Nature Will Releases Natura-1: A 50B MoE Model Achieving Zero-Shot Dexterous Manipulation via Million-Hour Human Data
Beijing-based embodied AI firm Nature Will released Natura-1, a 50B parameter Mixture-of-Experts (MoE) model pre-trained on 1 million hours of first-person human experience data. Taking language and vision as inputs, Natura-1 drives a 20-DoF dexterous hand within a 100ms control loop, demonstrating complex soft-object manipulation like making wontons and validating zero-shot generalization to unseen tasks.
Full Report
50B Parameters, One Million Hours, 100 Milliseconds, Zero-Shot Generalization: Natura-1 Has Arrived
For most people, wrapping wontons is simple: hold the wrapper with one hand, scoop the filling with the other, fold it in half, pinch the edges, and seal it. Few realize that throughout this process, the palm continuously supports a dough skin prone to deformation, fingers constantly sense the rebound of the filling and wrapper, and both hands must coordinate seamlessly at precise moments.
These everyday actions, designed for human hands, are extremely difficult for robots. But now, these challenges are being overcome. In a single-take, real-time video lasting 1 minute and 12 seconds, a pair of robotic dexterous hands successfully completed this entire process.
Driving these hands is Natura-1, recently released by Beijing-based embodied AI company Nature Will. This is a 50-billion-parameter Mixture-of-Experts (MoE) model, pre-trained on one million hours of first-person perspective human experience data. Taking language and vision as inputs, it drives a 20-degree-of-freedom (DoF) dexterous hand within a 100-millisecond control loop.
Wrapping wontons answers the question of "can it be done well." What Natura-1 aims to answer is a harder question: if presented with a task it has never practiced, can it still perform directly?
The Trade-off Between Dexterity and Generality
Let's look at dexterous manipulation first. The human hand has over twenty degrees of freedom, and everyday objects are often soft, deformable, and elastic. Every contact changes the object's state, requiring the robot to adjust while acting. However, the difficulty is evident: complexity grows exponentially with the number of degrees of freedom. Parallel grippers significantly simplify control but at the cost of struggling with many supporting, twisting, and bimanual coordination tasks.
Degrees of freedom and combinatorial complexity of parallel grippers vs. dexterous hands (AI-generated image)
According to the Nature Will team, previous work has largely focused on one end of the spectrum: either general policies targeting grippers or dexterous operating systems tailored for single tasks.
On the generality side, significant progress has been made with models like π0.7. However, achieving such generality with dexterous hands is primarily bottlenecked by data. Generalization relies on large-scale, diverse training data, yet real-world robot data is inherently expensive, and teleoperation data collection for dexterous hands is slow and difficult.
The Nature Will team also identified a more fundamental circular dependency: robots can only collect data in scenarios where they already know how to operate, yet they can only operate in scenarios where they have existing data. To break this cycle, data must come from sources outside the robot itself.
Consequently, more teams are turning their attention to an available data source: humans themselves.
In February of this year, NVIDIA and other institutions released EgoScale, which pre-trained a Vision-Language-Action (VLA) model using 20,854 hours of first-person human manipulation videos. They observed a log-linear relationship between data scale and validation loss. When deployed on a 22-DoF dexterous hand, the average success rate was 54% higher than the baseline without human data pre-training.
Scaling behavior of human pre-training: The function relating optimal validation loss at convergence to human data scale reveals a near-perfect log-linear Scaling Law (arXiv.16710).
Six months later, this magnitude was increased by approximately 50 times. On August 10, US robotics company Dyna Robotics released its world-action model DYNA-2. Its pre-training utilized over one million hours of first-person human videos, with no robot data involved during the pre-training phase. According to Dyna's report, as data scaled up from 1,000 hours to 1 million hours, the model's average normalized score across 14 robot tasks rose from 20% to 53%.
Scaling laws for the Dyna-2 world-action model, trained on a data ladder ranging from 1,000 to 1 million hours of human video. https://www.dyna.co/dyna-2
However, neither of these works bypasses the same bottleneck: translating human experience into specific tasks still requires additional adaptation. EgoScale relies on a small amount of human-machine alignment data for intermediate training, while DYNA-2 uses up to 10 hours of robot data for post-training on each evaluation task.
