Please Use the Robot Safely and in a Friendly Manner

[UN p.N]

UN Independent International Scientific Panel on AI, Preliminary Report (July 2026), page N.

[AI Index p.N]

Stanford HAI, The 2026 AI Index Report, Chapter 1, printed page N.

[MDS README]

Open Mobility Foundation, Mobility Data Specification README.

[ISO 13482]

ISO 13482:2014, abstract and scope as published on iso.org.

[ANSI on ISO 13850]

ANSI Blog on ISO 13850:2015: a secondary description of the standard, which is single-user licensed and not linkable.

[NTSB p.N]

NTSB, Midair Collision Over the Potomac River (AIR-26-02, February 2026), page N.

[NTSB About]

NTSB, “About the NTSB.”

[FERC]

Federal Energy Regulatory Commission, Reliability Explainer.

[USDA FSIS]

USDA Food Safety and Inspection Service, About FSIS.

[FAA AIR]

FAA Aircraft Certification Service (AIR) page.

[FAA drones]

FAA, How to Register Your Drone.

[INFERENCE]

Inferred from the record; the pages it rests on follow.

The line that stopped me

Unitree Robotics recently introduced a humanoid robot, the H2-X "Kung Fu Legend Guardian," in a LinkedIn post. The description promised "powerful performance to serve and guard you, with a maximum endurance of approximately 8 hours." It ended with one sentence: "Please use the robot safely and in a friendly manner."

My first reaction was excitement. I am genuinely thrilled by the possibility space we are standing in right now, and when I see a video like that, my mind fills with ideas for how I would use the technology. Then came a moment of sobriety, because of that last sentence. It hands the safety job to the user. How will others use this? Do they understand what safety means? How will robotics like this be governed?

A robotics and AI strategist replied to my post with sharper questions than mine. What would meaningful safeguards actually look like: mandatory operator certification, geofencing, tamper-resistant emergency stops? Who should own them, manufacturers or regulators? And voluntary guidelines, he pointed out, have a poor track record in other industries. In a follow-up he pushed further. How do you make geofencing and emergency stops hold up for general-purpose humanoids in unstructured homes? Who audits that those controls haven't been disabled? Should certification attach to the operator, the robot, or both?

I don't have complete answers. I do have a way of thinking about them, and this essay is my attempt to lay it out.

The gap underneath

When I compress everything I'm worried about into one idea, it is this: there are very large gaps between what the technology can do and what the people who will put it to use understand.

When we say "the public," we tend to picture consumers, everyday people. But the gap is not only theirs. It is businesses, executives, and enterprises. It may even exist inside the labs building these systems, whose people are solving hard mathematical and scientific problems while carrying forward the work and research of many others.

Capability also fills gaps quietly. An agent can now do a task for you, and you may never notice that something changed. Then something like this robot arrives, so far beyond what most of the population understands that the gap becomes visible. At the current rate, I don't think that gap closes for years. And I don't think it closes by itself. It closes when the people, homes, businesses, and institutions this technology lands in catch up with it, and when the safeguards we are discussing exist.

I also don't think this is only a robotics problem. People say the world wasn't ready for large language models. I suspect it isn't ready for what comes after them either, and that "what comes after" is bigger than robots.

There is a related question about who builds this technology. The UN's Independent International Scientific Panel on AI reports that "in 2025, institutions based in the United States produced 59 notable AI models, compared with 35 in China and just 13 in the rest of the world." [UN p.16] Stanford's 2026 AI Index reports the same figures for the United States and China. [AI Index p.14, p.16] Notable AI models are not robots, and I don't want to stretch that data further than it goes. But if the capacity to build this technology is concentrated in a few places, then decisions about how it is governed, and about who can afford it, may be concentrated too. [INFERENCE — UN p.16 and AI Index p.16: the counts show where notable models were produced, not who decides how the technology is governed or priced]

There is always an accountable human

One principle anchors the rest of my thinking: there is always an accountable human. As far as we know, no machine can activate itself. Someone started it, configured it, deployed it, or sold it.

The grave danger is people believing that machines run themselves, because that belief lets responsibility evaporate. So the question is never whether a human is accountable. The question is which human, or which organization, and whether we have decided that in advance or will find out only after something goes wrong.

That reframes the question of who should own safeguards. It isn't manufacturers or regulators. It is a set of roles. Each of us has an obligation to the structure of governance depending on where we sit.

What a stop function is for

My first instinct about emergency stops was biometrics. If a robot can recognize its operator, I thought, then the operator can start it and stop it. As I worked through the idea, I realized I had the stop function backwards.

