Zhiyuan Unveils Worken, a Metric for Effective Intelligent Work Hours, Along with GO-3 and Other Embodied Technologies

After GPT-6, do robotics companies still have value? AgiBot responds with the "Worken" compute ledger behind its 20,000-unit shipment volume.

English text is being prepared.

English text is being prepared.

In 2026, the embodied intelligence sector is sinking into an unprecedented mood of anxiety.

The embodied control capabilities demonstrated by GPT-6 Astra have made the argument that "physical AI is about to be breached by AI" increasingly loud; the S2 layer, responsible for perception and decision-making, is being taken over by general intelligence at a visibly accelerating pace. The industry has even seen a clear divergence in routes: a group of model companies have begun pivoting to selling data, and "selling shovels" has become the new safe zone.

And in the red-hot display scenarios, even if China's total robot production this year breaks 100,000 units, those that can truly replace real work on the assembly line and carry KPIs for humans are very few.

Everyone in the industry is waiting for an answer: as general large models keep devouring the upper layers of intelligence, where exactly is the value anchor for companies that make robots?

On October 29, at the Shanghai World Expo Center, at AgiBot Innovation Day (A.I. DAY), Zhihui Jun, Yao Maoqing, Xiong Yan, Wang Chuang and their AgiBot legion gave a complete response with a press conference lasting more than two hours. More than 1,000 research partners, developers, business partners and media attended on site, and more than 50 media outlets livestreamed online.

"China's total robot production this year is expected to break 100,000 units, but we must also be clear-headed: how many robots were produced does not equal how much productivity the real world has gained."

After this sentence, the venue fell silent.

This is also the "industry illusion" that AI Tech Review found in some industry surveys: news of mass production fills the air, but very few are truly and continuously investing in productivity. In response, AgiBot's solution is a brand-new unit of measurement: Worken (effective intelligent work hours):

Worken = number of active robots × autonomous working hours × output efficiency ratio

The three multipliers each conceal three pain points of the embodied industry: machines that are down, offline, or under maintenance do not count as active; teleoperation periods are not counted as autonomous working hours. A robot that is busily spinning around in front of the camera while three joystick operators stand behind it has its work hours directly zeroed out under this formula; the output efficiency ratio takes a standard worker's qualified output per unit time as 1.0, and anything slower than a human fails to meet the standard.

To explain this concept, Zhihui Jun traced a history of measurement: the steam age looked at "horsepower," the electric age at "kilowatt-hours," and the digital AI age uses "Token" to measure the amount of thinking. He concluded: "Token measures how much a machine has thought, Worken measures how much a machine has acted."

The reason the large-model industry can form today's commercial loop of billing by volume and pricing by effect is that token deserves the credit. It made "intelligence" a commodity that can be priced and traded for the first time. And the physical AI industry has long suffered from the lack of a unit of measurement: buyers and sellers cannot set prices, capital cannot calculate input-output, and Worken is precisely the key to untangling this dead knot.

Zhihui Jun did not stop at coining a new term; on site he directly cut to a real-time work-hour dashboard: the loading and unloading robots on the Longcheer production line, the service robots at Shangri-La hotels, and the cultural tourism robots at Chimelong parks—each machine's effective work hours contributed in the current month scrolled in real time. He joked: "Anyone who has worked a job has filled in timesheets. In the future, robot services will also be billed by Worken."

At the same event, the "2026 Embodied Intelligence Industry Innovation Report," jointly compiled by authoritative institutions such as the China Academy of Information and Communications Technology, was officially released, and Worken and related deployment-state examples were successfully included. The unit of measurement had just been released, and it already had authoritative endorsement.

When robots are billed by Worken, what customers buy is no longer a piece of equipment that peaks the moment it leaves the factory and depreciates year by year, but a continuously evolving unit of productivity. Another remark from Zhihui Jun highlighted the change in asset attributes: "Hardware will grow obsolete, but intelligence can appreciate."

