Author: Peng Zhao (Founder of Zhicifang, Co-Founding Partner of Yunhe Capital)
After writing the previous article "Re-evaluating Embodied AI: The Next Trillion-Level 'Excavator Index' Will Be Born in Factory Workshops", I reviewed Unitree Robotics' two rounds of inquiry replies to the Shanghai Stock Exchange (SSE).
In the documents, there is a set of obscured figures: among Unitree's humanoid robot revenue, scientific research and education account for the absolute majority (73.60%), commercial consumption accounts for 17.39%, while "industrial applications", which the capital market highly expects, account for only 9.01%.
What is more interesting is the internal composition of this 9.01%. The inquiry reply shows that 50% to 70% of the revenue comes from corporate tours. Revenue used for clear industrial scenarios such as intelligent manufacturing and intelligent inspection is only CNY 15.702 million.
During the same period, the total revenue of its humanoid robots was CNY 595 million. Dividing the two numbers equals 2.64%. I calculated this number myself.
I previously mentioned that what embodied AI lacks most right now is an "attendance sheet". Investors are eager to price the on-duty working hours of robots, but after going through all financial reports, this column simply cannot be found.
Coincidentally, just days after this viewpoint was raised, both the market and regulators responded.
The Information recently cited sources reporting that the China Securities Regulatory Commission (CSRC) has issued informal window guidance to some investment banks and enterprises: for humanoid robot startups to pursue an IPO, they must prove their ability to generate "recurring revenue" and be on a track of narrowing losses or achieving substantive innovation.
And as I am typing this, as of the market close on September 14, Unitree's stock price was reported at CNY 470, and its total market capitalization has quietly fallen below CNY 200 billion.
The reality of 2.64%, the market capitalization falling below the threshold, and the regulatory requirement for "recurring revenue" have collectively backed the embodied AI industry into a corner: When they truly stand on the assembly line to work, how exactly should their wages be calculated?
Therefore, this article will start from the underlying industrial logic and break it down step by step.
Three Types of Working Hours: Who Pays for the Robot "Apprenticeship"?
In July this year, there was a very interesting detail in Tesla's second-quarter shareholder materials: the production lines for Model S and Model X at the Fremont factory have been dismantled, and the freed-up space is being used to install the first-generation Optimus production line. These initial robots off the line will not immediately go to work to earn money; instead, they are sent to the internal Optimus Academy, specifically for training data collection and function development.
Elon Musk added a note to this move: this is the most difficult product Tesla has ever scaled, and everything on the robot is new.
A company has set up a school for its own robots, and the employer of the first batch of graduates is the company itself.
Following this phenomenon, if we analyze the current "on-duty working hours" of humanoid robots, we will find that it is not a unidimensional concept at all, but has been restructured into three distinct forms:
The first is apprenticeship working hours.
The robot works for its own manufacturer, or for a laboratory that bought it to train algorithms. It consumes R&D budgets or scientific research funds, and produces "data and capabilities", rather than any commercial delivery capability that customers are willing to pay for.
Looking back at the industry landscape in 2026 with this definition, the first employers of humanoid robots are almost entirely the manufacturers themselves. 73.6% of Unitree's revenue comes from scientific research and education; universities buy them as training equipment, and the results stay in the laboratory. The first batch of trials for XPENG IRON is also arranged in its own stores and parks.
The second is validation working hours.
Between apprentices and regular workers, there is an intermediate zone backed by policies and public funds. In June this year, the Ministry of Industry and Information Technology (MIIT) and the State-owned Assets Supervision and Administration Commission (SASAC) issued a notice on real-scene practical training, requiring the launch of "operation mode" in representative scenarios by the end of the year, driving the deployment of tens of thousands of units.
In this model, the scenario providers (such as various central state-owned enterprises) play the role of "sparring partners". They are responsible for opening up workstations, quantifying targets, providing operation process data, and cooperating with acceptance.
The third is customer working hours.
This is the true closed-loop of commercialization, where customers pay for the labor output of robots, but currently, this type of working hour is extremely scarce. In the two most detailed publicly available attendance records, the paying "customers" are actually "shareholders".
Therefore, when opening the commercially validated attendance sheet, the first page needs to record exactly who pays the hourly wage.
We also need to answer another question: Why is it difficult for humanoid robots to replicate the attendance sheet of quadruped robots?
Over the past three years, Unitree has reduced the proportion of scientific research and education revenue for quadruped robots from 68.61% to 31.58%, successfully transforming apprenticeship working hours into customer working hours. Many people optimistically infer from this: humanoid robots just started a few years late, and sooner or later, they will replicate this path.
However, a fact that cannot be ignored is that quadrupeds can achieve commercialization because their first industrial job comes with a traditional "attendance sheet".
According to data from DeepRobotics, the Jueying series has achieved a recognition accuracy of 96.5% in substation scenarios, working normally during rainstorms at night, with a mean time between failures (MTBF) exceeding 1,000 hours. Why is the locked scenario power inspection? Because inspection falls exactly within the robot's capability boundary: there are regulations, checkpoints, and cycles. The task was already highly standardized before the robot arrived. The quadruped robot simply took over a ready-made attendance sheet.
But what humanoid robots face is a completely unstructured reality. The first industrial jobs for humanoids are assembly, handling, or quality inspection, and the core requirement is "manipulation generalization", which is precisely the bottleneck. There are very few workshops that can produce a ready-made regulation to accommodate this uncertainty.
