50 Days After GPT-6 Astra's Release: The Capability Boundaries of the Embodied Industry and a Revaluation of the Industry Chain
Huxiu Digital Intelligence Frontline reviews the impact on the embodied industry in the 50 days since GPT-6 Astra's release, noting that this general-purpose model—not trained for robots—can already directly control robotic arms, crossing the boundary between general-purpose large models and embodied execution. Multiple evaluations show it has strong semantic understanding and generalization to unfamiliar scenarios, succeeding 19 out of 20 times at putting blocks into a bowl, but fine manipulation and real-time performance are limited: it succeeded only 2 times on a round puzzle, averaged 18.56 seconds per inference, and took over 4 hours and cost 55 dollars to put a pen into a pen holder. Three usage patterns have emerged in the industry: serving directly as a policy model, reviewing VLA candidate actions to raise the success rate from 5.83% to 25.83%, and building simulations and cleaning data; embodied financing fell back in September, and investors have added consideration of replacement risk, but the BMW production line's material feeding takt time is 37 seconds, and the short-term deployment pace remains unchanged.
Why read this
AI score 78
Reviewing the evaluation data and practitioner reactions in the embodied industry during the 50 days since GPT-6 Astra's release, to present the capability boundaries and industry chain revaluation after general-purpose large models took over robot execution.
Topics
Body · AI translation
[[**GPT-6 Astra can directly control robots to perform tasks, crossing the boundary between general large models and embodied execution, and disrupting the existing VLA and specialized policy paradigms; it generalizes strongly but is slow and expensive, and fine manipulation remains limited, so valuation and competition logic are being reassessed.**
**Key points**
**1. General large models directly take over robot execution**
In the past, large models only planned and VLA executed; Astra can control a robotic arm without robot-specific training. Unitree's Wang Xingxing said he previously did not believe enough in AI, and predicted that GPT-7 might break through embodied intelligence.
**2. Astra generalizes strongly, but precision, real-time performance and cost are weak points**
It is strong in unfamiliar scenes, touch and error correction, succeeding 19 out of 20 times at putting blocks into a bowl; for a round puzzle it succeeded only 2 times, single inference took 18.56 seconds, and putting a pen into a pen holder took 4 hours and cost 55 dollars.
**3. Astra can serve as an executor, a VLA supervisor and an R&D data tool**
Direct control is slow and expensive; reviewing VLA candidate actions can raise the success rate from 5.83% to 25.83% without additional data collection and retraining; it can also build simulations, generate 3D assets and clean data.
**4. Capital and the industrial chain are being reassessed, but production lines are unchanged in the short term**
Embodied financing fell back in September, and investors added "being replaced" to their considerations; OpenAI, ByteDance and others are stepping up robotics/data efforts, with hardware and data benefiting more. On BMW's production line the material feeding takt is 37 seconds, and Astra's latency and cost make it hard to apply, so short-term implementation is unchanged; the long-term market may expand. 2026-10-29 19:36 The Astra shockwave and the embodied industry's 50 days, a Shuzhi Frontline quick overview
A robotic arm costing a few hundred dollars holds a paintbrush and draws the Golden Gate Bridge on white paper.
It drew four times in total, changing to a new sheet of paper after each drawing; the camera recorded and sent back each attempt, and the AI summarized, reviewed and adjusted the force and position of the next stroke. On the first attempt, the lines were crooked; on the second, the bridge's outline emerged; on the third and fourth, the bridge towers and suspension cables became clearer bit by bit.
The whole process of the robotic arm drawing looks unremarkable, but what made the 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 OpenAI's 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 large models were connected to robots, they did the commanding work, understanding the task "draw a bridge" and breaking it into instructions; actually making the arm move and do things required another policy model specially trained on robot data.
In the video posted by Thijs Simonian, in the past it took a team to collect data, train models, and repeatedly debug before a single task could be accomplished; now Astra, as a model not built for robots, has the ability to directly take over the robot's "hand."
"Many robotics company founders are very panicked," an embodied industry practitioner told Digital Intelligence Frontline. After Astra, the paradigm of the embodied industry is changing. After Astra, the embodied intelligence technology paradigm will very likely become one combined with large language models, and embodied companies will have to face competition directly with general large model companies. This will reconstruct the previous industry logic.
Overseas, some comments assert 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 assumption that general models would find it difficult to swallow the embodied field; now that general large language models' performance on embodied tests has surpassed that of the previous top models, the valuation logic is bound to change.