It has been proven that human experience can be scaled to feed robots. The next question is: Can this experience transfer directly to new tasks on a high-degree-of-freedom 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-driven company centered on foundation models for the physical world. Its stated goal is Physical AGI (Artificial General Intelligence): training general-purpose foundation models capable of closing the loop between perception, prediction, reasoning, and action in the real world.
The Nature Will team told Synced that this vision stems from two key judgments:
- A general-purpose foundation model for the physical world is feasible;
- Human first-person view data is now ready for large-scale collection.
The methodology behind these judgments continues the approach established in LLM research, where models, data, and infrastructure must evolve together. The Nature Will team told Synced: "We followed the principle of scaling model size and data size in tandem, as seen in large language model research. Initially, we used smaller models and datasets to quickly validate the architecture. Once the direction was confirmed, we gradually expanded, rather than committing to a 50B parameter model from the start."
Natura-1 is Nature Will's first general-purpose foundation model for the physical world. Its training involves two steps: pre-training on 1 million hours of human experience data, followed by efficient post-training with limited robot data to adapt the world knowledge learned during pre-training to dexterous manipulation.
Unlike many systems that separate "understanding instructions," "predicting the world," and "generating actions" into different modules, Natura-1 is a single MoE Transformer jointly trained on language, images, video, and actions. It targets two capabilities:
- Dexterous manipulation, leveraging dexterous hands for multi-finger coordination, object pose adjustment, and complex contact interactions;
- Zero-shot instruction following, enabling the execution of operations based on natural language instructions in unseen scenarios and tasks, including handling spatial constraints, multi-object relationships, and negative conditions.
What Can Natura-1 Do?
The Nature Will team evaluated Natura-1 across three dimensions:
- Long-horizon tasks assess how precise and sustained a skill can be after training.
- Zero-shot tests evaluate whether the model can handle tasks it has never practiced.
- The final section records emergent capabilities, examining whether the model can devise solutions when tasks do not go smoothly.
Dexterity: Six One-Take Long-Horizon Tasks
In addition to making wontons, the Nature Will team released five other demos: brewing tea, preparing a yogurt bowl, packing a parcel, applying a screen protector, and writing calligraphy. The longest clip exceeds 3 minutes, all recorded at original speed, in one take, and without human intervention. Each segment contains multiple sub-tasks, requiring the robot to continuously handle liquids, granular materials, soft tools, and flexible components within a single workflow.
Taking the 3-minute screen protector application as an example, Natura-1 first wipes the screen, then aligns the pre-mounted protective film frame onto the phone, pressing it down firmly with both hands. Subsequently, one hand holds the phone and frame against a fixture while the other hand smoothly peels off the release liner. Finally, it uses its fingertips to push along the edges and corners to eliminate air bubbles.
When packing a parcel, it switches grip strategies between raffia grass, plush toys, and notebooks. One hand presses down on the spring-loaded box lid while the other operates the tape dispenser to seal the box, finishing with label application and barcode scanning.
These actions alter object shapes and contact conditions in real time as they are executed. The model must continuously correct itself based on real-time feedback, rather than following a fixed trajectory. More demos can be found in Nature Will's official article "The Will to Act | Natura-1 Physical General Foundation Model."
It is important to note that these six demonstrations showcase operational depth after task fine-tuning. Whether the capability can be transferred to unpracticed tasks depends on the zero-shot tests.
Zero-Shot: Unpracticed Tasks
What constitutes zero-shot? The Nature Will team told Synced: "By 'Zero-shot,' we mean that after completing pre-training and real-robot post-training, the model accepts instructions and executes them directly, without collecting specialized data or performing additional training for the test tasks."
