Natural Will Unveils 50B Physical World Foundation Model Natura-1

Natural Will, a startup founded in January 2026, stepped into the public spotlight for the first time in October 2026 with the release of Natura-1, a 50-billion-parameter physical world foundation model. The model was pretrained on one million hours of human first-person video and has demonstrated dexterous manipulation as well as zero-shot instruction following. According to the company, Natura-1 can draw on accumulated experience to act on unfamiliar objects and ad hoc instructions, and can handle longer-horizon tasks after adaptation.

In tests, Natura-1 combined previously learned skills to wipe an apple without specific training for the task. When staff brought a laptop it had never seen and asked it to open the screen and press the space bar, the model used existing object knowledge and motor experience to complete the operation. The team also observed emerging self-correction in the new version: after a failed grasp, the robot no longer kept moving toward the target as if holding the object; instead, it adjusted its grip in place until it succeeded.

Natural Will describes Natura-1 as a separate physical world foundation model direction that cannot be easily classified as either a VLA or a WAM. The company has chosen dexterous hands as the execution platform and built a data pipeline centered on human egocentric video, while also linking post-training with real-robot reinforcement learning. The team believes that as data and model size grow together, robots will gradually acquire capabilities that previously required task-by-task training.