What happened in the embodied intelligence industry 50 days after the release of GPT-6 Astra

By Xu Xin, Edited by Ren Xiaoyu | Digital Intelligence Frontline

A robotic arm costing a few hundred dollars holds a paintbrush and draws the Golden Gate Bridge on white paper.

It drew it four times in total, changing to a new sheet of paper after each attempt. A camera recorded and transmitted back the images of each attempt, and the AI summarized and reviewed them and adjusted the force and position of the next stroke. On the first attempt, the lines were crooked and twisted; on the second, the outline of the bridge emerged; on the third and fourth, the bridge towers and suspension cables gradually became clear.

The entire process of the robotic arm drawing looks unremarkable, but what made the robotic arm move was not any model specially trained for robots, but GPT-6 Astra, released by OpenAI on September 3.

After Thijs Simonian, an employee of the OpenAI robotics team, posted this real-world test video on X, it quickly ignited discussion about whether general large models can directly operate robots.

The reason Astra's video evaluation attracted so much attention is that it is crossing a dividing line in the robotics industry. In the past, when a large model was connected to a robot, it did command work: it understood the task of "draw a bridge" and broke it down into instructions; actually making the arm move and do things required another policy model specially trained on robot data.

But in the video posted by Thijs Simonian, what used to require a team to collect data, train models, and repeatedly debug in order to accomplish on a single task can now be done by Astra, a model not born for robots, which has the ability to directly take over the robot's "hands."

"Many founders of robotics companies are very panicked," a practitioner in the embodied AI industry told Digital Intelligence Frontline. After Astra, the paradigm of the embodied industry is changing. After Astra, the technological paradigm of embodied intelligence will very likely become one combined with large language models, and embodied AI will have to face competition directly from general large model companies. This will reconstruct the previous industry logic.

Overseas, some commentary asserted that "the OpenAI of the robotics field will be OpenAI itself, not Physical Intelligence or Skild."

A chain of reactions has already appeared.

The capital market has a keen sense of smell. An industry figure mentioned that the valuation logic of embodied companies was previously built on the idea that general models would find it difficult to swallow the embodied field. Now that general large language models outperform previous leading models on embodied tests, the valuation logic is bound to change.

In the latter half of the third quarter this year, the capital market was in a round of cooling. The secondary market tightened its review of humanoid robot listings, and because the commercialization progress of the embodied track had lagged behind its valuation logic, investors were already more cautious than in the first half of the year. After Astra, funds that had already tightened will only become more picky.

On the other hand, it is no longer news that leading big tech companies are entering the embodied field. The industry believes that companies on the embodied track will in the future have to face even fiercer market competition.

Recently, ByteDance's Seed Model team intensively updated dozens of robotics positions; OpenAI signed a 200,000-square-foot site in the Bay Area in March this year for its robotics business, and in September recruitment for robotics-related positions increased substantially; DeepSeek was also recently reported to have led the investment in Winverse AI, founded by former Tencent distinguished scientist Hu Han. Compared with these giants, embodied companies have yet to grow up, and their reserves of resources such as capital, computing power, and talent are far inferior to those of general large model companies. The severe tests ahead can be imagined.

Some people at leading domestic embodied companies have also taken the initiative to talk about Astra's impact. Wang Xingxing, founder of Unitree Robotics, which is known for its robot body capabilities, publicly discussed the shock brought by GPT-6 Astra twice in mid- and late October, and admitted that his biggest mistake in the past was that "my belief in AI was not enough."

Wang Xingxing mentioned that Unitree did not train the model at all, yet the GPT model could control the robot to autonomously identify the palette, change paint, clean the brush, and so on. If things go fast, by the end of this year or next year, robots may be able to complete tasks such as tidying up a room according to instructions.

He even predicted that when GPT-7 comes out next year, "it may directly break through the embodied intelligence problem."

In contrast to Unitree's frank acknowledgment of the impact, many companies have remained relatively silent amid this wave of discussion.

Hardware bodies, brains, data, simulation, production lines, the industry is still continuously observing and discussing the impact of Astra and subsequent general large model updates on the embodied circle.

However, just as the attitudes of practitioners in these different links differ, some views hold that Astra will deal a considerable blow to some companies, but in the long run, the process of applying and deploying embodied intelligence is expected to be greatly accelerated.

