Adoption Barriers: Technology, Trust, and Regulation

In 1865, the British Parliament passed the Locomotive Act, one of the first pieces of legislation ever written to govern a self-propelled machine. Its most famous provision required that any horseless vehicle traveling on a public road be preceded, at all times, by a person walking at least 60 yards ahead, waving a red flag to warn pedestrians and horse-drawn carriages of the approaching danger. Vehicles were limited to 4 miles per hour on open roads, 2 miles per hour in towns. The law stayed on the books until 1896, more than 30 years later. To modern eyes, the image is absurd: a man on foot, flag in hand, leading a machine that was built to move faster than he could walk. Yet the lawmakers who drafted the Locomotive Act were doing the only thing they could. They had no framework for a horseless vehicle, so they borrowed from the world they knew, horse-drawn transport, and stretched those rules to fit a technology that would soon make them irrelevant. The automobile had arrived. The institutions needed to govern it had not.

What followed in America was closer to anarchy than progress. In 1900, fewer than 10,000 cars existed in the entire world. A decade later, over 130,000 cars, 35,000 trucks, and 150,000 motorcycles were already on American roads, and almost nothing existed to manage them. There were no stop signs. No traffic lights. No lane markings. No speed limits in most states. No driver's licenses worth the name. Missouri required a license as early as 1901, but did not require drivers to pass an actual test until 1952. Connecticut became the first state to set a speed limit, 12 miles per hour in cities, but as late as 1930 a dozen states still had none. The first stop sign appeared in Michigan in 1915. By then, pedestrians were already dying in numbers that would shock the public into action. Automobile fatalities in the United States climbed from 36 in 1900 to over 1,500 by 1920, and past 12,000 by 1929. In Detroit, the Safety Council had church bells, school bells, and City Hall bells ring twice a day in memory of traffic victims. Teachers read the names of dead children aloud in classrooms. Newspapers printed "murder maps" showing where people had been killed. Wrecked cars were towed down Woodward Avenue with placards reading "Follow this one to the cemetery." Courts held serious debates about whether the automobile was inherently evil. The machines worked fine. Everything around them was missing.

The standards, insurance, licensing, liability, and safety engineering that make the automobile unremarkable today took decades to assemble, and almost every piece came after tragedy rather than before it. The first auto insurance policy was sold by Travelers in 1897 or 1898, covering a single driver in Dayton, Ohio, for $12.25. The policy was adapted from "teams" liability coverage written for horse-drawn vehicles; the insurer simply crossed out the word "team" and wrote "auto." Massachusetts did not make automobile insurance mandatory until 1927, and for over 30 years it remained the only state in the country to require it. Seatbelts told the same story, only slower. The first seatbelt patent was filed in 1885, three years before Karl Benz patented the automobile itself, yet when Ford offered seatbelts as an option in the 1950s, only 2 percent of buyers paid for them. Volvo engineer Nils Bohlin invented the three-point seatbelt in 1959, and the company made the patent free for all manufacturers to use, a decision Volvo later estimated saved more than a million lives. The federal government did not make seatbelts mandatory in new cars until 1968. Even then, by the mid-1980s, only 10 percent of American drivers actually wore them. It took Ralph Nader's 1965 book "Unsafe at Any Speed," a congressional investigation, and the passage of the National Traffic and Motor Vehicle Safety Act in 1966 to give the federal government real authority over vehicle safety. From the Model T rolling off Henry Ford's assembly line in 1908 to something resembling a coherent regulatory framework, roughly 60 years had passed. The institutions moved at the speed of consensus, litigation, and grief.

Humanoid robots are now entering a similar void. Tesla plans to produce 5,000 Optimus robots in 2025 and at least 50,000 in 2026. Figure AI sees a path to 100,000 units by 2029. Morgan Stanley Research projects more than one billion humanoid robots by 2050, part of a $5 trillion market. The money is real. The engineering talent is real. The production timelines are aggressive. And yet almost nothing exists to govern how these machines will operate among people. The emergency stop button, the simplest and most universal safety device in industrial robotics, shows how wide the gap is. On a traditional robot arm, pressing the e-stop kills power and the machine freezes in place. On a humanoid, cutting power means the robot collapses. A 60-kilogram machine with no ability to brake its own fall becomes a falling object, not a safe state. As Melonee Wise, Chief Product Officer at Agility Robotics, put it at Automate 2025: "We are silent with regards to safety. We don't even have the basic, that easy big red button e-stop capability that brings a robot to a stop."

