Just days ago, Fortune exclusively reported that OpenAI had disbanded its independent robotics team, a piece of news that immediately lit up the AI community.
Many people's first reaction was: even Altman is giving up on robots?
But on closer inspection, this is not "giving up" at all — it is "dimension elevation." The core logic here is that general-purpose large models can now directly control robots, eliminating the need for a separate "special forces" unit.
In an NVIDIA study, Astra achieved a 74.8% success rate on a simulated manipulation benchmark, while the best dedicated robotics model managed only 51.0%.
Anthropic went even further: Fable 5.2 scored 64.9% on RoboDojo's internal testing, more than double Astra's result.
The domestic embodied AI community's first reaction was: a dimensionality-reduction strike is here.
But hold on — no need to panic just yet.
We ran the numbers we could find, and the conclusion is actually the opposite — large model developers entering the field are not here to crash the party.
They are adding chairs to the embodied intelligence table, and the newcomers are bringing the largest compute resources and the deepest pockets.
The "GPT moment" for embodied AI will arrive sooner because of this.
However, these giants had barely sat down at the table before hitting a wall identical to the one from six years ago — data.
What is the same is that AI for the physical world is still being strangled by data. What is different is that this time, infrastructure companies specializing in data production have already "grown" in the market.
The playbook for this battle is completely different.
Is it a dimensionality-reduction strike or an expansion of the table?
Why are large models suddenly so good at controlling robots?
Simply put, large models are controlling robots by "writing code."
Where they are strong, they are truly very strong.
For example, understanding tasks, breaking down steps, and planning sequences — these "upper body" capabilities that embodied companies once treated as moats are now available off the shelf from general-purpose models.
But where they are weak is also obvious.
AA-Robotics published a set of figures: Astra takes an average of 30 minutes to complete a task, costing $16.
If you used that speed and cost to tighten a single screw in a factory, the factory manager would unscrew the robot first.
Why does this happen? Because cloud-based large models react too slowly.
Robot joint control frequencies range from hundreds to thousands of times per second, while a cloud-based large model takes several seconds per step.
For tasks that demand speed, precision, and hand-force feedback, top robotics teams have tested general-purpose large models and found success rates close to zero.
According to Shi Guanya of Carnegie Mellon University, slow response and network dependency are the two shortcomings of cloud-based large models, so robots still need small, specialized models on board.
So the more accurate picture is layered — large models handle the "thinking," responsible for long-horizon reasoning and planning; below that are trained action skills that the large model calls as tools; and at the bottom is millisecond-level motor control.
Large model developers have already conquered the top layer first.
To go deeper, what they lack is precisely what they do not have: motion data.
So the real change is that the list of companies building embodied foundation models has gotten longer.
Previously, the table was occupied by "native embodied players" like Physical Intelligence, Figure, and Generalist.
Now OpenAI has entered the game. Altman has stated that a self-developed humanoid robot demonstration will be ready by 2027.
Anthropic and Google's models are already on the leaderboard.
Meanwhile, domestic players are ramping up as well.
In August this year, DeepSeek contributed approximately 141 million yuan to participate in Unitree Robotics' IPO strategic placement. ByteDance had previously incorporated Seed Robotics into its multimodal interaction and world model system.
Ant Group's Lingbo, Xiaomi, and JD.com have each launched embodied models.
Globally, nearly a hundred companies are training embodied models.
What is different is that the newcomers have an essential distinction from the old players: models targeting trillions of parameters, budgets measured in tens of billions of dollars, and ready-made compute clusters.
The data battle for embodied intelligence has reached deep waters
Large language models have never lacked data in recent years.
Three decades of text accumulated on the internet can be used directly. The pre-training corpus for mainstream large models is about 100 trillion tokens, and everyone competes on compute and algorithms.
Embodied intelligence is exactly the reverse.
The physical world has no "internet." How humans grasp, twist, and with how much force — these operational processes have never been systematically recorded.
