Key Significance
- 2x DiT Expansion: Diffusion Transformer scale expansion from 16 to 32 layers
- Cosmos VLM Introduction: VLM changed from Eagle to Cosmos-Reason-2B, enhanced reasoning capability
- Relative Action Space: Improved generalization and adaptability with relative action space
- Sim-to-Real Performance Improvement: Improved zero-shot simulation-to-real-world transfer
- Loco-manipulation Support: Supports whole-body motion combining locomotion and manipulation
Overview
| Item | Details |
|---|---|
| Announced | September 29, 2025 (CoRL 2025, Seoul) |
| Type | Vision-Language-Action (VLA) |
| Parameters | 3B |
| VLM | Cosmos-Reason-2B |
| DiT | 32 layers (2x compared to N1.5) |
| Action Space | Relative Action Space |
| GitHub | NVIDIA/Isaac-GR00T |
| Hugging Face | nvidia/GR00T-N1.6-3B |
Key Improvements over N1.5
1. DiT Layer Expansion (16 → 32)
| Aspect | N1.5 | N1.6 |
|---|---|---|
| DiT Layers | 16 | 32 (2x) |
| Effect | - | Smoother, less jittery movements, easier adaptation to changing positions |
The larger 32-layer Diffusion Transformer, combined with state-relative action prediction, generates more flexible and adaptive motions.
2. Cosmos VLM (2B) Introduction
N1.6 uses NVIDIA Cosmos-Reason-2B VLM as its base VLM instead of Eagle.
| Aspect | N1.5 | N1.6 |
|---|---|---|
| VLM | Eagle 2.5 (1B) | Cosmos-Reason-2B |
| Parameters | ~1B | 2B (2x) |
| VLM Training | Fully frozen | Top 4 layers unfrozen |
| Adapter | 4-layer transformer | Removed |
Cosmos-Reason Key Features:
- Flexible resolution support: Can encode images at native aspect ratio without padding
- Deep thinking capability: Serves as the robot’s “deep thinking brain”
- Ambiguous instruction interpretation: Converts ambiguous instructions into step-by-step plans using prior knowledge, common sense, and physics
3. Relative Action Space
N1.6 predicts state-relative action chunks instead of absolute joint angles or EEF positions.
| Aspect | N1/N1.5 | N1.6 |
|---|---|---|
| Action Space | Absolute | Relative |
| Motion Characteristics | Fixed position based | Relative to current state |
Advantages:
- Smoother and more accurate motion generation
- Easier adaptation to changing positions
- Less jittery movements
Caveats:
- Error accumulation may occur on small datasets, affecting correction capability
Architecture
GR00T N1.6 Model Architecture (Source: NVIDIA Research)
Key Architecture Changes (N1.5 → N1.6)
| Component | N1.5 | N1.6 |
|---|---|---|
| Base VLM | Eagle 2.5 (frozen) | Cosmos-Reason-2B (top 4 layers unfrozen) |
| DiT Size | 16 layers | 32 layers |
| VLM Post-processing Adapter | 4-layer transformer adapter | Removed |
Benchmarks
Evaluation Environments
N1.6 is evaluated across various simulation and real robot environments:
| Evaluation | Description |
|---|---|
| LIBERO | Evaluation after 20-40k step post-training on LIBERO dataset |
| SimplerEnv | Evaluation after finetuning on Google Robot fractal dataset |
| BEHAVIOR1k | Post-training checkpoints provided |
| IsaacLabEvalTasks | Industrial manipulation tasks (Nut Pouring, Exhaust Pipe Sorting) |
Real Robot Demonstrations
Tasks demonstrated on NVIDIA Research page:
- T-shirt folding
- Object insertion
- Bimanual handoff
- Loco-manipulation with Unitree G1
Performance Characteristics
According to NVIDIA Research page:
- N1.6 converges faster than N1.5, generating smoother actions
- Requires more careful tuning to prevent overfitting
- 5-6% inter-experiment variance observed
Note: Specific benchmark numbers for N1.6 have not yet been published on the official research page. See N1 and N1.5 documents for their performance comparisons.
