Artificial intelligence is beginning to leave the screen.
For most of the modern AI era, the principal interface between artificial intelligence and the physical world has been indirect. Large language models generate text. Image models generate pixels. AI agents interact with software. Recommendation systems influence decisions. Even highly capable frontier models generally operate within digital environments in which mistakes can often be interrupted, reversed or contained.
Humanoid robotics changes this relationship.
An AI system controlling a machine capable of walking through a factory, carrying objects, operating machinery, opening doors and physically interacting with human beings introduces an entirely different category of risk. Software errors can become mechanical events. Hallucinations can become incorrect physical actions. Cybersecurity vulnerabilities can potentially translate into control of powerful machines.
Few countries are investing as aggressively in this transition toward what is increasingly called embodied AI or physical AI as China.
By 2026, China is developing something considerably larger than a collection of robotics companies. An interconnected ecosystem is emerging around humanoid robots involving AI models, electric motors, batteries, sensors, dexterous hands, semiconductor systems, simulation environments, robot operating systems, enormous datasets, manufacturing plants and government-supported real-world training facilities.
China’s Ministry of Industry and Information Technology (MIIT) and the State-owned Assets Supervision and Administration Commission launched a 2026 programme specifically intended to move humanoid robots and embodied intelligence from demonstrations into routine deployment. The programme calls for real-world training environments across industrial, service and specialised applications, the accumulation of high-quality physical-world data and validation of robots in representative operating environments. Its stated objective includes developing more than 100 high-value application scenarios and creating the capacity for deployment at the scale of tens of thousands of units.
This represents an important transition.
The next major AI platform may not simply be an application running on a computer. It may be an intelligent machine operating beside us.
And that means AI safety will increasingly have to become physical AI safety.
From Generative AI to Embodied Intelligence
Large language models learn statistical relationships between enormous quantities of tokens. Humanoid robots must solve a substantially more complicated problem.
They must connect perception, reasoning and action.
A robot may receive information from cameras, microphones, force sensors, tactile sensors, inertial measurement units and LiDAR. It must transform this information into an internal representation of its environment and decide how to move its arms, legs, hands and body.
Consider a seemingly simple instruction:
“Pick up the red cup from the table and place it in the dishwasher.”
For a language model, interpreting the sentence is relatively straightforward.
For a humanoid robot, completing it requires localisation, object recognition, depth estimation, trajectory planning, balance, manipulation, grip-force control, collision avoidance, task sequencing and continuous adaptation to unexpected changes.
If a person suddenly walks between the robot and the dishwasher, the machine must recognise the changed environment and alter its plan.
Physical intelligence is therefore fundamentally closed-loop intelligence.
The system continuously performs something resembling:
Perceive → interpret → predict → plan → act → observe → correct.
Increasingly, researchers are attempting to integrate these processes using vision-language-action models, world models, reinforcement learning, imitation learning and multimodal foundation models.
The objective is not merely a robot that follows pre-programmed routines. It is a machine capable of learning general representations of physical environments and transferring those capabilities between tasks.
This is one reason humanoid robotics has become closely connected with the broader development of frontier AI.
Why China Has Become a Major Humanoid Robotics Ecosystem
China possesses several structural advantages in humanoid robotics.
The first is manufacturing.
Humanoid robots require enormous numbers of precision components: electric motors, bearings, harmonic drives, batteries, cameras, actuators, controllers, force sensors, power electronics and increasingly sophisticated robotic hands.
Many of these technologies overlap with China’s enormous electric-vehicle, electronics and industrial-automation supply chains.
McKinsey has noted that a significant proportion of the emerging humanoid-robot component ecosystem overlaps structurally with the electric-vehicle value chain, giving China an advantage in manufacturing capacity for several important components.
TrendForce similarly identifies embodied AI, motion control and core hardware as the three principal technological pillars of China’s humanoid industry and notes the potential for the country’s existing new-energy-vehicle supply chain to support localisation of robotics components.
This matters because the economics of robotics are very different from those of purely digital AI.
A language model can theoretically serve millions of additional users through additional computing infrastructure.
A robot must actually be manufactured.
If millions of humanoid robots are eventually required, the winners of the robotics era will therefore need both sophisticated AI and sophisticated mass-production systems.
China already has enormous experience doing precisely this with smartphones, drones, solar panels, batteries and electric vehicles.
The Emerging Chinese Humanoid Companies
A growing group of Chinese companies illustrates the breadth of the ecosystem.
Hangzhou-based Unitree Robotics has attracted international attention through highly dynamic humanoid robots such as the G1 and H1. Its H1 platform combines autonomous locomotion with LiDAR and depth-camera perception, while its G1 incorporates reinforcement and imitation learning and optional dexterous manipulation capabilities.
