nature will train Natura-1 on one million hours of human experience

Nature Will unveils Natura-1, a 50-billion-parameter AI model trained on one million hours of human gestures to control robotic hands.

Folding a dumpling, handling a paintbrush, or retrieving an out-of-reach object: ordinary actions for a human, but ones that remain difficult to replicate for a robot. With Natura-1, Nature Will seeks to push past this limit by training an artificial intelligence model on everyday movements and experiences. The goal is to enable robots to execute naturally phrased instructions, including in situations they have never encountered before. One million hours to learn human actions

Natura-1 is based on a 50-billion-parameter Mixture-of-Experts (MoE) architecture, pretrained on one million hours of video recorded from a human perspective. These sequences document interactions with objects, hand movements, and the sequencing of actions across various environments.

In this way, the model learns to connect images, language, and physical actions. Following this learning phase, a limited amount of robotic data allows its knowledge to be adapted to artificial hands, each equipped with 20 degrees of freedom.

Nature Will states that scaling the training data from 44,000 to one million hours reduced the mean squared error of robotic action prediction by 67.7% on its validation set, without any additional training on this test data. Robots capable of improvising

In the demonstrations published by Nature Will, Natura-1 follows complex instructions without prior specific training on the presented situations. In particular, it can distinguish between multiple objects, follow an execution order, or simultaneously coordinate both of its hands.

Certain sequences also reveal unexpected behaviors. Faced with a ball of yarn placed in an out-of-reach basket, the robot uses a stick to pull the basket closer before retrieving the object. In another example, it slides a card toward the edge of a box in order to grasp it more easily.

These strategies were not explicitly described in the instructions. Nature Will considers them to be the first signs of an ability to solve novel physical problems, rather than merely replicating memorized movements. Complex, multi-minute manipulations

The company also presents six demonstrations of long-horizon tasks, recorded in continuous takes and carried out without human intervention during their execution. The robot prepares wontons, serves tea, packs a parcel, performs calligraphy, or applies a screen protector.

These six demonstrations, however, benefited from additional training tailored to the specific tasks involved. They therefore illustrate a capability distinct from zero-shot generalization: the ability to sustain a sequence of precise actions over time, despite the constraints associated with flexible materials, grip adjustments, and tool use.

To maintain this responsiveness, Nature Will developed an inference infrastructure capable of preserving a control loop under 100 milliseconds between observation and action. Robotics aims to learn from the real world

Nature Will's approach is rooted in one idea: robots will not be able to learn all human activities solely from demonstrations performed by other robots. Data derived from human experience offers far greater diversity, provided that this learning can be successfully transferred to machines with different physical constraints.

Natura-1 remains a research project, with results primarily documented through its developer's evaluations and demonstrations. Nature Will acknowledges that generalizing to a broader range of environments will still require advancements in data, hardware, and infrastructure.

Nevertheless, the project illustrates a major shift in physical AI: training models capable of understanding intent, observing their environment, and adapting their movements, without having to program every new situation separately.