Benchmarks hit 99% overnight, the robot's "GPT moment"?

For many people, the emergence of Astra means the "GPT moment" in the robotics field.

After Astra was released, the lab of Jim Fan, Nvidia's robotics head, quickly switched the backend model of its own robotics framework to Astra. "Overnight, all benchmarks were pushed to about 99%," Jim Fan said in a post.

The team of Xu Huazhe, founder of Pokebot and assistant professor at Tsinghua University's Institute for Interdisciplinary Information Sciences, also conducted tests on the very day Astra became usable. Xu Huazhe mentioned in a podcast interview that in the past, upgrades to large models on robotics tasks meant that what could not be done before still could not be done, while Astra brought "a change in the entire qualitative picture."

Wang Jiawei, chief scientist of Shenpu Intelligence, said that it was the first time he had seen a general-purpose language model achieve such a high success rate at grabbing blocks, "beyond everyone's imagination." Chen Boyuan, founder of Inverse Matrix Technology, also admitted candidly in a public speech in September that the emergence of Astra had greatly shocked him, showing everyone how a language model performs when it enters the real physical world.

The shock and impact it brings are real.

An embodied industry practitioner told Shuzhi Qianxian that the embodied industry had previously been mentally prepared for paradigm shifts in VLA and world models, but few had expected a general-purpose large model to directly strike with a dimensionality reduction attack.

Zhu Zheng, co-founder of Jijia Shijie, mentioned at the Bund Conference in September that after Astra, embodied and world model companies "are very likely to eventually be eaten by language model companies; this is a very realistic problem," and the time window left for them is only "a brief 1 to 2 years."

Shen Yujun, chief scientist of Ant Lingbo, also commented at the Bund Conference that companies doing embodied intelligence based on digital world models "will all find this terrifying," somewhat like the wave of post-training two or three years ago: as foundation model capabilities suddenly jumped from 60 points to 90 points, much of the work they had done became meaningless.

Wang Yunbo, associate professor at Shanghai Jiao Tong University, believes that Astra has changed the embodied brain. The "slow brain" of the embodied brain should be built on top of large language models; for the layer of understanding and planning, VLAs trained only on demonstration data are no longer the optimal solution.

Of course, not everyone is willing to define it with the "GPT moment."

SenseTime, together with teams including Daxiao Robotics, systematically tested Astra across 12 simulation benchmarks and 8 real-robot platforms. Their conclusion was that GPT-6-Astra has basically acquired embodied competence, but there is still an obvious gap before it reaches practical embodied intelligence in real environments.

Jia Peng, co-founder of Zhijian Dynamics, sees that it does not yet have productive value. He said at the Bund Conference that Astra already has generalized understanding of space and environments, but what he wants more is "a Claude Code moment, truly able to generate productivity." That step still requires three or four, or four or five, years.

Zhu Xing, CEO of Ant Lingbo, defined Astra as a tool at the Bund Conference, believing that GPT-6 is a tool for improving R&D efficiency, and that the more fundamental barriers are data know-how and pipelines.

Some practitioners also believe that Astra still cannot truly complete a closed decision-making loop in real physical environments.

Chen Boyuan tried using GPT-6 to operate a gripper to pick up a cup and found that "it might take 40 minutes to an hour to slowly pick it up," which made him reflect on the upper limits of language models. Chen Boyuan believes that 100 milliseconds feels like nothing in a chat window, but on a real robot it may correspond to three control cycles; the physical world is constantly changing, and the model cannot respond in time. Therefore, physical AI needs its own foundation model so as to better make decisions and close loops in the real world.

Chen Zhe, founding partner of Alphaist, observed that attitudes toward this model in the embodied circle are actually divided. Chen Zhe mentioned that at IROS, a top robotics conference held in Pittsburgh in late September and early October, someone surveyed more than 100 researchers; about one-third believed that "robotics research is over," while the largest group remained pragmatists who treat large models as tools.

Whatever the attitude, it is undeniable that the entire embodied industry is discussing GPT-6 Astra, and it is becoming an unavoidable presence in the embodied circle.

Where Astra is strong and where it is weak

Over the past 50-plus days, starting from the first set of real-robot test results released by third-party evaluation agency Robocurve on September 4 to the release of SenseTime's joint report in late October, various institutions, companies, and researchers have been taking Astra apart inside and out based on more than a dozen public benchmarks.