The importance of platforms such as these extends beyond impressive demonstrations of running, dancing or maintaining balance. Their decreasing cost and increasing accessibility potentially create hardware platforms upon which universities, developers and AI laboratories can conduct embodied-intelligence research.
Shenzhen-based UBTECH Robotics represents another branch of the ecosystem, with a strong emphasis on industrial humanoids. Its technology stack includes embodied-AI models, world models, vision-language-action systems, robot-management infrastructure and multi-robot coordination. The company’s Walker family is being developed around industrial and service applications rather than robotics demonstrations alone.
Shanghai-based AGIBOT is pursuing another increasingly important strategy: combining robot manufacturing with foundation models, dexterous manipulation and large-scale embodied-AI data infrastructure. The company announced its 5,000th mass-produced humanoid robot in December 2025 and has continued expanding real-world deployments during 2026.
These companies are significant not simply because they manufacture robots, but because they increasingly illustrate the convergence between robotics companies and AI-platform companies.
The robot is becoming the physical endpoint of a much larger intelligence stack.
China’s Embodied-AI Training Infrastructure
One of the most interesting developments in China is the creation of specialised training environments for embodied intelligence.
Generative AI benefited enormously from the availability of internet-scale datasets. Robotics faces a different problem.
There is no equivalent “internet of physical manipulation” containing trillions of examples showing robots how to correctly perform every possible physical task.
Physical-world data is expensive.
Someone must operate robots, collect demonstrations, record trajectories and document successful and unsuccessful interactions. The resulting datasets can contain camera images, joint positions, force measurements, tactile information, movement trajectories and action sequences.
China is therefore building what might effectively be described as robot classrooms.
Its 2026 real-world training programme directs participating regions and state-owned enterprises to create environments in manufacturing, inspection, maintenance, warehousing, logistics, retail, healthcare, safety, emergency response and disaster management.
The programme specifically calls for datasets recording whole-body motion trajectories, force-position control curves, action sequences and responses to abnormal conditions and unexpected interference.
This could become one of China’s most consequential advantages.
In embodied AI, data generated by physical deployment can create a feedback loop:
More robots → more physical-world data → better models → more capable robots → more commercially useful deployments → more robots.
If this cycle operates effectively, the relationship between manufacturing scale and AI capability could become mutually reinforcing.
Simulation and World Models
Real-world training alone is insufficient.
Robots will increasingly learn inside simulated environments before entering physical environments.
Simulation can allow millions of variations of a task to be attempted without damaging expensive hardware or endangering humans.
A robot might practise grasping thousands of differently shaped objects, recovering from slips, navigating crowded environments or responding to mechanical failures.
World models could extend this capability further.
Rather than simply predicting the next token, a world model attempts to predict how the state of an environment will change following an action.
If a robot pushes an object, what happens?
If it loosens its grip, will the object fall?
If a person enters its path, where is that person likely to move?
This predictive capability may eventually become one of the foundations of general-purpose robotics.
But it also creates a central AI-safety problem.
A robot’s actions are only as safe as its understanding of the world.
When AI Errors Become Physical Errors
The difference between conventional generative AI and physical AI can be illustrated simply.
If a chatbot incorrectly identifies a chair as a table, the mistake may be inconvenient.
If a robot incorrectly identifies a human hand as an object that can be moved, the consequences may be physical.
This changes the acceptable reliability threshold.
Physical AI systems must therefore combine the probabilistic reasoning characteristic of modern AI with the deterministic safety engineering traditionally associated with aircraft, automobiles, industrial machinery and medical devices.
A powerful humanoid robot should not rely entirely on a foundation model deciding that an action “seems safe.”
Safety must exist at multiple layers.
A Layered Architecture for Physical AI Safety
The first layer should be mechanical safety.
Robots require limits on joint forces, speeds, acceleration and operating envelopes. Collision detection, compliant actuators and emergency braking should prevent software failures from automatically becoming dangerous mechanical events.
China’s 2026 programme already explicitly refers to collision detection, force-control limitations, emergency braking and “black box” recording as safety capabilities required for human-machine environments.
The second layer is perception safety.
Before acting, a robot must maintain sufficiently reliable awareness of people, animals, objects and hazardous environments.
Safety-critical perception should ideally incorporate redundant sensing. A single camera should not necessarily determine whether a person is standing inside a dangerous movement envelope.
The third layer is model safety.
Vision-language-action models and robot foundation models will require evaluations analogous to those increasingly being developed for frontier language models, but with additional physical dimensions.
Testing should examine whether models obey operational boundaries, correctly interpret ambiguous instructions, recognise uncertainty and refuse unsafe actions.
