数百万小时、500亿参数，自然意志的Scaling信仰
Millions of hours, 500 billion parameters, the Scaling belief of natural will
Summary
Millions of hours of human video enable robots to learn zero‑shot generalization. By Guo Haiwei. Natura-1 is an embodied model pre‑trained on a million‑hour‑scale Ego‑Centric (note: first‑person video) dataset. In this …
Millions of hours of human video enable robots to learn zero-shot generalization.
By丨Guo Haiwei
Natura-1 is an embodied model pre‑trained on a million‑hour‑scale ego‑centric (note: first‑person video) dataset. Until now, there have been no more than three publicly known million‑hour‑scale ego data pre‑training embodied models worldwide. As a company that has existed for only nine months, Natural Will is probably the fastest team in the history of the embodied‑AI field to achieve this milestone.
Centered on Natura-1, the team works like an LLM lab. In interviews, several researchers told LatePost that they set scaling as their research goal, redesigning a research architecture for embodied cognition that both draws from and diverges from mainstream technical routes, rebuilding infra capabilities including hardware gloves and data pipelines, and even writing an independent training‑inference framework from scratch. “Our goal is to make everything scalable,” Natural Will said.
Just as professional boxers are weighed into weight classes before a match, in the AI domain data often implicitly carries many benchmark‑type signals. For example, larger data scales usually imply a higher ceiling for intelligence.
In Natural Will’s technical report “The Will to Act,” Natura-1 outperforms all publicly measurable benchmark models across three dimensions and twelve evaluation items.
In the technical report released by the company, Natura-1 achieved a 53.4% success rate on basic tasks in a self‑built real‑world evaluation; after adding memory, error‑correction, and other capabilities, the success rate increased to 59.1%. In the same test suite, π0.5 and GPT-6 Astra did not obtain any scoreable results with the dexterous hand configuration; when switched to a gripper, they achieved success rates of 13.5% and 16.7%, respectively.
As shown in the figure, the entire benchmark is divided into three core components: dexterous manipulation, instruction following, and long‑range task execution. Among the three scores, dexterous manipulation shows the largest lead; even in the long‑range operation and instruction‑following domains, where language models naturally excel, Natura-1 still demonstrates a notable advantage.
Because the industry currently lacks a credible third‑party benchmark, the top models available for the Natura team’s evaluation are limited; models such as π0.7, Helix‑02, Gen‑1.5, Dyna‑2, etc., cannot be accessed directly in the open market. The π0.5 we are evaluating now is last year’s open‑source model, and the recently hot‑talked GPT‑6 Astra is not a dedicated machine model. This set of results is better suited to illustrate Natura‑1’s strengths rather than being interpreted as an industry ranking.
Even harder to verify than benchmark scores is “zero‑shot”. “Zero‑shot generalization” means the model has not been specifically trained on the test task, nor has it undergone post‑training with data that combines the same objects, scenes, and instruction sets; it must complete the task directly from natural‑language instructions during testing—just as you can quickly pick up a video game you’ve never played, or a badminton enthusiast can instantly learn pickleball. Practitioners, including Wang Xingxing, publicly treat zero‑shot capability as an indicator of whether the embodied “real GPT moment” has arrived, believing it reflects the model’s true understanding of the world.
Take Natura‑1 as an example; the team wants to emphasize that zero‑shot generalization is a key characteristic of this model release.
In the technical report, the team released demos of six dexterous operations, covering wrapping dumplings, packing parcels, making tea, preparing yogurt bowls, applying a phone screen protector, and writing calligraphy. The six demos correspond to different model capability requirements such as flexible objects, delicate manipulation, tool coordination, long-range planning, fine assembly, and continuous force regulation.
The researchers at Natural Will told us they ensure each demo must be a continuous, real‑time, one‑shot coherent video. But they also realize this is not enough to fully allay market concerns, so the team decided to also produce a 29‑minute one‑shot zero‑shot generalization video, hoping to maximize market persuasiveness within the team’s capability range. At the same time, the team said that in the future they will let more and more people offline “play” Natura‑1 in various forms, not just watch a demo.
The third path beyond VLA and WAM
But natural will tells us that “benchmarks” and demos are actually the least important things about Natura-1. The R&D teammates have been trying to convey that Natura-1 is only the first step of a model with much larger scaling potential, so it is still an early‑stage model in its own development, and a 59% success rate still leaves a lot of room for improvement.
Natural Will founder Ding Ning told us that even before they ran Natura-1 internally, he wasn’t sure his team could actually pull it off—despite having already done everything they could: “All ablations (note: ablation experiments used to verify whether the model architecture holds) were performed,” every model was refined through evaluation, and they built the infrastructure and data pipelines from scratch. Yet even with data and experimental support, he still “couldn’t help but wonder whether it could really be done.” It wasn’t until the very last moment, when the model finally trained, that the team confirmed:
“Oh, such a model does indeed exist.”