02

The question posed six months ago has now been answered

The conception of Worken first originated at the APC2026 AgiBot Partner Conference six months ago. At that time, AgiBot founder Deng Taihua asserted: 2026 is the first year of embodied intelligence moving from the "development state" to the "deployment state." Zhihui Jun set the tone: "The real watershed for the industry is the shift from 'selling robots' to 'delivering results.'"

At that time, what supported this judgment was an 8-hour unedited livestream at the Longcheer factory: the Genie G2 worked continuously as a "formal employee," with a 100% success rate. That was a single-point proof that "it can work," but it could not yet prove that "it can work at scale."

Six months later, the granularity of implementation in the embodied industry and at AgiBot is completely different.

On the business side: 9,700 units shipped in the first half of 2026, ranking first globally with a share of over 43%; the 20,000th robot rolled off the line in September, and it took only half a year to go from 10,000 to 20,000 units; first-half revenue was close to 2 billion yuan.

Mass production of 10,000 units is the "life-or-death line" for the humanoid robot industry, testing comprehensive capabilities in supply chain management, yield control and cost dilution. AgiBot partner Wang Chuang laid out the cards for crossing this threshold: a 100% localized A-chain, a half-hour supply chain circle, and 92 automotive-grade mass production tests.

"Mass production is not just about building the robots and being done; it is about ensuring that every 10,000 units built are as reliable as the first one." Wang Chuang said bluntly, "There is no shortcut behind this; it is all dug out pit by pit on the production line."

On the scenario side, the "veteran actor" working on the Longqi production line in April has now been running stably for more than 7 months; over 6 days of livestreaming it accumulated more than 60 hours of work, producing 17,625 computers, with the robot performing 64,828 cumulative operations, a cycle time of 22.5 seconds per unit, comparable to human workers. The suspense over "can it withstand a full year of production capacity assessment" now has an answer.

Industrial depth is also widening: the AgiBot G2 has achieved full coverage of the tablet mass-production quality inspection section at the Longqi factory and has entered chip packaging and testing and CNC; the AgiBot G2 Max has been deployed for handling in food, FMCG and e-commerce; nearly 500 robots have been deployed at Chimelong in Hengqin, driving park visitor traffic up by nearly 40%; more than 120 robots have taken up inspection posts on the Guangzhou Metro; and at Kuanzhai Alley in Chengdu, the world's first VLA-model-driven bipedal humanoid retail operation has been deployed, with an accuracy rate of 98%.

There is another easily overlooked yet extremely critical number: the human-to-machine operations ratio.

The figure announced by Wang Chuang has already fallen from an initial 1:1 (one robot paired with one commissioning technician) to 10:1, with the goal pointing directly at 30:1. In AI Technology Review's view, the slope of this curve is the touchstone for whether "deployment state" is a real business or just a pile of money-losing custom projects.

"The laboratory is a greenhouse; the real world is the wilderness. The industry does not lack technical difficulties; what it lacks is productivity that can go on duty at scale."

In April, APC answered "what deployment state is"; in October, A.I. DAY answered "how to measure deployment state, how to do the accounting, and how to evolve it." It took AgiBot six months to go from proposing the concept to delivering the weights and measures.

The technology list at this launch event was extremely dense: the third-generation embodied foundation model GO-3 Preview, the four-layer AGILE motion engine, the omni-modal interaction large model WITA-Omni, the third-generation hardware architecture AHA3, and the heavy-load body G2 Max, among others.

But in AI Technology Review's view, rather than listing the parameters one by one, it is better to see clearly the structure these releases collectively point to: they are each providing insurance for the three multipliers in the Worken formula. (Leiphone)

GO-3 Preview insures the "output efficiency ratio." Yao Maoqing announced on site that GO-3 has topped all three authoritative leaderboards that recently evaluated GPT-6's embodied capabilities, including RoboDojo.