After multiple evaluations, the industry's understanding of Astra's strengths and weaknesses has become clearer.

The general industry consensus is that Astra is strong in semantic understanding and generalization to unfamiliar scenarios. Switch to an unfamiliar machine or an unfamiliar kitchen, and it can still get started directly; when it fails, it can review and correct itself, and it can even read tactile signals. This is precisely the generalization ability that specialized models found hardest to achieve in the past.

Xu Huazhe's team once conducted a flower-arranging experiment. They deliberately made the task harder: the lower end of the flower stem was split in two, and the forked stem was difficult to insert into the bottle mouth. Astra could first pinch the fork to narrow it, then push it in. "Essentially, this is because it already understands 3D space, and such understanding can already help with a great many non-dynamic physical-world manipulation tasks."

Xu Huazhe mentioned that what surprised him even more was Astra's ability to turn reflection into action. Given an unfamiliar robot body, it quickly discovers whether the robot moves fast or slow and how large its range of motion is; if it bumps into a table, it judges that "this table is quite high" and then raises its hand a little.

This generalization ability is exactly what specialized models lack most. With past VLA models, changing a green cup to a red cup could cause failure. Astra, however, needs only a few photos to reconstruct the entire laboratory as a simulation environment.

Even touch, regarded as a "forbidden zone" for general models, has had a breach opened in it.

According to a report by the Chinese edition of MIT Technology Review, in the TacTalent report released by the Tsinghua team, no policy model was fine-tuned at all; Astra was given only a few demonstrations, and then the tactile signals on the gripper were input as images together with the camera feed. On the UniVTAC benchmark, its average success rate rose from 76.3% for vision-only to 93.8%; among these, the key-pulling task rose from 30% to 90%. When the key gets stuck, it can determine the direction of obstruction based on changes in the tactile images on both sides of the gripper, and then adjust its action.

On two visuotactile benchmarks, its performance was 22.9 and 13.1 percentage points higher, respectively, than the best specialized models previously reported. "Without a breakthrough in foundation model capabilities, it would be hard to achieve such results relying only on demonstrations and a framework," said Yang Chen, a PhD student at Tsinghua and an author of the paper.

The industry has also seen the shortcomings of its current capabilities. For example, when encountering tasks that require millimeter-level contact, dynamic actions requiring real-time response, and tasks that are hard to explain in language, the success rate drops.

Take the evaluation of Astra by the third-party testing organization Robocurve as an example. When they used Astra for the task of putting blocks into a bowl, it almost never failed, succeeding 19 times in 20 tests. But when Astra controlled the robot to fit a round puzzle piece into a groove, it succeeded only 2 times. The gap occurs in those few millimeters that require contact and high-precision control.

Researchers from multiple institutions including the University of Hong Kong, Tsinghua, and Berkeley jointly tested Astra on 42 tasks in the RoboDojo benchmark and found that its capability distribution is "highly polarized": it generalizes well on tasks that require semantic understanding but do not demand high-precision control; once a task requires precision, dynamic control, or complex dual-arm coordination, its performance deteriorates markedly.

Another obvious shortcoming reported by the industry is that it is slow. A person in the embodied AI industry told Shuzhi Qianxian that most of the videos circulating online of Astra controlling robots have to be played at 16x or even 32x speed to be watchable. Meanwhile, Nvidia's Astranomy project measured that GPT-6 Astra requires an average of 18.56 seconds for a single inference under the direct policy mode of embodied intelligence.

Explaining the obvious latency, Xu Huazhe said that Astra's internal chain of thought is very long, and "the thinking time for complex tasks and simple tasks is basically similar." This means it is very smart on difficult problems, yet appears very dumb on simple matters.

Token consumption is also a problem. In a report jointly released by SenseTime and multiple companies and institutions, it was mentioned that for the task of "putting a pen into a pen holder," which a human can complete in a few seconds, Astra took more than 4 hours in total and cost $55. A report by the Galaxy General team showed that on 50 RoboDojo test instances, direct control by Astra consumed more than 1.13 billion tokens cumulatively.

To some extent, the conclusion of the report jointly issued by SenseTime and multiple institutions precisely summarizes Astra's current capabilities: it already has the ability to complete embodied tasks, but there is still a gap before it reaches practical embodied intelligence.

