GR00T N1

NVIDIA Isaac GR00T N1 is an open foundation model for generalized humanoid robot reasoning and skills. The cross-embodiment model accepts multimodal input (vision and natural language) and outputs continuous control actions for manipulation tasks across diverse environments.

Architecture

GR00T N1 employs a dual-system architecture inspired by human cognition: System 2 is a Vision-Language Model (based on NVIDIA-Eagle with SmolLM-1.7B) responsible for reasoning and planning; System 1 is a Diffusion Transformer generating continuous robot motion trajectories. Both systems are tightly coupled and jointly optimized during post-training.

Training

The model was trained on a pyramidal data mix: internet-scale web data and human videos at the base, synthetic data generated on NVIDIA Omniverse in the middle, and real teleoperated robot data at the peak. Combining 750K synthetic trajectories (generated in 11 hours via the Isaac GR00T Blueprint) with real data yielded a 40% performance gain over real-data-only training.

Availability

The GR00T-N1-2B weights are publicly available on Hugging Face. PyTorch fine-tuning and inference scripts are on GitHub (NVIDIA/Isaac-GR00T). Minimum post-training configuration: a single NVIDIA RTX A6000 or RTX 4090. Inference is supported on NVIDIA Jetson AGX Orin.