But more important than the leaderboards is the ability to evolve after deployment: the collective experience learning module CEL, jointly developed with Tsinghua University, has robots specifically "practice" on failure cases, and in actual tests the task success rate of cluster training with 100 robots was raised from the 50% level to 98%; when encountering an unexpected "hook obstruction" in a real convenience store scenario, it can derive new strategies, raising the single-attempt success rate from 20% to more than 80%.

As Yao Maoqing put it incisively: "The growth path of embodied models is not only trained out of data, but also 'worked' out in real scenarios."

AGILE and WITA-Omni insure "autonomous working hours." Once a machine goes down or calls for human help, its working hours are immediately reset to zero. To this end, the AGILE motion engine gives robots "closed-loop recovery" capability through a four-layer architecture. Zhihui Jun said bluntly that this completely bids farewell to the industry-wide common failing of many dancing robots today that "imitate blindly with their eyes closed."

On the interaction side, another AgiBot partner, Xiong Yan, shared what was also the most "human" segment of the whole event. WITA-Omni models vision, hearing, speech and motion synchronously, achieving "listen while thinking while speaking," with response speed compressed to as fast as 400 milliseconds, already entering the natural rhythm range of everyday human conversation.

AHA3 and G2 Max insure the "number of active robots." The three intelligences are physically unified at the hardware level, with controller volume reduced by 50% and latency reduced by 50%; the world's first subcutaneous joint liquid cooling system, with heat dissipation capacity reaching 7 times that of traditional air cooling, allowing continuous torque output at the joints to reach 3 times that of air cooling; the G2 Max has a dual-arm peak load of 50 kilograms and dual-battery hot swapping that sustains 8 hours of all-weather operation.

Also released at the same time was the new-generation dexterous hand family OmniHand 3 series, among which OmniHand 3 is the first industrial-grade dexterous hand that has already been operating continuously on a production line.

The maturity of this product line directly determines the upper limit of fine manipulation. There is also the Guinness World Record for the "most humanoid robots dancing in unison" set by 110 Expedition A3 units. Zhihui Jun called it "an important starting point for the industry to move toward scale."

Beyond maxing out software and hardware configuration, AgiBot also left behind a more lethal follow-up move: releasing the industry's first embodied native operating system, Lingqu OS 1.0, and promising full open sourcing by the end of the year. Together with the "Yuansheng Plan" that links more than 80,000 developers over half a year, AgiBot is trying to monopolize the infrastructure of the physical world.

Taken apart one by one, they are technologies; taken together, they are a precision device for "making effective working hours continuously increase."

The most noteworthy thing about this launch event is that AgiBot did not evade the shadow hanging over the industry: will large models completely "break through" the embodied track? Do body companies that make robots still have value?

He further explained: a general large model takes an extremely long time for a single inference and has extremely high computing cost, "making it difficult to deploy in real industrial scenarios that require low latency and high precision." AgiBot's choice is to stand on the body, invest heavily in self-developed large models, and at the same time precipitate capabilities into "production-grade technology," and announce that GO-3 and the entire embodied AI technology framework will be open sourced within the year, turning the architecture into an ecosystem.

Zhihui Jun, for his part, laid out two possible futures in response to the question the whole industry cares about:

The first: general foundation models are becoming increasingly powerful and trending toward openness, so anyone can obtain advanced intelligence. Then the focus of competition becomes who can connect intelligence to the real world at lower cost and with higher reliability;

The second: models still rely heavily on training with data from different embodiments and different scenarios, and there is no universal model. Then "whoever owns large-scale real deployments and can stably obtain high-quality data" wins.

And both of these futures are home turf for embodiment manufacturers.

"The two technical paths seem different, but ultimately they point to the same conclusion: scaled embodiment advantages will become increasingly important."

This is a rather elegant argument. In the view of AI Technology Review, "connecting intelligence to the physical world" has never been a software problem, but a comprehensive engineering problem of reliability, cost, supply chain, deployment networks, and operations and maintenance systems, and this is precisely the part that large-model companies with internet DNA find hardest to make up.

