A company founded nine months ago enables robots to step out of the "rehearsed world".

In the past nine months when embodied intelligence has been exceptionally bustling, one team has remained particularly low-profile.

While peers compete for attention with successive robot demos, it has placed its bets on a slower and more costly undertaking: training a general-purpose model for the physical world.

This company, called Natural Will, was founded in January 2026. From the very beginning, it set its sights on building a general foundational model for the physical world, with the goal of enabling the same model to continuously adapt to new tasks.

It was not until October 2026 that the team stepped onto the stage with its product for the first time, releasing Natura-1, a 50B-level general foundational model for the physical world. It completed pre-training using millions of hours of first-person human data, and has currently demonstrated dexterous manipulation and zero-shot instruction following capabilities. When faced with unfamiliar objects and ad-hoc instructions, it can mobilize existing experience to organize movements, and can even handle more complex long-horizon tasks after adaptation.

At present, Natura-1 has also begun to absorb real-world experience.

01 Robots Learn to Solve New Problems

It is no longer news to see robots in daily life. Take the common coffee-making robots in shopping malls as an example: the robotic arm takes the cup from a fixed position, adds ingredients in a pre-set order, and then places the finished product at the pickup port. As long as the cups and ingredients stay in familiar positions, it can work stably.

Troubles often start with an accident. When the cup deviates from its original position, the coordinates written in the program will no longer work. The robot's movements are still accurate, but the result is that it accurately grabs at thin air.

Team members once summed up this difference in one sentence: "If there are no variables, there is no need for modeling, and programming is sufficient."

This is indeed the case. Pre-programmed robots act according to the process written in advance by engineers, and are prone to failure once the environment changes. Robot models have an extra step of judgment: they observe the current environment, understand the instructions given by humans, and then organize movements according to changes on site.

It is the moment when the program deviates from reality that the model comes into play.

The Zero-Shot Instruction Following emphasized by Natura-1 solves this type of change, that is, the model skips the special training for the current task, directly learns autonomously under unfamiliar instructions, and executes movements.

For example, the training data has taught the robot to pick up apples, but has not taught it to wipe objects, nor has it arranged the task of wiping apples. But when the staff gives this instruction temporarily, the model can quickly and autonomously learn the new combined movement of wiping apples.

In another test, the Natural Will team first asked the robot to move objects on the left and right sides, and it completed two grabs in sequence. After adding the word "simultaneously" to the instruction, the robot's movement changed: one hand grabbed the object and stayed in place, the other hand then completed the grab, and finally both hands moved together.

In another test, the team temporarily brought a laptop and asked the robot to open the screen and press the space bar. There had never been computers and keyboards in past training, and the team did not have high expectations originally. Surprisingly, the robot quickly lifted the screen, moved its hand above the keyboard, and finally extended its index finger to press the space bar.

This is undoubtedly an interesting experiment, which shows that the model can already bring the accumulated object knowledge and movement experience from the past to unfamiliar tasks. After several months of training, the progress of Natura-1 has become increasingly obvious. It can not only handle more temporary instructions in a zero-shot manner, but also complete continuous operations with longer steps after task adaptation.

These two examples seem very small, but behind them is a long-standing problem faced by embodied intelligence: robots can memorize a set of movements in training, but new problems and new requirements will appear in the real world every day. Only when the model understands the relationship between objects and movements can it reorganize the experience it has learned.

For robot companies, this capability will also change the project delivery method.

Natural Will has long connected post-training and real-machine reinforcement learning, so it can quickly use grippers to complete tasks such as making dumplings, but the team believes this is not the end point. In their view, "The world is built for hands", and dexterous hands are the future. From a practical perspective, the knife holders and trays that seem to be casually placed on the table in many robot demonstrations have actually gone through several rounds of testing. If they are moved a few centimeters to the side, the success rate may drop significantly.