It is worth noting that some views in the industry clearly hold that the current shortcomings of general large language models represented by Astra in embodied settings may be only temporary, because general large models such as GPT-6 Astra were not specifically optimized for physical manipulation. In the future, as relevant data are completed and enough physical contact data are obtained, the shortcomings may be remedied.

What to use Astra for

Now that Astra has become an unavoidable presence, a new question naturally arises. In a robotic system, which layer should Astra be placed in, and what can companies use a general model like Astra to do?

Chen Zhe divided the current uses of Astra into three categories: first, directly serving as a policy model to complete robot tasks; second, supervising, assisting in, and correcting existing VLA action models to improve task success rates; third, serving as an embodied researcher to autonomously build simulation environments, debug motion policies, and realize Physical RSI. A mid-October embodied intelligence industry research report by Debon Securities also adopted this division.

In fact, directly serving as a policy model to complete robot tasks is exactly the use shown in the most popular bridge-drawing video on X. Although there are obvious problems of being slow and expensive, compared with the previous role of large models as planners, it is one step further.

Xu Huazhe gave an analogy. In the past, when you told a robot, "I want to eat tomato and egg stir-fry," the large model was responsible for planning "open the refrigerator door, take out the tomatoes, take out the eggs," while how exactly to open the refrigerator and crack the eggs was left to the embodied model below. Now, Astra can directly say "move your hand 2 centimeters to the right," and if it fails to grasp, "lean forward a bit more." The execution policy beneath Astra has become more like a reflex system.

Using Astra as a supervisor differs from the previous model of layered large and small brains, in which the large model gives the overall goal and the VLA executes. Astra has to check step by step, letting the VLA first propose actions, then observing, reviewing, and correcting them when necessary. Judging from the evaluation results currently published on the market, the effects and costs brought by this usage are both obvious.

A joint report released by SenseTime and several institutions shows that, under the previous layered approach in which only Astra issued subgoals and the VLA executed them, the success rate was just 5.83%; when it was changed to having Astra review the 4 candidate actions proposed by the VLA before execution, the success rate rose to 25.83%. More critically, adding the Astra layer requires no additional data collection or retraining.

Having Astra act as supervisor also comes with costs in money and time. For example, in Nvidia's tests, after adding Astra's review, a single episode consumed about 11 million tokens and cost $17.72, and speed slowed accordingly.

At present, the industry is also paying close attention to Astra's capabilities in simulation, and many teams use it to build simulation environments and debug motion policies.

Zhu Wentao, an assistant professor at Ningbo Eastern Institute of Technology, mentioned that a student in his team tried giving just a single instruction, and Astra ran autonomously for a day and a half in the Isaac Lab simulation, training a policy that lets a dexterous hand twirl a pen. Two weeks ago, Nvidia's GEAR lab also open-sourced the ASPIRE framework, in which a large model writes control code itself, reads execution logs, revises and retries, and stores validated corrections as code in a skill library for direct use in subsequent tasks, accelerating embodied R&D.

Beyond the three usages above, some teams have observed that Astra can be used directly for data annotation and cleaning.

Wang Xiaogang, chairman of Daxiao Robotics, said at this year's Bund Summit that Astra can output valuable intermediate data, and such data may help foundation models accelerate their evolution. Wang Jiawei of Shenpu Intelligence also revealed on a podcast that they tried using Astra to annotate and clean self-collected data and found that it "does better than humans and can work 24 hours a day," at a cost similar to manual labor. In addition, Yuance Future also said that in internal tests using it to generate 3D assets, "the results were fairly good."

Some major companies in the industry have already turned this into dedicated tools. This month, Nvidia open-sourced the data engine Flora and estimated the current price of data annotation. It concluded that annotating one hour of robot video with Astra at the highest reasoning intensity costs about $88.5, while Flora, specially fine-tuned on top of open-source models, is estimated to cost only about $0.72, can annotate 200,000 hours per day, and scores slightly higher on quality than Astra (the two are measured on somewhat different bases, and Flora is currently limited to research use). It is foreseeable that the threshold for annotation will be lowered accordingly.

Whether building simulations, generating 3D assets, or annotating data, all rely on Astra's understanding of three-dimensional space. At present, the industry has differing judgments on the source of this capability.

