Natura-1发布：从百万小时人类经验中学会零样本泛化与灵巧操作

Natura-1 Release: Learning Zero-Shot Generalization and Dexterous Manipulation from Millions of Hours of Human Experience

Wrapping wontons is simple for most people: one hand holds the skin, the other hand picks up the filling, fold, pinch the edges, and seal. Few realize that throughout this process the palm is constantly supporting a sheet of dough that can deform at any moment, the fingers continuously feel the rebound of the filling and the dough, and the two hands must hand off at the right moment.

These everyday motions, designed for human hands, are extremely challenging for robots. But now those challenges are being overcome. In a 1 minute 12 second, single‑take, real‑time video, a pair of robotic dexterous hands successfully completed the entire process.

Driving these hands is the newly released Natura-1 from Beijing embodied‑intelligence company Nature Will. It is a 50‑billion‑parameter mixture‑of‑experts (MoE) model, pretrained on 1 million hours of first‑person human experience data, taking language and vision as input, and driving a 20‑degree‑of‑freedom dexterous hand within a 100‑millisecond control loop.

Bao Wonton answered “whether it can be done well.” Natura-1 wants to answer an even harder question:If we switch to a task it has never practiced, can it still start right away?

Dexterity and generality are hard to achieve together

First, look at dexterous manipulation. The human hand has more than twenty degrees of freedom, and everyday objects are often soft, deformable, and elastic; each contact changes the object’s state, so robots must adjust on the fly; but the difficulty is obvious: complexity grows exponentially with the number of degrees of freedom. Parallel grippers can greatly simplify control, at the cost that many supporting, twisting, and bimanual coordination actions become hard to accomplish.

Degrees of freedom and combination counts of parallel grippers and dexterous hands, generated by AI

From the Nature Will team’s perspective, past work has largely been split: either a gripper‑centric general strategy or a dexterous operating system built for a single task.

On the general side, there have already been important advances such as π0.7, but achieving this kind of generality on dexterous hands is first blocked by data. Generalization relies on large‑scale, diverse training data, and real‑world data is already expensive; collecting tele‑operated data for dexterous hands is both slow and difficult.

The Nature Will team also highlighted a more fundamental loop: robots can only collect data in places where they already know how to work, and they can only operate in places where data already exists.To break this loop, data must come from outside the robot.

As a result, more and more teams are turning their attention to an existing data source:humans themselves.

In February this year, Nvidia and other organizations released EgoScale, which pre‑trains a VLA model on 20,854 hours of first‑person human operation video. They observed a logarithmic linear relationship between data scale and validation loss; after deploying it to a 22‑degree‑of‑freedom dexterous hand, the average success rate was 54% higher than the baseline without human data pre‑training.

The scaling behavior of human pre‑training: the functional relationship between optimal validation loss at convergence and the scale of human data, revealing an almost perfect logarithmic linear scaling law, arXiv:2602.16710

Six months later, this scale was increased by about 50 times. On August 10, U.S. robotics company Dyna Robotics released the world‑action model DYNA‑2, which was pre‑trained on more than 1 million hours of first‑person human video, and the pre‑training phase contained no robot data at all; according to Dyna’s report, the data was scaled stepwise from 1,000 hours to 1 million hours, and the model’s average normalized score on 14 robot tasks rose from 20% to 53%.

Scaling law of the Dyna‑2 world‑action model, trained on a ladder of human video data ranging from 1,000 hours to 1 million hours. https://www.dyna.co/dyna-2

However, both of these works share the same bottleneck:Mapping human experience onto specific tasks still requires additional adaptation.EgoScale relies on a small amount of human‑machine alignment data for intermediate training, and DYNA‑2 uses up to 10 hours of robot data for fine‑tuning on each evaluation task.

Human experience has been shown to be scalable for feeding robots, and the next question is:Can these experiences transfer directly to new tasks on a high‑DOF dexterous hand without per‑task adaptation?