Therefore, behind a smooth video of a few minutes, there is often one or two weeks of debugging. As the number of projects increases, the number of on-site debugging personnel also needs to increase. Every time a robot is sold, the company takes on an additional delivery workload.

General-purpose models are trying to change this situation. They first master a wider range of operational experience, and only need to supplement a small amount of data when entering new scenarios, allowing engineers to focus their energy on longer-tail links.

Natural Will later chose to use dexterous hands to test making wontons. The team found a simple wrapping method online, asking the robot to place the wonton wrapper on the palm of one hand, and then pinch it closed with the other hand. However, the small model often pinched empty, and sometimes used too much force to poke through the wonton wrapper.

According to past experience in making demonstrations, this means that the process needs to be redesigned. But the next day, after a version of the model trained with more data was put into operation, the robot aligned with the wonton wrapper by itself and controlled the finger strength. The movement that was originally expected to be modified repeatedly succeeded in this way.

This is also the reason why Natural Will chose dexterous hands. Grippers are mature, but their movements are mainly opening and closing. When facing scissors, they need supporting modified tools; when handling soft objects, the two directions of opening and closing are difficult to express the subtle cooperation between fingers.

Dexterous hands allow robots to work in the way humans do. Zero-shot instruction following enables them to understand new tasks that appear temporarily, while dexterous manipulation is responsible for turning understanding into movements. After the two capabilities are combined, robots can get rid of the working mode of matching one program to one task, and have the opportunity to turn customized deliveries one after another into products that can be replicated on a large scale.

02 Where Do Millions of Hours of Data Come From

After the dexterous hand route was determined, the team soon faced the next challenge: data collection.

Repeatedly collecting data around the same task will indeed make the model more and more proficient, but the next task still needs to be prepared from scratch. If general capabilities depend on the model having seen a sufficiently rich variety of operations, the data source must be expanded for a long time.

Natural Will identified Ego first-person perspective data as the core training data very early. Compared with robot teleoperation, the collection process of Ego data is lighter. A mobile phone fixed on the human head can record the whole process of both hands completing real work. The collection scenarios can be expanded quickly, and the cost is easier to control. Based on this, Natural Will has gradually built a set of data processing pipelines. Today, the pre-training data used by Natura-1 has reached millions of hours, mainly from human operations in the real world.

Ego first-person perspective videos record human hands. Information such as how fingers bend, how the palm holds the object, and how the contact point changes is all hidden in the video. If a gripper is used in the end, most of these details can only be compressed into one opening and closing action.

In the team's exact words, "Information can be reduced from much to little, but it cannot be increased from little to much."

Human hand movements can be simplified into gripper movements, but gripper data cannot make up for the cooperation of five fingers. Since Natural Will chose human operation data, dexterous hands have become a more matching carrier.

In terms of model architecture selection, the team initially tested different solutions for two or three weeks, one of which was close to the common VLA in the industry, and the other paid more attention to action generation efficiency. In the end, the team did not hesitate too much, and quickly established the route of "general foundational model for the physical world".

It is not a simple Action Model (world action model). As the training scale expanded, Natural Will incorporated language understanding and visual information into the same model. The team believes that Natura-1 can hardly be fully classified into the existing categories of VLA or WAM, so it is defined as an independent route of general foundational model for the physical world.

This is not a simple naming game. The computing power and capital of start-ups are limited. Every additional training of the model consumes real money. With the same batch of data input, an architecture with higher learning efficiency can obviously avoid many detours.

After rounds of training, the parameter scale of Natural Will's model has reached 50B, making it a physical world foundational model with a particularly large parameter scale in the current public disclosures of the industry. When the data volume is only tens of thousands of hours, the performance of the 5B and 50B models is close. But after the data increases to a larger scale, the 50B model has room to absorb more experience.

In other words, 50B and millions of hours are a set of supporting inputs. Millions of hours provide the model with broader experience, and 50B parameters are responsible for accommodating these experiences. In the team's view, the Scaling Law that has been verified by large language models will also take effect in the physical world: as the data volume and model scale grow synchronously, robots will gradually acquire new capabilities that previously required item-by-item training.

