Author: Peng Zhao
On November 2, the final day of the 2026 World IoT Expo, I hosted the "2026 Wuxi Liangxi Science and Technology City Robotics Industry Development and Project Matching Conference."
An investor from Beijing took the stage, first sighing about how early winter arrives in the north, then pivoting sharply: the landscape for embodied AI has changed. Two years ago, funding in this sector totaled just over CNY 3 billion; by the first half of this year, it had surged to CNY 93.5 billion, attracting approximately CNY 500 million per day.
The hype rose quickly and receded just as fast.
After autumn set in, enthusiasm plummeted, with September's funding volume dropping by 40%. He remarked, "It took the automotive industry 200 years to reach its current state, yet people expect robots to complete that journey in five years."
Once the fever breaks, where will the value of embodied AI settle?
Many immediately think of the "brain," but this is precisely where disagreements are most acute at present.
This investor remains wary of focusing on the brain. This is due to AI's alarming "cost collapse": capabilities developed today for CNY 10 million can be replicated by others for CNY 1 million ten months later. In a recent trial, a general-purpose large model completed most designated robotic manipulation tasks directly without specialized training. However, Xinghaitu, also present at the expo, offered a contrasting judgment: core competitiveness still lies in the model.
These two views are not contradictory; they simply belong to different cycles.
During the rising tide, capital pays for the "ceiling," betting on how smart the model can become. When the market cools, customers focus only on the "floor": Can the machine avoid breaking down? Can data be used legally? Who bears responsibility when things go wrong?
These three life-or-death bottom lines are exactly the three barriers embodied AI must currently cross: the Physical Barrier, the Data Barrier, and the Identity Barrier.
The Physical Barrier: The Decisive Factor Lies in Production Lines, Not Brains
In the early days of railways, cast iron rails were extremely prone to wear, requiring frequent replacement on busy lines, which made rail companies wealthy. It wasn't until 1856, when Briton Henry Bessemer invented a new steelmaking process, that strong, wear-resistant steel rails became affordable enough for railways to escape the cycle of constant replacement, eventually becoming an indispensable mode of transport worldwide.
Today's robots resemble those early railways: Affordable to buy, but too costly to fix.
On one hand, the price curve is steeply declining. According to Zhongyi Group, in their Shenzhen projects, the price of a humanoid robot has dropped from around CNY 500,000 last year to approximately CNY 100,000. Core components are also rapidly becoming cheaper; Pacini, a tactile sensor manufacturer, reduced the price of a single sensor from tens of thousands of yuan to just over CNY 100.
But on the other hand, the reliability curve has failed to keep pace. Zhongyi, which specializes in robot leasing and maintenance, deeply understands this challenge. They report that current robot bodies "break very easily." On-site installation, delivery, and repair cost CNY 500 per hour. At the current rate, the annual maintenance burden borne by body manufacturers for a delivered humanoid robot reaches CNY 40,000–50,000.
Purchasing a unit costs CNY 100,000, while annual maintenance requires nearly half that amount again. The faster the overall price drops, the higher the proportion of maintenance costs, creating a paradoxical "maintenance inversion."
A deeper issue is that system reliability is multiplicative.
Engineering often uses "nines" to describe reliability: one nine is 90%, four nines is 99.99%. Currently, success rates for embodied AI in many tasks do not even reach one nine. Entering factories requires four nines, homes require six nines, and sensitive fields like healthcare require eight nines. People typically interpret these figures as requirements for the "brain," but the "body" is also a multiplier.
Suppose the model achieves a task success rate of 99.99%, but the machine spends 5% of its time idle awaiting repairs. For the customer, this is still only a 95% available machine. No matter how many "nines" the brain accumulates, they are eroded by the shortcomings of the body.
Fortunately, a shift is already underway.
A joint module company in Suzhou noted that customers' primary concern has shifted from price to safety, demanding that joints maintain temperatures below 60°C after two hours of continuous operation. They will launch a joint assembly line with over 90% automation by the end of the month. A startup specializing in motors for dexterous hands is using multi-layer PCB boards to print stators, replacing manual winding, which struggles to ensure consistency.
The turning point for the widespread adoption of embodied AI will likely not be a single astonishing invention, but rather improvements in manufacturing consistency: replacing winding with printing, manual assembly with automated production lines, transforming robots from "handcrafted items" into robust "industrial products."
After all, reliability is a manufacturing problem.
