Natura-1: 50B trained on 1M hours of human video drives 20-DoF hands at 100ms (www.naturewill.ai)

Natura-1, announced by Nature Will, is a groundbreaking general foundation model designed for physical interaction, utilizing a 50-billion-parameter mixture-of-experts architecture. Trained on over one million hours of egocentric human video, this model can interpret and execute complex natural language instructions, driving dexterous 20-degree-of-freedom hands with a rapid 100-millisecond control loop. Natura-1 demonstrates the ability to generalize across various tasks and environments, including intricate activities like folding materials and performing calligraphy, showcasing its capacity for zero-shot learning. This development is significant for the AI/ML community as it marks a pivotal advance in the integration of natural language processing with physical manipulation, progressing toward general physical intelligence. Natura-1's innovative training approach and robust data infrastructure allow it to learn from diverse human experiences, thus overcoming limitations faced by traditional robot systems. The model's architecture facilitates high-efficiency training and execution, and its successful benchmark performance against existing models illustrates its superior dexterity and adaptability. Ultimately, Natura-1 paves the way for scalable solutions in robotics, emphasizing the importance of human-like dexterity for future AI applications in the physical world.