Natura-1

One model, two capabilities

Nature Will was founded in January 2026 by Ding Ning, an assistant professor in the Department of Electronic Engineering at Tsinghua University. It is a research-oriented company centered on foundational models of the physical world. Its self‑set goal isPhysical AGI:to train a universal foundational model that can perform perception, prediction, reasoning, and action loops in the real world.

The Nature Will team told Machine Heart that this stems from two judgments:

1. A universal foundational model of the physical world may be feasible;

Human first‑person perspective data now meets the conditions for large‑scale collection.

The methodology behind these two judgments continues from LLM research, namely models, data, and infrastructure need to develop together. The Nature Will team told Machine Heart: “We followed the approach in large language model research of scaling model size and data size in tandem. Early on we used smaller models and datasets to quickly validate the architecture, and after confirming the direction we gradually expanded, rather than deciding from the start that we must build a 50B model.”

Natura-1 is the first general‑purpose foundational model for the physical world from Natural Will.Its training consists of two steps: first pre‑training on 1 million hours of human experience data, then performing a single round of efficient fine‑tuning with limited robot data to adapt the world knowledge learned during pre‑training to dexterous manipulation.

Unlike many systems that split “understanding instructions,” “predicting the world,” and “generating actions” into separate modules,Natura-1 is a single MoE Transformer jointly trained on language, images, video, and actions.It targets two capabilities:

• Dexterous manipulation, leveraging a dexterous hand to achieve multi‑finger coordination, object pose adjustment, and complex contact;

• Zero‑shot instruction following, handling unseen scenes and tasks by directly executing natural‑language commands, including spatial constraints, multi‑object relationships, and negation conditions.

What can Natura-1 do?

The Nature Will team evaluated Natura-1 from three perspectives:

• Long‑term task: Assess how precise and durable a skill can become after training.

• Zero‑shot test: Evaluate whether the model can tackle tasks it has never practiced.

• The final part records what can be called emergent ability, i.e., whether the model can devise its own solutions when a task does not go smoothly.

Dexterity: A six‑segment, single‑take long‑term task

In addition to the dumpling-wrapping demo, the Nature Will team released five more demos: making tea, preparing a yogurt bowl, packing a parcel, applying a phone screen protector, and practicing calligraphy,the longest one exceeds 3 minutes, all captured at real speed, in a single continuous shot, with no human intervention. Each segment contains multiple sub‑tasks, and the robot must continuously handle liquids, granules, soft tools, and flexible components within the same workflow.

Take the 3‑minute phone‑screen‑protector demo as an example: Natura‑1 first wipes the screen, then places the alignment frame pre‑loaded with the protective film onto the phone, aligning both hands and pressing firmly; next, one hand holds the phone and frame against the fixture while the other hand smoothly pulls off the release liner, and finally it pushes out bubbles along the edges with its fingertips.

When packing a parcel, it switches its grasping method between raffia, plush toys, and notebooks, presses down on the spring‑back box lid with one hand, uses the other hand to operate the tape dispenser to seal the box, and finally applies the label and scans the barcode.

These actions are performed while simultaneously changing the object’s shape and contact conditions, and the model must continuously adjust based on real‑time feedback rather than following a fixed trajectory. More demos can be found in the official Nature Will article “The Will to Act | Natura‑1 Physical Generalist Foundation Model”.

It should be emphasized that these six demonstrations show the operational depth after task fine‑tuning. Whether the capability transfers to unseen tasks depends on zero‑shot testing.

Zero‑shot: Unseen tasks

What counts as zero-shot? The Nature Will team told Machine Heart: “When we say Zero-shot, we mean that after the model completes pre‑training and fine‑tuning on real devices, it no longer collects additional data or undergoes extra training for the test task, but directly receives instructions and executes them.”

The team built a benchmark covering 12 capability dimensions, where the test tasks and their scene‑action combinations never appear in the post‑training data; next to each test video, the most similar demonstration from the training set is placed side by side for external comparison. For example, a task requires putting an apple into a bowl, placing a bottle on a tray, and then carrying the tray away, while the training set only contains isolated actions like “pick up fruit” and “carry a loaded tray.” Natura‑1 combined them, and it didn’t bump into the other tray while carrying the tray.

