NVIDIA Isaac GR00T

NVIDIA open foundation model for humanoid robot reasoning and skills

License

Apache-2.0

Overview

NVIDIA Isaac GR00T N1 (now at version N1.7) is an open foundation model for generalist humanoid robot reasoning and skills. This cross-embodiment Vision-Language-Action (VLA) model takes multimodal input including language and images to perform manipulation tasks in diverse environments. Features a dual-system architecture inspired by human cognition: System 1 (fast-thinking action model for reflexes) and System 2 (slow-thinking vision-language model for deliberate reasoning). Trained on real robot data, synthetic trajectories, and internet-scale video content. Supports multiple robot embodiments including GR1, OXE_DROID, and AGIBOT_GENIE1.

NVIDIA Isaac GR00T N1 (now at version N1.7) is an open foundation model for generalist humanoid robot reasoning and skills. This cross-embodiment Vision-Language-Action (VLA) model takes multimodal input including language and images to perform manipulation tasks in diverse environments. Features a dual-system architecture inspired by human cognition: System 2 is a vision-language foundation model based on NVIDIA Cosmos-Reason-2B for deliberate reasoning, while System 1 is a diffusion transformer head with 32 layers that denoises continuous actions. Trained on thousands of hours of teleoperated data from bimanual YAM arms, AGIBot Genie1, Galaxea R1 Pro, and Unitree G1 whole-body locomanipulation tasks, plus synthetic trajectories and internet-scale video content.

Why it matters

GR00T N1 is NVIDIA's open foundation model specifically designed for humanoid robots. Its dual-system architecture — fast reflexive actions plus slow deliberate reasoning — mirrors human cognition and represents a major step toward general-purpose humanoid intelligence.

Problem it solves

Humanoid robots need both fast reactive skills (catching a falling object) and slow deliberate planning (figuring out how to set a table). GR00T provides a unified model architecture that handles both, eliminating the need to build separate systems for each capability.

Key takeaways

• Dual-system architecture: System 1 (fast reflexes) + System 2 (deliberate reasoning)
• Cross-embodiment VLA model works across different humanoid platforms
• Trained on real robot data, synthetic trajectories, and internet-scale video
• Supports multimodal input: language instructions and visual observations
• Open model weights available for research and commercial use

Who it is for

• Humanoid robot developers
• Embodied AI researchers
• Companies building general-purpose robots
• Research labs working on cross-embodiment transfer

Limitations

• Optimized for NVIDIA hardware (Isaac Sim, Jetson)
• Humanoid-focused — may not transfer well to non-humanoid platforms
• Requires Isaac Sim ecosystem for full feature set
• Large model size demands significant compute for deployment