Zhiyuan "Robot General Mobilization": Why Haven't We Got Our Own WALL-E Yet?

Fast Reading

• Peng Zhihui proposed a new concept of Worken (Effective Intelligent Working Hour), which consists of the number of active robots, autonomous working duration and output efficiency ratio, aiming to measure how much effective labor robots actually create in real scenarios.
• A robot being powered on for 8 hours does not mean it works effectively for 8 hours; completing one operation at a press conference does not mean it can work continuously and stably on the production line for a whole month. There is still a long way to go from the moment a robot is manufactured to when it becomes a real labor force.
• The shift from "being able to perform on stage" to "being able to take up a formal post" changes the competitive logic of the entire industry. Technical capabilities must ultimately pass the tests of cost and efficiency before they can turn into services that customers are willing to purchase continuously.

In the 2008 Pixar animated film WALL·E, WALL·E is a garbage cleaning robot. After humans left the 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, but only needed to complete the pre-set work day after day.

The most surprising capability of WALL·E is not that it has emotions, but that it never stops 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 been running, jumping and somersaulting on stage, looking more and more like agile humans. But when they enter factories, hotels and shopping malls, they face a whole different set of challenges: how to adapt to changing environments, handle unexpected situations, work continuously, and complete tasks without the close care of engineers.

In the last week of October, Agibot held the 2026 Agibot Innovation Day in Shanghai. Peng Zhihui, Co-founder, President and CTO of Agibot, proposed a new concept: Worken (Effective Intelligent Working Hour) — a portmanteau of Work and Token. Composed of the number of active robots, autonomous working duration and output efficiency ratio, it is designed to measure how much effective labor robots actually create in real scenarios. Simply put, Agibot aims to create a "working hour sheet" for robots.

Behind this working hour sheet is a shift in evaluation criteria.

At present, most enterprises are accustomed to demonstrating the degrees of freedom, motion capabilities and model parameters of robots. These indicators answer the question of "how capable the robot is", but fail to answer the question of "how useful it is".

A robot being powered on for 8 hours does not mean it works effectively for 8 hours; completing one operation at a press conference does not mean it can work continuously and stably on the production line for a whole month. There is still a long way to go from the moment a robot is manufactured to when it becomes a real labor force. Worken attempts to include the actual output on this path into the measurement scope, and provides a new idea for billing by tasks, services and even labor outcomes in the future.

On the Innovation Day, Agibot launched a number of products in one go: the third-generation hardware architecture AHA3, the embodied foundation model GO-3 Preview, the end-to-end full-modal interaction model WITA-OMNI, the full-stack self-developed embodied intelligence native operating system Lingqu OS 1.0, as well as products including the heavy-load robot Sprite G2 Max and the new generation of dexterous hands, covering multiple software and hardware fields from ontology and models to operating systems.

All these products point to one common goal: enabling robots not only to complete a single action, but also to understand tasks, execute operations, and replicate the skills they have learned to different devices and scenarios.

Agibot has already deployed a batch of robots into real scenarios. According to data disclosed by the company, as of September 2026, a total of 20,000 Agibot general-purpose AI robots have rolled off the production line, and they are operating in places such as factories, metro stations, theme parks and hotels.

The real industry needs to answer more specific questions: Can a robot truly take over a job and make the job generate economic value? How much manpower can robots save? How much does equipment depreciation, deployment and integration, maintenance and operation cost? Is re-debugging required when changing scenarios?

The shift from "being able to perform on stage" to "being able to take up a formal post" changes the competitive logic of the entire industry. Technical capabilities must ultimately pass the tests of cost and efficiency before they can turn into services that customers are willing to purchase continuously.

On the day of the Innovation Day, Snow Leopard Finance and other media communicated with multiple senior executives of Agibot. From models, ontology and data to deployment, operation and maintenance and business models in real scenarios, their answers show what a leading robotics company is solving and what problems remain unsolved as it moves toward the real world.

We have reorganized and aggregated the interview content by theme, and made slight edits on the premise of retaining the original intention of the interviewees. The following is the full record of the conversation:

Robot Value Begins to Be Measured by "Working Hour Sheet"

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

Peng Zhihui (Co-founder, President and CTO of Agibot): In one sentence, we have 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: if you only have good ontology but no intelligence, it is just a remote-controlled toy; if you have a strong model but no reliable ontology, it cannot be implemented either.

