Imagine a world where you can just tell a robot what you want it to do, even with objects it’s never seen, and it just… does it. While huge companies like Figure are pouring billions into building full-blown humanoid robot brains, a lesser-known startup called Nature Will is making waves with a different approach. They’ve just unveiled something they call a “physical general foundation model” for robot hands, claiming it gives robots instant intelligence for dexterous tasks.
This isn’t just a small step; Nature Will’s Natura-1 is a massive 50-billion-parameter model. It’s a mixture-of-experts trained on a mind-boggling amount of data – over a million hours of human experience, plus language, images, and video. What does that mean for you? It means a robot hand powered by Natura-1 could understand plain language commands and execute them zero-shot, tackling tasks with impressive speed and flexibility.
This release comes right in the middle of a serious debate in the robotics world: will these big AI models eventually get so good that much of the specialized, expensive training for robots becomes obsolete? Experts like Ashish Kapoor from General Robotics see a three-layer cake of robot intelligence, and they believe this “middle layer” – where robot foundation models like Natura-1 sit – might be in for some disruption. Natura-1 isn’t trying to be a full general AI like GPT-6 that also controls robots; it’s laser-focused on hands, which are, let’s face it, pretty crucial for getting anything done.
While there are still big questions about its real-world success rates and rigorous testing, if Natura-1 truly delivers, it could drastically lower the entry barrier for advanced robot manipulation. This would mean that the real value shifts to making robots super specific and useful for individual jobs and companies, rather than just building the foundational brains. It’s a potential game-changer that could redefine how we build and deploy intelligent robots. Read the full story on Forbes.