On this benchmark, π0.5 and OpenAI’s GPT-6 Astra (acting as a control policy by writing code) both failed to score on the dexterous hand platform; after switching to the gripper platform, they scored 13.5 and 16.7 respectively, while Natura-1 achieved 53.4 points (Base) and 59.1 points (Agentic).

Moreover, Natura-1 also shows a clear advantage in execution time.

The team also tested the open commands that researchers thought of on the fly, with many deliberately tricky tasks. For example, when placing a toothbrush case and a water bottle into a basket, the commands required both hands simultaneously, then left then right, then right then left—Natura‑1 executed each as instructed.

Regarding “unseen,” the team clarified: “The pre‑training data is very broad and may already contain related objects and manipulation experience, so ‘unseen’ here mainly refers to post‑deployment fine‑tuning.” Their example is pressing the space bar on a laptop: the robot’s fine‑tuning data doesn’t include this task, but similar actions are likely to appear in human videos.

This points out the meaning of zero-shot on this path:Can the experience from human videos be directly applied to a real robot when there is no corresponding robot data?In the team’s words: “We care not only about whether swapping an object still allows grasping, but also about whether, when faced with tasks that haven’t been specifically trained for, we can organize existing capabilities to complete the task.”

Unexpected emergence

The team also observed some behaviors that were not part of the test plan.

In a deliberately arranged scenario, a basket containing yarn is placed on a high platform out of reach, and a long pole passes through the handle. The instruction is only a single sentence “Try to grasp the yarn (Try to grasp the yarn)”. Natura-1 grabs the pole and pulls the basket closer, then steadies the pole with one hand while extracting the yarn with the other. In another instance, it failed to pinch a playing card from the edge, so it instead pushed the card to the edge of the box, leaving part of it hanging, and then used both hands to lift it. According to the team,these intermediate steps are not present in the robot demonstration data.

Team recollection: when the data scale was small, the model didn’t catch it, but the hand kept moving according to the original plan; after scaling up, it began to adjust in time and try again. The Nature Will team told Machine Heart: “We didn’t deliberately collect such failure‑recovery demonstrations during real‑device training, but the model still exhibited recovery ability. Our judgment is that large‑scale human pre‑training data contain abundant adjustment and retry experiences, and the model learned how to cope with changes during operation from them.”

The team also admits: “This is a phenomenon we observed when scaling up training; the exact source and mechanism of the capability still require further study.”

Where does the ability come from?

From the Nature Will team’s perspective, long-range dexterous manipulation and zero-shot generalization “won’t happen by accident.” Human experience is abundant, but it is only raw material; turning it into robot actions requires data, models, and infrastructure to be designed together:

• Data provides language and motion supervision for each video segment

• Models learn as much as possible from each trajectory

• Infrastructure must both support large‑scale training and perform inference within 100 ms

Data

The training data for Natura-1 isthe filtered 1 million hours of first‑person human experience, containing about 61.7 million clips, 318,000 tasks, and 14,000 scenes, primarily collected in a controlled manner and supplemented with first‑person videos from the internet.

The team uses a structured task taxonomy and agent workflow analysis to cover gaps, shifting the collection priority toward scarce areas.Collection priority is redirected to under‑represented regions. Raw recordings must undergo cleaning, deduplication, motion reconstruction, and hierarchical language annotation before becoming “video‑language‑action” trajectories; in internal benchmarks, our proprietary motion reconstruction pipeline achieves joint position errors 41% lower than existing open‑source methods.

Natura-1 data system diagram

Choosing human data and choosing a dexterous hand is the same decision. The Nature Will team tells Machine Heart: “Many tools and operation methods in home and industrial settings are designed around the human hand, and we want robots to adapt to these tools rather than redesigning tools and processes for each new environment.” On the other hand, first‑person human data is easier to collect at scale, and “using a dexterous hand preserves richer hand‑motion information, reducing the loss that occurs when human hand actions are compressed into simple gripper opening/closing signals.”