In the latter half of the third quarter this year, coinciding with a round of cooling in the capital market, the secondary market tightened its review of humanoid robot listings, and coupled with the fact that commercialization progress in the embodied track had lagged behind the 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, and the industry believes that companies in the embodied track will in the future have to face even more intense 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 for its robotics business as early as March this year, and in September robot-related hiring 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, so severe tests can be expected ahead.
Some people at leading domestic embodied companies have also taken the initiative to talk about Astra's impact. Unitree Technology, which excels in robot body capabilities, saw its founder Wang Xingxing publicly discuss the shock brought by GPT-6 Astra twice in mid- and late October, and admit that his biggest mistake in the past was "not believing in AI enough."
Wang Xingxing mentioned that Unitree did not train the model at all, yet the GPT model could control the robot to autonomously recognize the palette, change paints, clean brushes, 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 been relatively silent amid this wave of discussion.
Hardware bodies, brains, data, simulation, and production lines—the industry is still continuously observing and discussing the impact of Astra and upcoming general large model updates on the embodied sector.
However, just as the attitudes of practitioners in these different segments 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.
1. 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 lead, 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 Pokeba Robotics and assistant professor at Tsinghua University's Institute for Interdisciplinary Information Sciences, also ran tests on the very day Astra became available. Xu Huazhe mentioned in a podcast interview that in the past, upgrades to large models on robotic tasks meant that what could not be done before still could not be done, whereas Astra "changed the entire qualitative picture."
Wang Jiawei, chief scientist of Shenpu Intelligence, said that it was the first time he had seen a general language model achieve such a high success rate in grasping blocks, "beyond everyone's imagination." Chen Boyuan, founder of Inverse Matrix Technology, also admitted frankly in a public speech in September that the emergence of Astra gave him a great shock and showed everyone the performance of language models entering the real physical world.
The shock and impact it brings are real.
A practitioner in the embodied industry told Digital Intelligence Frontline that the embodied industry had previously been mentally prepared for paradigm changes in VLA and world models, but few had expected that a general large model would directly deliver a dimensionality-reduction strike.
Zhu Zheng, co-founder of Jijia Vision, mentioned at the Bund Summit 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."
Ant Group Lingbo's chief scientist Shen Yujun also commented at the Bund Summit that companies doing embodied work based on digital world models "will all find this terrifying"—somewhat like the wave of post-training work two or three years ago: as foundation model capabilities suddenly jumped from 60 points to 90 points, much of their work became meaningless.
Wang Yunbo, an associate professor at Shanghai Jiao Tong University, believes Astra has changed the embodied brain. The "slow brain" of the embodied brain should be built on top of large language models, at the level of understanding and planning; VLAs trained solely on demonstration data are no longer the optimal solution.
Of course, not everyone is willing to define it with a "GPT moment."
SenseTime, together with teams including Daxiao Robotics, systematically tested Astra across 12 simulation benchmarks and 8 real-robot platforms. Their conclusion: GPT-6-Astra has basically acquired embodied competence, but there is still a clear gap before it reaches practical embodied intelligence in real environments.
Jia Peng, co-founder of Zhijian Dynamics, sees it as not yet having productivity value. At the Bund Summit he said Astra already has generalized understanding of space and environments, but what he wants more is "a Claude Code moment, something that can truly generate productivity." That step will still take three or four, or four or five, years.
Zhu Xing, CEO of Ant Group Lingbo, defined Astra at the Bund Summit as a tool, arguing that GPT-6 is a tool for improving R&D efficiency, and that the more fundamental moat is data know-how and pipelines.
Some practitioners also believe Astra cannot yet truly complete a 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 lift it," which made him reflect on the ceiling 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 the loop in the real world.
Chen Zhe, founding partner of Alphaist, observed that attitudes toward this model within the embodied community are actually divided. Chen Zhe mentioned that at IROS, the 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 the 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 community.
2. Where Is Astra Strong, and Where Is It Weak?
Over the past 50-plus days, starting from September 4, when the third-party evaluation firm Robocurve released its first set of real-machine test results, through the release of SenseTime's joint report in late October, various institutions, companies, and researchers have been taking Astra apart from the inside out based on more than a dozen public evaluation benchmarks.
After multiple evaluations, the industry's understanding of Astra's strengths and weaknesses has become clearer.