Picture a living room with two small children, four adults, and a robot moving through the space. A single remote control won't do, because it can be lost. A stop that only the right person can trigger will fail exactly when the right person isn't there. ISO 13850, the machinery standard for emergency stops, describes the function as available and operational at all times and activated by a single human action, and the machine must not be able to operate again until the function is reset. [ANSI on ISO 13850 — a secondary description of the 2015 edition; the standard is licensed to one user and cannot be linked] In a home, "a single human action" has to be available to whoever is there. [INFERENCE — ANSI on ISO 13850: the standard asks for a single human action; whose action it must be, in a home, is the author's inference]

So the logic runs the other way. Authentication belongs on who can enable, configure, or change the robot. Stopping it has to be easy for everyone.

That points to layers: a physical stop on the robot itself, a way for anyone in the room to trigger it, and a home environment that responds. The environment should be smart, but the stop cannot depend on the environment working. If the network or the sensors fail, the default should be to halt.

Biometrics also expose the trust problem I come back to below, because they ask people to hand over some of the most personal data they have.

What the scooters taught me

I have watched geofencing fail in a small way, and I learned more from it than I expected.

Years ago, when I worked for a city, an e-scooter company was coming to town. The president of the local university was wary of the rollout and did not want scooters lying around his campus. A lot of work went into geofencing: the scooters were supposed to stop working at the edge of a defined perimeter. It did not work perfectly. I remember being in the city manager's office when the university president called to ask why scooters were all over his campus.

Several lessons from that story apply to robots.

Responsibility was shared. Contractually, the company was responsible for the geofence. But the complaint came to the city, and the city and the university had to light the fire, create the urgency, and follow up. When you hire a contractor, you hope your role stays minimal, maybe a phone call or an email. In business, that is not always the case. You can delegate the work, but you cannot delegate the accountability.

A good stop is graduated, and it has to be a safe stop. Many scooter systems use slow zones, no-ride zones, and no-parking zones, so the response fits the setting. A scooter that cuts out in the middle of a ride can throw a rider. A legged robot that loses power may fall, possibly on someone. For a humanoid, a controlled stop is a harder design problem than a kill switch.

Boundaries are fuzzy. Location drifts near buildings, and indoors, satellite positioning is unreliable. A robot in a home would need another way of knowing where it is, plus conservative margins around anything that matters.

Setting the rules and following them can be separate jobs. The Mobility Data Specification, created by the Los Angeles Department of Transportation and now maintained by the Open Mobility Foundation, gives cities a standard interface to publish local rules for mobility operators and to analyze the data operators send back. [MDS README] The documentation describes policy information that "may be used to determine compliance." [MDS README — Policy API] In my scooter story, oversight came through a phone call. A data feed means the overseer can see problems without waiting for a complaint. [INFERENCE — MDS README describes the Policy API and data ingestion; what that spares an overseer is the author's inference]

Ownership changes everything. The scooters belonged to the company, which controlled them. A humanoid in a home belongs to the household. That changes who can disable a control, and it makes the question of who audits those controls much harder, because the party being protected from may also be the party holding the controls.

Operator, robot, or both?

I think the answer is both. When I think about the legal structure around purchasing a firearm, I don't mean the danger. I mean the parameters: who may purchase, what is registered, and where liability sits. Combine that with how we handle cars, where the vehicle is registered and the driver is licensed. The object is registered and the person is licensed.

I am less sure how that works for something that lives in a home. We already attach requirements to certain things people bring into their homes, and drones above a certain weight are registered too. [FAA drones — "All drones must be registered, except those that weigh 0.55 pounds or less"] But I don't know how this part plays out for a household robot, and I would want people with legal and regulatory expertise in that conversation.

What can actually be audited

Who audits is where I think the conversation needs the most care, because audit means different things depending on where oversight exists.

Imagine the consumer in this scenario. Something goes wrong in their home, and their humanoid hurts someone. They are sued, or they are in trouble. I suspect they will turn around and take issue with the manufacturer, and what happens next depends on how tight the warranties and agreements are. That is not an audit. Evidence in that situation is subpoenaed as part of legal proceedings. It happens after the harm, it concerns one incident, and it only works if records exist. If a manufacturer's logs of commands, stop events, and changes were never kept, there is nothing to subpoena.

Move up a level and you find a different kind of audit: ongoing oversight of producers, done before and between incidents. NERC, with the Federal Energy Regulatory Commission overseeing it, develops and enforces mandatory reliability standards for the power grid in the United States. [FERC — certified NERC as the Electric Reliability Organization in 2006; NERC develops standards "subject to FERC approval and oversight"; FERC must approve standards before they become mandatory and enforceable] The USDA inspects meat and poultry. [USDA FSIS — Food Safety and Inspection Service acts under the Federal Meat Inspection Act and the Poultry Products Inspection Act] The FAA certifies and oversees aircraft manufacturers [INFERENCE — FAA AIR: the page describes oversight of design, production and airworthiness certification of civil aviation products; "manufacturers" is the author's shorthand], while the National Transportation Safety Board, an independent agency, investigates accidents. [NTSB About] Separating the body that sets the standard from the body that investigates is a useful idea for robots.

