Nature Will unveils Natura-1, a 50B model trained on 1 million hours of human experience for dexterous zero-shot robotics

Beijing-based embodied AI company Nature Will has introduced Natura-1, its first general foundation model for the physical world, built to connect large-scale human experience with real-world robot control. The model uses a 50-billion-parameter mixture-of-experts architecture and was pretrained on 1 million hours of first-person human experience data before being post-trained with a limited amount of robot data. According to the company, Natura-1 takes language and visual input and runs a 100-millisecond control loop for a 20-degree-of-freedom dexterous hand. The release is accompanied by a set of real-time, single-take demonstrations, including wrapping wontons, making tea, preparing a yogurt bowl, packing parcels, applying a phone screen protector, and writing calligraphy. Nature Will says the system is designed not only for fine-grained dexterous manipulation, but also for zero-shot instruction following on tasks that were not specifically collected or trained for during robot post-training. The company also shared benchmark results, scaling data, and infrastructure details, arguing that large-scale first-person human video can transfer to high-DoF robotic hands with less task-specific adaptation. It positions Natura-1 as an early attempt to combine dexterity and generalization in one model, while acknowledging that the source of some observed recovery and replanning behaviors still needs further study.