The organizations that write safety standards know this, and they have started working. The International Organization for Standardization published a working draft in May 2025, ISO 25785-1, the first standard aimed at what it calls "industrial mobile robots with actively controlled stability," a category that covers bipedal humanoids, quadrupeds like Boston Dynamics' Spot, and wheeled balancing robots. The working group is led by representatives from Agility Robotics, Boston Dynamics, and the Association for Advancing Automation. But the standard covers only industrial environments, and only the robots themselves, not their integration into applications. Homes, hospitals, retail floors, and restaurants have no humanoid-specific standard even in development. The IEEE Humanoid Study Group published a strategic framework in late 2025, assembling more than 60 experts from industry, academia, and regulatory bodies to map the gaps. Their report identified three areas that need to be addressed before humanoids can deploy safely at scale: classification (what counts as a humanoid, and what rules apply to it), stability (how to measure and mitigate fall risk), and human-robot interaction (how to ensure predictable behavior around untrained people). ASTM International is building a companion classification framework that would sort humanoids along five axes: physical capabilities, behavioral intelligence, operational context, stability profile, and level of human contact. None of these frameworks are finished. The IEEE report is a roadmap, not a standard. ISO 25785-1 is a working draft with a typical three-to-four-year path to ratification, and Aaron Prather, who leads ASTM's robotics standards program, estimates 18 to 36 months before the first ratified standards based on this collective work are published. That puts enforceable, internationally recognized humanoid safety standards somewhere around 2027 or 2028. Until then, manufacturers are building, investors are funding, and customers are piloting, while the most basic questions, how to safely stop a walking robot, who is liable when it falls on someone, what level of autonomy requires what level of oversight, have no settled answers.

Humanoid robots will close this institutional gap faster than the automobile did. The pace of standardization is quicker now, the organizations are coordinating earlier, and the lessons of the autonomous vehicle sector, which spent a decade learning that a demo is not a deployment, are there for anyone willing to absorb them. But the comparison to the automobile also falls short in one important way. A car in 1910 did what its driver told it to do. It had no trained behavior, no capacity to act on a decision its operator had not made. A humanoid robot operating on learned AI behaviors is something else entirely. When it moves through a warehouse, picks up a box, steps around an obstacle, or reaches toward a shelf, it is executing decisions shaped by training data, neural networks, and policy models that no single engineer fully understands and no legal framework knows how to assign responsibility for. The licensing, insurance, liability, and safety standards that the automobile eventually received all rested on a straightforward assumption: a human operates the machine, and the human is accountable. That assumption breaks down the moment a robot acts on behaviors trained by engineers on another continent. Every institutional layer that humanoid robots need, safety standards, insurance products, liability precedent, business models that align the interests of manufacturers, operators, and workers, must be built not just quickly, but differently. The projections from Tesla, Figure AI, and Morgan Stanley all rest on an unstated belief: that these foundations will materialize on a timeline that matches the production ramp. The automobile's history, and the harder problem of machine autonomy, suggest they will not.

The Regulatory Void#

The standards bodies are working. But standards alone do not create a regulatory regime. They must be adopted by governments, embedded in law, referenced by courts, and enforced by agencies. And on this front, the world's three major economic blocs are moving in very different directions, at very different speeds, with very different ideas about what regulation is for. Europe is regulating before the industry has arrived. China is building the industry and writing the rules at the same time. The United States is doing neither. Each approach carries its own risks, and none of them is producing the one thing that would actually help: a coherent international framework that would let a humanoid robot certified in one jurisdiction operate in another.

