VLA Gallery / All models / Humanoid VLA

GR00T N1

NVIDIA · 2025-03

VLAHumanoid VLA2.2Bflow matching

Attaches pseudo-actions from latent actions (LAPA) and an inverse dynamics model (IDM) to action-unlabeled human video and neural-generated video, pretraining them together with real-robot and sim data.

Paper

GR00T N1: An Open Foundation Model for Generalist Humanoid RobotsarXiv 2503.14734

Overview

Category

Humanoid VLA

Type

VLA

Released

2025-03

Organization

NVIDIA

Open weights

Yes

Parameters

2.2B (VLM 1.34B; DiT not disclosed)

Architecture

Backbone LLM

SmolLM2 · Eagle-2 VLM (uses layer-12 representation)

Vision encoder

SigLIP-2

Action head

flow matching DiT (4 steps) (flow matching)

Chunk

16 steps

Control rate

120 Hz

Architecture (original summary)

System 2: Eagle-2 VLM (1.34B, 10 Hz on L40) / System 1: flow-matching DiT, action chunk H=16·denoising K=4, 120 Hz continuous motor actions

Training

Pretraining accelerators

Up to 1,024 GPUs (GPU-hours based on H100)

Pretraining duration

Not disclosed

Pretraining steps & batch

200,000 steps · batch 16,384

GPU/TPU-hours

50,000 (Stated · H100-hours ("roughly 50,000"))

Fine-tuning

Default 60k steps·batch 1,024 (Table 6: 20,000~60,000 steps, batch 128 or 1,024), GPU·time Not disclosed; on 1× A6000, batch up to 200 (adapter+DiT)·16 (including vision encoder)

Data type

Real-robot teleoperation / human video / simulation·synthetic

Data scale

Total 592.9M frames·8,375.7 hours (robot 262.3M·3,288.8 h, human 181.3M·2,517.0 h, sim 125.5M·1,742.6 h, neural 23.8M·827.3 h)

Main datasets

GR-1 Teleop 88.4 h; AgiBot-Alpha 1,979.4 h (140,000 trajectories); DROID 428.3 h; Ego4D 2,144.7 h; GR-1 Simulation 1,742.6 h; GR-1 Neural Videos 827.3 h

Tasks & evaluation

Sim (100 demos per task) RoboCasa 32.1%·DexMimicGen 66.5%·GR-1 Tabletop 50.0% (average 45.0%); real GR-1 full data average 76.8%, 10% data 42.6%

Robot

Fourier GR-1 humanoid (real · sim), cross-embodiment incl. Franka · bimanual sim; upper body · tabletop work, no locomotion

Inference

Inference device & speed

16-step chunk 63.9 ms on L40 bf16; S2 10 Hz·S1 120 Hz

Inference memory

Not disclosed

Weights bf16 (calc.)

4.4 GB

Where it runs

Not disclosed

Reinforcement learning & base

Reinforcement learning

Not used · flow matching imitation learning (including sim·neural trajectories)

Base control

Fixed · Upper-body control only (arms·hands·waist·neck), no locomotion; RoboCasa also outputs only EE·gripper

Notes

Key features

Attaches pseudo-actions from latent actions (LAPA) and an inverse dynamics model (IDM) to action-unlabeled human video and neural-generated video, pretraining them together with real-robot and sim data. Also released pretraining GPU-hours and a per-dataset hours table.

Notes

Pretraining steps corrected in verification to the Table 6 value 200,000; the Total row was also added in verification (NEW). Data-generation compute is separate: neural trajectories on 3,600 L40s·about 105k L40 GPU-hours (~1.5 days); DexMimicGen generated 780,000 trajectories (equivalent to 6,500 h of human demonstrations) in 11 h.

Verification

Corrections applied