By the end of 2024, the world's largest real-machine dataset was only about 3,000 hours. In the second half of 2025, Generalist trained GEN-0 with approximately 300,000 hours. This year, leading players like Dyna and AgiBot have exceeded one million hours in training sets.
A several-hundred-fold increase in two years sounds impressive.
But compared to the 10 billion hours of language models, it is still three to four orders of magnitude short.
Does feeding data in actually work? The effect is immediate.
Dyna Robotics announced in August that when pre-training data increased from 1,000 hours to 1 million hours, the model's average score across 14 robotic tasks rose from 20% to 53%.
The industry estimates that training a trillion-parameter embodied foundation model requires at least 100 million hours. By 2030, approximately 20 such foundation models will be in training globally, with total demand reaching billions of hours.
Venture capital firm Stellaris has a similar estimate: in the next two to three years, Frontier Labs will need 100 million to 1 billion hours of first-person perspective data, corresponding to $1.5 billion to $50 billion in expenditure.
However, the trouble is that this data is not simply "filmed."
In April this year, Build AI open-sourced approximately 1 million hours of factory first-person video, free for commercial use.
Logically, the data drought should have eased. But leading manufacturers are still scrambling for data worldwide.
This is because what robots need to learn is the trajectory of "where the hand is, which direction it faces, and how it moves" in 3D space at every millisecond.
Going from 2D video to 3D trajectories requires millisecond-level alignment of multiple cameras, millimeter-level hand pose reconstruction, and conversion into joint actions for different robots.
Embodied data must be manufactured.
Moreover, model developers cannot realistically collect it themselves.
OpenAI's San Francisco lab has about 100 data collectors, producing at most 200,000 work hours per year. At the industry-standard rate of 1 to 4 hours of effective data per 8 hours of collection, annual effective output is only tens of thousands to 100,000 hours — two orders of magnitude short of tens of millions of hours.
So no matter how much money you have, you still have to buy externally.
The industry reality is that model companies' demands for data quality, granularity, and scarce scenarios will only grow higher.
The data platform opportunity is larger than Scale AI, and the barriers are higher
Every AI wave produces large data companies.
The autonomous driving wave produced Scale AI. They raised $1 billion in a Series F in 2024 at a $13.8 billion valuation, which rose to $29 billion last year after Meta invested $14.3 billion for a stake.
The large language model wave produced Surge AI and Mercor. The former never raised funding but surpassed $1 billion in revenue in 2024, while the latter's annualized revenue exceeds $2 billion.
Domestically, UniPat, an evaluation data company less than a year old, was reported to be receiving a proposed $300 million lead investment from Alibaba at a $2.5 billion valuation.
For the embodied wave, our judgment is that even larger data platforms will emerge.
There are three reasons.
First, the data barriers for embodied intelligence are higher. The raw corpora of the previous two waves were either readily available on the internet or generated by fleets themselves, and data companies sold annotation and expert hours. For embodied intelligence, even the pre-training corpus itself must be produced from scratch at a cost.
Second, embodied data is harder to obtain, scarcer, and a higher investment priority. In autonomous driving, vehicles hit the road first, and drivers generate data every day. But robots cannot enter scenarios without good models, and without entering scenarios there is no data. Only with data first can there be models, and then applications.
This turns data companies from supporting service providers into front-loaded infrastructure.
Finally, the data paradigm will change, but demand will persist. This is not a one-time deal.
Foundation model pre-training is only the first order. Each scenario that gets deployed requires additional real-machine data for post-training.
After deployment, there is still data回流 (回流), supplementary collection, and retraining. Once models open up applications, vertical small models, integrators, and solution providers will continue to buy.
Let us estimate — the scale is staggering.
At current market rates, first-person perspective data costs 100 to 200 yuan per hour. Ten million hours would be 1 to 2 billion yuan, while IDC predicts China's entire AI basic data services market will be 7.8 billion yuan in 2026.
