AgiBot Releases AHA3 Architecture, GO-3 Preview, and Lingqu OS 1.0, Proposes Worken Effective Intelligent Work Hours

Fast Reading

Peng Zhihui proposed a new concept, Worken (effective intelligent work hours), which consists of the number of active robots, the length of autonomous working time, and the output efficiency ratio, attempting to measure how much effective labor robots actually create in real environments.

A robot being powered on for eight hours does not mean it works effectively for eight hours; completing an operation at a launch event does not mean it can work continuously and stably on a production line for a month. From being manufactured to becoming a true labor force, robots still have a long way to go.

From "able to take the stage" to "able to take the job," what changes is the competitive logic of the entire industry. Technological capability must ultimately pass the test of cost and efficiency before it can become a service that customers are willing to keep buying.

In the 2008 Pixar animated film WALL·E, WALL·E is a garbage-collecting robot. After humans left Earth, it worked alone on the desolate planet for 700 years, compressing garbage into cubes and stacking them one by one. It did not need to understand human society; it only needed to complete its pre-set work day after day.

WALL·E's most astonishing ability is not that it has emotions, but that it never stopped working over the long years.

The real-world robotics industry is still far from this seemingly simple goal. In the past few years, humanoid robots have run, jumped, and done somersaults on stage, looking more and more like agile humans. But when they enter factories, hotels, and shopping malls, they face a different set of tests: how to adapt to changing environments, handle surprises, work continuously, and complete tasks without engineers closely attending to them.

In the last week of October, AgiBot held 2026 AgiBot Innovation Day in Shanghai. Peng Zhihui, AgiBot co-founder, president, and CTO, proposed a new concept: Worken (effective intelligent work hours)—Work+Token. It consists of the number of active robots, the length of autonomous working time, and the output efficiency ratio, attempting to measure how much effective labor robots actually create in real environments. Simply put, AgiBot wants to create a "timesheet" for robots as well.

Behind this timesheet is a shift in evaluation standards.

At present, most companies are accustomed to showcasing robots' degrees of freedom, motion capabilities, and model parameters. These metrics answer how "strong" a robot is, but they have not yet answered how "useful" it is.

A robot being powered on for eight hours does not mean it works effectively for eight hours; completing an operation at a launch event does not mean it can work continuously and stably on a production line for a month. From being manufactured to becoming a true labor force, robots still have a long way to go. Worken attempts to bring the actual output along this path into the measurement scope, and also provides a new way of thinking for future billing by task, service, or even labor outcome.

At Innovation Day, AgiBot unveiled multiple products in one go: the third-generation hardware architecture AHA3, the embodied foundation model GO-3 Preview, the end-to-end omni-modal interaction model WITA-OMNI, the fully self-developed embodied intelligence native operating system Lingqu OS 1.0, as well as the heavy-duty robot Genie G2 Max, a new-generation dexterous hand, and other products, covering multiple software and hardware areas from the body and models to the operating system.

These products all point to one goal: enabling robots not only to complete individual actions, but also to understand tasks, execute operations, and replicate learned skills across different devices and scenarios.

AgiBot has already sent a batch of robots into real scenarios. According to data disclosed by the company, as of September 2026, AgiBot's general-purpose AI robots had cumulatively rolled off the production line in 20,000 units, appearing in factories, subways, theme parks, hotels, and other venues.

The real industry needs to answer more specific questions: Can a robot truly take on a job and make that job generate economic value? How much manpower can robots save? How much do equipment depreciation, deployment integration, maintenance, and operations cost? Does changing scenarios require reconfiguration?

From "able to take the stage" to "able to take the job," what changes is the competitive logic of the entire industry. Technological capability must ultimately pass the test of cost and efficiency before it can become a service that customers are willing to keep buying.

On the day of Innovation Day, Xuebao Finance and others spoke with several AgiBot executives. From models, bodies, and data to deployment, operations and maintenance, and business models in real scenarios, their answers show what a leading robotics company is solving as it moves toward the real world, and which problems remain unsolved.

We reorganized and grouped the interview content by theme, making slight edits while preserving the interviewees' intended meaning. The following is a transcript of the conversation:

Robot value begins to be measured with a "timesheet"

Q1: How would you summarize AgiBot's core progress this year? What stage has physical AI reached now?

