GR00T (Generalist Robot 00 Technology)
Home > Models > GROOT
Key Significance
- First Open Humanoid Foundation Model: First open VLA for humanoid robots with GR00T N1
- Dual-System Architecture: Inspired by human cognition - separates System 2 (VLM, planning/reasoning) and System 1 (Diffusion, real-time motion)
- Demonstrates Power of Synthetic Data: Generates 780K trajectories (6,500 hours) in just 11 hours via Omniverse, 40% performance improvement over real data only
- NVIDIA Ecosystem Integration: Hardware-software vertical integration from Isaac Sim/Lab, Omniverse, to Jetson Thor
- Common Platform for Entire Industry: Collaboration with major humanoid companies including Figure AI, Boston Dynamics, Unitree
- Jetson Thor Dedicated Hardware: 800 TFLOPS Blackwell-based robotics computing platform announced
- Representative Case of Simulation-Based Scaling: Presents possibility of large-scale training without real data collection
- Continuous Improvement (N1.5/N1.6): Language instruction compliance 46.6%→93.3% (N1.5), 2x DiT and Cosmos VLM introduction (N1.6)
GR00T N1 Architecture: System 2 (VLM) + System 1 (Diffusion Transformer) Dual-System Structure
GR00T N1.6: Various Manipulation Task Demonstration
Overview
GR00T (Generalist Robot 00 Technology) is a general-purpose humanoid robot foundation model announced by NVIDIA at GTC 2024 and open-sourced with the N1 series in 2025.
| Item | Details |
|---|---|
| First Announced | March 18, 2024 (GTC 2024) |
| N1 Released | March 2025 (GTC 2025) |
| Company | NVIDIA |
| Paper | arXiv:2503.14734 |
| GitHub | NVIDIA/Isaac-GR00T |
| License | Apache 2.0 (Open Source) |
Versions
Project GR00T (2024.03)
First announced at GTC 2024. Presented vision for foundation model for humanoid robots.
- Announced goals: natural language understanding, human motion imitation, simulation-based learning
- Jetson Thor computing platform announced simultaneously
GR00T N1 (2025.03)
World’s first open humanoid robot foundation model.
| Item | Details |
|---|---|
| Type | Vision-Language-Action (VLA) |
| Architecture | Dual-system architecture |
| Action Generation | Diffusion Transformer |
| VLM | Eagle2-1B |
| DiT | 16 layers |
| Training | 1K H100 GPUs, 250K steps |
GR00T N1.5 (2025.05)
Significantly improved language instruction compliance.
| Item | Details |
|---|---|
| Key Improvement | Frozen VLM (VLM fixed during training) |
| Loss Function | FLARE Loss (Flow-based Action Reconstruction) |
| Language Compliance | 46.6% → 93.3% (+100%) |
| VLM | Eagle2-1B (Frozen) |
| DiT | 16 layers |
Major Improvements:
- Preserves language understanding by freezing VLM
- Optimizes balance between action prediction and language compliance with FLARE loss
- More than 2x improvement in language instruction compliance
GR00T N1.6 (2025.06)
Model scale expansion and real-world performance improvement.
| Item | Details |
|---|---|
| Key Improvement | 2x larger DiT, Cosmos VLM |
| VLM | Cosmos-2B (2x compared to Eagle2-1B) |
| DiT | 32 layers (2x compared to 16) |
| Action Space | Relative Action Space |
| Real World | Improved sim-to-real transfer performance |
Major Improvements:
- Diffusion Transformer expanded to 32 layers (2x)
- Enhanced visual understanding with Cosmos-2B VLM
- Improved generalization with Relative Action Space
- Better sim-to-real transfer performance
Version Comparison
| Item | N1 | N1.5 | N1.6 |
|---|---|---|---|
| VLM | Eagle2-1B | Eagle2-1B (Frozen) | Cosmos-2B |
| DiT Layers | 16 | 16 | 32 |
| Language Compliance | 46.6% | 93.3% | 93%+ |
| Action Space | Absolute | Absolute | Relative |
| Key Contribution | First open model | Language compliance improvement | Scale expansion |
Architecture
GR00T N1 is a Dual-System architecture inspired by human cognitive principles.
