Zhiyuan Innovation Day proposes Worken effective intelligent work hours metric, 9,700 units shipped in first half

On October 19, the Ministry of Industry and Information Technology disclosed a set of figures at a press conference held by the State Council Information Office: in the first half of this year, China's humanoid robot whole-machine products had already surpassed 400 models, with output exceeding 40,000 units, and the full-year figure is expected to exceed 100,000 units.

Ten days later, on October 29, at the Shanghai World Expo Center, AgiBot once again verified this industry progress. AgiBot co-founder, president and CTO Peng Zhihui presented the first-half report card at the 2026 AgiBot Innovation Day: in September, the 20,000th robot officially rolled off the production line; shipments in the first half reached 9,700 units, with a global market share of over 43%; the company's first-half revenue was close to 2 billion yuan.

Placed at the start of the year, such figures alone would have been enough to carry a launch event. But this time Peng Zhihui put the emphasis on the latter half. "How many robots we produce does not equal how much productivity has been added to the real world."

In his view, this is the problem the industry most needs to solve together right now: scale manufacturing is being achieved rapidly, while the ability to work steadily on the job has not yet been verified. And customers' mindset is shifting from "I'm willing to try your product and see whether you can build it" to "can I actually afford to keep your product running."

Eight hours powered on does not equal eight hours of work

Peng Zhihui mentioned in his speech: in the past, looking at a robotics company meant looking at what it could do and whether the moves in its videos were cool enough; now customers are starting to ask three other things: how much work can it stably do, how well can it do it, and how much does it cost to do that work.

"A robot being powered on for 8 hours does not mean it has created 8 hours of value. A high success rate on a single task does not mean it can work continuously and stably for a month like a human." How exactly should the value a robot creates for customers be measured? To answer this question, AgiBot created a new yardstick — Worken (effective intelligent work hours).

The formula is not complicated: Worken (effective intelligent work hours) = number of active robots × autonomous working hours × output efficiency ratio.

The three multipliers answer three very plain questions. "Number of active robots" looks at how many robots are truly on duty, rather than lying in a warehouse; "autonomous working hours" is used to distinguish whether a robot is a passive tool or an autonomous agent. Peng Zhihui gave an example: a robot may look very smart, but if it makes a mistake every 10 minutes of work and needs a human to take over each time, then the autonomous work hours it creates are still very limited; "output efficiency ratio" calculates an equivalent account, "whether the work done by this one machine is worth one person, or whether it takes two to match one, or even whether one machine can match several people in the future."

The significance of Worken's birth is that, beyond the number of robots manufactured, a new dimension has finally emerged that can measure whether robots can autonomously and continuously complete valuable work, and how much output they actually produce.

At the launch event, Peng Zhihui showed us the number of Worken achieved by AgiBot robots in multiple deployment scenarios, including loading and unloading on the Longcheer production line, hotel service staff at Shangri-La, visitor services in Chengdu's Kuanzhai Alley, and security-check guidance at the Guangzhou Metro, among others. Among them, the Chimelong entertainment performance robots led all scenarios with 33,480 Worken.

Peng Zhihui said that if measurement is unclear, buyers and sellers cannot set prices, input and output cannot be calculated clearly, and the industry cannot form a consensus. He hopes that in the future, robot services will be billed by Worken hours, with robots of different intelligence levels doing different tasks at different rates.

AgiBot partner, senior vice president and president of the general business unit Wang Chuang described this model in more detail: in the future, customers may no longer rent robots by the day, but buy a "productivity package," reserve tasks, and be charged according to Worken consumed, "what serves you is not one robot, but a combination of robots across different occasions."

20:1

How many engineers must stand behind one robot on duty was previously a number the industry was unwilling to disclose. But Wang Chuang said frankly in an interview after the event: "When Chimelong's first batch of 300 robots entered the site, several hundred machines were paired with several hundred engineers, close to one to one. Now, that ratio has dropped to three to one; by the end of the year we want to reach ten to one, next year down to twenty to one, and in the longer term thirty to one."

If the human-to-machine ratio cannot be brought down, Worken cannot be calculated. Senior engineers are very scarce, and if experts must be piled onto every scenario, that is essentially doing individual projects rather than a scalable business. Wang Chuang said: "20 to 1 is a critical point, that is, 5%. When a position only requires an additional 5% of the manpower budget, for many customers that is an acceptable point."

After withdrawing operations and maintenance experts, how to ensure stable robot operation is one of the issues customers care about most. AgiBot's answer is a set of automated operations and maintenance systems. Experts provide remote support at headquarters, on-site operations staff execute according to templates prepared by the experts, and underneath is a one-stop operations platform that automates all four steps from parameter inspection and site onboarding to online management and problem closure.

