EN

NVIDIA

GR00T N1.7

GR00T N1.7

Checkpoint

Open 3B reasoning VLA for cross-embodiment and humanoid robot control.

Specifications

Architecture components: Flow-matching action expert
Architecture components: Action Cascade reasoning
Backbone: Cosmos-Reason2-2B / Qwen3-VL with flow-matching action head
Frameworks: NVIDIA Isaac GR00T / LeRobot
Input modalities: Language instruction
Input modalities: RGB camera images
Input modalities: Robot state / proprioception
Languages: English
Model type: Vision-language-action model
Output action space: Continuous robot action chunks
Parameter count: 3B parameters
Pipeline task: Robotics
Training data scale: 20K+ hours of human egocentric video

Lineage

Unknown

Architecture & I/O

trajectory · continuous

Unknown

Continuous robot action chunks

Source · nvidia

Default checkpoint

nvidia/gr00t17-lerobot-libero_goal-640

openApache License 2.0

Other releases

5

published checkpoints

nvidia/GR00T-N1.7-3B nvidia/gr00t17-lerobot-libero_10-640 nvidia/gr00t17-lerobot-libero_goal-640 nvidia/gr00t17-lerobot-libero_object-640 nvidia/gr00t17-lerobot-libero_spatial-640

Supported robots

No verified robot compatibility recorded.

Benchmark results

LIBERO · LIBERO-Goal

Success rate: 97 percent

official · reported · Protocol LeRobot 400-episode protocol

At least 50 evaluation episodes for this LIBERO suite.

Source · Hugging Face LeRobot

Training data

finetune

GR00T LIBERO Spatial (no no-ops)

open

Source · Hugging Face LeRobot

finetune

GR00T LIBERO Object (no no-ops)

open

Source · Hugging Face LeRobot

finetune

GR00T LIBERO Goal (no no-ops)

open

Source · Hugging Face LeRobot

finetune

GR00T LIBERO-10 (no no-ops)

open

Source · Hugging Face LeRobot

Resources

documentation · LeRobot GR00T documentation project page · NVIDIA GR00T N1.7 release repository · NVIDIA Isaac GR00T repository hub · nvidia/gr00t17-lerobot-libero_goal-640 model card weights · nvidia/gr00t17-lerobot-libero_goal-640 model card

Submit info