Europe has moved first, and it has moved with characteristic thoroughness. The EU AI Act, adopted in June 2024 and phasing in through August 2027, classifies AI systems by risk level and imposes strict obligations on those deemed high-risk. A humanoid robot with self-evolving AI behaviors operating as a safety component of machinery falls squarely into the high-risk category, triggering requirements for conformity assessment, risk management documentation, human oversight provisions, and explainability of AI-driven decisions. Running alongside the AI Act is the new EU Machinery Regulation, which takes full effect in January 2027, replacing a directive from 2006 that was written before autonomous mobile robots existed in any meaningful commercial form. The updated regulation introduces "high-risk machinery products," a category that includes machines using self-evolving AI algorithms where outcomes may not be fully predictable at design time. These machines will require third-party conformity assessment rather than self-declaration. Add the Cyber Resilience Act, which mandates lifetime cybersecurity protections for any product with a digital element, and the revised Product Liability Directive, which now explicitly includes software and AI, and you have a tripartite regulatory stack that is, on paper, the most comprehensive in the world. The problem is that Europe has been here before. GDPR was supposed to make Europe the gold standard for data privacy; instead, it became a compliance burden that advantaged large American tech platforms with the legal departments to absorb it, while European startups struggled under the weight. The pattern is familiar: Europe regulates, the rest of the world innovates, and by the time the rules are clear, the market has been won elsewhere. The humanoid sector risks the same outcome. While Brussels layers conformity assessments, documentation requirements, and phased implementation deadlines across three overlapping regulations, Chinese manufacturers are shipping robots and American startups are iterating in the field. Europe has no equivalent of China's government-brokered pilot programs that place humanoids directly into state-owned factories, and no equivalent of the permissive American environment where companies can test, fail, and adapt without waiting for a regulatory green light. The EU framework has no serious sandbox mechanism, no controlled space where humanoid companies could deploy robots under supervised conditions, collect real-world performance data, and feed those findings back into the standards process. Without sandboxing, European companies are stuck. They need real-world deployment data to demonstrate compliance, but the compliance requirements make it difficult to deploy in the first place. The certifiable path to market that Europe offers is real, and companies that meet EU requirements will carry credibility globally. But certifiability means little if the companies that need it most cannot afford the years and the millions of euros required to obtain it, while competitors in Shenzhen and San Francisco are already learning from deployed robots on factory floors.

China has taken a different path. Where Europe leads with regulation, China leads with industrial policy, and the gap between the two approaches is widening fast. The Ministry of Industry and Information Technology published its Guiding Opinion on the Innovation and Development of Humanoid Robots in November 2023, calling for China to establish a preliminary innovation system by 2025 and to deeply integrate humanoid robots into the real economy by 2027. In March 2025, embodied AI appeared for the first time in the Government Work Report, China's most authoritative policy document, signaling top-level political commitment. The effects are hard to miss. Chinese firms are projected to manufacture more than 10,000 humanoid robots in 2025, accounting for over half of global output. Investment in robotics and embodied intelligence in the first seven months of 2025 reached $3.4 billion, 42 percent more than the United States and five times Europe's total. Six of the country's 11 major humanoid manufacturers launched mass-production initiatives in 2024, aiming for over 1,000 units per year each. UBTECH won a $12.7 million single-company order from a Shanghai-based EV startup, the largest known humanoid robot procurement deal at the time. The standardization effort is keeping pace with production. In November 2025, the MIIT published a proposed roster for a new Standardization Technical Committee for Humanoid Robots, naming 65 members drawn from government, leading startups like Unitree and AgiBot, and institutions including Tsinghua University. The founders of the country's top robotics firms were appointed as deputy chairs, a deliberate signal that Beijing wants the people building robots to write the rules for them. China's first national standards, covering perception, decision-making, motion control, and task execution, were approved for development in early 2025. The approach is unmistakable: build first, standardize alongside, and use national standards to consolidate a domestic industry before international frameworks are finalized. There is an obvious logic to this. Real-world deployment generates the data and experience that standards need to be grounded in reality rather than theory. But the pressure to hit production targets set by industrial policy can crowd out the slow, unglamorous work of testing edge cases, documenting failure modes, and establishing the kind of safety evidence that customers outside China will demand. State-brokered pilot programs in state-owned factories create a protected environment where safety shortcuts may go unreported and where commercial incentives are distorted by political ones. And national standards written by domestic manufacturers for domestic conditions may not translate well to international markets, particularly in Europe, where conformity assessment is about to get much stricter. China is building the largest humanoid fleet in the world. Whether those robots will be trusted, purchased, and insured outside China depends on safety evidence that mass production alone cannot provide.

The United States, the country that incubated most of the world's leading humanoid startups, has done almost nothing at the federal level to prepare for what those startups are building. OSHA still lacks any robot-specific regulations. No federal humanoid legislation has been introduced. There is no national standards body dedicated to humanoid safety, no certification pathway, no government-brokered pilot program. The country relies on voluntary consensus standards and product liability law, which in practice means the rules are written by courts after someone gets hurt rather than by agencies before deployment begins. Defenders of this approach argue that regulatory restraint is what allows American companies to move fast, attract capital, and iterate freely. And to a point, they are right. Figure AI, Agility Robotics, Apptronik, and Tesla's Optimus program all benefit from an environment where they can test and fail without navigating a European-style compliance maze. But the absence of a framework is not a deliberate strategy. It is a vacuum. Individual states will likely begin introducing their own humanoid legislation, and the variation will be wide. Innovation-friendly states like Texas and Nevada will take a different approach from California or New York. The result will be a domestic patchwork that forces companies to manage multiple compliance regimes within a single country, the kind of fragmentation that the EU was specifically designed to prevent. More damaging still is what the vacuum means internationally. An American company that has never had to meet any formal safety standard will find itself locked out of Europe after 2027, when the Machinery Regulation and AI Act take full effect. A Chinese manufacturer with national standards behind its robots will have a stronger story to tell global customers than an American competitor whose only credential is that nobody told them they couldn't ship. The country with the most advanced humanoid technology may end up with the weakest hand at the table when international standards are negotiated, simply because it never bothered to develop a domestic position. For the global humanoid sector, the three-way fragmentation creates a problem that is more than bureaucratic. A humanoid designed for the EU market carries documentation and safety features that are unnecessary in China and nonexistent in the US. A robot built to Chinese national standards may fail European conformity assessment. An American robot with no formal certification may be the most capable machine in the room and the hardest to deploy outside its home market. Until these regimes converge, or at least develop mutual recognition, the regulatory landscape is fractured along the same geopolitical lines that already divide the global technology economy.