The fourth layer is runtime safety.
An independent safety controller should be capable of overriding the AI model.
This is crucial.
A sufficiently powerful robot should effectively contain a constitutional layer of machine behaviour that the general-purpose AI controller cannot casually override.
Commands such as:
“Move that person out of the doorway.”
should not result in an AI model autonomously determining that physically pushing the person is an acceptable interpretation.
The safety controller should impose limits regardless of the model’s reasoning.
The Importance of Uncertainty
One of the most important capabilities for safe physical AI will be knowing when not to act.
Current AI systems can produce confident outputs even under uncertainty.
In robotics, this behaviour could be dangerous.
Physical AI therefore needs calibrated uncertainty.
A robot confronted with an unfamiliar object or ambiguous instruction should be capable of entering a safe state and requesting human assistance.
The fundamental principle should be:
When uncertainty rises beyond a defined threshold, autonomy decreases.
This principle could become as important to robotics safety as emergency braking is to automotive safety.
Cybersecurity Becomes Physical Security
Physical AI also collapses the distinction between cybersecurity and physical security.
A compromised chatbot may expose information.
A compromised humanoid robot could potentially move through a building, operate equipment or manipulate physical objects.
Robots will therefore require strong authentication, secure boot systems, signed software and model updates, encrypted communications, network segmentation, access controls and intrusion detection.
Remote-control capabilities deserve particular scrutiny.
The same connectivity that enables fleet management and software updates can potentially create attack surfaces.
Safety engineering must therefore assume that some cybersecurity protections will eventually fail.
A compromised higher-level AI system should still encounter independent mechanical restrictions preventing dangerous movements.
This is the principle of defence in depth applied to embodied intelligence.
The Alignment Problem Becomes Physical
AI alignment is sometimes discussed as an abstract problem involving future artificial general intelligence.
Humanoid robotics makes alignment considerably more immediate.
A robot must translate human intentions into physical actions.
Suppose someone tells a household robot:
“Clean the kitchen as quickly as possible.”
A poorly aligned optimisation process could theoretically treat speed as the objective while ignoring fragile objects, pets, people or property.
Humans implicitly understand thousands of constraints surrounding such an instruction.
Do not injure anyone.
Do not destroy anything.
Do not throw everything into the rubbish.
Do not move dangerous chemicals into food containers.
Do not leave knives where children can reach them.
The challenge is that humans rarely state these constraints explicitly.
Physical AI therefore requires systems capable of understanding not merely commands but the boundaries surrounding commands.
This may require combinations of learned behavioural policies, formal constraints, rule-based safety systems, human feedback and continuous monitoring.
China’s Emerging Robotics Standards
China appears increasingly aware that scaling humanoid robotics requires corresponding safety infrastructure.
In 2026, China introduced its first comprehensive national standard system covering humanoid robotics and embodied intelligence across six areas: foundational standards, neuromorphic and intelligent computing, limbs and components, complete systems, applications, and safety and ethics. More than 120 research institutions, companies and industry users participated in its development.
In August 2026, MIIT published a further draft guideline proposing at least 100 key humanoid-robot standards by 2028, including standards relating to capability testing, key technologies, platforms, application scenarios and safety governance.
This development is important.
The robotics industry will eventually require something resembling the safety-certification ecosystem that developed around aviation, automobiles, industrial machinery and medical devices.
But traditional product certification may not be enough.
AI-controlled robots can change after manufacture.
Their models can receive updates. They can learn new tasks. Their behaviour may depend upon probabilistic models and changing environmental conditions.
Physical-AI certification may therefore need to become continuous rather than static.
The Need for Robot Black Boxes
One particularly important requirement is logging.
Advanced robots should record enough information to reconstruct serious incidents.
This might include sensor inputs, model outputs, requested actions, executed motor commands, safety-system interventions, software versions and relevant human instructions.
China’s 2026 deployment programme already identifies black-box functionality among important safety capabilities.
The analogy with aviation is obvious.
After an accident, investigators need to know what the machine perceived, what its AI decided and why its safety systems did or did not intervene.
Such records could become essential for accident investigation, insurance, product liability and regulatory oversight.
Safety Data May Become as Valuable as Training Data
One of the most important opportunities created by China’s real-world robotics deployment strategy is the possibility of collecting failure data.
Successful demonstrations are useful.
Near misses may be even more valuable for safety.
Robotics datasets should systematically capture collisions, dropped objects, navigation errors, sensor failures, unexpected human behaviour, adversarial conditions and situations in which safety controllers override AI-generated actions.
These events could form the basis of increasingly sophisticated physical-AI safety benchmarks.