A single order from one top customer would equal one-eighth to one-quarter of that market.
But this money is not easy to earn, because buyers are not purchasing "hours."
A cautionary tale circulates in the industry: a certain company bought 1 million hours of data in the first half of the year. When they validated it upon delivery, they were stunned — only 10% to 20% was usable.
On the surface, the listed price was only 100 yuan per hour, but factoring in the waste, the true cost came to 500 to 1,000 yuan.
This perfectly illustrates what "the cheaper, the more it traps you" means.
The potential pitfalls in this data are numerous.
For example, scenarios may be too fake — folding clothes at a desk in a data collection factory cubicle. Or the footage may be too poor — dropped frames, overexposure, hands constantly leaving the frame.
Or processing may be too crude — multiple cameras not aligned to the millisecond, trajectory errors of several centimeters.
Clearly, embodied intelligence data is moving from the sale and resale of low-quality data to a stage where customers genuinely demand precision and scenarios.
In essence, this will also eliminate companies without technical capabilities.
And because data has become extremely scarce, top buyers mainly look at three things.
First, is the category coverage complete? Industry consensus is that million-hour-level human first-person perspective data forms the base, tens of thousands of hours of real-machine data handle action adaptation, and a small amount of on-site data enables deployment. A supplier that only does one type can only participate in one segment.
Second, can they provide stable delivery capability? Customer demand-side data standards are relatively unified. Most companies fail to meet quality standards after sample submission, or their delivery capacity cannot keep up, so they cannot handle large orders.
In this situation, data companies that can stably supply and deliver are more likely to build stickiness. Whoever can prove they can handle large orders without dropping the ball will be firmly bound to the big players, and no one can poach them away.
Third, can they close the loop? When a model team discovers a capability gap, the supplier must be able to collect supplementary data for specified tasks and prove that the supplementation actually works.
Everyone claims to serve top customers — what is the difference?
More and more companies in China are working on embodied data.
But on closer inspection, the track is far less crowded than it appears.
Why? Because the barriers are actually very high.
Doing embodied data requires three capabilities simultaneously: hardware manufacturing (devices must be self-developed, mass-producible, and consistent), algorithm processing (calculating 3D trajectories from raw video with sufficient precision), and scenario operations (being able to enter real environments like factories, supermarkets, and hospitals, and building a collection ecosystem network).
These three capabilities actually belong to different DNA.
This is why more companies appear to be doing data, yet data has become extremely scarce.
Because companies that can truly deliver end-to-end and get top customers to sign off on acceptance are actually few and far between.
From public information, Mifeng Technology is currently one of the few companies that has achieved scale in all three areas simultaneously.
Mifeng is regarded as the "physical AI data infrastructure platform that best understands model and scenario needs," providing one-stop real-machine, simulation, and body-free training data.
Its approach differs from most data companies.
The common industry practice is to first obtain a batch of devices or scenario resources, collect a batch of data, then go find customers.
Mifeng does the opposite: it starts from the needs of leading model customers at home and abroad — what precision, what scenario mix, what format — then works backward to design hardware and pipelines.
This reversal creates a widening gap.
The core issue on the collection side is whether consistency can be maintained after large-scale deployment.
In data collection, one device running well is nothing special. The hard part is 20,000 devices spread across dozens of cities and thousands of collectors, with every unit's precision, timing, and format matching up.
The common industry situation is: send out 100 devices, 20 or 30 will have problems, and a large amount of work hours is consumed fixing devices and adjusting parameters.
Mifeng's approach is to mass-produce collection devices to consumer-grade hardware standards — millimeter-level recording precision, wireless, lightweight, quick battery swap, quick-release components.
These seemingly unremarkable designs solve the most fatal industrial-grade problems: can devices run stably in distributed scenarios, can they be replaced immediately if broken, and can each unit's data align with others.