Peng Zhihui (AgiBot co-founder, president, and CTO): In one sentence, it has entered the Y moment, truly moving from the development state to the deployment state. No scenario can be solved by a single point of technology alone: if the body is good but there is no intelligence, it is a remote-controlled toy; if the model is strong but there is no reliable body, it cannot be deployed either.

Among AgiBot's R&D team, 30% work on body hardware and 70% on intelligence R&D. Physical AI has already passed the stage of a "usable body," and vertical-domain intelligence is being deployed, such as the 24/7 quality inspection workstation at the Longqi factory. We are moving from the second stage to the third stage, and the "ChatGPT moment" will appear in the third stage.

Q2: Why are full-stack and "three intelligences in one" mandatory choices for AgiBot?

Peng Zhihui: In the deployment state, the gap cannot be broken down into separate problems of body, algorithm, or data; it must be solved in a coupled way. The body is the lower bound, while the model and intelligence capabilities are the upper bound. What we care about is not just running, jumping, and punching, but whether it can operate continuously, reuse the technology foundation, and reduce customers' integration and maintenance costs. Rolling 10,000 units off the line is only a victory for manufacturing; getting 10,000 units on duty is the victory of physical AI.

Yao Maoqing (AgiBot co-founder, co-president, and president of the embodied business system): A stable body, multi-robot consistency, data collection, model training, and scenario integration are naturally coupled. These links are subject to the "barrel effect": if any one is missing, the complete closed loop may be unable to advance.

Xiong Yan (AgiBot partner, co-president, president of the Central R&D Department, and president of the interaction business system): The Central R&D Department is responsible for shared technologies such as hardware, joints, and the public software platform, while also balancing business needs and long-term reserves. The "three intelligences" may also merge in the future; for now they are developed separately, and specialized models can all serve as teachers for comprehensive models in the future.

Q3: How does Worken become the measure of robot productivity?

Peng Zhihui: Worken is Work plus Token, and the Chinese name is set as "Zhiyuan." In the past, people looked at spec sheets; in the next stage, they will look at work-hour sheets: being powered on for 8 hours does not equal creating 8 hours of value, and one success does not equal continuous stable work for a month. What Worken answers is how much sustained, effective intelligent labor a robot can contribute in a real environment.

Wang Chuang (AgiBot partner, senior vice president, and president of the general business department): The main metrics are the actual number of deployed units, the working hours per unit, and the density of labor value. We have already counted the Worken created by more than 300 robots at Chimelong, 60 at Guangzhou Metro, and 3 in hotels, but the value density of different scenarios still needs to be refined. For example, a hotel robot may be able to do two of ten things, but it can work continuously, so we need to leverage strengths and avoid weaknesses.

Q4: In the future, will you charge by body, by service, or by Worken?

Peng Zhihui: It is not an either-or relationship. Customers who need support from technical and operations teams are suitable for purchasing equipment; customers who only need stable results are suitable for service delivery; and some customers hope to operate jointly with us. Our core is to make technology reusable, callable, and deliverable over the long term.

Wang Chuang: At present, we mainly sell the full-stack capability as a package, and we also have leasing, maintenance, and development agreements. In the future, it may become purchasing a productivity package: reserve tasks, and after completion, deduct according to the Worken consumed. AgiBot in the future will be a company that continuously creates productivity, not merely a maker of robot bodies.

Q5: How do you choose scenarios? Among industrial, commercial, and home, which comes first and which comes later?

Wang Chuang: There are three criteria: technically, it must be "within reach with a stretch"; it must have scale potential; and it must be able to accumulate along the main technical line. Industrial applications mostly use wheeled robots, while retail, guidance, and reception mostly use bipedal robots. Industry has higher requirements for cycle time, stability, safety, and cost-effectiveness; if service robots can bring traffic, emotional value, and revenue, their ramp-up may actually be faster.

Peng Zhihui: The home is not the starting point; it should be the endpoint. Home tasks vary greatly and have long chains, and they must also meet safety and consumer cost requirements. General-purpose robots are better suited to work that requires a certain degree of generalization, does not have such high cycle-time requirements, and is boring and repetitive; they can also collaborate with humans and traditional automation.

