brains/gr00t-n1-6 · model record

Isaac GR00T N1.6

3B-parameter VLA that swaps in an internal Cosmos-2B VLM, doubles the diffusion transformer to 32 layers and predicts state-relative action chunks. Pretraining adds several thousand hours of teleoperation data; NVIDIA demonstrates it on Unitree G1 whole-body loco-manipulation, AgiBot Genie-1 and bimanual YAM arms.