With this playbook, they launched in April, reached mass production in June, and by the end of August had produced 20,000 units.
This speed is unmatched in the industry.
Once devices are deployed widely enough, capacity follows.
One million hours of body-free data was completed in three months, with daily output exceeding 10,000 hours.
Scale AI's disclosed figures are just over 1,000 hours collected per day, with 150,000 hours delivered in all of 2025 — a full order of magnitude behind Mifeng.
Moreover, the core issue in the data pipeline is not whether processing is possible, but efficiency.
As mentioned earlier, across the industry, only 10% to 20% of 1 million collected hours is often usable.
Mifeng's efficiency rate exceeds 90%.
They achieve this by obsessing over details and widening the gap at every step.
On the device side, hand spatial positioning precision reaches millimeter level, and multi-sensor signal alignment reaches sub-millisecond level.
The quality inspection process has three layers: real-time interception at the collection end catches 60% of obvious problems; the cloud uses vision-language models to automatically make semantic judgments; and finally, manual spot-checking keeps deviation within 1%.
Among these, trajectory reconstruction is the step that most tests algorithmic prowess in the pipeline.
Mifeng's self-developed HandPose algorithm achieves sub-centimeter precision, and in real scenarios such as occlusion, interaction, and rapid movement, reconstruction error is significantly lower than the industry's next-best solution.
This precision is currently in the industry's first tier.
Furthermore, the entire pipeline processes over 10,000 hours daily, and the conversion from raw data to trainable assets runs automatically, not manually one by one.
This step's efficiency directly determines how much usable data can be delivered to customers in the same amount of time.
The closed-loop capability is the most easily overlooked yet highest-barrier layer.
Most suppliers' service endpoint is "quality inspection passed, delivery signed." But after data is delivered, suppliers often do not know how model training performs, which data is useful, which is noise, or where the gaps are — and they have no ability to verify it.
To this end, Mifeng developed an end-to-end model and does its own real-machine deployment to test data effectiveness, study mix ratios, and develop supplementary collection strategies.
In tests across some scenarios, each round of data closure improved task success rates by 5 to 10 percentage points.
This way, when a customer's model team discovers a capability gap, Mifeng can collect targeted supplementary data for specified tasks and prove that the supplement actually works.
This is no longer a simple data transaction but a relationship of joint iteration.
From public information, Mifeng currently works with dozens of customers, covering large model platforms, embodied and world model companies, and robot body manufacturers.
Tencent Robotics X and Ant Lingbo have publicly used its data, and AgiBot's GO3 foundation model training data includes a share from Mifeng.
Overseas, it has become a core supplier to multiple leading tech companies.
Of course, to leap from one million hours to 100 million hours in the future, linear replication of data collection factories will not be enough.
Mifeng is betting on a crowdsourcing route — Mifeng Pai, launched in September, allows ordinary people to wear devices and collect data while working. It currently has over 200,000 registered users, covers more than 50 cities, and over 50 scenario partners including hotels, supermarkets, hospitals, and factories have joined its data alliance.
Next, the most worthwhile thing to watch is whether crowdsourcing can scale up another order of magnitude while maintaining an efficiency rate above 90%.
Large models entering the field is good news for the embodied industry
Back to the opening question: is large models entering the field a "dimensionality-reduction strike" against embodied AI?
But analysis shows the conclusion is actually the opposite: this is good news for the embodied industry.
Going forward, the table is bigger, there is more money, and the entire industry's progress is accelerated.
Originally, the industry predicted that embodied intelligence's GPT-3 moment would arrive around 2030.
But the speed of leaderboard updates in recent months proves this moment will come earlier.
And each additional player with a trillion-parameter large model that enters will widen the market's data gap further.
Because the faster models run, the more data they need.
OpenAI disbanded its robotics team twice. The first time was because there was no data. This time, the model is ready, but what is most lacking is still data.
The difference is that this time, someone has already turned data production into a serious business.
Comments