VLA Gallery / All models / Humanoid VLA

GR00T N1.5

NVIDIA · 2025-05

VLAHumanoid VLA3Bflow matching

Freezes the VLM in both pre- and post-training and adds a FLARE (future latent representation alignment) loss (coefficient 0.2) to flow matching, learning also from human egocentric video.

Paper

No paper · Related GR00T N1: An Open Foundation Model for Generalist Humanoid Robots

Overview

Category

Humanoid VLA

Type

VLA

Released

2025-05

Organization

NVIDIA

Open weights

Yes

Parameters

3B (VLM 2.1B)

Architecture

Backbone LLM

Eagle 2.5 VLM 2.1B (internal LLM not stated, frozen)

Vision encoder

Built into Eagle 2.5

Action head

flow matching DiT (4 steps) + FLARE (flow matching)

Chunk

Not disclosed

Control rate

Not disclosed

Architecture (original summary)

Frozen Eagle 2.5 VLM (2.1B, grounding-tuned) + adapter MLP with added layer normalization + DiT action head (flow matching, 4 denoising steps)

Training

Pretraining accelerators

1K H100

Pretraining duration

Not disclosed

Pretraining steps & batch

250K steps · batch 16,384

GPU/TPU-hours

Not disclosed

Fine-tuning

1 H100 or L40 node recommended, max batch·20k steps recommended; LoRA on A6000 or 2 RTX 4090; post-training on 1K G1 teleoperation episodes (GPU·time not disclosed)

Data type

Real-robot teleoperation / simulation·synthetic

Data scale

Time not disclosed; G1 post-training teleoperation 1K episodes

Main datasets

Internal GR-1 data, OpenXE, sim GR-1 (DexMG), DreamGen neural trajectories, AgiBot-Beta

Tasks & evaluation

Language Table 93.2%; Sim GR-1 Language 54.4%; RoboCasa (30 demos) 47.5; real GR-1 language instructions 93.3% (N1 46.6%); Unitree G1 seen objects 98.8%·novel 84.2% (N1 44.0%); DreamGen 12 new verbs 38.3%

Robot

Fourier GR-1 (real), Unitree G1 (real), sim cross-embodiment

Inference

Inference device & speed

H100: VLM 23.18 ms + action head 4 × 6.18 = 24.7 ms, total 47.88 ms; Jetson Thor TensorRT FP16/FP8+FP4: DiT 5.46/3.41 ms, ViT 5.21/4.10 ms, LLM 8.53/5.81 ms

Inference memory

Not disclosed

Weights bf16 (calc.)

6 GB

Where it runs

Not disclosed

Reinforcement learning & base

Reinforcement learning

Partial (RL low-level controller) · VLA by imitation learning. Uses Decoupled WBC (lower-body RL·upper-body IK) to run on Unitree G1

Base control

Fixed · No mention of locomotion commands (upper-body manipulation such as GR-1·G1 pick·place)

Notes

Key features

Freezes the VLM in both pre- and post-training and adds a FLARE (future latent representation alignment) loss (coefficient 0.2) to flow matching, learning also from human egocentric video. Generalizes to new motions beyond teleoperation via DreamGen neural trajectories.

Notes

Release month corrected in verification to 2025-05 (NVIDIA press release 2025-05-18, Computex); research page 2025-06-11. The HF card describes it as SigLip2·T5·flow matching transformer, differing from the research page (Eagle 2.5). The '36 hours' of development in the press release is data generation·development time, not training compute, and the same press release names AeiRobot as an early adopter. Nvidia License (non-commercial).

Verification

Corrections applied