30% of Agibot's R&D team works on ontology hardware, and 70% works on intelligence R&D. Physical AI has passed the stage of "usable body", and domain-specific intelligence is being implemented, such as the 24/7 quality inspection station at Longcheer 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 capability and "three-intelligence integration" mandatory options for Agibot?

Peng Zhihui: The gap in the deployment state cannot be split into separate ontology, algorithm or data problems, which must be solved in a coupled manner. Ontology sets the lower limit, while model and intelligence capabilities set the upper limit. We do not only care about running, jumping and boxing, but also about continuous operation, reuse of the technology base, and reduction of customer integration and maintenance costs. 10,000 units rolling off the production line is only a victory for the manufacturing industry, while 10,000 units being put into actual work is the victory of physical AI.

Yao Maoqing (Co-founder, Co-President, President of Embodied Business System of Agibot): Stable ontology, multi-machine consistency, data collection, model training and scenario integration are naturally coupled. These links follow the "wooden barrel effect": if one link is missing, the complete closed loop may not be able to move forward.

Xiong Yan (Partner, Co-President, President of Central R&D Department, President of Interactive Business System of Agibot): The Central R&D Department is responsible for shared technologies such as hardware, joints and public software platforms, while taking into account business needs and long-term technology reserves. The "three intelligences" may also be integrated in the future. Now they are developed separately, and professional models can all act as tutors for the comprehensive model in the future.

Q3: How can Worken become the measurement standard for robot productivity?

Peng Zhihui: Worken is a combination of Work and Token, and its Chinese name is officially set as "Agi Worken". In the past, people looked at parameter sheets, and in the next stage, we need to look at working hour sheets: 8 hours of power-on does not mean creating 8 hours of value, and one success does not mean continuous and stable work for a whole month. Worken answers the question of how much continuous and effective intelligent labor robots can contribute in real environments.

Wang Chuang (Partner, Senior Vice President, President of General Business Department of Agibot): It mainly depends on the actual number of deployed units, the working duration of each unit and the labor value density. 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 in different scenarios still needs to be refined. For example, a hotel robot may be able to do 2 out of 10 tasks, but it can work continuously, so we need to leverage its strengths and avoid its weaknesses.

Q4: Will future billing be based on ontology, services or Worken?

Peng Zhihui: It is not an either-or relationship. Customers who need the support of technical and operation teams are suitable for purchasing equipment, customers who only need stable results are suitable for service delivery, and there are also customers who want to co-operate with us. Our core is to make the technology reusable, callable and deliverable in the long term.

Wang Chuang: Currently, we mainly sell the full set of full-stack capabilities, and we also have leasing, maintenance and development agreements. In the future, it may become purchasing productivity packages: pre-set tasks, and deduct fees according to the consumed Worken after the tasks are completed. Agibot will be a company that continuously creates productivity in the future, not just a manufacturer of robot ontology.

Q5: How to select scenarios? Which comes first among industry, commerce and household use?

Wang Chuang: There are three criteria: the technology should be "achievable with a little extra effort", the scenario has large-scale potential, and it can precipitate the main technical line. Wheeled robots are mostly used in industry, while biped robots are mostly used in retail, tour guiding and reception. Industry has higher requirements for rhythm, stability, safety and cost performance; if service robots can bring traffic, emotional value and revenue, their deployment volume may grow faster instead.

Peng Zhihui: Household use is not the starting point, but the end point. Household tasks vary greatly, have long chains, and need to meet safety and C-end cost requirements. General-purpose robots are more suitable for jobs that require certain generalization, do not have very high rhythm requirements, and are boring and repetitive, and they can also collaborate with humans and traditional automation.

Why Do We Need an Embodied Brain Besides General-Purpose Models

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

Yao Maoqing: General models are good at high-level reasoning and planning, but they are slow in action and consume a lot of tokens. In the future, VLA, world model and control layer will still be highly coupled with ontology and tasks, and every layer requires investment from embodied companies.

Peng Zhihui: General foundation models will penetrate into the embodied field, and embodied models will also penetrate into general models. If the general foundation model finally covers embodied capabilities, the competition will focus on who can apply the capabilities to physical scenarios through the ontology; if not, the data for domain-specific training will still rely on the ontology. For full-stack teams, both results have advantages.

Q7: What thresholds still need to be crossed for cross-scenario generalization?

