GR00T (Project GR00T)

NVIDIA's Humanoid Robot Foundation Model Series

Author’s Note

  • First open humanoid VLA. GR00T N1 opened the era of open foundation models for humanoid robots.
  • Practical Dual-System architecture. A representative example of human cognition-inspired System 1/2 separation applied to real robot VLA.
  • Demonstrated synthetic data potential. Showed the viability of a training pipeline that reduces teleop dependency through Isaac Sim physics simulation + Neural Trajectory generation.

Overview

ItemDetails
Initial AnnouncementMarch 2024 (GTC 2024)
N1 Open-Source ReleaseMarch 2025 (GTC 2025)
CompanyNVIDIA
GitHubNVIDIA/Isaac-GR00T
LicenseApache 2.0

Version History

ItemN1N1.5N1.6
Announced2025.03 GTC2025.05 Computex2025.09 CoRL
Parameters2.2B3B3B
VLMEagle2-1BEagle 2.5 (Frozen)Cosmos-Reason-2B
DiT Layers161632
Action SpaceAbsoluteAbsoluteRelative
Key ImprovementFirst open-sourceFLARE, 2x language complianceScale-up, Loco-manipulation

Dual-System Architecture

GR00T N1 Architecture

GR00T N1 Architecture: System 2 (VLM) + System 1 (DiT) Dual-System Structure

Inspired by human cognition principles (Kahneman, 2011):

SystemRoleImplementation
System 2 (Slow)Environment understanding, planningVision-Language Model
System 1 (Fast)Convert plans to precise motionsDiffusion Transformer

References


See Also

GR00T Series