The Insurance Puzzle#

Insurance is the business of pricing risk. Actuaries build their models on historical data: how often a particular event occurs, under what conditions, and at what cost. For automobile insurance, more than a century of accident data, across billions of miles driven, millions of claims filed, and tens of thousands of court rulings, feeds models that can price a policy for a 22-year-old male driver in Houston, Texas, to within a few dollars of the expected loss. For humanoid robots, that data does not exist. No insurer in the world has a loss history for a bipedal machine that walks through a warehouse, picks up objects, and makes decisions shaped by neural networks trained on synthetic data. The closest analogue is industrial robotics, where the risk profile is well understood because the machines are bolted to the floor, separated from workers by cages or light curtains, and programmed to repeat the same motion thousands of times. A humanoid shares almost nothing with that profile. It moves freely. It operates near people. Its behavior changes as its AI models are updated, sometimes over the air, sometimes without the operator knowing exactly what changed. An insurer trying to price a humanoid policy has no frequency data, no severity data, and no established liability precedent. The technology is changing quarter to quarter, and the actuarial tables are blank.

Even if an insurer could price the risk, the liability chain for a humanoid accident has no settled answer. A logistics company leases a humanoid robot from a manufacturer. The robot runs software built on a foundation model licensed from a separate AI company. The foundation model was fine-tuned by the manufacturer using training data collected in Chinese factories and American warehouses. A third-party integrator configured the robot for the specific warehouse where it operates. One morning, the robot miscalculates a step, falls forward, and strikes a human worker, breaking the worker's arm. Is the manufacturer liable, for building a robot that fell? The AI company, whose foundation model generated the movement plan? The integrator, who configured the robot for that environment? The logistics company, which chose where to deploy it? In traditional product liability, the manufacturer bears responsibility for a defective product. But a humanoid's behavior is not a fixed product attribute. It emerges from the interaction of hardware, software, training data, and deployment conditions that no single party controls. The EU's revised Product Liability Directive, which took effect in 2024 and now explicitly covers software and AI, creates a framework where any party in the chain could be held liable if their contribution to the product is defective. American courts, without equivalent legislation, will work through these questions case by case, building precedent the way automobile liability was built: expensively, and after people have already been hurt.

The first humanoid-specific insurance products are appearing anyway, and China is leading. China Pacific Insurance launched what it called the country's first dedicated humanoid robot insurance product in September 2025, a policy named Ji Zhi Bao that covers the full commercial chain from production and sales to leasing and end-use. It bundles property damage coverage for the robot itself with third-party liability for injuries or damage the robot causes. PICC Property and Casualty followed weeks later with a similar product combining body loss insurance and third-party liability that explicitly covers natural disasters, accidental damage, cybersecurity incidents, pedestrian injuries from delivery robots, and medical accidents involving robotic surgery. Ping An rolled out a comprehensive financial solution in November 2025 that integrates insurance with credit and IPO services for robotics firms. In November, Huazhong University of Science and Technology purchased what was described as the first embodied-intelligence robot insurance policy in Hubei province, covering two 60-kilogram humanoids at a premium of about $700 per robot per year, with a maximum payout of roughly $69,000 per incident. The maintenance cost for a single humanoid repair can run from $4,000 to more than $40,000, which means the policy covers roughly one and a half major repairs before the limit is exhausted. In the United States, the picture is thinner. Koop Technologies, a Pittsburgh-based insurtech founded in 2020 and backed by Hyundai, has built a platform for underwriting autonomous vehicles and robotics using real-time telemetry data pulled directly from robot fleets via API. Koop launched the first robotics-specific errors and omissions policy through Lloyd's of London in 2022, and it is licensed in more than 30 states. But these are commercial products for fleet operators and developers. The scenario approaching fastest is different: a consumer who buys a $16,000 Unitree G1 or a $4,900 Unitree R1 and brings it into a home where no commercial policy applies and no homeowner's insurance was written with a walking robot in mind.

