Summary
NVIDIA released GR00T N1.7 in early access at GTC 2026 on March 16, 2026 — an open, commercially licensed VLA model for humanoid robots. Pretrained via EgoScale on 20,854 hours of egocentric human video spanning 20+ task categories, it provides cross-embodiment generalization with a new Cosmos-Reason2-2B/Qwen3-VL backbone. Commercial licensing is enabled for the first time in the GR00T series, opening production deployments.
Key Contributions
- EgoScale pretraining on 20,854 hours of egocentric human video across manufacturing, healthcare, retail, and home domains
- New VLM backbone (Cosmos-Reason2-2B / Qwen3-VL) for enhanced multimodal reasoning
- Cross-embodiment VLA supporting diverse humanoid configurations without architecture changes
- First GR00T release with a commercial open license
Significance
GR00T N1.7 is the most capable open humanoid robot foundation model to date and the first commercially licensed release in its series, significantly lowering the barrier for production-scale humanoid deployments. NVIDIA simultaneously previewed GR00T N2 (based on the DreamZero world action model architecture), signaling a shift toward world-model-based robot policies.