GPT-6 Astra Shakes Up Embodied Intelligence: Data Orders Rebound, Hardware Companies Revalued, Brain Companies Under Pressure
After OpenAI released GPT-6 Astra, general-purpose large models can directly control robots to perform tasks, and the embodied intelligence supply chain has been re-stratified accordingly. In NVIDIA's tests, Astra achieved a 66.3% success rate in pick-and-place without fine-tuning, higher than π0.5; in Galbot's RoboDojo tasks, Astra's direct control success rate was 26%, rising to 48% when combined with π0.5. A data services company received more orders in two weeks than in the previous two months; Unitree's stock price rose more than 70% in the last two weeks of October, with its market value once approaching 300 billion yuan, while companies building embodied brains face competition on the same stage as major foundation model players. The article also points out Astra's shortcomings: direct action output takes an average of 18.56 seconds per inference, far slower than π0.5's tens of milliseconds; a successful action costs about $1.96, and planning plus execution costs about $17.72.
Why read this
AI score 75
Using measured data from NVIDIA and Galbot as well as changes in supply chain orders, the article breaks down the gains and losses at each link after general-purpose models enter the embodied intelligence field.
Topics
Body · AI translation
[[**GPT-6 Astra enables general-purpose large models to directly control robots, putting embodied "brain" companies under pressure, data procurement rebounds, hardware/body companies see value reassessed, and the industry shifts toward a division of labor where general models plan and embodied models execute.**
**Key Points**
**1. General models can directly control robots, putting pressure on the valuations of embodied "brain" companies**
In NVIDIA's tests, Astra achieved a 66.3% success rate on pick-and-place without fine-tuning, higher than π0.5; Galbot reached 48% when combined. The advantage zone of specialized models is being compressed.
**2. Data collection orders rebound, data service providers benefit first**
After Astra's release, one data service company received orders in two weeks exceeding the previous two months; a Nomura report said OpenAI procured about 10 million hours of embodied data, accounting for about 70% of the U.S. market.
**3. Hardware/body companies see value reassessed, general models fill in "brain" capabilities**
Unitree used Astra with zero training to complete painting and playing Go; in the last two weeks of October its stock price rose more than 70%, with market value once approaching 300 billion yuan; body motion control, cost, and reliability have become new anchors.
**4. General models still have shortcomings, and the industry may move toward layered division of labor**
Astra's direct action inference takes 18.56 seconds, far slower than π0.5's tens of milliseconds; direct action costs about $1.96, while planning plus execution costs about $17.72. In the short term, a division of labor may form between general planning and embodied execution. 2026-11-05 09:35 Embodied intelligence has changed Dingjiao One Quick Overview
This article comes from the WeChat public account: Dingjiao One, author: Dingjiao One team, editor: Wei Jia
Data collection orders surged, and embodied brain companies panicked.
Dingjiao One (dingjiaoone) original
An embodied intelligence data service company founded about a year and a half ago has received data procurement orders in the past two weeks exceeding the volume of the previous two months, which is like a shot in the arm against the backdrop of tightening financing in the entire embodied industry.
"The orders not only come from the embodied industry; some foundation model giants that previously only asked for quotes without placing orders have also started ordering," the company's founder told "Dingjiao One."
The change began with the release of GPT-6 Astra. In early September, after OpenAI released its latest-generation model, a large number of tests using Astra to control robotic arms for various tasks appeared on social media. Many embodied industry practitioners found that general multimodal foundation models also possess spatial understanding capabilities and can directly control robots to perform tasks, with success rates even exceeding some embodied models.
As test videos spread more widely, voices such as "general large models will sooner or later 'eat' embodied models" and "large language models will devour embodied models; embodied intelligence is merely a subset of general large models" began to circulate in the industry. Unitree Technology founder Wang Xingxing even said bluntly that after GPT-7 comes out next year, it may directly break through the embodied intelligence problem. But there are also different voices in the industry. Some believe that current general large models are still far from delivering a crushing blow to embodied intelligence, and that there are still obvious shortcomings in many areas such as long-horizon task success rates, latency, and cost.