This is a uniquely architected model; the colleagues responsible for the model architecture at Natural Will told us that it both draws from and differs from today’s two mainstream embodied brain tech routes: VLA (vision‑language‑action model) and WAM (world‑action model). Natural Will hopes Natura‑1 can become an alternative path for scaling intelligence beyond those two routes. The architecture lead said that the differences in this architecture can be understood from at least two perspectives.
First, Natural Will states that Natura‑1 is a model without “master‑servant” hierarchy.
The so‑called master‑servant relationship refers to the status differences among internal modules: WAM’s core is video generation, VLA’s core is the language model, with the primary modality handling the most critical reasoning while other modalities are more subordinate to the “core.” NVIDIA scientist Jim Fan once joked that the essence of VLA is LVA (language‑vision‑action). Because there is a core, scaling becomes a data‑growth game that focuses on the “master model.”span>
The Natural Will team tells us that, in their view, Natura-1 is the third path. Here, the status of each modality is equal, with no need for one to obey another. The advantage of Natura-1 is that, because there is no master model, “all modalities can be organized in the sequence in the most scaling way,” says the architect of the Natural Will model.
When we were building this model, “there was nothing to refer to,” so at the beginning the team also “didn’t know what the model would look like.”
Secondly, according to Natural Will, this is a model that is closer to “end‑to‑end.”
Although past explorers of VLA and WAM also discussed end‑to‑end approaches, they often chose a “brain + cerebellum” hierarchical architecture in practice, separating task planning from motion control. Because all modalities in Natura-1 are equal, it can simultaneously model continuous video and actions, intent understanding, memory, cerebellar motor control, and even some agentic capabilities. This makes Natura-1 closer to the “end‑to‑end” concept than previous VLA and WAM, allowing it to handle motor control within a single model.
For example, Natural Will incorporated dexterous hand actions into Natura-1, rather than, as some current industry players do, building a separate small brain for a dexterous hand.
Because “action is the most important thing,” a Natural Will researcher told us later, “if you change the action, it’s catastrophic—basically equivalent to wasted training.”
In Natural Will’s view, dexterous hand manipulation can help an embodied brain bridge human experience with purposeful actions, and the latter is the core of the general embodied intelligence they are building. Therefore, Natura-1 itself is a foundational model for dexterous robotic hands.
“Hand actions are inherently diverse; even we find it hard to describe the types of dexterous hand motions, and they are purely experiential, beyond the scope of language,” says the head of the Natural Will model architecture. Describing fine manipulation solely with language loses information about finger contact, force, and continuous motion—“this intelligence lies outside its descriptive range.”
Natura-1 is the first general action model officially released by Natural Will, and it is the first model explicitly targeting dexterous hands.
The team once considered naming this “Model” — because “there is only one model.” Since Natura-1 was defined by the team as a model beyond VLA and WAM, Natural Will has been contemplating a new name. However, the team ultimately abandoned that unconventional name and chose to define it as “a universal foundational model for the physical world.”
Richard Sutton, known as the “father of reinforcement learning,” presented a famous conclusion in his well‑known essay “The Bitter Lesson”: stuffing human domain knowledge into intelligence is always effective in the short term, but becomes a bottleneck in the long run. Only sufficient generality can yield massive power. Ilya Sutskever attributes this to an aesthetic and taste principle, saying good AI should be “beauty, simplicity, elegance.” History of technology also shows that the simpler the architecture, the better it withstands massive data scaling.
Because Natural Will has not disclosed its architecture details, we cannot find a suitable way to independently verify whether they truly adopted a design that aligns with “technical aesthetics” and better suits scaling. However, according to the Natura-1 technical report, the team emphasized that the model’s scaling potential has not yet been fully realized in this version.
The figure below shows the scaling validation data curve for Natura-1:
The chart contains two lines: the dashed line is the power‑law fit representing the predicted values when the model scales normally, and the solid line is the actual measured values. The two lines align closely across all four validation metrics, indicating that the model’s scaling process is proceeding smoothly; even at the million‑hour data point, no substantive bottleneck has been encountered so far.
The lead architect of the Natural Will model told us that Natura-1 currently has 50 billion parameters. However, he believes that even this version has not hit a scaling bottleneck; continuing on the current trajectory, the next “conservative” phase is expected to deliver another exponential iteration. If everything goes well, Ding Ning said the model could be released next year.
Make Everything Scalable
Several researchers from Natural Will told us that belief in embodied scaling is almost the starting point of this company—because most of them transitioned from various research fields to work on embodied brains at Natural Will. Some were doing AI for Science during their undergraduate years, others previously worked on schema development in the LLM domain. The core reason they are willing to pursue embodiment is that they have seen ego‑centric data collection rapidly addressing industry data gaps.
The founding team of Natural Will said they realized around January that “the data became scalable,” meaning “it can be captured in any scenario and any action worldwide.” And “when this becomes scalable, it’s hard to believe that a general model does not exist.”
