Zhiyuan's Peng Zhihui: 10,000 Humanoid Robots Rolling Off the Line Is Only a Manufacturing Victory; 10,000 Robots Truly "On the Job" Is the Victory of Physical AI

As embodied intelligence is gradually deployed in various real-world scenarios, companies' accounting logic is changing—from evaluating "can a robot be built" in the past to thinking "can I actually afford to keep a robot."

"In the early stage of industry development, everyone competed over robots' paper specifications and flashy demos; after entering a new stage, what matters more is the value of real work hours." At the 2026 AgiBot Innovation Day held on October 29, Peng Zhihui, co-founder, president and CTO of AgiBot, said in an interview with Securities Times reporters and other media that robot power-on time does not equal effective output, and a single task's success rate does not equal long-term stable work. The core of AgiBot's newly released metric unit Worken (effective intelligent work hours) is to measure whether a robot can autonomously and continuously complete valuable work and how much actual output it produces.

Peng Zhihui pointed out that he hopes robot services will be billed by Worken work hours in the future, but robots with different levels of intelligence doing different tasks will have different Worken rates. He said the charging method is not an "either-or" relationship: "You can sell both the robot body and services, just as the software industry has PaaS (Platform as a Service) and SaaS (Software as a Service); different customers need different delivery methods."

In the interview, Peng Zhihui divided the development of physical AI into five stages: having a usable body, entering the deployment stage in vertical-domain scenarios, forming a sufficiently generalized embodied foundation model, co-evolution as a group, and ultimately achieving self-iterative upgrading of intelligent systems. He believes AgiBot has already passed the first stage, with mass-produced bodies achieving continuous delivery in strict scenarios such as industry. It is currently advancing toward the second stage (vertical-domain intelligent deployment), for example, at the Longcheer 3C factory, where robots are replacing quality inspection workstations on a 7×24 basis.

In Peng Zhihui's view, only when humanoid robots reach the third stage of single-machine cross-scenario generalization may the key node of the "ChatGPT moment" arrive.

However, to truly move toward the third stage of "forming a sufficiently generalized embodied foundation model," physical AI still faces four major pain points. Peng Zhihui pointed out: first, real-time performance—trillion-parameter cloud large-model inference takes a long time and is difficult to match the high requirements of robots' physical operations; second, reliability—the "hallucinations" of large models are harmless in the text field, but in physical environments they can cause production downtime and physical losses; third, lack of physical common sense—digital models lack real perception of gravity, friction and so on; fourth, the physical closed loop—language models only output results, while robots must verify the actual impact of actions. These pain points urgently need to be addressed by the industry.

Physical AI has been a hot topic in the industry in recent years, and it is also a war of attrition over computing power and data. Peng Zhihui revealed that AgiBot's training cluster computing power this year has reached nearly 10,000 cards, a huge scale in the industry. "If Token consumption becomes the bottleneck of the deployment stage, that is actually a good thing—it shows the business model has been proven." He admitted that consumption varies greatly across scenarios: high-frequency interaction scenarios such as the Chimelong embodied intelligence theme park consume a lot, while factory operations consume very little. AgiBot is optimizing costs through device-cloud collaboration and model distillation, so that most interactions are closed-loop on small on-device models.

As for the eventual large-scale popularization of physical AI, Peng Zhihui believes the home is the endpoint, not the starting point. Because home scenarios have extremely high requirements for safety, cross-task generalization and consumer-side costs, industrial and commercial services will be the first to achieve deployment-stage implementation. The ultimate test of physical AI is always in real scenarios. Peng Zhihui emphasized: "The rollout of 10,000 humanoid robots is only a victory for manufacturing; 10,000 humanoid robots being able to 'take up their posts' is the victory of physical AI."