The Data Barrier: The Obstacle Lies in Ownership, Not Collection
Even if both the body and the brain pass inspection, when robots truly enter the workforce—especially in unfamiliar environments—they must first solve a critical question: Where does the corpus and knowledge used to teach them come from?
Currently, the industry primarily acquires data through two paths.
The first path is the "Digging Wells" model: Collecting actions via teleoperation of robots or having humans wear devices. This type of data is high-quality with clear sources, but it is extremely costly, narrow in coverage, and difficult to scale.
The second path is the "Desalinating Seawater" model: There are billions of cameras globally, with about one-third installed in commercial and industrial scenarios. A single camera records tens of thousands of hours of footage annually, containing vast amounts of human work processes, including factory line operations, supermarket stocking, kitchen prep, etc. The marginal cost of this data is nearly zero. The difficulty lies in the high redundancy and sparse effective information; training models directly on this raw data would leave them "thirstier," much like a human drinking seawater.
However, technological breakthroughs have occurred. Associate Professor Feng Shuo from Tsinghua University's Department of Automation proposed the "Dense Learning" method (featured on the cover of Nature magazine in 2023), which can "desalinate" high-value segments from surveillance video for robot training. As he puts it, this has transformed data acquisition from a "collection problem" into a "production problem."
If the technology to "desalinate seawater" has been solved, where does the real obstacle lie?
It lies in data ownership and compliance.
Data obtained by "digging wells" naturally belongs to the digger; however, the owners of the "seawater" are the scenario providers. Those cameras and sensors spread across factories, supermarkets, office buildings, and visible kitchens were originally installed for security and regulatory purposes.
Cheap data lies dormant within the inventory built by the IoT industry over decades, but what is truly expensive is issuing them a legal "birth certificate."
Who has the right to "desalinate" this data? How should the benefits be distributed? The industry has yet to provide answers.
Tesla's path cannot be replicated here.
Tesla's rapid evolution in autonomous driving relies on millions of car owners returning data during daily driving almost for free. Owners cooperate because they first bought a good car.
A Tesla without intelligence is still a drivable electric vehicle; but a robot without intelligence lacks even a basic business model.
The robotics industry does not yet have a cohort of "car owners" who can continuously contribute compliant data.
The Identity Barrier: The Passport Depends on Rules, Not Performance
Returning to the opening statement: It took the automotive industry 200 years to reach its current state.
Over those 200 years, half of the effort was spent building cars, and the other half on building rules.
Cars need license plates, drivers need licenses, roads have traffic laws, accidents have insurance, vehicles undergo periodic inspections, defects trigger recalls, and manufacturers guarantee parts supply after production ceases. It is this system that transformed a high-speed moving metal object capable of causing harm into a daily tool trusted by households.
Today, robots taking public roles is no longer hypothetical. But for robots to truly get "certified to work," the industry must first establish a "Vehicle Management Office."
Comparing item-by-item with automotive systems, robots lack several key elements:
License plates correspond to identity registration. Each robot should have a verifiable unique identity. IoT has long been doing this; from device coding to the AI-eSIM demonstrated by China Mobile at this expo, there is a ready technical foundation for device identity management.
Licenses correspond to scenario-specific qualifications. The aforementioned "nines" can serve as benchmarks for qualification levels: four nines allow entry onto production lines, six nines permit access to homes, and eight nines are required for hospitals.
Periodic inspections correspond to regular third-party testing. These check the hardware body, i.e., the manufacturing reliability discussed earlier.
Archives correspond to data "birth certificates" and behavior logs. These ensure every action is traceable, addressing data compliance and trustworthiness.
Insurance corresponds to liability backstop mechanisms. This answers the question: "If something goes wrong, who is ultimately responsible?"
As can be seen, this "Vehicle Management Office" essentially institutionalizes the requirements of the previous two barriers: Inspections manage the body, archives manage data, and licensing/insurance manage identity and responsibility.
Fortunately, the bricks and tiles for building this "Vehicle Management Office" are being prepared. Only when the other half of the institutional puzzle falls into place will robots truly receive their "passport" to integrate into human society.
Conclusion
Whether looking at automobiles a century ago or IoT over the past few decades, their widespread adoption proves one thing: Technology determines how fast products can run, while engineering and rules determine how far an industry can go.
Embodied AI must complete three steps: maturing hardware through manufacturing systems, ensuring data traceability through rights confirmation and compliance, and connecting social trust through institutional standards.