation models. The regulatory calendar adds urgency: the Machinery Regulation generally applies from 20 January 2027. Regulation (EU) 2026/1744, the Digital Omnibus on AI, establishes 2 August 2028 for the relevant high-risk product-AI provisions and moves machinery to a sectoral approach. Its Article 3 requires AI-related machinery safety requirements to be added through delegated acts under the Machinery Regulation, applicable by that date. This does not suspend the machinery safety obligations that apply in January 2027.

The most revealing insight from my conversation with Goel was that intelligence is only one part of the problem. The more capable a robot becomes, the more decisions it makes, and the higher the stakes of deciding when it may act.

When the safety of a movement depends on context, the robot's interpretation of that context becomes part of the safety case, which is precisely why it cannot be its only judge. Robotics has spent decades teaching machines complex tasks. Its next challenge is to set the conditions under which they can be given more freedom to act, while the people responsible for them keep the authority to say no. That freedom will be earned one safety case at a time, on factory floors and in warehouses, when the world refuses to behave like a demonstration.

Edge AI Is Starting to Move Beyond Pilots. NVIDIA Explains Where Scale Is Emerging First., ApexTransform, October 2026.

NVIDIA Is Moving AI Agent Security Into the Infrastructure Layer. The Next Question Is Authority., ApexTransform, October 2026.

Building Trust Into AI Systems: Inside NVIDIA's Halos for Physical AI Safety, IEEE Computer, 2026.

The Platform War Behind Robotics: How NVIDIA Plans to Become the Infrastructure of Physical AI, ApexTransform.

Why AI Agents Need a Chain of Authority, Not Just a Human in the Loop, Forbes, September 2026.

Why Physical AI Forces Executives to Rethink Safety, Accountability and Trust, Forbes, February 2026.

Reporting: one-on-one interview with Amit Goel at NVIDIA GTC Berlin, October 2026, complemented by the NVIDIA and Agile Robots announcements of October 21, 2026.

PUT THIS INSIGHT TO WORK

Are you governed to decide?

Clarify decision rights, autonomy limits and investment conditions with the AI Board Decision Kit.

Open the AI Board Decision Kit →

See both steps of the decision journey →