Beyond general models, why is an embodied brain still needed?

Q6: Will general large models "eat" embodied models?

Yao Maoqing: General models are good at high-level reasoning and planning, but they act slowly and consume a lot of tokens. In the future, VLA, world models, and the control layer will still be highly coupled with the body and tasks, and each layer will require investment from embodied companies.

Peng Zhihui: General foundation models are penetrating into embodied AI, and embodied models will also penetrate into general models. If general foundation models ultimately cover embodied capabilities, competition will depend on who can use capabilities in physical scenarios through the body; if not, data for vertical-domain training will still depend on the body. For a full-stack team, both outcomes have advantages.

Q7: What thresholds must cross-scenario generalization still cross?

Peng Zhihui: First, real-time performance: a cloud-based general model may take more than ten minutes to complete a simple task, but a robot must operate in real time. Second, reliability: if language is wrong, it can be corrected, but if an action is wrong, it means downtime losses. Third, physical common sense: digital models cannot feel gravity or friction. Fourth, task closure: the real physical impact of the output must be verified.

Q8: What architecture and training data does GO-3 Preview use?

Yao Maoqing: It uses the WRAM (World Reasoning Action Model) architecture, with training redone, while continuously adding data at different checkpoints. Data sources include existing collection, Mifeng procurement, and supplementary channels. There is both existing high-end stereo data and low-cost monocular data such as GoPro; the training set has now exceeded one million hours. (Editor's note: GO-3 Preview is the world's first embodied model released by AgiBot to achieve full-link large-scale learning across pre-training, post-training, and post-deployment.)

Q9: How large is the investment behind one million hours of data?

Yao Maoqing: The industry has rapidly moved from a few thousand hours in 2024 to one million hours, and is now pushing toward ten million hours. Ego data varies by monocular, stereo, and scenario; much of it costs one to two hundred yuan per hour, with expensive data at three to four hundred yuan, so one million hours is about 200 million yuan. Computing costs are even greater: training a model of tens of billions of parameters requires several thousand cards, and one thousand cards for a year costs one to two hundred million.

Peng Zhihui: This year, AgiBot should reach a training cluster of nearly ten thousand cards. If tokens become a deployment bottleneck, that is actually a good thing—it means users are willing to use it. Through device-cloud collaboration and model distillation, most interactions can be closed locally, reducing costs.

Q10: Does the drop in data prices mean data is no longer scarce?

Yao Maoqing: Prices for homogeneous data such as monocular, indoor bare-hand, and clothes-folding data will decline, while prices for real, segmented-scenario stereo, tactile, and wrist-camera data will remain firm. Low-quality data includes scene distortion, dropped frames and overexposure, hands out of frame, multi-camera setups failing to achieve millisecond-level synchronization, and insufficient 3D and 6D trajectory accuracy. Some companies buy one million hours and find only 10% or 20% usable, while top suppliers can achieve efficiency of 80% or even 90%; mainstream demand is shifting toward stereo, tactile, and full-body data.

Q11: Why does WITA-OMNI (AgiBot's end-to-end full-modality interaction model) unify speech, vision, facial expressions, and actions?

Xiong Yan: We return all modalities to the raw stage for unified pre-training; adding modalities later causes loss of intelligence and makes it difficult to align conversation timing. People do not interact in a question-and-answer fashion; they respond with facial expressions and body language, which is naturally full-duplex interaction. Embodied robots also need vision to judge "is this sentence addressed to me?" At Chimelong Paradise, when a child's voice is quiet, the robot proactively greets them, kneels on one knee, or bends down to shorten the interaction distance—something fixed speakers and ordinary models cannot do.

Q12: What is the most important improvement in the official version of WITA-OMNI?

Xiong Yan: First is robustness—it can handle anything said to it. For example, in a live demonstration with about one second of reverberation, the model still relied on its rejection capability to withstand confusing sounds. Second is emotion: understanding emotions and giving emotionally intelligent responses. Interaction speed and intelligence were already on the right path in the Preview stage.

