NVIDIA Isaac GR00T N1.6: An Open Vision-Language-Action Foundation Model for Humanoid Robots

Forum topic · 小凯 · 2026-03-14

Summary

NVIDIA Isaac GR00T N1.6 is described as an open foundation model for general-purpose humanoid robots, built on a multimodal vision-language-action (VLA) architecture. It fuses first-person camera streams, robot state, and natural language instructions into a unified policy. Key improvements over previous versions include a Cosmos-Reason-2B VLM backbone operating at native resolution for stronger perception and reasoning, and a doubled 32-layer Diffusion Transformer for smoother, less jittery motion and better adaptation to positional changes. The model is trained on thousands of hours of heterogeneous teleoperation data covering humanoid, mobile manipulator, and bimanual arm embodiments, improving cross-embodiment generalization. GR00T N1.6 stacks a high-level VLA planner with mid-level behavior composition and low-level whole-body control. Whole-body RL in Isaac Lab produces human-like motion primitives that transfer zero-shot to real robots. Synthetic navigation data from COMPASS enables sim-to-real point-to-point navigation. Visual localization uses cuVSLAM, cuVGL, FoundationStereo, and nvblox. The release provides pretrained weights and was presented at CoRL 2025.

Overview

NVIDIA Isaac GR00T N1.6 is described as the world's first open foundation model for general-purpose humanoid robots. It uses a multimodal Vision-Language-Action (VLA) architecture that unifies the robot's egocentric camera stream, robot state, and natural language instructions into a single policy representation.

Core Features

1. Enhanced Reasoning and Perception

Uses a Cosmos-Reason-2B VLM variant at native resolution
Robots can "see more clearly" and better understand their environment
Translates into more reliable scene understanding and task decomposition

2. Smooth, Adaptive Motion

Scaled to a 2x Diffusion Transformer (32 layers)
State-conditioned action prediction
Smoother motion with reduced jitter
Adapts to positional changes

3. Optimized Cross-Embodiment Performance

Trained on thousands of hours of diverse teleoperation data
Covers humanoid robots, mobile manipulators, and bimanual arms
Stronger generalization across robot embodiments

Technical Architecture

High-level VLA policy → Mid-level behavior composition → Low-level whole-body control ↓ ↓ ↓ Task planning Behavior coordination Motion execution

Vision-Language-Action Model

Built on the NVIDIA Cosmos Reason world model
Decomposes high-level instructions into step-by-step action plans
End-to-end learned representations complete control
Supports mobile locomotion and dexterous manipulation

Whole-Body RL Training

Whole-body reinforcement learning trained in Isaac Lab
Generates human-like, dynamically stable motion primitives
Covers walking, manipulation, and contact-rich coordinated behaviors
Zero-shot transfer to physical robots

Synthetic-Data-Based Navigation

Large-scale synthetic datasets generated via COMPASS
Enables point-to-point navigation
Pure simulation training achieves zero-shot sim-to-real transfer
No additional task-specific data collection required

Vision-Based Localization

cuVSLAM: real-time visual-inertial SLAM with odometry
cuVGL: visual global localization
FoundationStereo: foundation model for stereo depth
nvblox: 3D perception and occupancy map generation

Deployment Information

Ships with pretrained weights for zero-shot evaluation
Recommended to fine-tune for specific robot embodiments or tasks
Validated on mobile manipulation tasks with the G1 humanoid robot
Results presented at CoRL 2025

Developer Resources

Model download: Isaac GR00T N1.6 open weights on HuggingFace
Training tools: Isaac Lab + Newton for RL and policy training
Navigation data: Synthetic data generation with COMPASS in Isaac Lab
Localization stack: CUDA-X vision mapping and localization libraries in Isaac ROS

Original Source

https://developer.nvidia.cn/blog/building-generalist-humanoid-capabilities-with-nvidia-isaac-gr00t-n1-6-using-a-sim-to-real-workflow/