The answer may not be unique, but the impact of Astra on the industry is real. Lingyu Intelligence CEO Jin Ge told "Dingjiao One" that the release of Astra itself is exciting news for the embodied industry. It first validated a new feasible direction and proved that data from the physical world and the non-physical world complement each other: "Under the Scaling Law, training physical data into large language models can likewise improve their non-physical capabilities."
This gives new value to embodied physical data, while also placing higher demands on embodied "brain" companies. An embodied intelligence investor told "Dingjiao One" that after Astra's release, the value of some embodied model startups and small and medium-sized companies will be called into question: "If a company does not have more than 1,000 compute cards, it will be difficult for them to survive in future competition with foundation model giants."
With the emergence of Astra, the industry has to answer whether embodied intelligence is an independent business or a branch on the capability tree of large language models that is destined to be covered.
01. Giants enter, is embodied intelligence being "dimensionally reduced"?
"The barbarians at the gate have arrived," an embodied industry practitioner told "Dingjiao One." This was the first reaction of many people around him to GPT-6 Astra.
In his view, once AI giants with money, compute cards, and talent reserves extend their capabilities into the field of embodied intelligence, it is like a group of barbarians standing outside the door who may break in at any time. What is even more unsettling is that Astra has made many practitioners realize that the know-how accumulated in this industry may not be as deep as everyone thought.
The core change brought by Astra is a cross-domain capability improvement: general large language models can handle robot manipulation and navigation tasks previously undertaken mainly by specialized physical models such as VLA and WAM, and they also demonstrate zero-shot generalization and in-context learning capabilities.
NVIDIA's research team tested Astra in mid-October. In object pick-and-place tests, without special fine-tuning using robot data, Astra could directly control the robot to perform tasks based on visual information and task instructions. The success rate of directly outputting actions relying on the Astra model reached 66.3%, far higher than the specialized model π0.5 of embodied intelligence company Physical Intelligence.
Chinese embodied AI company Galbot also ran a similar test. Among the selected RoboDojo tasks, Astra's success rate in directly controlling the robot was 26%, while π0.5's success rate was only 15.67%. If the two are combined, with Astra making decisions and π0.5 handling execution, the success rate rises further to 48%.
The robot's "brain" is usually divided into three layers. System 2 is responsible for understanding scenes, planning tasks, and reasoning; System 1 receives instructions from System 2 and, combined with environmental feedback, plans and adjusts actions; and finally System 0 handles low-level motion control. From the above experimental results, it can be seen that the capabilities of large language models are already better able to handle System 2 and part of System 1, while the advantage zone of embodied proprietary models has been compressed to part of System 1 and System 0.
Whether embodied intelligence can be deployed in the real world depends on the brain—this has long been a consensus in the industry, and the emergence of Astra has accelerated the brain's development process. At Unitree's partner ecosystem conference not long ago, Wang Xingxing's judgment was that some capabilities of GPT-6 can already solve part of the robot planning problem. As the capabilities of general large models improve, physical AI may see a major breakthrough in as soon as one to two years, or at the slowest two to three years.
He represents the more radical view in the industry. There is also a more conservative camp. Wang Qian, founder and CEO of the robotics company X Square Robot, believes that it is unlikely for foundation models to generalize to embodied intelligence through general intelligence in other domains. General large models cannot perform inference and deployment on robots at a reasonable speed; a purely embodied model is still needed to do this.
From the perspective of embodied intelligence companies, Astra has compressed the development window and living space of some companies, while at the same time providing new development paths and ideas for others. From the perspective of large model companies, the opportunities brought by Astra far outweigh the threats.
A recent embodied industry research report by Nomura Securities shows that from September 2025 to August-September 2026, OpenAI purchased about 10 million hours of embodied data, accounting for about 70% of U.S. market supply.