For a start-up company, this route means heavier upfront investment. Many robot companies first carry out projects and then use revenue to support R&D; Natural Will chose to build a data system first, and then train large models. The former is easier to get orders immediately, while the latter bets on the repeated use of the same set of models in more scenarios, which has greater industrial scale effect.

New data improves the model, the improved model returns to the real machine for verification, and the errors in the test provide clues for the next round of training. Natural Will hopes to form a set of continuously evolving physical intelligence infrastructure through this, so that the capabilities of robots can continue to grow with the accumulation of data.

03 After Stepping Out of the Demo

After the model scale expanded, Natura-1 had another pleasant surprise: the model began to learn to correct errors by itself.

A test required the robot to pick up the mobile phone on the table and then place it in a designated position. After the early version missed the grab for the first time, the robotic arm would still continue to move towards the target position, as if the mobile phone was already in its hand. When the arm reached the end point, it sometimes found that the task was not completed, and then returned to the original position to start over.

The new version responds earlier to this situation. When the fingers touch the mobile phone but do not grasp it stably, the robotic arm will stay in place and readjust the angle. Sometimes it will even try several times in a row until the mobile phone truly leaves the table.

Humans can naturally interact with the model in language, and then the robot executes the movements. If it fails, it will try and correct errors by itself.

A similar situation occurred when using scissors. After the first finger deviated from the scissor loop, the robot would continue to fine-tune downward to let the finger find the entrance, and then pick up the scissors. The team introduced that only a small number of consecutive failure samples were added during training, and the model was more likely to have learned this retry behavior from millions of hours of human operations. Just like people sometimes slip their hands when holding a mobile phone, the next reaction is usually to grab it again in place.

This also involves the memory capability of Natura-1. Some early robot models only generate movements based on the current picture, and each frame is like a fresh start. The Natural Will team described this state as "Every day is the first day".

Natura-1 refers to the pictures it has seen before. It knows that it has grabbed once just now, and can compare the changes before and after the movement. For short tasks of a few seconds, the role of this memory may be limited; but after the task is extended, the memory capability becomes more and more indispensable.

The emergence of error correction capabilities brings robots closer to real work, but Natural Will believes that another standard is equally important: an occasional success only means that the model has the possibility to complete the task. Only when the task reaches a level of stable reproduction can it be regarded as a mastered capability.

After the model is released, Natural Will plans to invite external personnel to the site for testing. Visitors can temporarily input instructions, or take out items they carry with them to let the robot operate on the spot. Compared with pre-prepared demonstrations, this kind of test is closer to the customer environment that the model will face in the future. Of course, it will also expose its boundaries more directly.

Fortunately, the future is getting clearer and clearer. The team gave an example of stir-fried pork with green peppers: the robot recognizes green peppers and knows how to use a knife, but to independently complete a dish, it not only needs to control the cutting size, handle various changes during cooking, but also accurately put the used knife back into the knife holder. Once the knife falls, the subsequent steps need to be adjusted accordingly. The team's goal is that after a robot enters an unfamiliar kitchen, it can also complete the whole set of operations according to the new recipe.

Of course, there are dozens of uncertainties between a short instruction and finishing a dish. The model may produce errors at every step, and the hardware also needs to run for a long time. In a specific environment, factories will compress the time to a fixed beat, while families will put forward stricter requirements for safety.

This is also the current distance between all embodied models and commercialization. What is surprising about Natura-1 is that it has already demonstrated the ability to handle unfamiliar instructions, and has begun to actively retry after failure. In the next stage, Natural Will will put the model into more complex long-horizon tasks, and test the zero-shot capability with the completion rate of the whole set of tasks.

By then, millions of hours of data will become the basic standard configuration of general physical models, and tens of billions or even larger parameter scales will also become a business with more imagination.