GR00T N1.7 (NVIDIA)

Key Claims

• Open VLA model for humanoid manipulation — multimodal input (language + images) to generate robot actions
• 20K hours of EgoScale human video used in pretraining — larger scale than Ψ₀ (800h) but similar recipe philosophy
• Successor to GR00T N1.6 with improved generalization and language-following
• Integrated with NVIDIA's Isaac Sim ecosystem — training data flows from Isaac Sim simulation to real deployment

Why This Matters

GR00T is NVIDIA's commercial bet on being the "Android of humanoids" — a foundation model every humanoid OEM can use. Companies building humanoids without internal AI talent (the majority of the industry) will default to GR00T unless they have a compelling reason to build in-house.

Compared to Ψ₀ and π₀.₅: GR00T brings compute-stack integration and brand-scale distribution; the research labs bring novel recipes. The commercial question for the humanoid market is whether recipe innovation (Ψ₀'s 10× data efficiency) beats distribution muscle (NVIDIA's ecosystem).

Notes

First-pass stub. Version numbering suggests rapid iteration — N1 → N1.6 → N1.7 in recent months. Schedule future discovery passes around NVIDIA GTC announcements.

Source: NVIDIA Isaac-GR00T GitHub