The broad industry consensus is that Astra is strong in semantic understanding and generalization to unfamiliar scenarios. Swap in an unfamiliar machine or an unfamiliar kitchen, and it can still get started directly; when it fails, it reviews and corrects itself, and it can even read tactile signals. This is precisely the generalization capability that specialized models have 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 opening. Astra could first pinch the fork to narrow it, then push it in. "Fundamentally, this is because it already understands 3D space, and that 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. Give it an unfamiliar robot body, and it will quickly figure out whether this robot moves fast or slow and how large its range of motion is; if it bumps into a table, it will judge that "this table is quite tall" and then raise its hand a little higher.
This generalization capability is exactly what specialized models lack most. Past VLA models could fail if a green cup was swapped for a red one. Astra, by contrast, needs only a few photos to reconstruct an entire laboratory as a simulation environment.
Even touch, regarded as a "forbidden zone" for general models, has been breached.
According to a report by the Chinese edition of MIT Technology Review, in the TacTalent report released by the Tsinghua team, without fine-tuning any policy model, they gave Astra only a few demonstrations and then fed the tactile signals on the gripper into the model as images together with the camera feed. On the UniVTAC benchmark, its average success rate rose from 76.3% with 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 visual-tactile 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 capability, it would be very difficult to achieve such results with demonstrations and frameworks alone," said Yang Chen, the paper's author and a Tsinghua doctoral student.
The industry has also seen the shortcomings of its current capabilities. For example, when it encounters tasks requiring millimeter-level contact, dynamic actions requiring real-time response, or tasks that cannot be clearly explained in language, the success rate drops.
Take the third-party evaluation agency Robocurve's assessment of Astra as an example: when using Astra to do the task of putting blocks into a bowl, it almost never missed—it succeeded 19 times out of 20 tests. But when Astra controlled the robot to fit a round puzzle piece into a groove, it succeeded only 2 times. The gap occurred in those few millimeters that require contact and high-precision control.
Researchers from the University of Hong Kong, Tsinghua, Berkeley, and several other institutions jointly tested Astra on 42 tasks in the RoboDojo benchmark and found that its capability distribution is "highly polarized": for tasks that require semantic understanding but do not demand high-precision control, it generalizes very well; 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. And NVIDIA's Astranomy project measured that GPT-6 Astra, under Direct Policy mode for directly generating actions in embodied intelligence, requires an average of 18.56 seconds per inference.
As for why there is such an obvious latency, Xu Huazhe explained that Astra's internal chain of thought is very long: "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 things.
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 dollars. 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 in total.
To some extent, the conclusion of the report jointly issued by SenseTime and multiple institutions accurately summarizes Astra's current capabilities: it already possesses the ability to complete embodied tasks, but it is still some distance from practical embodied intelligence.
English text is being prepared.
III. What is Astra used for?
Now that Astra has become an unavoidable presence, a new question naturally emerges. In a robotic system, at which layer should Astra be placed, and what can companies use a general model like Astra to do?
Chen Zhe divided Astra's current uses into three categories: first, directly acting as a policy model to complete robot tasks; second, supervising and assisting in correcting existing VLA action models to improve task success rates; third, serving as an embodied researcher that autonomously builds simulation environments, debugs motion policies, and realizes Physical AI self-evolution (Physical RSI). Debang Securities' embodied intelligence industry research report in mid-October also followed this division.
In fact, directly acting as a policy model to complete robot tasks is exactly the use in the most popular bridge-drawing video on X. Although there are obvious problems of being slow and expensive, compared with the previous model in which large models played the planning role, it is another step forward.
Xu Huazhe gave an analogy. Previously, when telling 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 underneath. Now, Astra can directly say "move your hand 2 centimeters to the right," and if it fails to grasp, "move a little closer forward." The execution policy under 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 must check step by step, letting the VLA first propose actions, then observing and reviewing them, and correcting them when necessary. Judging from the evaluation results currently published on the market, the effects and costs brought by this use are both very obvious.
A joint report released by SenseTime and multiple institutions shows that under the previous layered approach, with only Astra issuing sub-goals and the VLA executing, the success rate was only 5.83%; when changed to having it review the 4 candidate actions proposed by the VLA before execution, the success rate rose to 25.83%. More importantly, adding the Astra layer does not require collecting additional data or retraining.
