Zhiyuan's Peng Zhihui: Physical AI Still in Early Stage, Proposes Worken Effective Intelligent Work-Hour Framework

Download the app

+1 Peng Zhihui of AgiBot: Physical AI is still in its early stage; only when tens of thousands of humanoid robots are on the job will physical AI have won. The Paper reporter Yu Yan 2026-10-30 07:29 Source: The Paper ∙ Sci-Tech 101 > Listen to full text Font size

On October 29, at the 2026 AgiBot Innovation Day held in Shanghai, Peng Zhihui, co-founder, president and CTO of AgiBot, said in an interview with The Paper Tech (www.thepaper.cn) that the overall development of physical AI is still in its early stage. AgiBot has already passed the first stage, that is, possessing robot bodies that can work over the long term, stably and at low cost; it is now advancing deployment in vertical segmented scenarios in the second stage.

In his view, only in the third stage, when humanoid robots achieve single-machine cross-scenario generalization, may the key node of the "ChatGPT moment" arrive, though he did not give a specific time frame.

In the first half of 2026, AgiBot's overall revenue was close to 2 billion yuan, and robot deliveries reached 9,700 units; in September, the cumulative 20,000th general-purpose AI robot rolled off the production line. Peng Zhihui says AgiBot is in the second stage of physical AI. Source: The Paper reporter Yu Yan (02:32)

In the past, people focused on spec sheets; in the future, more attention will be paid to "work-hour sheets"

Over the past year, competition among humanoid robot hardware manufacturers has largely centered on hardware performance, shipment volumes and flashy demos.

But in this interview with The Paper Tech, Peng Zhihui pointed out that a robot being powered on for 8 hours does not mean it has created 8 hours of value. He believes that in the future more attention needs to be paid to the "work-hour sheet" of humanoid robots, and AgiBot has for the first time proposed a measurement framework called Worken (effective intelligent work hours). What Worken tries to answer is: in real environments, how much sustainable and effective intelligent labor can a robot actually contribute.

According to the calculation framework proposed by Peng Zhihui, Worken = number of active robots × autonomous working hours × output efficiency ratio.

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

Peng Zhihui believes that the body is the lower limit of a humanoid robot, while models and intelligence capabilities are the upper limit. "What we focus on is how to simply make technology and capabilities reusable enough to deliver to customers; this is our long-term competitiveness."

In the past two years, the concept of physical AI has become a focus of the industry. The large-scale promotion of this concept came from Nvidia founder and CEO Jensen Huang in June 2024. Jensen Huang believes that the next wave of artificial intelligence is physical AI that can understand the laws of physics and work in the real world.

The "Embodied Intelligence Development Report (2026)" recently released by the China Academy of Information and Communications Technology gives a definition of the concept of physical AI: physical AI aims to break through the limitation of traditional models that "can only represent, lacking physical constraints," and refers to embedding the laws of physics into the entire process of perception, reasoning, generation and action, so that artificial intelligence has greater authenticity, reliability and deployability.

At this Innovation Day event, Peng Zhihui divided the technological evolution of physical AI into five leaps: having a usable body, entering the deployment state in vertical-domain scenarios, forming a sufficiently generalized physical AI foundation model, evolving from single machines to groups, and ultimately realizing self-elevation of intelligent systems.

In an interview with The Paper Tech, he said that AgiBot has already passed the first stage and is advancing the second stage, that is, entering the deployment state in segmented vertical-domain scenarios, carrying out horizontal replication and expanding into more scenarios.

The third stage is when humanoid robots have a generalized embodied foundation model, robots break free from customized development, quickly adapt to various new tasks at lower cost, and achieve single-machine cross-scenario generalization. Peng Zhihui believes that only when the third stage is reached can the "ChatGPT moment" of physical AI ultimately be realized.

He pointed out that as the foundational capabilities of large language models become stronger and stronger, generalization in single scenarios is gradually taking shape. But to truly enter the physical world and achieve cross-scenario generalization, four major pain points still need to be overcome: real-time performance, reliability, physical common sense and task closed loops.

Taking GPT-6 Astra as an example, he said that a cloud-based large-parameter model may take more than ten minutes or even half an hour to complete a simple task, making it difficult to meet the real-time operation requirements of robots; in addition, general models still have hallucination problems, and the physical world's tolerance for errors is far lower than that of the digital world.

"If a language model says one or two wrong sentences, it is fine to correct them, with no substantive impact, but when a humanoid robot makes a wrong movement while working, it will cause downtime and huge losses," Peng Zhihui said.

Beyond technology, whether customers are willing to pay also determines whether robots can achieve larger-scale deployment.

Peng Zhihui clearly pointed to the customer side as the bottleneck for large-scale deployment: "It is not a question of whether we are willing to roll it out, but whether customers in this market are willing to pay. We are fully capable of manufacturing tens of thousands of units a year; there is no problem with that."

Peng Zhihui also emphasized that, in addition to Worken, supporting toolchains and platform-based operations and maintenance capabilities are needed. Only when robot output can be measured, anomalies can be tracked and scheduled, experience can be reused through the platform, operating costs are reduced, and the break-even point of return on investment (ROI) is ultimately reached, will customers be willing to continue purchasing and enterprises be able to expand deployment scale.

A full-stack approach is a mandatory choice for humanoid robotics companies

As general-purpose large models such as GPT-6 Astra continue to penetrate the physical world, where lies the value of embodied intelligence companies developing their own brain models, and is it still necessary to invest substantial resources in developing their own "brains"?

In response, Peng Zhihui believes that general-purpose large models and embodied intelligence models will penetrate each other in the future.

He stated that AgiBot does not wish to be merely a supplier of robot hardware bodies, but rather to continuously invest in full-stack capabilities including models, software, and bodies, so as to avoid core competitiveness becoming overly dependent on external model suppliers.

On the same day, AgiBot also released the new-generation embodied intelligence foundation model GO-3 Preview, as well as the AHA3 hardware platform and the industrial robot G2 Max.

Peng Zhihui said that GO-3 Preview is the embodied model with the largest number of parameters released in the industry to date.

According to disclosures by Yao Maoqing, co-founder, co-president, and president of the embodied intelligence system at AgiBot, based on the evaluation results presented by the company, GO-3 Preview topped all three embodied intelligence evaluation leaderboards: RoboDojo, RoboTwin, and RoboColiseum.

Peng Zhihui defined the full-stack path as a mandatory choice rather than an option for AgiBot. He believes that if general-purpose models possess sufficiently strong embodied capabilities in the future, the focus of competition will be who can efficiently deploy intelligence into the real world; if vertical-scenario training is still required, companies that possess robot bodies and real operational data will also have an advantage.

Peng Zhihui also disclosed the structure of R&D investment: 30% of the entire R&D team works on body hardware-related R&D, and 70% works on intelligence brain-related R&D. From this investment structure, AgiBot regards models and intelligence capabilities as the focus of long-term competition. Peng Zhihui believes that traditional equipment tends to depreciate gradually with use, but in the era of physical AI, robots may exhibit a pattern in which "hardware depreciates, but intelligence ultimately appreciates." Responsible editor: Huan Yanhong; image editor: Zhu Weihui; proofreader: Liu Wei. The Paper news tip line: 021-962866. The Paper, unauthorized reproduction prohibited. +1 Favorite I want to report #PhysicalAI #HumanoidRobots #AgiBot View more