Nature Will Releases First Physical World Foundation Model: Natura-1

On October 13, Nature Will released its first foundation model designed for the physical world, Natura-1. The model was trained on one million hours of first-person perspective videos of humans interacting with different scenes and objects, enabling it to handle previously unseen objects and tasks. Natura-1 supports fine-grained instructions specifying an object's final position and orientation, maintains control when grasping methods change mid-way, executes different tasks with both hands, and tracks moved or opened objects during long tasks to determine the next step. A previous news flash reported that Nature Will released the 50B-parameter MoE model Natura-1; this piece provides a supplement on the same model's physical-world capabilities.