Natural Will releases Natura-1: After learning 1 million hours of human motion, the robot can perform tasks it has never practiced.

Beating AI News Flash: Embodied intelligence company Nature's Will has released Natura-1, a MoE model with 50 billion total parameters, aiming to enable robots to perform operations according to instructions without repeatedly training for each new task.

In the past, training robots mainly relied on manually remote-controlling real machines to collect motion data, which was costly and limited in scenarios. Natura-1 takes a different route: it first pretrains on 1 million hours of first-person human video, then adapts dexterous hands with a small amount of robot data. This batch of data contains about 61.7 million operation segments, covering more than 14,000 scenarios.

Nature's Will first reconstructs the 3D motion of fingers and the body from the footage, then annotates task goals and operation steps layer by layer. The model places language, video, and action into the same Transformer, jointly learning how the hand should move next and how the画面 will change after the action. Every step of the entire operation can provide training signals, without having to extract only individual action clips. During execution, the model can also adjust actions by referring to previous frames and real-time feedback, directly controlling a dexterous hand with 20 degrees of freedom per hand. The team says the latency from acquiring frames to outputting actions is controlled within 100 milliseconds.

After completing general robot adaptation, Natura-1 can combine existing skills without additional training for new tasks. For example, if the training demonstration only taught pulling out a tissue and placing it on a tray, the robot can nevertheless follow instructions to pull out a tissue, wipe away water stains, and then throw it into a trash can. When faced with the requirement to exclude six colors, it can select bottles that meet the conditions. When grabbing a flat playing card, it will also first push the card to the edge of the box, then lift it from below. These demonstrations use the same set of model weights, without separate fine-tuning for each task. The team also released a nearly 30-minute single-take video of daily interactions.