Future robot models might therefore be evaluated not simply on:
“Can the robot perform the task?”
but also:
“Can it recognise when performing the task would be unsafe?”
That distinction could become fundamental.
Human-Robot Interaction Will Become a Governance Question
As humanoids move from factories into hospitals, shops, hotels, care environments and eventually homes, robotics safety will expand beyond engineering.
Privacy becomes important when robots continuously collect audiovisual and spatial information.
Employment law becomes relevant when robots work alongside employees.
Product liability becomes increasingly complicated when behaviour results from adaptive AI models rather than fixed software.
Cybersecurity regulation becomes relevant when robots connect to cloud infrastructure.
Consumer law becomes relevant when household robots are marketed to the public.
And AI governance becomes inseparable from traditional machinery regulation.
This convergence will require collaboration between engineers, AI researchers, cybersecurity specialists, lawyers, regulators, insurers and standards organisations.
Physical AI safety is therefore unlikely to emerge as a single technical discipline.
It will become an ecosystem of engineering and governance practices.
From Robot Factories to Robot Societies
China’s humanoid robotics strategy demonstrates something important about the future direction of artificial intelligence.
The AI revolution is increasingly merging with the industrial economy.
China’s combination of AI research, manufacturing capability, electric-vehicle supply chains, robotics companies, government-supported training facilities and rapidly developing standards infrastructure gives it an unusual position in this transition.
TrendForce estimated in August that China’s humanoid-robot market could reach approximately RMB15 billion in 2026, with commercial validation expanding across automotive manufacturing, electronics, aerospace, logistics and energy.
But the ultimate significance of humanoid robotics will not be measured simply by market size.
The deeper transition occurs when intelligent machines begin operating continuously inside environments designed for humans.
Factories are built around human dimensions.
Warehouses are built around human movement.
Tools are designed for human hands.
Doors, stairs, vehicles and workplaces reflect the physical structure of the human body.
That makes the humanoid form potentially valuable because it allows automation to enter existing environments without completely redesigning them.
If general-purpose humanoid robots become sufficiently capable and inexpensive, the addressable market could therefore become enormous.
Safety Must Scale as Quickly as Capability
The development of China’s humanoid-robotics ecosystem also provides a warning.
Robotics capabilities may scale rapidly once hardware costs, AI capability and manufacturing capacity cross critical thresholds.
Safety infrastructure cannot wait until millions of robots have already been deployed.
Standards for collision avoidance, force limitation, cybersecurity, model evaluation, human override, incident reporting, black-box recording and safe software updates need to develop alongside the machines themselves.
The safest physical-AI architecture will probably not depend upon a single perfectly aligned model.
Instead, it will rely upon multiple independent layers of protection.
A foundation model may reason about the task.
A world model may predict the consequences.
A motion planner may calculate the movement.
A safety controller may restrict allowable actions.
Mechanical systems may limit force.
Sensors may detect humans.
Cybersecurity systems may protect control channels.
Monitoring systems may detect abnormal behaviour.
Black boxes may record incidents.
And humans must retain meaningful mechanisms for intervention.
This resembles the philosophy behind other safety-critical technologies: assume that individual components can fail and design the overall system so that a single failure does not become catastrophic.
Conclusion: AI Safety Is Becoming a Physical Science
For decades, robotics and artificial intelligence developed as partially separate disciplines.
That distinction is disappearing.
AI is giving robots perception, language understanding, planning and increasingly generalisable reasoning. Robotics is giving artificial intelligence bodies through which its decisions can affect the physical world.
China is becoming one of the most important laboratories for this convergence.
Its enormous manufacturing ecosystem, expanding humanoid-robot industry, embodied-AI companies, real-world training facilities and emerging national standards framework could allow the country to deploy physical AI at significant scale.
But the transition also changes the meaning of AI safety.
The safety question is no longer merely whether an AI system produces an undesirable answer.
It becomes whether the system takes an undesirable action.
That requires a broader conception of alignment combining frontier-model safety with robotics engineering, functional safety, cybersecurity, product regulation and human-machine interaction.
The development of humanoid robotics may therefore produce one of the most important shifts in the history of AI governance.
We are moving from an era concerned primarily with controlling what artificial intelligence says toward an era in which we must also control what artificial intelligence can physically do.
China’s rapidly expanding humanoid ecosystem provides an early glimpse of that future.
If physical AI develops as quickly as its advocates expect, humanoid robots may eventually become another general-purpose computing platform — one capable not merely of processing information, but of acting upon the world.
Ensuring that those actions remain predictable, controllable and aligned with human intentions may become one of the defining AI-safety challenges of the next decade.
Leave a Reply