Infrastructure

“Make everything scalable” is the technical philosophy of the Nature Will team:

• Data side, the data lake‑warehouse manages dozens of PB of data, and the Kubernetes‑based parallel pipeline can produce 350,000 hours per week, with throughput scaling almost linearly with compute power. The team explains that the design goal of this pipeline is: “When the data volume increases tenfold, by proportionally increasing resources about tenfold, it can still complete processing in roughly the original time without rebuilding the entire system.”

• Training side, the in‑house framework achieves over 50% model compute utilization on hundreds of GPUs, and the custom attention operator’s throughput is more than six times that of the tuned FlexAttention.

• Inference side, the service framework rewritten in Rust compresses end‑to‑end latency to 100 ms, meeting real‑time control on actual devices.

The Nature Will team says: “The thinking behind this aligns with the infrastructure building for large language models: Data and model scale will keep growing, and the system must be prepared for this growth in advance.”

Model

Natura-1 arranges the entire task segment into a sequence of language, image, video, and action tokens, modeling them autoregressively. Its backbone is a standard MoE Transformer.

Its core idea istreat video and action as two projections of the same physical state: action labels are scarce, while videos are abundant; the model uses past videos as memory to predict future videos, and trains video and action simultaneously with flow matching, obtaining more supervision from each trajectory. To prevent the model from only looking at the visual and ignoring commands, pretraining also leverages fine-grained descriptions to let language “carry weight”, so that different intents correspond to different prediction results.

The scaling curve shows that training experience increased from about 44,000 hours to 1,000,000 hours, and the model’s motion prediction error (MSE) on human hold‑out data decreased by 47.7%; the same set of weights on real‑robot hold‑out data saw an error reduction of 67.7%. These two figures are prediction errors under Teacher Forcing, not task success rates, but they demonstrate that the growth of human experience does indeed transfer to the robot..

Natura-1 Scaling Curve

Conclusion

Natura-1 has already provided early evidence that dexterity and generality can coexist. As for the next stage, the Nature Will team told Machine Heart: “Our core goal is to continue improving the model’s generalization across different tasks and scenarios, and make it more reliably handle complex real‑world tasks.”

The team says the goal is to move from “being able to complete partial steps” to “stably and efficiently completing the entire task.” To that end, they plan to incorporate tactile information into the data and model, pre‑train on a larger‑scale dataset, reduce the adaptation cost when transferring to robots, and enable the model to run on cheaper hardware. The team expects this to be a long‑term investment.

Placed in a broader industry context, the trajectory that Natura-1 follows is shifting from a bet by a few teams to a growing consensus. In June this year, a study called HumanScale compared two types of pre‑training data at the same data scale and found that first‑person human video can outperform tele‑operated real‑world data, with the most pronounced advantage in out‑of‑distribution generalization. In August, Dyna Robotics, which released DYNA‑2, announced that the next step is to scale training up to 10 million hours.

When a million hours are no longer scarce, the focus of competition shifts: can the same human video be more accurately reconstructed into actions and annotated with language; how much adaptation is needed for the model to be deployed on robots; and what kind of ontology should be used to incorporate these experiences.Natura-1 is betting on the most challenging combination, using a highly dexterous, high‑DOF hand to embody human experience and letting a single model tackle tasks it has never practiced.How far this path can go depends both on whether the dexterous hand’s stability and cost can keep up, and on whether the scaling curve of human experience can continue to extend.

In a world shaped by human hands and meant for human hands, how far can robots go by learning from human hands? Nature Will has delivered the first testable answer.

Reference links

Official blog: https://www.naturewill.com/blog/natura-1

Official website: https://www.naturewill.com

Official tweet: https://x.com/naturewill/status/2109858475333570565

© THE END

For reprints, please contact this public account for permission

Submit contributions or request coverage: liyazhou@jiqizhixin.com