GR00T N1.6

NVIDIA · 2025-09

VLAHumanoid VLA3Bflow matching

Removes N1.5's 4-layer adapter after the VLM, unfreezes the top 4 VLM layers, and doubles the DiT (32 layers).

Paper

Backbone LLM: Cosmos-Reason-2B variant (README) / Eagle-Block2A-2B-v2 (N1.7 README)
Vision encoder: VLM built-in
Action head: flow matching DiT 32 layers (4 steps) (flow matching)
Chunk: Not disclosed
Control rate: Not disclosed
Architecture (original summary): Internal Cosmos-Reason-2B variant VLM (research page calls it a Cosmos-2B variant), top 4 layers trained + 32-layer DiT; state-relative action chunks for most embodiments

Pretraining accelerators: Not disclosed
Pretraining duration: Not disclosed
Pretraining steps & batch: 300K steps · batch 16,384
GPU/TPU-hours: Not disclosed
Fine-tuning: 10K~30K steps · batch 1K or less; one H100 or L40 node recommended (A6000 possible, takes longer); single-GPU example 2,000 steps · batch 32
Data type: Real-robot teleoperation / simulation·synthetic
Data scale: N1.5 mix + thousands of additional hours of teleoperation; base checkpoints (N1.5·N1.6) 10k+ hours of robot data
Main datasets: N1.5 mix + YAM bimanual, AgiBot Genie-1, sim Galaxea R1 Pro (BEHAVIOR), Unitree G1 whole-body loco-manipulation teleoperation
Tasks & evaluation: Improvement over N1.5 on sim manipulation benchmarks and real YAM·AgiBot Genie-1·Unitree G1 (numbers not disclosed); G1 loco-manipulation pick-and-place fine-tuned checkpoint released
Robot: Cross-embodiment: Unitree G1 (whole body), AgiBot Genie-1, bimanual YAM, sim Galaxea R1 Pro, etc.

Inference device & speed: 4 denoising steps·single view·torch.compile: RTX 5090 37 ms (27.3 Hz), H100 38 ms (26.3 Hz), RTX 4090 44 ms (22.8 Hz), Thor 105 ms (9.5 Hz)
Inference memory: Not disclosed
Weights bf16 (calc.): 6 GB
Where it runs: Not disclosed

Reinforcement learning: Partial (RL low-level controller) · VLA by imitation learning. Combined with Decoupled WBC (lower-body RL·upper-body IK) for G1 whole-body loco-manipulation
Base control: Wheeled·Humanoid legged (Commands/latents → RL locomotion controller) · G1: VLA outputs navigate_command · base_height_command → RL lower-body WBC handles locomotion; sim R1 Pro (BEHAVIOR) outputs base actions directly

Key features: Removes N1.5's 4-layer adapter after the VLM, unfreezes the top 4 VLM layers, and doubles the DiT (32 layers). Combined with Decoupled WBC (RL lower body·IK upper body) for Unitree G1 whole-body loco-manipulation.
Notes: Release month corrected in verification to 2025-09 (press release 2025-09-29, 'available soon'); WBC repository 2025-11-12, research page 2025-12-15. Backbone naming conflicts among NVIDIA primary sources (README Cosmos-Reason-2B, research page Cosmos-2B, N1.7 README says it replaced Eagle-Block2A-2B-v2). '10k+ hours' is a sentence describing the N1.5·N1.6 base checkpoints together. Hardware guide: 8× RTX PRO 6000 Blackwell server (starter), DGX B300, deployment Jetson AGX Thor.
Verification: Corrections applied