Autonomous vehicles offer some guidance on what comes next, and the guidance is not encouraging. More than a decade into the deployment of self-driving cars, insurance remains one of the sector's unresolved problems. Liability has shifted unevenly from drivers toward manufacturers and software providers, but no jurisdiction has a clean framework for apportioning fault when an algorithm makes a decision that a human would have made differently. Repair costs for autonomous vehicles run roughly double those of conventional cars because of the sensors, cameras, and computing hardware built into every panel. Waymo, Cruise, and other operators have largely self-insured or negotiated bespoke policies with specialty carriers, arrangements that work for a handful of fleets but do not scale to millions of units. Humanoid robots will inherit all of these problems. They will also bring new ones. A self-driving car operates on roads with lanes, signals, and maps. A humanoid operates in a warehouse, a hospital corridor, or a living room where the environment changes every time someone moves a chair. A car stays on the road. A humanoid can fall, and a 35-kilogram machine toppling onto a person can break bones. In September 2025, security researchers published that the Unitree G1 collects and transmits sensor data without notifying the operator, and disclosed a wormable vulnerability that allows an attacker in close physical proximity to gain full control of Unitree's quadruped and humanoid robots over Bluetooth. An infected robot could in turn compromise other robots nearby. No insurance product on the market today covers a scenario in which a hacked humanoid injures someone in a private home. Until insurers can price the layered risks of hardware failure, software defect, cybersecurity breach, and unpredictable human-robot interaction, the gap in coverage will slow adoption where it matters most: consumer markets and healthcare, where a single injury lawsuit can destroy a company's market access overnight.

The Reality Check#

The automobile took 60 years to acquire the institutional framework it needed. Humanoid robots will not get 60 years. The production ramps are too aggressive, the capital investments too large, and the competitive pressure between nations too intense for the industry to wait while standards bodies deliberate, insurers gather loss data, and courts build precedent one lawsuit at a time. But the alternative, deploying millions of autonomous machines into homes, hospitals, and workplaces without enforceable safety standards, clear liability rules, or insurance products that can price the risk, is not a shortcut. It is a way to guarantee the backlash that kills an industry. The autonomous vehicle sector learned this lesson expensively. Uber's self-driving test vehicle killed a pedestrian in Tempe, Arizona, in 2018. Cruise pulled its fleet from San Francisco after a robot taxi dragged a pedestrian 20 feet in 2023. In both cases, the technology worked most of the time. What failed was everything around it: the oversight, the accountability, the institutional willingness to admit that moving fast had outrun the ability to move safely. Humanoid companies watching those outcomes should be taking notes.

The institutional work ahead is unglamorous. Writing standards, negotiating liability frameworks, building actuarial models from scratch, designing insurance products for risks that have never been priced before: none of this makes for a compelling investor pitch or a viral demo video. But every projection of a multi-trillion-dollar humanoid market rests on the assumption that this work will get done. Tesla's plan to produce tens of thousands of Optimus robots, Figure AI's path to 100,000 units, Morgan Stanley's vision of a billion humanoids by 2050, all require that someone, somewhere, builds the institutional scaffolding that lets a robot certified in Shenzhen operate in Stuttgart, that gives a warehouse operator in Ohio a policy to cover the machine working alongside their employees, and that tells a court in any jurisdiction who bears responsibility when something goes wrong. The three-way regulatory fragmentation between Europe, China, and the United States is not a temporary inconvenience. Unless it converges, it becomes the defining constraint on the global humanoid industry, more limiting than battery life, more limiting than dexterity, more limiting than the cost of actuators.

And yet the prize for getting this right is larger than the market projections suggest. For two centuries, industrial civilization has systematically organized human beings around the needs of machines, converting people into predictable, replaceable, optimizable components. Humanoid robots offer the possibility of reversing that relationship: machines that conform to human environments, perform the work we mechanized ourselves to do, and free people to do what machines cannot. That possibility is real. But it lives on the far side of the institutional work this chapter describes. Without safety standards, the robots cannot be trusted. Without insurance, they cannot be deployed at scale. Without liability frameworks, no company will risk putting them in a home or a hospital. The utopian potential of humanoid robots is not a fantasy. It is an engineering and institutional problem, and the engineering is the easier half.*