NVIDIA Isaac GR00T N1.6

2025-12-15 · Publisher : NVIDIA · Domain : Robotics

NVIDIA Research publishes the GR00T N1.6 report, an open ~3-billion-parameter humanoid VLA built on a 2B Cosmos VLM and a diffusion transformer twice as deep as N1.5's.

What changes

  • About 3 billion parameters (GR00T-N1.6-3B).
  • Base VLM: an internal Cosmos-2B variant.
  • 32-layer diffusion transformer, vs 16 in N1.5.
  • Added data: YAM arms, AgiBot Genie-1, simulated Galaxea R1 Pro, Unitree G1.
  • Announced September 29, 2025 at CoRL; technical report published December 15.

### Background On December 15, 2025, NVIDIA Research's GEAR lab published the technical overview of Isaac GR00T N1.6, an improved version of GR00T N1.5. The model had been announced on September 29, 2025 at CoRL; its weights went up on Hugging Face in early December. ### Specifications - **Size**: about 3 billion parameters (GR00T-N1.6-3B). - **Base VLM**: an internal Cosmos-2B variant supporting flexible resolution, trained on general vision-language and embodied reasoning tasks. - **Architecture**: a diffusion transformer twice as large (32 layers vs 16), removal of N1.5's adapter in favor of unfreezing the VLM's top 4 layers, and state-relative action prediction. - **Data**: several thousand additional hours of teleoperation on bimanual YAM arms, AgiBot Genie-1, simulated Galaxea R1 Pro and Unitree G1 loco-manipulation. - **Results**: N1.6 outperforms N1.5 in simulation and on real robots (results published by NVIDIA). ### Why it matters N1.6 integrates Cosmos reasoning and loco-manipulation, so humanoids can move and manipulate at the same time. NVIDIA consolidates an open base that robot makers can post-train on their own hardware. ### Key takeaway N1.6 was showcased at CES 2026 and integrated into LeRobot; N1.7 and the N2 preview followed at GTC 2026.

Links

Tags : VLA, Humanoïdes, NVIDIA, Poids ouverts