The prerequisite for quadrupeds to take over the power grid's attendance sheet is that power inspection has long been transformed by the IoT (Internet of Things) into checkpoints, defects, and cycles. For humanoid robots to enter the workshop, first, someone must turn those complex unstructured workstations into equally standardized data.
Robot companies may not be good at this; instead, those working on industrial IoT are more proficient.
Three Columns of the Attendance Sheet: The Obscured Hard Logic of the Industry
If we are truly to price the "work" of robots, there are three main columns of decisive indicators on the attendance sheet.
First Column: Manual Intervention Rate
Existing public data almost entirely presents itself in the glamorous guise of "success rate". In June, Zhiyuan held a six-day live broadcast at the Longqi factory, where 8 G2 robots performed 64,828 operations, and the official success rate was 99.99%.
Calculated by this proportion, about 6 to 7 failures occurred in six days. Then the question arises: who rushed up to handle it after the failure? How long was the production line delayed? This is the hidden corner behind the success rate, which directly determines exactly how many "nannies" need to stand by a robot.
Second Column: Fault Recovery Time
Current indicators are all competing on "Mean Time Between Failures" (MTBF), but this is tricky because it only records the span between two failures, not how long each downtime actually lasts.
Equipment in traditional factories allows a maximum of 4% to 5% fault downtime per day. Jammed? Clearing the material or restarting can recover it. But whether the fault recovery time of embodied AI is within the factory's downtime tolerance remains unanswered.
Third Column: Line Change and Redeployment Cost
This should have been the absolute home ground for humanoid robots, yet it lacks quantitative data the most. The cycle for traditional robotic arms to retrofit production lines is often half a month, while humanoid robots can switch workstations in a few hours. It can be said that the CNY 100 billion valuation of the entire embodied AI industry is partly built on this column.
The good news is that technology is indeed sprinting. In a blog post in early September, NVIDIA introduced the S1 model released by Skild AI: the operator only needs to record a video as a prompt, and the model can directly generate the robot's actions, eliminating the need to update weights and perform post-training for specific tasks. According to publicly relayed data, its multi-step task success rate is about 66% (compared to only 9% for the baseline system).
But this absolutely does not mean that "apprenticeship working hours" have been eliminated; it just means the unit price has been lowered. The flip side of the coin is the law of cost conservation: a 66% success rate means that the remaining one-third of tasks require recovery or manual takeover and fallback.
This is the most ruthless game on the attendance sheet: when the "deployment cost" in the third column decreases, the "intervention and recovery costs" in the first and second columns will inevitably rise. Fixing one problem only creates another; the compression of costs in any column may ultimately be squeezed into the other two.
The Ultimate Question: Who Will "Keep Attendance" for the CNY 100 Billion Scale?
For the three columns of pricing-deciding indicators mentioned earlier (intervention rate, recovery time, line change cost), none actually requires inventing new sensors.
The moments of manual intervention are usually saved in the logs of emergency stop buttons, safety light curtains, and remote control takeovers; how long each downtime lasts is faithfully recorded by the clocks of the production line PLC and MES systems; and exactly how many engineer hours it takes to change a line is clearly calculated in the integrator's project schedule.
The data is all there. What is lacking now is the step of extracting them from three mutually unconnected systems, aligning them on the same timeline, and then issuing a report as an absolutely neutral third party.
The current embodied AI track is playing a game of "being both the player and the referee". When the robot body manufacturers themselves take a 99.99% success rate to negotiate piece-rate billing with customers, I am afraid no leasing company or insurance company would dare to give a fair quote based on this.
To make the commercial mechanism work, an independent "third-party AIoT platform" may emerge in the future to completely integrate the three layers mentioned in the previous article: reading, reconciliation, and pricing:
1. Reading Layer (Tamper-proof Black Box): Attach an edge gateway with independent timing and direct data transmission to the robot body. IoT manufacturers have long made the exact same things for excavators and elevators; now it is just a matter of changing the installation location.
2. Reconciliation Layer (Piece-rate Counter Connected to MES): Interconnect operation counts with the factory's MES system. The G2 on the Longqi production line has already been connected to the MES. This is just the daily routine for industrial software manufacturers; the only difference is that today's connected object has changed from a pick-and-place machine to a walking robot.
3. Pricing Layer (Data-based Financial Settlement): Only by obtaining desensitized and authentic data from the first two layers can leasing companies and insurance companies launch "operation-volume-based settlement" contracts and policies.
With this third-party refereeing mechanism, embodied AI may truly receive its own "work badge" and step onto the assembly line to earn real money. When thousands of real, tamper-proof "attendance sheets" converge in the cloud, that trillion-level "excavator index" will finally land in factory workshops.
Conclusion
The current robot industry has not yet generated its own "excavator index". The "refereeing power of production working hours", which lies between the power of acceptance and the power of pricing, is currently a vacuum. Who will blow the "referee's whistle" for this CNY 100 billion track?
Industrial Internet platforms have reconciliation capabilities but cannot access the underlying interfaces of robots; robot body manufacturers have the interfaces, but their data is not trusted by financial institutions.
Perhaps, the one to break this deadlock will not be the next robot body manufacturer valued at CNY 100 billion, but the third party that can keep the "attendance sheet" on the assembly line.