Training
Pretraining
| Item | N1.6 |
|---|---|
| Pretraining Steps | 300K |
| Global Batch Size | 16,384 |
| Post-training Steps | 10K-30K (batch size ≤1K) |
Pretraining Data Distribution
GR00T N1.6 Pretraining Data Distribution (Source: NVIDIA Research)
N1.6 was trained with thousands of hours of new teleoperation data compared to N1.5.
Main Data Sources
| Data Source | Platform Type | Description |
|---|---|---|
| Bimanual YAM Arms | Bimanual manipulator | Precise bimanual manipulation task data |
| AGIBot Genie1 | Semi-humanoid | Various manipulation task data |
| Simulated Galaxea R1 Pro | Humanoid | Synthetic data based on BEHAVIOR suite |
| Unitree G1 | Humanoid | Whole-body loco-manipulation data |
Finetuned Checkpoints
N1.6 provides finetuned checkpoints for various tasks/environments.
| Checkpoint | Robot | Task |
|---|---|---|
| GR00T-N1.6-bridge | WidowX | Manipulation |
| GR00T-N1.6-fractal | Google Robot | Manipulation |
| GR00T-N1.6-BEHAVIOR1k | Galaxea R1 Pro | Loco-manipulation |
| GR00T-N1.6-G1-PnPAppleToPlate | Unitree G1 | Loco-manipulation (Pick & Place) |
Full list and usage: GitHub - Isaac-GR00T README
Post-training Notes
- 5-6% performance variance observed even with identical settings, seed, and dropout
- N1.6 converges faster than N1.5, increasing overfitting risk
- Careful hyperparameter tuning required
Supported Robots
Robot platforms validated on N1.6:
| Platform | Type | Documentation |
|---|---|---|
| Bimanual YAM robot | Bimanual manipulator | - |
| AGIBot Genie-1 | Semi-humanoid | AGIBot |
| Unitree G1 | Humanoid | Unitree Humanoid |
| Fourier GR-1 | Humanoid | - |
Version Comparison Summary
| Feature | N1 | N1.5 | N1.6 |
|---|---|---|---|
| Announced | 2025.03 GTC | 2025.05 Computex | 2025.09 CoRL |
| Model Size | 2.2B | 3B | 3B |
| Base VLM | Eagle2-1B (trainable) | Eagle 2.5 (frozen) | Cosmos-Reason-2B (top 4 unfrozen) |
| DiT Layers | 16 | 16 | 32 |
| Action Space | Absolute | Absolute | Relative |
| Pretraining Steps | 250K | 250K | 300K |
| Key Feature | Basic VLA, synthetic data | FLARE, language understanding | Scale-up, Loco-manipulation |
Reference: NVIDIA Sim-to-Real Workflow
Note: This section describes NVIDIA’s robotics workflow that can be used with N1.6, not features of the N1.6 model itself.
Details: Building Generalist Humanoid Capabilities with GR00T N1.6 (NVIDIA Developer Blog)
The sim-to-real workflow introduced in NVIDIA’s developer blog includes three components:
| Component | Description |
|---|---|
| Whole-Body RL | Dynamically stable motion primitives trained via RL in Isaac Lab/Sim |
| COMPASS Navigation | Synthetic-data-trained navigation achieving zero-shot sim-to-real transfer |
| Vision-Based Localization | CUDA-accelerated libraries: cuVSLAM, cuVGL, FoundationStereo, nvblox |
References
GR00T N1.6
News
Base
See Also
GR00T Series
- GR00T - Series Overview
- GR00T N1 - First Open-Source Model
- GR00T N1.5 - Language Understanding Improvement
Related Models
- Cosmos - N1.6’s VLM (Cosmos-Reason-2B)
- Eagle - N1, N1.5’s VLM
- DreamGen - Synthetic Data Generation Pipeline
Related People
- Jim Fan - NVIDIA GEAR Lab, GR00T Research Lead