The January 2025 midair collision near Reagan National Airport shows both sides of this. [NTSB p.xxi] The NTSB investigated and determined that the probable cause included the FAA's placement of a helicopter route close to a runway approach path, its failure to regularly review that route and the available data, and its failure to act on recommendations to reduce the risk of a midair collision. [NTSB p.299] An independent body could say that about the regulator because it does not answer to it. [INFERENCE — NTSB About: "congressionally mandated independence"] But the NTSB's own chair has said that its recommendations are voluntary, and that unless they are acted on they are "simply words on a page." [NTSB p.312] That is the gap I keep coming back to, between knowing a risk and anyone being required to close it.

An audit needs criteria, meaning standards to measure against, and an organization with the authority to enforce them. Standards for this field do exist. ISO 13482, for example, sets safety requirements for personal care robots, including mobile servant robots, though it excludes industrial robots and medical devices. [ISO 13482] What I could not determine is whether a general-purpose humanoid falls within it, or whether any regulator audits producers against it for use in homes. Without that, we are left with after-the-fact investigation, which is the weakest form of oversight.

If a regulator did audit producers, here is what I would want examined:

• whether the stop and zone controls exist and perform the way they were certified
• whether those controls were disabled or modified, and by whom
• how software updates are approved, since a robot's behavior can change after it is sold
• the testing evidence behind the safety claims
• how incidents are recorded and reported

A layered structure

When I try to picture what the structure of governance could look like, the most helpful visualization is layers. At the top, a baseline applies to anyone who engages in an activity: if you do this, you are governed by these provisions. Within each domain, more specific rules apply, and the specifics vary by domain the way they already do. Documentation standards in medicine differ from documentation standards in accounting, but an underlying structure runs through both, and it asks the same questions: is it complete, traceable, retained, and attributable to someone?

There are precedents. The European Union's AI Act sorts systems into tiers by risk. The United States has relied more on rules sector by sector. Neither is a perfect model, but both show that a baseline plus domain rules is workable.

I see two places where this kind of structure is most likely to fail.

The first is the trigger. If the baseline applies based on a capability, someone has to define where the line sits, and capabilities keep moving. A definition that is too narrow gets outpaced. One that is too broad catches things nobody intended.

The second is who gets heard. Leaders and researchers in each domain advocate for the specialized rules, the way we lobby for regulation now. That is how rules get the details right, and it is also how rules can end up shaped by the industry they are meant to constrain. That connects back to the point that voluntary guidelines have a poor track record in other industries. The real question is whether the people most affected, including households and disadvantaged populations, have a voice when the specifics are written.

The trust gap

Everything above runs into a trust gap. For safeguards like these to work, people usually have to give something up: data about their homes, their routines, and even their bodies. Some of us in leadership roles are not being responsible with our language, our actions, or our work. Some of us aren't doing enough education. Some of us aren't taking equity and disadvantaged populations seriously enough. A very large trust gap stands between those failures and people willingly sharing the data that safe environments would need.

That makes me wonder whether we can design safeguards that don't depend on closing that gap first, for example by keeping verification local to the home so that safety doesn't require surrendering data. I don't know if that is feasible, and I would like to hear from people who do.

What I know and what I don't

I can't tell you how to make a geofence hold in a living room. That is an engineering problem for people closer to the hardware than I am. I can't tell you which regulator should own home robots, and I'm not sure anyone can yet.

What I can say is this. A one-line instruction that tells users to behave safely is not a safeguard. Safety has to be designed in, with a stop that anyone can use, controls that can't be quietly disabled, records that can be examined, and a named, accountable human at every layer. Closing the gap between capability and readiness is the real work, and it cannot be left to the people on the least-informed end of it.

Sources

• Independent International Scientific Panel on AI, United Nations. Preliminary Report: Evidence-based assessment of opportunities, risks and impacts of artificial intelligence (July 2026).
• Stanford HAI. The 2026 AI Index Report, Chapter 1: Research and Development.
• Open Mobility Foundation. Mobility Data Specification (README).
• ISO. ISO 13482:2014, Robots and robotic devices: Safety requirements for personal care robots.
• ANSI Blog. ISO 13850: Safety of machinery, emergency stop.
• National Transportation Safety Board. About the NTSB.
• Federal Energy Regulatory Commission. Reliability Explainer.
• U.S. Department of Agriculture, Food Safety and Inspection Service. About FSIS.
• Federal Aviation Administration. Aircraft Certification Service (AIR).
• Federal Aviation Administration. How to Register Your Drone.
• National Transportation Safety Board. Midair Collision Over the Potomac River, Washington, DC, January 29, 2025 (AIR-26-02, February 2026). Probable cause, p. 299; Chairwoman Homendy's concurring statement, pp. 308–313.