GR00T N1.7

NVIDIA · 2026-03

VLANew in 20263Bflow matching

An open reasoning VLA pretrained on 20K hours of EgoScale human video and robot demonstrations.

Paper

Backbone LLM: Cosmos-Reason2-2B (Qwen3-VL architecture)
Vision encoder: Qwen3-VL vision
Action head: flow matching DiT 16 layers (4 steps) (flow matching)
Chunk: 40 steps
Control rate: Not disclosed
Architecture (original summary): Cosmos-Reason2-2B VLM (Qwen3-VL architecture) + flow matching DiT action head (AdaLN, 16 layers; 32 layers in N1.6); state · action dim 132, 40-step horizon; relative end-effector action space shared by humans and robots

Pretraining accelerators: 64 GB200 nodes × 4 GPUs per node (original: '64 GB200 nodes * 4 gpus per node')
Pretraining duration: 120 hours
Pretraining steps & batch: Not disclosed
GPU/TPU-hours: 30,720 (Converted · GB200-hours (64 nodes × 4 GPUs × 120 hours))
Fine-tuning: Not disclosed (user fine-tuning supported)
Data type: Human video (EgoScale egocentric) + real-robot teleoperation + simulation·synthetic
Data scale: 20K h of human video (README; secondary: about 32,000 h); 21.6M data points, 13 datasets (including Bridge 60,096 trajectories, DROID about 76,000 trajectories, LIBERO 130 tasks)
Main datasets: 13 sources incl. EgoScale human video, Bridge, DROID, LIBERO
Tasks & evaluation: Simulation DexMG 9·RoboCasa 24·Digital Cousin GR-1 24 tasks; real-robot grocery packing·new objects·multi-robot collaboration (no numbers on the card); 61% improvement in some simulations (secondary); G0.5 paper R1lite/R1pro 24.4%
Robot: Humanoid · cross-embodiment incl. GR-1 (simulation); real-robot model not stated on the card

Inference device & speed: TensorRT full pipeline: H100 80GB·RTX Pro 6000 27.9 ms (35.9 Hz), AGX Thor 93.8 ms (10.7 Hz); Jetson Orin 216.5 ms (4.6 Hz) is TensorRT DiT-only; 4 denoising steps·1 camera
Inference memory: Not disclosed
Weights bf16 (calc.): 6 GB
Where it runs: Not disclosed

Reinforcement learning: Not used · Supervised learning on human video·robot data (disclosed scope)
Base control: Humanoid legged (Commands/latents → RL locomotion controller) · G1 (SONIC tag): VLA outputs latent action tokens → SONIC learned WBC decodes them into whole-body joints incl. legs for locomotion · manipulation

Key features: An open reasoning VLA pretrained on 20K hours of EgoScale human video and robot demonstrations. Supports task- and subtask-level reasoning and finger-level dexterous control.
Notes: Corrections applied: the Jetson Orin value is in TensorRT DiT-only mode, a different condition from other devices (full pipeline). Early access 2026-03-16 (press release), forum 2026-04-17, GA date 2026-07-07† (secondary). License: weights NVIDIA Open Model License, code Apache 2.0 (internally conflicts with the 'Apache 2.0 commercial use' sentence in the README overview). Pretraining energy 41,288 kWh, 16.949 tCO2e. GPU·TPU-hours converted as nodes×GPUs×hours. Data hours (h) are the README human video portion (robot demonstrations separate).
Verification: Corrections applied