AgiBot has already achieved takeover by this system on the Longcheer production line. This production line entered scaled production in March 2026 and has now run stably for more than 7 months. In the first half of the year, AgiBot conducted a 6-day live broadcast of the entire production line; the robots accumulated more than 64 hours of operation and 64,824 operations, producing 17,625 tablet computers, with a success rate of 99.99% and an average cycle of 22 seconds per unit, reaching the same level as humans. AgiBot's R&D team had long since withdrawn from the Longcheer site, and basic software operations and maintenance had been fully handed over to Longcheer's own team.

The reason Chimelong was able to become the Worken champion is directly related to having the highest number of active robots, and the rapid growth of that number comes from customers' repurchase rate.

Chimelong's first batch of more than 300 robots was distributed across more than 100 locations, and the task at each location was completely different: some played three-in-a-row with people, some played rock-paper-scissors, some played table tennis, some led large groups of visitors on guided tours, and some competed with children in obstacle crossing. Xiong Yan, AgiBot partner, co-president, president of the Central R&D Department, and president of the Interaction Business System, revealed that one guided-tour location quickly went from one unit to three, "when customers find that a robot can successfully complete a work task, they are willing to repurchase."

At present, Chimelong has deployed nearly 500 robots. During the National Day holiday, the park received about 10,000 visitors per day on average, foot traffic at the Spaceship Park increased by nearly 40%, and by the end of the year Chimelong plans to increase its robots to 1,000.

General Intelligence and Embodied Intelligence

As for the most discussed topic recently: will general large models eat embodied intelligence? The answer given by Yao Maoqing, AgiBot co-founder, co-president, and president of the Embodied Business System, is: no, at least not now.

Yao Maoqing cited an NVIDIA report: among 120 simulation tasks, GPT-6 Astra's average success rate was 66.3%, with its capabilities concentrated in semantic understanding, action planning, and decision-making. "The stronger the general model, the higher the starting point of the embodied industry will be; this is a clear positive." But on the other hand: Astra's average inference is very slow, taking more than 18 seconds per run, and it cannot perform high-dynamic and fine manipulation accurately. In links involving force-tactile feedback, the success rate of large models is basically completely zero, because they have inherently never learned manipulation capabilities.

Xiong Yan interpreted the irreplaceability of embodied models from another angle. "A very simple action may require minute-level reasoning, cost hundreds of dollars per call, and still need fine-tuning when entering a vertical domain. Training costs, data costs, and fine-tuning costs along the chain are all very expensive. I don't believe anyone would deploy a 5-trillion model on a robot's body; ultimately, implementation still has to consider practicality."

In Yao Maoqing's view, the relationship between general intelligence and embodied intelligence is more like a layered architecture. The general model is responsible for long-horizon planning and reasoning at the System 2 level; in the middle are individual skills trained by VLA and world models, which are invoked as tools; at the bottom is real-time motor control at 100 to 1,000 Hz, tightly coupled with the body and the task. Conversely, if a general model wants to reach expert level in the physical world, it likewise cannot avoid massive high-quality post-training: "without massive trial and error in the physical world for specific tasks, it cannot even do the planning layer well."

Only Doing 0 to 1

AgiBot only does 0-to-1 scenarios and lets partners do 1-to-N. This is the division of labor Wang Chuang repeatedly mentioned in the interview: "We cannot obtain a lot of data, and we do not know a lot of industry know-how. There is no need for us to earn every dollar."

Lingqu OS 1.0 is the foundation for realizing this division of labor. It is the industry's first embodied-intelligence native operating system, supporting plug-and-play for more than 100 types of peripherals across 15 categories, with end-to-end latency jitter of less than 4 microseconds under 1,000 Hz high-frequency tasks. AgiBot plans to fully open-source it by the end of 2026, contributing the 200 million lines of code and more than 5,000 ecosystem software packages that have been validated in mass production inside the company.

The human-to-machine ratio must continue to be pushed down from 20:1 to 30:1. This cannot be achieved by AgiBot simply hiring more engineers; it can only be done by these partners breaking through non-standard scenarios one by one. The partner ecosystem released by AgiBot at the beginning of the year also delivered a report card at this conference: the Yuansheng Plan has linked 255 universities, 511 enterprises, and more than 80,000 developers, and has invested more than 300 million yuan this year. Wang Chuang said that the charm of an ecosystem is that you can never imagine how many surprises these partners can bring you.

In Peng Zhihui's view, the sign of robots truly becoming widespread is that AgiBot almost no longer needs to provide services. Enterprises can buy them themselves and distribute them themselves. Wang Chuang expressed the same view: just as intelligent driving looks at increasingly longer takeover mileage, robots should look at the growing proportion of cases that "do not require AgiBot services."