AgiBot's current 20,000-unit shipment volume, 99.99% stability, 92 automotive-grade tests, and 10:1 human-to-machine ratio are the moat that blocks pure-algorithm companies. General intelligence is not the gravedigger of the deployed state, but an ammunition depot; thinking is left to large models, but the barrier to action will always belong to companies that grow close to the physical world.

Finally, I want to mention the moment at this event that was easiest to overlook: at the launch event, AgiBot partner Xiong Yan, during a demonstration, asked Expedition A3, "Who is the most handsome at AgiBot?" The robot answered without hesitation: "I am AgiBot's face value representative." Then came the follow-up, "Besides you?" Expedition A3 did not give a mechanical answer, but instead replied, "Of course you, your blue POLO shirt looks very spirited."

In the live demonstration, Expedition A3 did not interrupt in the noisy venue, waited for the right moment, and could automatically tell who was speaking to it; what was harder was the flexibility of language: Xiong Yan took it on site from Mandarin seamlessly to Cantonese, Sichuan dialect, and Hokkien, and even completed simultaneous interpretation.

Xiong Yan's explanation was very practical: "Dialect recognition is not showing off skill; it is a work qualification certificate." WITA-Omni supports more than 100 languages and 29 dialects, and has topped four authoritative embodied interaction leaderboards including Daily-Omni, and will complete full deployment in China in December. When these capabilities land at the guidance desk of Peking University People's Hospital, in Kuanzhai Alley, and at Chimelong's information points, they become solid on-duty working hours.

This moment is worth recording because behind it stands the most "human" model of this launch event: the WITA-Omni full-modal interaction large model. It does not simply stitch together modules such as speech, vision, and motion, but models all inputs and outputs synchronously in the same model, which is how it can praise someone so specifically and with emotional intelligence.

In addition to being able to observe and having emotional intelligence, it can also make the robot "listen, think, and speak at the same time," with response speed compressed to as fast as 400 milliseconds and 600 milliseconds on average, already entering the natural rhythm range of everyday human conversation.

Xiong Yan's on-site demonstration was more like a pop quiz: taking Expedition A3 from Mandarin seamlessly to Cantonese, Sichuan dialect, and Hokkien, and even completing simultaneous interpretation; in the noisy venue, the robot did not interrupt, waited for the right moment, and could distinguish who was speaking to it and who was just chatting with others nearby. WITA-Omni supports more than 100 languages and 29 dialects, and has topped four authoritative embodied interaction leaderboards including Daily-Omni, and will complete full deployment in China in December.

Similar texture was everywhere throughout the event: the 22.5-second takt time on the Longcheer production line, the failure case of a convenience store hook being blocked, Chimelong's 5G solution of "physical isolation between the human and robot networks," and the operations platform's design of "one-click reuse for newly added sites." None of these can be planned in a laboratory; they grow in the wild.

Six years ago, this industry competed over whose robot could do flips; three years ago, over whose demo video was smoother; this April, AgiBot said the competition had changed to "whose robot can work continuously"; and this October, it simply issued a timer for the competition.

As Zhihui Jun said in his opening: "AgiBot got first place the first time it entered the competition, but what makes this result more meaningful is commercial deployment. Our robots can both enter the arena and go into factories and shopping malls."

The 46 sports meet gold medals are the face; the real-time scrolling work-hour dashboard is the substance. Whether Worken can become an industry-recognized measure ultimately depends on joint adoption by buyers and sellers, and that takes time to verify. But at least in this autumn overshadowed by the "breakthrough theory," AgiBot has done something the industry should have had someone do long ago:

When everyone was arguing about whether robots are smart enough, it was the first to step forward, clock the robots in, and tell the embodied industry: the second half is not about who is better at performing intelligence, but about who is better at delivering work hours. (Leiphone)