This data is the source of Astra's embodied "brain" capabilities. In addition to large-scale data procurement, OpenAI also plans to enter embodied intelligence itself. In June this year, OpenAI CEO Sam Altman posted a recruitment notice on social media, formally announcing its entry into the physical robotics track. He said the company is forming a new team called "OpenAI Robotics" and is publicly recruiting full-stack hardware, operations, systems, and machine learning engineers.
DeepSeek is also preparing embodied intelligence-related business and investments. Previously, according to media reports, Liang Wenfeng paid close attention to the embodied intelligence field and exchanged views with many entrepreneurs in the field, discussing directions such as robots, large models, and interaction with the real world. DeepSeek's plan to enter embodied intelligence may include forming its own team to conduct research and development, or entering through investment in related companies.
DeepSeek previously participated in Unitree's IPO allocation. Last month, it was also reported to have agreed to lead the seed round of Winverse AI, an embodied model startup founded by former Tencent scientist Hu Han, with an amount of tens of millions of dollars.
In addition, frontier general model companies such as Anthropic and Kimi are all brewing moves into embodied intelligence and physical AI and have begun related personnel recruitment. Embodied intelligence has changed from a proprietary track for robotics companies into the terminus on every AI company's roadmap.
02. Data collection "orders surge," embodied hardware makers rally, model companies hurt
The first to benefit from the "Astra effect" are the embodied data collection providers at the very upstream of the industrial chain.
The founder of the aforementioned embodied data service company told "Dingjiao One" that shortly before Astra was released and sparked discussions about embodied models, many data companies were not doing well. As the financing environment tightened, data procurement orders also decreased. Some model companies could not get investors' money and correspondingly reduced data procurement.
"Ego Centric data originally priced at 300 yuan per hour was even bargained down to a few dozen yuan," he said.
This is not actually because demand for data across the industry is decreasing. The general consensus in the embodied industry is that for embodied intelligence models to reach capabilities like GPT-3.5 in large language models, roughly 100 million hours of data are needed, while the amount of effectively cleaned data in the industry is still at the million-hour level.
Low data effectiveness, severe homogenization, and cutthroat competition among data collection companies have driven data prices lower and lower. But the emergence of Astra has changed the structure of data demand to some extent. General large language model companies have seen the value of embodied data. Whether they do not plan to enter the embodied intelligence industry and merely want to improve the model's overall intelligence level through training on physical data, or want to lay a data foundation for entering the embodied intelligence industry in the future, both have created new demand.
Compared with data collection companies, model companies in the middle of the industrial chain need to face changes in the competitive environment.
An embodied intelligence investor told "Dingjiao One" that OpenAI's launch of Astra has the most obvious impact on companies in the primary market that build embodied "brains" or implement scenarios. As the physical understanding capabilities of large language models gradually mature, the value of companies engaged in pure general embodied brain research and development will shrink.
Previously, a core reason why embodied "brain" companies could command high valuations was that investors believed general AI and physical AI were two independent, parallel development paths, and that a new OpenAI could emerge in the embodied model field. But if OpenAI itself can possess a certain degree of embodied model capability, that narrative will be broken.
However, in today's embodied intelligence industry, almost no company does only a single link: robot body companies will lay out data collection and model training-related businesses, and data companies also do model training. So the changes Astra brings to company valuations and the competitive landscape have not yet surfaced in a short time; it is more advancing the industry's overall development at the technical level.
The clearest beneficiaries of technological progress are downstream body companies, especially those robot body companies that already have mass-production and cost advantages; they can quickly fill in their "brain" capabilities by connecting to general large language models.
According to Wang Xingxing, after Astra was released, Unitree internally tried several robot experiments with this model. For example, using only prompts, having the robot draw, or having the robot play Go under zero-training conditions, Astra completed the tasks smoothly.
In the last two weeks of October, Unitree's stock price rose more than 70% cumulatively, and its market value at one point approached 300 billion yuan. An important reason behind this is that Astra's emergence raised market expectations for progress in embodied large models. When large language models begin trying to solve the problems that embodied large models need to solve, the value of OEMs is re-anchored in motion control, body reliability, cost control, and other aspects, while the "brain" instead becomes a commodity that can be plugged in and out at any time.
