9-month-old startup 自然意志 releases Natura-1, a 50B physical-world model
Chinese embodied-AI startup 自然意志, founded in January 2026, has introduced Natura-1, a 50B-parameter general-purpose physical-world foundation model pretrained on one million hours of human first-person (ego) footage, according to 36Kr. The company says the model shows dexterous manipulation and zero-shot instruction following, at times learning new compound actions on the spot — such as wiping an apple, or opening a laptop and pressing the space bar even though laptops and keyboards were absent from training. It also demonstrates self-correction, pausing and re-adjusting its grip after a failed grasp instead of continuing toward the goal, and uses memory of earlier frames rather than acting frame by frame. After release, 自然意志 plans to let outside visitors issue ad-hoc commands and test the robot with their own objects, judging zero-shot ability by completion rates on full long-horizon tasks.
If the general physical-world model approach holds, robot deployment could shift from per-project on-site tuning toward small-data adaptation and replication — though the reported capabilities so far rest on the company's own demonstrations.