The team built a benchmark covering 12 capability dimensions. The test tasks, along with their scenario and action combinations, were absent from the post-training data. Next to each test video, the closest demonstration from the training set is displayed side-by-side to allow external assessment of the differences. For instance, one task requires placing an apple in a bowl, putting a bottle on a tray, and then carrying the tray away. In the training set, only discrete actions like "pick up fruit" and "carry a loaded tray" existed. Natura-1 combined these actions, successfully carrying the tray without touching another tray nearby.
On this benchmark, π0.5 and OpenAI's GPT-6 Astra (acting as a control policy via code generation) failed to score on the dexterous hand platform. When switched to a gripper platform, they scored 13.5 and 16.7 respectively. Natura-1 achieved scores of 53.4 (Base) and 59.1 (Agentic) on the dexterous hand.
Moreover, Natura-1 demonstrates a clear advantage in execution time.
The team also tested the model with open-ended instructions spontaneously devised by researchers, many of which were intentionally tricky. For example, when asked to place a toothbrush holder and a water bottle into a basket, the instructions specified different hand coordination patterns: placing both simultaneously, left-hand first then right-hand, or right-hand first then left-hand. Natura-1 successfully executed all these variations as requested.
Regarding "unseen" tasks, the team provided a specific definition: "Since the pre-training data covers a wide range of scenarios, it may already include relevant objects and operational experiences. Therefore, 'unseen' here primarily refers to tasks not covered during post-training on real robots." They cited pressing the spacebar on a laptop as an example: while this task was absent from the robot's post-training data, similar actions likely appeared in human video datasets.
This highlights the significance of zero-shot capabilities in this research direction: can experiences derived from human videos be directly transferred to real-world robots without corresponding robotic data? As the team stated, "Our focus is not just on whether the model can grasp a different object, but also on whether it can organize existing capabilities to complete tasks that have not been specifically trained for."
Unexpected Emergence
The team also observed behaviors that were not part of the original test plan.
In one deliberately set-up scenario, a basket containing yarn was placed on a high platform out of reach, with a long rod inserted through its handle. The only instruction given was, "Try to grasp the yarn." Natura-1 grabbed the rod to pull the basket closer, stabilized the rod with one hand, and retrieved the yarn with the other. In another instance, after failing to pick up a playing card from its edge, the model pushed the card toward the edge of a box until part of it overhung, then used both hands to lift it. According to the team, none of these intermediate steps existed in the robot demonstration data.
The team recalled that when the data scale was smaller, the model would fail to grasp an object but continue moving its hand according to the original plan. However, as the scale increased, the model began to adjust in real-time and retry. The Nature Will team told Synced (Machine Heart): "We did not explicitly collect failure-recovery demonstrations during real-robot post-training, yet the model still exhibited recovery capabilities. We believe that large-scale human pre-training data contains rich experiences of adjustment and retrying, from which the model learned how to handle variations during manipulation."
The team also acknowledged: "This is a phenomenon we observed when scaling up training; the specific sources and mechanisms of these capabilities require further investigation."
Where Do the Capabilities Come From?
According to the Nature Will team, long-horizon dexterous manipulation and zero-shot generalization "do not happen by chance." While human experience data is abundant, it serves merely as raw material. Transforming this into robotic actions requires co-designing data, models, and infrastructure:
- Data provides language and action supervision for each video segment
- Models learn as much as possible from every trajectory
- Infrastructure must support large-scale training while ensuring inference completes within 100 milliseconds
Data
Natura-1’s training data consists of 1 million hours of curated first-person human experience, comprising approximately 61.7 million clips, 318,000 tasks, and 14,000 scenes. The dataset is primarily composed of controlled collections, supplemented by first-person videos from the internet.
The team uses a structured task taxonomy and agent workflow analysis to identify coverage gaps, shifting data collection priorities toward underrepresented areas. Raw recordings undergo cleaning, deduplication, action reconstruction, and hierarchical language annotation to transform into "video-language-action" trajectories. On internal benchmarks, their proprietary action reconstruction pipeline achieves 41% lower joint position error compared to existing open-source methods.
