Let's play a little game: imagine having to explain to someone how to pick up a cup, move it, and put it down somewhere else without spilling the contents. It seems simple, but if you try to describe each movement precisely, you'll realize that it would take pages and pages of instructions. That's why humanoid robotics has always been so complex. Until yesterday. With the announcement of the NVIDIA Isaac GR00T N1, robots can now understand and replicate complex movements simply by observing human actions, thanks to a cognitive architecture that simulates both the intuitive and rational processes of the human brain.
NVIDIA and robots, artificial intelligence becomes concrete
I really like how Jensen Huang, CEO of NVIDIA, put it during his presentation at GTC 2025: “The era of generalist robotics is here.” This is not an empty slogan, but a statement of fact. For the first time, we have an open and customizable foundational model that allows humanoid robots to reason and act in the real world.
Il GR00T N1 (Generalist Robot 00 Technology) is not just another lab experiment. It is a technology available now to developers around the world, designed to transform industries suffering from labor shortages (estimated at over 50 million people globally).
A two-speed brain
What makes the GR00T N1 truly extraordinary is its architecture inspired by human cognition: a System One which works as a “fast action” model (similar to our reflexes) and a System One which operates as “slow thinking”, capable of reasoning about the surrounding environment and planning complex actions.
This duality allows robots to perform tasks that are seemingly simple for us humans but incredibly complex for machines: grasp objects, move them with one or both arms, and transfer them from one arm to the other. They can even perform multi-step operations that require extended context and combinations of different skills.
NVIDIA and Robots, From Simulation to Reality
NVIDIA didn't just develop the model, it created an entire ecosystem for humanoid robots. In collaboration with Google DeepMind e Disney Research, the company is developing Newton, an open-source physics engine optimized for robot learning.
Using components of the NVIDIA Isaac GR00T Blueprint, it was possible to generate 780.000 synthetic trajectories (equivalent to 6.500 hours, or nine continuous months, of human demonstrations) in just 11 hours. The combination of synthetic and real data improved the performance of GR00T N1 by 40% compared to using real data alone.
An open future
The real change, as you may have guessed, is in the open approach. The training data and evaluation scenarios of GR00T N1 are available for download at hugging face e GitHub. This will allow developers and researchers from around the world to participate in the creation of increasingly capable and useful humanoid robots.
Companies like 1X Technologies, Agility Robotics, Boston Dynamics and others have already had early access to this technology, and the results are impressive. Bernt Bornich, CEO of 1X Technologies, said that with a minimal amount of post-training, they were able to fully implement the system on their robot NEO Gamma.
We are no longer in the realm of promises: the era of generalist robotics has truly begun.
