Explainers
NVIDIA Isaac GR00T explained: the open models and tools behind many humanoids
Short answer: Isaac GR00T is NVIDIA's platform for humanoid robot developers: a family of open foundation models that turn instructions and camera input into robot actions, plus tools for collecting data, training in simulation and running the result on NVIDIA's Jetson Thor computer. NVIDIA announced GR00T N1, which it called the world's first open humanoid robot foundation model, on 18 March 2025. Its developer page now lists GR00T 1.7 as the latest open vision-language-action model in the family.
How the model works. NVIDIA said GR00T N1 uses a dual-system design. "System 2" is a slower vision-language model that reasons about the scene and the instruction and plans actions. "System 1" is a fast action model that turns the plan into precise, continuous movements, and NVIDIA said it was trained on human demonstration data and a large amount of synthetic data. Figure's Helix uses a similar split; see what is a VLA model.
GR00T 1.7. According to NVIDIA's developer page, GR00T 1.7 is a 3-billion-parameter model with a Cosmos backend and a commercial licence, available on GitHub, Hugging Face and the LeRobot library. NVIDIA says it can run on Jetson Thor at 25 Hz using TensorRT with NVFP4 quantisation.
The workflow. NVIDIA describes three steps: capture demonstrations by teleoperation in simulation or on a real robot; post-train GR00T on real or simulated demonstrations and evaluate in Isaac Lab-Arena; then deploy the trained policy on Jetson Thor for on-robot inference. It also offers a GR00T-Ready evaluation for robot makers and a Halos safety offering for humanoids.
Why it matters. Because the models are open, smaller robot makers and labs can start from a pre-trained base instead of training from scratch. Many companies in our coverage build on NVIDIA hardware or software; see the NVIDIA hub and our report on NVIDIA's Agile ONE SN reference humanoid.