Schematic of the Natura-1 Data System
Choosing human data and choosing dexterous hands are essentially the same decision. The Nature Will team told Synced: "Many tools and operational methods in home and industrial scenarios are designed around the human hand. We want robots to adapt to these tools, rather than requiring us to redesign tools and workflows every time we enter a new scenario." On the other hand, first-person human data is easier to scale for collection, while "using dexterous hands preserves richer hand motion information, reducing the information loss that occurs when compressing human hand movements into simple gripper open/close actions."
Infrastructure
"Make everything scalable" is the core technical philosophy of the Nature Will team:
- Data Side: A data lakehouse manages tens of petabytes of data. Kubernetes-based parallel pipelines produce 350,000 hours of data per week, with throughput scaling nearly linearly with compute resources. The team explains that the design goal of this pipeline is: "When data volume increases tenfold, by adding approximately ten times the resources, processing can still be completed in nearly the original time frame without needing to rebuild the entire system."
- Training Side: Their proprietary framework achieves over 50% Model FLOPs Utilization (MFU) on hundreds of GPUs. The throughput of their custom attention operators is more than 6 times that of tuned FlexAttention.
- Inference Side: A service framework rewritten in Rust reduces end-to-end latency to under 100 milliseconds, meeting the requirements for real-time control on physical robots.
The Nature Will team states: "The underlying logic aligns with the infrastructure development of Large Language Models (LLMs): as data and model scales continue to grow, systems must be prepared for this growth in advance."
Model
Natura-1 arranges entire task segments into a single sequence comprising language, image, video, and action tokens, employing autoregressive modeling with a standard Mixture-of-Experts (MoE) Transformer backbone.
Its core concept is to view video and actions as two projections of the same physical state. Action labels are scarce, whereas video data is abundant. The model uses past videos as memory to predict future videos, training both video and action simultaneously via flow matching to extract more supervision from each trajectory. To prevent the model from focusing solely on visual inputs while ignoring instructions, pre-training leverages fine-grained descriptions to ensure language carries significant weight, mapping different intents to distinct prediction outcomes.
Scaling curves show that as training experience grows from approximately 44,000 hours to 1 million hours, the action prediction error (MSE) on held-out human data decreases by 47.7%. With the same weights, the error on held-out real-world robot data decreases by 67.7%. These metrics represent prediction errors under Teacher Forcing, not task success rates, but they demonstrate that increased human experience effectively transfers to robotic performance.
Natura-1 Scaling Curves
Conclusion
Natura-1 has provided early evidence that dexterity and generality can coexist. Regarding the next phase, the Nature Will team told Synced: "Our most core objective is to continue improving the model's generalization capabilities across different tasks and scenarios, and to enable it to reliably complete complex tasks in the real world."
The team states that the goal is to move from "completing partial steps" to "reliably and efficiently completing entire tasks." To achieve this, they plan to incorporate tactile information into their data and models, pre-train on larger-scale datasets, reduce adaptation costs when transferring to robots, and enable models to run on more affordable hardware. The team anticipates that this will require long-term investment.
In the broader industry context, the approach taken by Natura-1 is shifting from a bet made by a few teams to a growing consensus. In June of this year, a study named HumanScale compared two types of pre-training data at the same scale and found that first-person human videos outperformed real robot data collected via teleoperation, with the most significant advantage observed in out-of-distribution generalization. Dyna Robotics, which released DYNA-2 in August, stated that their next step is to expand training scale to 10 million hours.
As millions of hours of data become less scarce, the focus of competition is also shifting: Can the same human videos be more accurately reconstructed into actions and annotated with language? How much adaptation is required when deploying models onto robots? And what kind of embodiment should be chosen to leverage these experiences? Natura-1 is betting on one of the most challenging combinations: using high-degree-of-freedom dexterous hands to embody human experience, while enabling the same model to directly handle tasks it has never encountered during training. How far this path can go depends not only on whether the stability and cost of dexterous hands can keep pace, but also on whether the scaling curve for human experience can continue to extend.
In a world shaped by and designed for human hands, how far can robots go by learning from them? Nature Will has delivered its first verifiable answer.
References
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
Please contact our official account for authorization before reprinting.
For submissions or media inquiries: [email protected]