Astra as a supervisor also has costs in money and time. For example, in Nvidia's tests, after adding Astra review, one episode consumes about 11 million tokens and costs $17.72, and the speed also slows accordingly.
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 the team tried giving only one instruction, and Astra ran autonomously in the Isaac Lab simulation for a day and a half, training a policy for a dexterous hand to spin a pen. Two weeks ago, Nvidia's GEAR lab also open-sourced the ASPIRE framework, in which the large model writes control code itself, reads execution logs, modifies and retries, and stores verified effective corrections in a skill library in code form for direct invocation by subsequent tasks, accelerating embodied R&D.
In addition to the three uses 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 Conference that Astra can output valuable intermediate data, which 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 roughly comparable to manual labor. In addition, Yuance Future also said that internal tests using it to generate 3D assets produced "fairly good results."
Some major tech companies in the industry have already turned this into a dedicated tool. Nvidia open-sourced the data engine Flora this month and estimated the current price of data annotation. It concluded that using Astra to annotate one hour of robot video at the highest reasoning intensity costs about $88.5, whereas Flora, which is specially fine-tuned on top of an open-source model, is estimated to cost only about $0.72 and can annotate 200,000 hours per day, with a quality score slightly higher than Astra's (the two are not measured on exactly the same basis, and Flora is currently limited to research use). It is foreseeable that the barrier to entry in the annotation step will be lowered accordingly.
Whether building simulations, generating 3D assets, or annotating data, all of it relies on Astra's understanding of three-dimensional space. At present, the industry has differing judgments about where this capability comes from.
One view holds that it comes from specialized 3D data. OpenAI may have hired a large number of data annotators and done a large amount of modeling and trajectory annotation.
Another view attributes it to spillover from coding ability. Li Tianyu of Yuance Future believes that it uses programming to do reinforcement learning in a purely cloud-based virtual environment, "forcing out some understanding of the 3D world." Overseas developer George Wing is even more radical, believing that Astra's ability to control a real machine to complete tasks may have emerged from reinforcement learning at scale on code.
It can be said that the general-purpose large models represented by Astra are opening a new door for the embodied field.
IV. Aftershocks
Astra's impact is also spreading outward along the industry chain.
English text is being prepared.
At present, leading hardware manufacturers are seen by many as the 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," and that returns in the robotics field will ultimately accrue to the "category-leading super-humanoid robot platform." This judgment is quite similar to the trajectory of the large-model industry over the past few years: leadership in frontier models often cannot last more than a few months before being caught up, 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 a 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 customers, besides embodied companies and world-model companies, general large-model companies are their third category of customers. The latest versions of leading large models have acquired certain capabilities for decision-making and planning in the physical world, also because they have begun training this data into the multimodal capabilities of their foundation models.
Large-model companies are no longer limited to buying data. The factory building OpenAI leased in the Bay Area has a power capacity of more than 14,000 amperes, far exceeding ordinary industrial buildings, and is regarded by outsiders as the prototype of a robot data factory; its robot data collection operations job responsibilities explicitly include coordinating the daily operation of workstations, operators and materials. The data and evaluation positions newly released 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; even if others want to collect it, they will also need a long time to collect it. But no one can say clearly how long this time gap can be maintained.
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: the real-to-sim work done by Astra mainly reconstructs scenes in terms of geometry and appearance, and "the physics engine has not become more accurate."
The reaction of capital has been very direct. In September, embodied financing fell significantly, exactly in the same month as Astra's release; but as a result of multiple factors such as tighter IPOs and pressure to verify orders, "the risk of being replaced by general large models" is becoming a new consideration for investors.
The impact has not yet reached the production lines. This year, embodied intelligence companies have been accelerating the data closed loop in factory and logistics scenarios, and what industrial customers value most is cycle time and stability. 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. At present, although Astra performs extremely well in understanding and planning complex tasks, on real production lines, high latency and cost remain thresholds it is difficult for it to overcome right now. This means that projects already running in factories and warehouses will not change their pace because of it in the short term.
The impact brought by Astra is still continuing.
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 implementation of physical AI, and the volume of the entire market will instead be enlarged. AI industry signal channel: Frontier Technology
Has published 20 articles on Huxiu. This content is authorized by the author for publication, and the views represent only the author himself and do not represent Huxiu's position. If you have objections or complaints about this manuscript, please contact tougao@huxiu.com. Everyone who is changing and wants to change the world is on the Huxiu APP[ ]]]