Q13: How does real data form a flywheel? What if industrial data cannot leave the factory?

Xiong Yan: Chimelong Paradise already has more than one million interactions per month. Data from users who are dissatisfied or leave without completing a task is more valuable; it is checked, labeled, and fed back into training, so after a period of deployment the robot becomes noticeably more capable. Full-modality native training and real deployment data are both moats for the model.

Yao Maoqing: Industrial data involves process and commercial information, so we strictly comply with customer rules: it does not leave the factory or the workshop. Training servers and software platforms can be deployed on edge devices inside the factory to complete model upgrades. Data permitted to leave the factory must also be authorized; pre-training values diversity more, and a single workstation may not need hundreds of hours of data.

Operations, repeat purchases, and the ecosystem determine the commercial loop

Q14: From a few hundred units to tens of thousands, what really constrains scale-up?

Wang Chuang: Chimelong's more than 300 robots are distributed across more than 100 locations, with completely different tasks and environments; unified access, fault management, content updates, and issue closure are all very complex. Initially the ratio of robots to engineers was almost 1:1, later reduced to 3:1, with the goal of 10:1 by the end of this year and 20:1 next year. 20:1 means adding only 5% to the manpower budget, which many customers can accept. We also build dedicated networks for robots: carbon-based humans have human networks, silicon-based robots have their own networks, with higher real-time requirements.

Peng Zhihui: Customers will shift from "can you build it" to "can I afford to keep it." Equipment health, software versions, spare parts, remote diagnostics, and task optimization determine the full lifecycle cost. If every robot sold comes with a nanny, the day of large-scale deployment will never arrive.

Q15: What is the most critical quantitative standard for industrial scale-up?

Yao Maoqing: The core is deployment efficiency—how many person-days it takes to deploy a similar scenario. Training, testing and release qualification, software releases, OTA, data loops, and issue feedback must be turned into a toolchain so they can be replicated to new customers with higher efficiency. If the whole industry sees more than 10,000 robots in China working stably and for long periods in factories, that will be an important turning point. Manufacturers, secondary developers, and integration teams must also earn reasonable profits rather than falling into low-quality price wars.

Q16: How do Longcheer and Chimelong achieve replication and repeat purchases?

Yao Maoqing: Longcheer has expanded from tablet inspection loading and unloading to other categories, and has moved into handling and depalletizing. The R&D team withdrew long ago, basic operations were handed over to the customer, and sensor, wiring harness, and motor issues were resolved through version upgrades. More than three other 3C companies are working with us, and some have already completed line integration.

Xiong Yan: At Chimelong, one guide location went from one robot to three—that is a repeat purchase. Once customers see that the robot can successfully complete tasks, they are willing to buy again.

Wang Chuang: What customers are most satisfied with is the speed of iteration: software, algorithms, structure, and models keep improving, and the robot's growth is visible to the naked eye.

Q17: Will rapid new product launches sacrifice testing and lifecycle?

Yao Maoqing: We do not launch products just for the sake of launching products. Genie G2 has been refined for more than a year, and its predecessor design was prepared for a market lifecycle of about three years. G2 Max targets heavy-load needs: G2 has a 5 kg single-arm payload and 10 kg dual-arm payload, while G2 Max achieves a rated payload of 40 kg. The IPD process has strict reviews and release qualification tests; if key tests fail, the schedule is extended—we will not rush out a force-ripened product for speed.

Q18: Why must scale-up rely on the ecosystem? What is the signal of widespread adoption?

Wang Chuang: We cannot possibly master all data and industry know-how, and replication from 1 to N must also rely on partners. After Lingqu OS was open-sourced, Lingchuang, Lingxi, and Genie Studio let developers build skills, which are then orchestrated by Agents. What is opened up is not just code but also interests and the market: the humanoid robot market is large, and there is no need to earn every single yuan.

The sign of widespread adoption is robots everywhere, with customers placing orders at scale and purchasing and distributing on their own, with almost no need for AgiBot to provide services.

This article is from the WeChat public account "Snow Leopard Finance" (ID: xuebaocaijingshe), author: Chen Xuning, editor: Huang Yuntao, published by 36Kr with authorization.