One view holds that it comes from specialized 3D data. OpenAI may be hiring large numbers of data annotators and doing extensive modeling and trajectory annotation.

Another view attributes it to spillover from programming ability. Li Tianyu of Yuance Future believes that it uses programming to do reinforcement learning in purely cloud-based virtual environments, "forcing out some understanding of the 3D world." Overseas developer George Wing is even more radical, believing that Astra's ability to control real machines to complete tasks may be an emergent result of reinforcement learning at scale on code.

It can be said that the general large models represented by Astra are opening a new door for the embodied field.

Aftershocks

Astra's impact is also spreading outward along the industrial chain.

Different companies have reacted very differently to this wave. Unitree's Wang Xingxing made intensive statements, moving the timetable for general-purpose robots at the industrial expo held in October from "two to three years" mentioned in August to "one to two years." Some companies with embodied layouts, such as SenseTime and Daxiao Robotics, have released evaluation reports, open-source models, and frameworks, embracing the wave. Many other companies remain silent.

At present, leading hardware manufacturers are seen by many as relatively certain beneficiaries. One saying is that no matter who makes the model, it ultimately has to run on a real robot.

Dhanush Radhakrishnan, CEO of overseas biomimetic humanoid robot company Clone Robotics, believes that the "half-life of frontier models is only 6 months," while returns in the robotics field will ultimately go to "the category-leading superhumanoid robot platform." This judgment is quite similar to the trajectory of the large-model industry in recent years: leadership in frontier models often lasts no more than a few months before being caught up with, and the most stable winner is Nvidia, which sells GPUs.

The importance of data has been reconfirmed. Astra's shortcomings in fine manipulation point precisely to the lack of physical contact data, and it is no longer news that large-model companies have begun buying and collecting data in large quantities.

In September this year, Yao Maoqing, CEO of the embodied data crowdsourcing platform Mifeng Technology, said in an interview that among their clients, besides embodied companies and world-model companies, general large-model companies are their third category of customers. And the latest versions of leading large models have acquired some ability to make decisions and plans in the physical world precisely because they have begun training such data into the multimodal capabilities of their foundation models.

Large-model companies are no longer limited to buying data. The factory OpenAI leased in the Bay Area has a power capacity of more than 14,000 amperes, far exceeding ordinary industrial buildings, and is seen by outsiders as the prototype of a robot data factory; its robot data collection operations positions explicitly list coordinating the daily operation of workstations, operators, and materials. The data and evaluation positions newly posted by ByteDance's Seed team require building a complete closed loop spanning collection, processing, training, evaluation, and model feedback.

An embodied industry practitioner told Shuzhi Qianxian that embodied companies have invested heavily in data in the past, and even if others want to collect data, it would take them a long time to do so. But no one can say clearly how long this time gap can last.

Whether simulation benefits is a matter of differing views. Chen Zhe believes that if general models need to do reinforcement learning in virtual worlds, whoever can provide more and more realistic 3D scenes will have bargaining power. Xu Huazhe poured cold water on this, saying that the real-to-sim work Astra does mainly reconstructs scenes in terms of geometry and appearance, and "the physics engine has not become more accurate."

The capital market's reaction was very direct. In September, embodied financing dropped noticeably, exactly in the same month Astra was released; but this was the result of multiple factors including cooling, tighter IPOs, and pressure to validate orders, and "the risk of being replaced by general large models" is becoming a new consideration for investors.

The impact has not yet been transmitted to production lines. This year, embodied intelligence companies have accelerated the data closed loop in factory and logistics scenarios, and industrial customers value cycle time and stability most. For example, when BMW's Spartanburg plant was testing Figure 02, it set the total cycle time at 84 seconds, of which loading time was only 37 seconds. Although Astra currently performs extremely well in understanding and planning complex tasks, on real production lines, high latency and cost remain thresholds it cannot yet overcome. As a result, deployment projects already running in factories and warehouses will not change pace because of it in the short term.

Astra's impact continues.

Some practitioners believe that for the entire industry, the impact needs to be viewed separately from the perspectives of "proportion" and "volume." The proportion that embodied intelligence companies get in the industrial chain may be squeezed, but progress in general models will accelerate the deployment of physical AI, and the overall market volume will instead be enlarged.