Peng Zhihui: First is real-time performance: a cloud general model may take more than ten minutes to complete a simple task, but robots must operate in real time. Second is reliability: if you say a wrong word, you can correct it, but if you make a wrong action, it will cause shutdown losses. Third is physical common sense: digital models cannot perceive gravity and friction. Fourth is task closed loop: the real physical impact caused by the output must be verified.

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

Yao Maoqing: It adopts the WRAM (World Reasoning Action Model) architecture, the training is redesigned, and data is continuously added at different check points. Data sources include existing collected data, data purchased from Mifeng and supplementary channels. There are both high-end binocular data and low-cost monocular data from devices such as GoPro. At present, the training set has exceeded millions of hours. (Editor's note: GO-3 Preview is the world's first embodied model released by Agibot that realizes full-link large-scale learning from pre-training, post-training to post-deployment.)

Q9: How much is the investment behind millions of hours of data?

Yao Maoqing: The industry has rapidly developed from thousands of hours in 2024 to millions of hours, and is now impacting the level of tens of millions of hours. Ego data varies with different monocular, binocular and scenarios. Many types of data cost 100 to 200 yuan per hour, and expensive ones cost 300 to 400 yuan per hour, so millions of hours of data cost about 200 million yuan. The computing power cost is even higher: training a tens-of-billion-parameter model requires thousands of GPUs, and 1,000 GPUs cost 100 to 200 million yuan a year.

Peng Zhihui: Agibot will reach a training cluster of nearly 10,000 GPUs this year. If tokens become the deployment bottleneck, it is a good thing instead, which means users are willing to use the service. We can realize most interactions are closed-loop locally through end-cloud collaboration and model distillation to reduce costs.

Q10: Does the fall in data price mean that data is no longer scarce?

Yao Maoqing: The price of homogeneous data such as monocular data, indoor bare-hand data and clothes-folding data will go down, but the price of binocular, tactile and wrist camera data in real segmented scenarios remains firm. Low-quality data includes data with distorted scenarios, frame drop and overexposure, hands out of the frame, multi-cameras not synchronized at millisecond level, and insufficient accuracy of 3D and 6D trajectories. Some companies buy 1 million hours of data but only 10% to 20% of it is usable, while the effective rate of leading suppliers can reach 80% or even 90%. The mainstream demand is moving towards binocular, tactile and full-body data.

Q11: Why does WITA-OMNI (the end-to-end full-modal interaction large model developed by Agibot) need to unify voice, vision, expression and action?

Xiong Yan: We make all modalities return to the original stage for unified pre-training. Adding modalities later will cause intelligence loss, and it is difficult to align the conversation timing. Human-to-human communication is not just question-and-answer, but also responses with expressions and body language, which is naturally full-duplex interaction. Embodied robots also need to use vision to judge "whether this sentence is spoken to me". In Chimelong Paradise, when children speak in low voices, the robot will take the initiative to greet them, kneel on one knee or bend down to narrow the interaction distance, which fixed speakers and ordinary models cannot achieve.

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

Xiong Yan: The first is robustness: it can respond to any sentence. For example, there is about 1 second of reverberation in the real scene demonstration, and the model can still resist confusing sounds through its rejection recognition capability. The second is emotion: it can understand emotions and give emotionally intelligent responses. The interaction speed and intelligence level have been on the right track in the Preview stage.

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

Xiong Yan: There have been more than 1 million interactions in Chimelong Paradise in one month. The data of users who leave because of dissatisfaction or uncompleted tasks is more valuable, which will be checked, labeled and returned for training. After the robot is deployed for a period of time, it will become obviously more capable. Full-modal native training and real deployment data are both the moat of the model.

Yao Maoqing: Industrial data involves process and commercial information, so we strictly abide by customer regulations, and ensure data does not leave the factory or workshop. We can deploy training servers and software platforms on the edge side inside the factory to complete model upgrades. Data that is allowed to leave the factory must also obtain authorization; pre-training pays more attention to diversity, and a single post does not necessarily require hundreds of hours of data.

Operation and Maintenance, Repurchase and Ecosystem Determine the Commercial Closed Loop

Q14: From hundreds of units to tens of thousands of units, what really restricts large-scale deployment?

Wang Chuang: More than 300 robots in Chimelong are distributed in more than 100 points, with completely different tasks and environments. Unified access, fault management, content update and problem closed loop are very complex. At first, the ratio of robots to engineers was almost 1:1, then