Although the shock brought by Astra has affected every link in the industrial chain, it is still too early to say at this stage that Astra will completely overturn the embodied intelligence industry.
03. Under large models, what kind of embodied companies thrive?
Zero-shot generalization, spatial understanding, and deep reasoning—these obvious advantages are not yet enough to support Astra in swallowing and replacing embodied models, and its shortcomings are equally obvious.
First is the latency problem. The videos on social platforms of Astra operating a robotic arm that are relatively popular are usually played at increased speed. At normal speed, the "slowness" problem of general large models would be exposed.
When large language models control robots, they often need to stop and think about what to do next. In Nvidia's test report, when Astra directly outputs end-effector targets on RoboLab-120, the average inference time per step reaches 18.56 seconds, and in Agent mode it still requires 13-18 seconds of thinking time. By contrast, the embodied proprietary model π0.5 needs only about tens of milliseconds.
If robots enter production scenarios or home service scenarios, the impact of this latency and pause will be magnified exponentially.
In addition, large language models run in the cloud, which requires considering network latency and stability. However, there are different voices in the industry on this point.
One view holds that robot operations have high requirements for real-time response. If the "brain" runs in the cloud and the execution layer is local, it will lead to unstable robot operation. Especially when doing long-horizon tasks, if a subtask fails due to signal fluctuation, the entire process may get stuck.
But another view holds that the robot's brain should run in the cloud. Lingyu Intelligence's Jin Ge told "Dingjiao One" that the ceiling of end-side computing power is not high. Taking Thor, the most advanced end-side chip in the industry, as an example, it takes about more than 100 milliseconds to compute one frame, while Nvidia's A100 chip from several years ago takes only more than 40 milliseconds to compute one frame.
In addition, if the large language model is only responsible for System 2 and part of the decision-making links of System 1, it actually does not require particularly high immediacy.
Besides the shortcoming of latency, cost is also an important factor that must be taken into consideration. Also in Nvidia's experiment, a successful direct action output consumes an average of 160,000 tokens, equivalent to about $1.96; if Astra is used for planning and the embodied model executes actions, it consumes an average of 11 million tokens per attempt, about $17.72; and further, having Astra generate a strategy to adjust actions based on interaction history to achieve one successful action requires 30.9 million tokens, about $36.40.
This is only the cost of one action. From a commercial perspective, the cost-effectiveness is very low.
At present, the most ideal application of large language models in the embodied intelligence field is to work in layers with embodied models. In the planning layer and semantic layer—task decomposition, scenario understanding, and instruction following—general models do these better and more cheaply. Millisecond-level motion control, force and tactile feedback, body adaptation, and the most critical execution layer of real-machine data closed loops are handed over to embodied models.
Under this new division of labor, the value chain of the embodied intelligence industry will also be redistributed. Wang Qian of Zhiyuan once proposed a view: he believes that today's intelligence body exists in data, and the model is more of a distiller and a container—during training, it is a distiller; during deployment, it is a container.
This means that high-quality embodied data is one of the decisive factors determining the unique competitive advantage of future embodied large models. In addition, data such as touch, force control, and force feedback that large language models cannot obtain through computation and reasoning will be more valuable.
More than a month after Astra's release, the industry has already voted with its actions: data orders have returned, the value of body companies has been re-anchored, and companies building "brains" have to face head-on competition with major foundation model companies.
Some windows are being closed, while at the same time some new windows are being opened, and the pace of the entire industry's development has become faster.
*The title image and illustrations in this article are all from Unsplash. AI industry signal channel: Frontier Technology
Has published 581 articles on Huxiu. This content comes from the internet and the original link; the views represent only the author and do not represent Huxiu's position. If there are copyright issues, please contact hezuo@huxiu.com, and we will verify and handle them promptly. Everyone who is changing and wants to change the world is on the Huxiu APP[ ]