System 2 (Slow Thinking)
- Based on Vision-Language Model
- Develops action plans by understanding environment and instructions
- Careful and methodical decision-making
System 1 (Fast Thinking)
- Based on Diffusion Transformer
- Converts plans into precise continuous motions
- Corresponds to human reflexes/intuition
Training
Data Sources
| Source | Description |
|---|---|
| Human Demonstration Data | Actual human actions (teleoperation) |
| Synthetic Data | Generated by NVIDIA Omniverse/Isaac Sim |
| Human Video | For motion pretraining |
GR00T Training Data Distribution: Combination of real and synthetic data
Power of Synthetic Data
Efficiency of synthetic data generation through Omniverse:
| Item | Value |
|---|---|
| Generated Synthetic Trajectories | 780,000 |
| Real-time Equivalent | 6,500 hours (9 months continuous) |
| Generation Time | 11 hours |
| Performance Improvement | +40% over real data only |
Training Infrastructure
| Item | Details |
|---|---|
| GPU | 1,000x H100 |
| Training Steps | 250K steps |
| Framework | Isaac Lab + Omniverse |
Capabilities
Performable Tasks
- Object grasping
- Moving objects with single/bimanual
- Object transfer between arms
- Multi-step tasks requiring long context
- Combination of general skills
Key Features
| Feature | Description |
|---|---|
| Natural Language Understanding | Understands and executes language instructions |
| Motion Imitation | Learns by observing human behavior |
| Generalization | Easily generalizes to common tasks |
Hardware: Jetson Thor
New computing platform designed for GR00T execution.
| Item | Spec |
|---|---|
| GPU Architecture | NVIDIA Blackwell |
| AI Performance | 800 TFLOPS (8-bit FP) |
| Transformer Engine | Built-in |
| Design | Modular, optimized for performance/power/size |
Ecosystem: Isaac Platform
GR00T is part of NVIDIA’s Isaac robotics platform.
| Component | Role |
|---|---|
| Isaac Sim | Simulation environment |
| Isaac Lab | Reinforcement learning framework |
| Omniverse | Synthetic data generation |
| GR00T | Foundation model |
Industry Partners
NVIDIA is collaborating with major humanoid robot companies:
- 1X Technologies
- Agility Robotics
- Apptronik
- Boston Dynamics
- Figure AI
- Fourier Intelligence
- Sanctuary AI
- Unitree Robotics
- XPENG Robotics
Cross-Embodiment
GR00T is designed to support various robot forms:
| Supported Robot | Type |
|---|---|
| Fourier GR-1 | Humanoid |
| Unitree G1/H1 | Humanoid |
| Agility Digit | Humanoid |
| ALOHA | Bimanual manipulator |
| Franka | Single arm |
Benchmarks
LIBERO Benchmark (Simulation)
| Task | N1 Success Rate |
|---|---|
| LIBERO-Object | 96.7% |
| LIBERO-Spatial | 92.5% |
| LIBERO-Goal | 85.0% |
| LIBERO-Long | 78.3% |
Language Instruction Compliance (Language Following)
| Version | Compliance Rate |
|---|---|
| N1 | 46.6% |
| N1.5 | 93.3% |
| N1.6 | 93%+ |
Impact
Significance of GR00T series:
N1 (2025.03)
- First open humanoid foundation model
- Proved utilization of large-scale synthetic data through simulation
- Separated planning and execution with Dual-system architecture
N1.5 (2025.05)
- Solved language understanding preservation problem by freezing VLM
- Achieved balance between action learning and language compliance with FLARE loss
N1.6 (2025.06)
- Improved expressiveness through model scale expansion
- Enhanced generalization ability with Relative action space
- Provides common platform for entire humanoid industry
References
GR00T N1
GR00T N1.5
GR00T N1.6
Project GR00T
- NVIDIA Newsroom - Project GR00T
- NVIDIA Developer - Isaac GR00T
- GitHub - Isaac-GR00T
- IEEE Spectrum Article
See Also
Related People
- Jim Fan - NVIDIA GEAR Lab, GR00T Research Lead