In the past year, the global amount of data has expanded from hundreds of thousands of hours to at least several million hours. Natural Will estimates that it will reach tens of millions next year, and future models may consume hundreds of millions of hours of data, requiring tens of thousands of GPUs for training. This is the opportunity for the current generation of entrepreneurs at Natural Will.
Data capability is seen by Natural Will as an even more important competitive factor.
Natural Will has built its own data pipeline. Today you see that Natura-1 is a model trained on a scale of millions of hours of data, but behind it Natural Will has processed data on the order of several million hours in total. Moreover, as models iterate and data explodes, the scale of data processing will continue to rise.
“LLMs generally can’t store that much data,” and “ingesting massive video data is an inefficient activity,” the Natural Will team said. “In the past, only those working on autonomous driving would store and handle such large volumes of data.”
“Data is a headache,” the model architecture lead said. “If we could buy good vendor data we would, but many external data providers basically don’t meet our requirements.” Although there are many data vendors today and the total hourly volume of data collection on the market keeps rising, truly qualified data are scarce. This has caused the team to spend a great deal of time and effort on it, “but objectively, the data that this industry can produce can’t keep up with our demand or the speed at which we consume it.” One of the bottlenecks is the vendors’ data processing capability. Raw data cannot be fed directly into the model; it must go through multiple processing steps.
“But the resources consumed by data processing are absurd,” a researcher said. Natural Will built a filtering pipeline for Natura‑1 consisting of more than 20 operators; once data arrive, they are first filtered by operators and then labeled, which requires different compute capabilities. “Extracting hand pose needs a lot of GPU, decoding and format conversion need a lot of CPU, and linking the operations together demands heavy communication handling.”
“Some big companies don’t have as much dirty, tedious work as we do; they outsource a lot of it,” Natural Will told Latepoint. “But sometimes the success or failure of a model hinges on that dirty, tedious work.” And data, as this kind of ‘dirty, tedious work,’ is what Natural Will is building as a competitive edge for a brain‑in‑a‑body company.
The hiring principle at Ziran Yizhi is: smart, good at math, have conviction and truly be willing to dig in to write code, clean data, and do the dirty, hard work. In contrast, having published papers or having embodied‑brain development experience is not the most important here, and in some cases it can even be a “negative point.”
“Everyone comes here not to write a world‑changing paper, but to build a world‑changing model,” Ding Ning said.
Ziran Yizhi told Wan Dian that they are probably one of the few teams in the embodied industry that built training and inference frameworks from scratch. The model team deeply customized high‑performance CUDA operators for different series of GPUs, achieving co‑optimization of the framework and low‑level operators. From roughly DAY 1, Ziran Yizhi spent about a month writing an initial training framework, then spent several months iterating on it.
The training framework is considered a crucial part of scaling; it determines computational efficiency, and once problems arise they are hard to recover from, so many teams opt for mature open‑source community frameworks. But precisely because of that, “if you want to seriously get scaling right,” you also need to redesign it.
“Open‑source infra must be compatible with different algorithms and modules, so it inevitably becomes bloated,” the model‑framework lead told Wan Dian. Even for LLMs, top teams won’t use open source.
Take a certain high‑profile domestic AI large‑model company as an example: the model architecture may change between each generation, so the infra team will build it themselves and won’t use community tools. “The nature of infra is that you can’t rely on someone else’s ecosystem for patchwork; it simply can’t be accelerated to the extreme,” the lead told Wan Dian.
More importantly, “LLM training frameworks are mature, but none of them are designed and optimized for embodied scaling.” Coupled with the team’s innovative model architecture, the gap between the infra needed for LLMs and the algorithms desired by Natural Will is too large, “so you definitely have to write your own infra from scratch.”
In the Natura‑1 technical report, Natural Will wrote: “Infrastructure makes experience usable; data turns it into signal; the model turns signal into intelligence. None of these can scale independently—yet when they scale together, they amplify each other.”
This may seem like common sense in the LLM era, but turning common sense into practice is never simple.
Today, no one can definitively say how large an embodied brain should be. The industry consensus is that it should be huge—and larger than today’s LLMs. Scaling is reshaping the development logic of embodied brains, with more and more companies at home and abroad pursuing bigger models and higher intelligence. This spawns different data strategies, varied model technologies, differentiated infra infrastructure, attracts talent from more diverse backgrounds, and ultimately gives rise to organizations with a different character than before.
A researcher told Wan Dian that, in his view, Natural Will is a lab that doesn’t publish papers, with a flat, pragmatic, low‑ego structure, where sometimes researchers and leaders treat each other as equals. Someone who has closely interacted with Natural Will told Wan Dian that the people in the company “all look very similar”—it feels like a “delicately youthful cohort.” The visual presentation of the newly released website and technical report was drawn by Ding Ning and the R&D staff themselves.
Model illustration designed by the technical team
It’s like “The Will To Act” is more like a footnote of a group of young developers, and within a larger vision, the million‑hour embodied model is just a starting point.
Cover image source: Natural Will
– FIN –
Get MirrorSpeed
Get MirrorSpeed