Zhiyuan's Peng Zhihui responds to GPT-6 Astra: the stronger the large model, the more important the robot body

On October 29, AgiBot held its fourth Innovation Day at the Shanghai World Expo Center. In the main forum and the group interview afterward, GPT-6 Astra was mentioned repeatedly.

This general-purpose large model was recently shown to be able to directly control robots to complete certain operational tasks. The market's question is straightforward: if the foundation model is strong enough, is there still value in embodied companies developing their own brains, and will the body be reduced to contract-manufactured hardware?

The answer from Peng Zhihui (ZhiHuiJun), co-founder, president and CTO of AgiBot, is quite the opposite.

His reasoning is that no matter whether large models ultimately take over embodied intelligence entirely, intelligence must land on a reliable body to get work done. There are not many companies that can build this body at low cost, with high consistency, and at large scale.

AgiBot is currently the fastest-moving one. In 2025 AgiBot shipped more than 5,000 units, accounting for 40% globally according to Omdia data; in the first half of 2026 it shipped 9,700 units, with a share of over 43% according to Counterpoint. In September, its 20,000th robot rolled off the line, and first-half revenue approached 2 billion yuan.

01 The stronger the large model, the more important the body

Why the body is more important: Peng Zhihui's reasoning splits into two scenarios.

One is that a general foundation model truly covers embodied capabilities. Everyone gets the same intelligence, and what is compared is who can install it into the real world at lower cost and higher reliability, with scaled body manufacturers holding the advantage.

The other is that the foundation model cannot achieve true generalization, and each vertical domain still has to train its own model. Training data comes from deployment, deployment relies on the body, and body manufacturers still hold the advantage.

Both paths lead to the same conclusion: scalable, high-quality bodies will become increasingly important.

AgiBot partner Yao Maoqing cited an Nvidia report in the main forum: Astra's average success rate across 120 simulation tasks is 66.3%, with capabilities concentrated in semantic understanding and motion planning. But a single inference takes more than 18 seconds, and for highly dynamic fine motions involving force and tactile feedback, AgiBot's tests show a success rate of basically zero.

Peng Zhihui added another piece of information: Astra, with nearly a trillion parameters, runs in the cloud, and completing a simple task may take more than ten minutes or even half an hour, while robot operations are real-time.

Yao Maoqing's judgment is that large models will increasingly handle System 2 long-horizon planning, the skills trained by VLA are tools it invokes, and further down is 100 to 1,000 Hz real-time motor control; these layers are all deeply coupled with the body.

The role of the body is also reflected in post-training. In the lab, AgiBot networked 100 real robots and synchronized them in real time with the training cluster behind them; the robots send data back while working, and model weights are then distributed to each one. With the same training duration, comparing 4 robots and 100 robots, task success rate rose from about 50% to 98%. Yao Maoqing's point is that post-training also has scaling, and the scale refers to the number of real machines.

This path is hard for pure model companies to replicate.

AgiBot is working on both directions. GO-3 Preview is the embodied model with the largest parameter count released in the industry so far, topping three evaluation leaderboards including RoboDojo; at the same time, AgiBot also uses Astra as a base to create a self-evolving agent for motion intelligence.

In Peng Zhihui's view, models are a mandatory choice. If you do not build models, you will eventually be replaced by automakers that develop their own models, just like traditional autonomous driving solution companies.

This logic has a premise. Astra's slowness comes from parameter scale and cloud inference, and Peng Zhihui himself listed the solutions: layered design, model distillation, caching systems. If these problems are solved first by large model companies, the window for body manufacturers will be shorter than expected.

02 From spec sheets to work-hour sheets

Customers' questions have already changed.

In the past, when looking at robots, people looked at specs and at videos of running, jumping and punching. Now customers ask how much work it can do stably and how much each job costs.

AgiBot's answer is Worken, a combination of Work and Token, whose Chinese name is simply Zhiyuan, measuring a robot's effective intelligent work hours. The calculation splits into three multipliers: active units on duty, autonomous working time, and output energy-efficiency ratio. If a robot needs human takeover every 10 minutes of work, its autonomous work time is very limited.

On charging methods, Peng Zhihui did not give a single answer. Selling the body, selling services, and co-creating operations with customers are all possible, depending on whether the customer wants equipment or results.

What really constrains scale is operations and maintenance.

Peng Zhihui's exact words were, "If every machine sold still requires a nanny for the robot, that day will never come." Customers will be willing to buy at scale only when their question changes from "can it be built" to "can it be afforded."

Longcheer is currently the most complete sample. This 3C production line began scaled production in March 2026 and has run stably for more than seven months; during a six-day livestream, the operation success rate was 99.99%, with an average cycle of 22 seconds per item, on par with manual labor. Yao Maoqing revealed that AgiBot's R&D team had long withdrawn from the site, basic operations and maintenance were handed over to Longcheer's own team, and more than three other companies in the 3C field are cooperating.

Service scenarios are moving even faster. Nearly 500 robots have been deployed at Hengqin Chimelong Park, receiving more than 10,000 visitors per day on average, with iterations of a minor version every week and a major version every month after opening.

The human-to-machine ratio is another line. At first one engineer was assigned per robot, later 3 to 1, now 10 to 1, with the goal of 20 to 30 to 1. If the human-to-machine ratio cannot come down, every scenario requires piling up experts, and the business can only be done project by project.

Yao Maoqing gave two industry-level yardsticks. First, how many person-days it takes to deploy a new scenario; drawing an analogy to autonomous driving, he said a good team can get a vehicle model done with 10 people in 3 months, while a poor one needs 100 people working for a year. Second, when China's factories will have more than 10,000 robots operating stably over the long term.

Worken is still a metric defined by AgiBot itself; whether it can become an industry standard depends on whether customers and peers are willing to settle accounts by it.

03 Where the moat lies

Peng Zhihui's judgment on the price war is clear: without sufficient in-house development of the robot body and vertical integration capability, in the end all you can do is compete on price and cost.

AgiBot's hardware path has taken three steps. The new-generation AHA3 architecture, for the first time, puts three kinds of intelligence—motion, operation and interaction—into a single SoC, reducing the control system's volume by 50% and communication latency by 90%. The in-house development rate of core components exceeds 90%, but the most critical chip is still not self-developed; the next-generation AHA4 will raise that share.

Money and people are tilting toward intelligence. The R&D team has 30% working on the body and 70% on intelligence. This year's training compute is approaching a 10,000-card cluster; Peng Zhihui said that at this scale they are already competing for compute with general large-model companies.

He is not worried about Token costs on the deployment side. Factory work consumes almost no Tokens; interaction-heavy scenarios like Chimelong consume a lot, but small on-device models can close the loop on most conversations, and large models can then be compressed through distillation. In his view, Tokens truly becoming a bottleneck would actually be a good thing, because it would mean customers are willing to use it.

On the order of scenarios, he judges that the home is the endpoint. Safety, cross-task generalization, long-horizon tasks and consumer pricing all impose much higher requirements, so industrial and commercial scenarios go first.

According to the five stages of physical AI proposed by AgiBot, a usable body already exists, vertical-domain intelligence is being deployed, and the next step is to form a sufficiently generalized foundation model. Peng Zhihui believes the ChatGPT moment will appear at the third stage, "and it should be very soon."

Before that moment arrives, large-model companies and body companies will meet on the same question: who can make robots truly take up their posts.

The market carries risk, and investment requires caution. This article does not constitute personal investment advice, nor does it take into account the special investment objectives, financial situation or needs of individual users. Users should consider whether any opinions, views or conclusions in this article fit their particular circumstances. They invest at their own risk.